ECE 636: Systems identification · 2011-12-05 · ECE 636: Systems identification Lectures 23‐24...
Transcript of ECE 636: Systems identification · 2011-12-05 · ECE 636: Systems identification Lectures 23‐24...
ECE 636 Systems yidentificationLectures 23‐24
Identification of closed‐loop systemsNonlinear systemsNonlinear systems
Applications
bull Model order selectionbull Start from simple models and try successively more complex modelsbull Compare model prediction to observations (we shouldnrsquot get zero error) ˆ( ) ( ) ( )my t G q u tΝ= θ
bull Residual testingbull PEM residuals should be white for
bull Whiteness testing Residual autocorrelation function0
ˆN =θ θ 1ˆ ( ) ( ) ( )
N
t tτ
ϕ τ ε ε τminus
= +sum
ˆ ˆˆ( | ) ( ) ( | )N Nt y t y tε = minusθ θ
bull Whiteness testing Residual autocorrelation function
Cross‐correlation function between inputresiduals
Statistical hypothesis testing
1( ) ( ) ( )
tt t
Nεεϕ τ ε ε τ=
= +sum
1
1ˆ ( ) ( ) ( )N
ut
t u tN
τ
εϕ τ ε τminus
=
= minussum2ˆ (0)ϕ λrarrStatistical hypothesis testing (0)
ˆ ( ) 0 0εε
εε
ϕ λϕ τ τ
rarrrarr ne
2 22
1
ˆ ( )ˆ (0)
m
mN
εετεε
ϕ τ χϕ =
rarrsum ˆ ( ) (01)ˆ (0)
N Nεε
εε
ϕ τϕ
rarr
Open loop systems any τ Closed‐loop systems only for τgt0
( ) ( ) ( ) 0u E t u tεϕ τ ε τ= + =
12
ˆ ( )ux ετ
ϕ τ= (0 1)N x Nrarr[ ]12ˆ ˆ(0) ( )uu
xτεεϕ ϕ τ
(01)N x Nτ rarr
bull Cross‐validationbull Sufficient data pointsbull System (almost) time‐invariant
M d l ibull Model comparison
( )
St ti ti l it i
1 2 2 1
2 2
( ) ( ) ( )
MSE MSE p pFMSE N pminus minus
=minus 2 1 2p p N pF minus minus 2 1
2p pχ minus
2ˆ pbull Statistical criteria 2ˆ( ) ( )1 N NpAIC p VN
= +θ
1 ˆ( ) ( )1 N Np NFPE p Vp N
+=
minusθ
bull Recursive identificationU d t th ti t t h ti i t i t f
lnˆ( ) ( )1 N Np NMDL p VN
= +θ
ˆ( 1)θˆ( )θbull Update the estimates at each time point ie get frombull Recursive least‐squares
( 1)t minusθ( )tθ
ˆ ˆ( ) ( 1) ( ) ( )t t t tε= minus +θ θ K( 1) ( ) ( ) ( 1)( ) ( 1)1 ( ) ( 1) ( )
T
T
t t t tt tt t t
minus minus= minus minus
+ minusP φ φ PP P
φ P φˆ( ) ( ) ( ) ( 1)Tt y t t tε = minus minusφ θ
1 ( ) ( 1) ( )t t t+φ P φ( 1) ( )( ) ( ) ( )
1 ( ) ( 1) ( )T
t tt t tt t tminus
= =+ minus
P φK P φφ P φ
bull Time‐varying systemsbull Modified cost function 2
1( ) ( )
tt s
ts
V sλ εminus
=
=sumθ
λ forgetting factor (between 0 and 1) Smaller λ faster adaptation but more sensitivity to noise
ˆ ˆ( ) ( 1) ( ) ( )t t t tε= minus +θ θ K( 1) ( ) ( ) ( 1)1( ) ( 1)
( ) ( 1) ( )
T
T
t t t tt tt t tλ λ
⎡ ⎤minus minus= minus minus⎢ ⎥+ minus⎣ ⎦
P φ φ PP Pφ P φ
ˆ( ) ( ) ( ) ( 1)Tt y t t tε = minus minusφ θ( ) ( ) ( )⎣ ⎦φ φ
( 1) ( )( ) ( ) ( )( ) ( 1) ( )T
t tt t tt t tλminus
= =+ minus
P φK P φφ P φ
ˆ (0) 0=θ(0) 0(0) ρ
==
θP Ι
Closed‐loop systemsbull Many true systems (industrial physiological) operate with feedback
bull Often the open‐loop system may be unstable so experiments are performed in closed‐loop fi ti ( i i d t i l t )configurations (eg in industrial systems)
bull We have already seen (nonparametric identification ndash HW 2) that the presence of feedback may affect our estimates considerablybull Eg the estimation of G from measurements of uy is significantly affected by the fact that the noise d i t i l l t dand input signals are correlated
(t) f l t i t i t i th l tHs(q-1)
e(t)1 1 2 2
1 1
( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( )s sy t G q u t H q e t E e t
u t F q y t L q v t
λminus minus
minus minus
= + =
= minus +
v(t) reference valuesetpoint or noise entering the regulatorbull Goal Identification of Gs(q-1) Hs(q-1) bull Cases
bull v(t) observable non‐observableF( 1) k k
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
bull F(q-1) knownunknownbull Assumptions
bull Gs(0)=0 (no instantaneous effects of u on y)bull The subsystems between ve and y are stable ie LHs and[1+GF] 1 t bl
F(q-1)
-
[1+GF]‐1 are stablebull v(t) is stationary and persistently exciting of sufficient orderbull v(t) and e(s) are independent for any ts
Closed‐loop systemsbull The general relations that describe this system are
1( ) ( ( ) ( ))s sy t G Lv t H e t= +( ) ( ( ) ( ))1
( ) 1 ( ) ( )1 1
s ss
ss
s s
yFG
FG Fu t Lv t H e tFG FG
+
⎡ ⎤= minus minus⎢ ⎥+ +⎣ ⎦
Hs(q-1)
e(t)
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
F(q-1)
-
Closed‐loop systemsbull Often control systems applications aim to minimize the difference between a reference signal v(t) and y(t) which implies small u(t) values ndash not ideal for identification purposesbull Spectral analysis
W h b f (HW2) th t th t i ti t f G ( 1) i th f d ibull We have seen before (HW2) that the nonparametric estimate of Gs(q‐1) in the frequency domain (assuming L=1) ie
i ld bi d ti t I th l th bi i
ˆ ( ) ( ) ( ) ( ) ( )ˆ ( ) ˆ ( ) ( )( )yu s vv zz
vv zzuu
G FGω ω ω ω ωω
ω ωωΦ Φ minusΦ
= =Φ +ΦΦ
yields biased estimates In the general case the bias is
where
1 ( ) ( ) ( )ˆ ( ) ( )( ) ( ) ( )
s zzs
vv zz
G FG GFω ω ωω ωω ω ω
+ Φminus = minus
Φ +ΦHs(q-1)
e(t)
For e(t)=0 no bias in the estimateFor v(t)=0 we have
1 1( ) ( ) ( ) ( )sz t F q H q e tminus minus=
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
1ˆ ( )( )
GF
ωω
minusF(q-1)
-
Closed‐loop systemsbull Modified spectral analysis
bull Assuming that the external input signal ν(t) is measurable ( ) ( ) ( ) ( ) ( )vy s vu s veG Hω ω ω ω ωΦ = Φ + Φ =
Therefore we can obtain the following (unbiased) estimate
Ti d i t l h f id tifi ti f l d l t
( ) ( )s vuG ω ω= Φ
ˆ ( )ˆ ( ) ˆ ( )vy
vu
Gω
ωω
Φ=Φ
bull Time domain ndash two general approaches for identification of closed‐loop systems arebull Direct identification We treat the system as an open‐loop systembull Indirect identification We assume that the external input reference signal v(t) is measurable and that the feedback law F is known
Hs(q-1)
e(t)
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
F(q-1)
Closed‐loop systemsbull Direct identification We use the measurements of u(t) and y(t) and the general model structure we have used before
1 1 2 2ˆ ˆˆ( ) ( ) ( ) ( ) ( ) ( )y t G q u t H q e t E e t λminus minus= + =θ θ
bull The goal is to fine the parameter vector that minimizes the prediction error eg
[ ]22
1 1
1 1 ˆ( ) ( ) ( ) ( | )N N
Nt t
V t y t y tN N
ε= =
= = minussum sumθ θ θ 1 1 1ˆˆ( ) ( )[ ( ) ( ) ( )]t H q y t G q u tε minus minus minus= minusθ
bull The question is whether for this parametrization the closed loop system is identifiable ie whether there exists θ such that
ˆ ˆs sG G H H= =
bull This depends on the parametrization and the true system form (eg feedback law)
bull Example Let the system
Model
2 2( ) ( 1) ( 1) ( ) ( )( ) ( )y t ay t bu t e t E e tu t fy t
λ+ minus = minus + == minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +
( ) ( ) ( 1) ( )y t a fb y t e t+ + minus =
If we use eg least squares we can estimate only the linear combination a+fb
Closed‐loop systemsbull Assume we have two different regulators f1 and f2 which are active for fractions γ1 and γ2 of the total time ie
bull As before the system is2 2( ) ( 1) ( 1) ( ) ( )
( ) ( )y t ay t bu t e t E e t
fλ+ minus = minus + =
and the model( ) ( )iu t f y t= minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +( ) ( ) ( 1) ( )
ˆˆ( ) ( ) ( ) ( 1)i iy t a bf y t e t
t y t a bf y tε
+ + minus =
= + + minus
therefore
bull Cost functionN
( ) ( ) ( ) ( 1)i i it y t a bf y tε + +2
2 22
ˆˆ( ) ( ) 11 ( )
i ii
i
a bf a bfE ta bf
ε λ⎡ ⎤+ minus minus
= +⎢ ⎥minus +⎢ ⎥⎣ ⎦
2
1
2 22 2 21 1 2 2
1 22 2
1ˆˆlim ( ) ( )
ˆ ˆˆ ˆ( ) ( )1 ( ) 1 ( )
N
N Nt
V a b tN
a bf a bf a bf a bfa bf a bf
ε
λ γ λ γ λ
rarrinfin=
=
+ minus minus + minus minus= + +
minus + minus +
sum θ
bull We can see that ‐minimum
1 21 ( ) 1 ( )a bf a bf+ +
2ˆˆ( ) ( )N NV a b V a b λge =
Closed‐loop systemsbull Global minimum Solution of
bull unique solution for f1nef2
bull Indirect identificationbull Knowledge of v(t) and F(q‐1) L(q-1) is requiredbull Using one of the examined methods (LS PEM) we identify the total system between v ybull Using knowledge of L F we can get estimates of Gs HsUsing knowledge of L F we can get estimates of Gs Hs
Hs(q-1)
e(t)1( ) ( ( ) ( ))
1
( ) 1 ( ) ( )
s ss
s
y t G Lv t H e tFG
FG Fu t Lv t H e t
= ++
⎡ ⎤⎢ ⎥
bullComparing this to the general LTI structure
we can get estimates of
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
( ) 1 ( ) ( )1 1
ss
s s
u t Lv t H e tFG FG
= minus minus⎢ ⎥+ +⎣ ⎦
1 1ˆ ˆ( ) ( ) ( ) ( ) ( )y t G q v t H q e tminus minus= +θ θ1we can get estimates of
and use knowledge of LF to obtain GsHs
F(q-1)
11
11
ss
ss
G G LFG
H HFG
=+
=+
Nonlinear systems
S(τ) y(t)x(t) S(Ax1+Bx2)neAS(x1)+BS(x2)
bull Most real‐life systems are nonlinear linearity is an approximation that may or may not yield satisfactory resultsbull Nonlinear models identification methods
N t i (V lt Wi i )bull Nonparametric (VolterraWiener series)bull Parametric (nonlinear ARX ARMAX models nonlinear differential equations)bull Block diagrams Neural network models
yu z
A b f th l i t ti t th t d l f IO d t
Gyu z
21 2
( ) ( ) ( )( ) ( ) ( )z t g t u ty t c z t c z t
=
= +
bull As before the goal is to estimate the system model from IO databull More complicated problem ndash number of free parameters much higherbull No simple relations exist in the frequency domain as for linear systems (the output of a nonlinear system to a sinusoidal signal is not a sinusoidal signal with the same frequency but hibit h i t l )exhibits harmonic components also)
bull Generally more difficult to obtain theoretical results for convergence consistency etc compared to linear systems
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
Volterra‐Wiener seriesbull Generalization of convolution sum bull Goal estimate Volterra kernels kn from IO measurements
system memory
k 1(m
)k 2
(m1m
2)k
system memory
k
131313m
m1m2 m1 m2
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Estimationbull Least‐squares we can write the above relation as a linear regression problems ‐ very large number of free parameters (order ΜQ) bull Cross‐correlation method Estimate high‐order cross‐correlation functions between input output eg
As we saw for linear systems also long data records are typically needed to obtain accurate estimates for these cross‐correlation functions
1 2 1 2( ) ( ) ( ) ( )xxy E y t x t x tϕ τ τ τ τ= + +
Nonlinear systems⎫⎧
sum sum sum=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Efficient Volterra kernel estimationF ti i R d ti f f tbull Function expansions Reduction of free parameters
bull One of the most widely used basis set is the Laguerre basis( ) 2 1 2
0( ) (1 ) ( 1) (1 ) 0 1
jj m j m m
jm
jb a
m mτ τ
τ α α α αminus minus
=
⎛ ⎞⎛ ⎞= minus minus minus lt lt⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠sum
03
04
05α =01α =02α =04
1 1 1
1
1 10 0
( ) ( ) ( )q
q
L L
q q j j j j qj j
k m m c b m b m= =
= sum sum
V + b
0
01
02y=Vc+ε vj=xbj
cest=[VTV]-1VTy
bull Orthonormal basis in [0infin) ndash well‐
-03
-02
-01Orthonormal basis in [0 infin) wellconditioned estimation of causal systems
bull Exponential decay suitable for finite memory systems
0 5 10 15 20 25 30 35 40 45 50-05
-04
Time lags
y ybull α critical for model accuracy
parsimony
Nonlinear systems
bull Polynomial activation ffunctionsbull Free parameters ~ QMbull Equivalent to Volterra model
M
sum=
minus=j
jii jnxwnv0
)()(miiiiiii zazaazf 110 )( +++=
H
sum=
=H
iijijijmiimm m
wwwacjjjk1
21 )(11
Nonlinear systems
bull Combination of neural networks with functionexpansions (Laguerre‐Volterra networks)
x(n)
expansions (Laguerre Volterra networks)
sum=
=Q
m
mkkmkk nucnuf
1 ))(())((
m 1
bull Drastic reduction of free parameters
vL(n)b0 bj bL
v0(n) vj(n)
hellip hellip
wwK0
w1jw1L
Drastic reduction of free parameters (Q+L+1)∙K+2 bull K number of hidden units f1 fK
hellip
w10wKL
K0 wKj
+
y(n)
y0
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
bull Model order selectionbull Start from simple models and try successively more complex modelsbull Compare model prediction to observations (we shouldnrsquot get zero error) ˆ( ) ( ) ( )my t G q u tΝ= θ
bull Residual testingbull PEM residuals should be white for
bull Whiteness testing Residual autocorrelation function0
ˆN =θ θ 1ˆ ( ) ( ) ( )
N
t tτ
ϕ τ ε ε τminus
= +sum
ˆ ˆˆ( | ) ( ) ( | )N Nt y t y tε = minusθ θ
bull Whiteness testing Residual autocorrelation function
Cross‐correlation function between inputresiduals
Statistical hypothesis testing
1( ) ( ) ( )
tt t
Nεεϕ τ ε ε τ=
= +sum
1
1ˆ ( ) ( ) ( )N
ut
t u tN
τ
εϕ τ ε τminus
=
= minussum2ˆ (0)ϕ λrarrStatistical hypothesis testing (0)
ˆ ( ) 0 0εε
εε
ϕ λϕ τ τ
rarrrarr ne
2 22
1
ˆ ( )ˆ (0)
m
mN
εετεε
ϕ τ χϕ =
rarrsum ˆ ( ) (01)ˆ (0)
N Nεε
εε
ϕ τϕ
rarr
Open loop systems any τ Closed‐loop systems only for τgt0
( ) ( ) ( ) 0u E t u tεϕ τ ε τ= + =
12
ˆ ( )ux ετ
ϕ τ= (0 1)N x Nrarr[ ]12ˆ ˆ(0) ( )uu
xτεεϕ ϕ τ
(01)N x Nτ rarr
bull Cross‐validationbull Sufficient data pointsbull System (almost) time‐invariant
M d l ibull Model comparison
( )
St ti ti l it i
1 2 2 1
2 2
( ) ( ) ( )
MSE MSE p pFMSE N pminus minus
=minus 2 1 2p p N pF minus minus 2 1
2p pχ minus
2ˆ pbull Statistical criteria 2ˆ( ) ( )1 N NpAIC p VN
= +θ
1 ˆ( ) ( )1 N Np NFPE p Vp N
+=
minusθ
bull Recursive identificationU d t th ti t t h ti i t i t f
lnˆ( ) ( )1 N Np NMDL p VN
= +θ
ˆ( 1)θˆ( )θbull Update the estimates at each time point ie get frombull Recursive least‐squares
( 1)t minusθ( )tθ
ˆ ˆ( ) ( 1) ( ) ( )t t t tε= minus +θ θ K( 1) ( ) ( ) ( 1)( ) ( 1)1 ( ) ( 1) ( )
T
T
t t t tt tt t t
minus minus= minus minus
+ minusP φ φ PP P
φ P φˆ( ) ( ) ( ) ( 1)Tt y t t tε = minus minusφ θ
1 ( ) ( 1) ( )t t t+φ P φ( 1) ( )( ) ( ) ( )
1 ( ) ( 1) ( )T
t tt t tt t tminus
= =+ minus
P φK P φφ P φ
bull Time‐varying systemsbull Modified cost function 2
1( ) ( )
tt s
ts
V sλ εminus
=
=sumθ
λ forgetting factor (between 0 and 1) Smaller λ faster adaptation but more sensitivity to noise
ˆ ˆ( ) ( 1) ( ) ( )t t t tε= minus +θ θ K( 1) ( ) ( ) ( 1)1( ) ( 1)
( ) ( 1) ( )
T
T
t t t tt tt t tλ λ
⎡ ⎤minus minus= minus minus⎢ ⎥+ minus⎣ ⎦
P φ φ PP Pφ P φ
ˆ( ) ( ) ( ) ( 1)Tt y t t tε = minus minusφ θ( ) ( ) ( )⎣ ⎦φ φ
( 1) ( )( ) ( ) ( )( ) ( 1) ( )T
t tt t tt t tλminus
= =+ minus
P φK P φφ P φ
ˆ (0) 0=θ(0) 0(0) ρ
==
θP Ι
Closed‐loop systemsbull Many true systems (industrial physiological) operate with feedback
bull Often the open‐loop system may be unstable so experiments are performed in closed‐loop fi ti ( i i d t i l t )configurations (eg in industrial systems)
bull We have already seen (nonparametric identification ndash HW 2) that the presence of feedback may affect our estimates considerablybull Eg the estimation of G from measurements of uy is significantly affected by the fact that the noise d i t i l l t dand input signals are correlated
(t) f l t i t i t i th l tHs(q-1)
e(t)1 1 2 2
1 1
( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( )s sy t G q u t H q e t E e t
u t F q y t L q v t
λminus minus
minus minus
= + =
= minus +
v(t) reference valuesetpoint or noise entering the regulatorbull Goal Identification of Gs(q-1) Hs(q-1) bull Cases
bull v(t) observable non‐observableF( 1) k k
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
bull F(q-1) knownunknownbull Assumptions
bull Gs(0)=0 (no instantaneous effects of u on y)bull The subsystems between ve and y are stable ie LHs and[1+GF] 1 t bl
F(q-1)
-
[1+GF]‐1 are stablebull v(t) is stationary and persistently exciting of sufficient orderbull v(t) and e(s) are independent for any ts
Closed‐loop systemsbull The general relations that describe this system are
1( ) ( ( ) ( ))s sy t G Lv t H e t= +( ) ( ( ) ( ))1
( ) 1 ( ) ( )1 1
s ss
ss
s s
yFG
FG Fu t Lv t H e tFG FG
+
⎡ ⎤= minus minus⎢ ⎥+ +⎣ ⎦
Hs(q-1)
e(t)
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
F(q-1)
-
Closed‐loop systemsbull Often control systems applications aim to minimize the difference between a reference signal v(t) and y(t) which implies small u(t) values ndash not ideal for identification purposesbull Spectral analysis
W h b f (HW2) th t th t i ti t f G ( 1) i th f d ibull We have seen before (HW2) that the nonparametric estimate of Gs(q‐1) in the frequency domain (assuming L=1) ie
i ld bi d ti t I th l th bi i
ˆ ( ) ( ) ( ) ( ) ( )ˆ ( ) ˆ ( ) ( )( )yu s vv zz
vv zzuu
G FGω ω ω ω ωω
ω ωωΦ Φ minusΦ
= =Φ +ΦΦ
yields biased estimates In the general case the bias is
where
1 ( ) ( ) ( )ˆ ( ) ( )( ) ( ) ( )
s zzs
vv zz
G FG GFω ω ωω ωω ω ω
+ Φminus = minus
Φ +ΦHs(q-1)
e(t)
For e(t)=0 no bias in the estimateFor v(t)=0 we have
1 1( ) ( ) ( ) ( )sz t F q H q e tminus minus=
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
1ˆ ( )( )
GF
ωω
minusF(q-1)
-
Closed‐loop systemsbull Modified spectral analysis
bull Assuming that the external input signal ν(t) is measurable ( ) ( ) ( ) ( ) ( )vy s vu s veG Hω ω ω ω ωΦ = Φ + Φ =
Therefore we can obtain the following (unbiased) estimate
Ti d i t l h f id tifi ti f l d l t
( ) ( )s vuG ω ω= Φ
ˆ ( )ˆ ( ) ˆ ( )vy
vu
Gω
ωω
Φ=Φ
bull Time domain ndash two general approaches for identification of closed‐loop systems arebull Direct identification We treat the system as an open‐loop systembull Indirect identification We assume that the external input reference signal v(t) is measurable and that the feedback law F is known
Hs(q-1)
e(t)
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
F(q-1)
Closed‐loop systemsbull Direct identification We use the measurements of u(t) and y(t) and the general model structure we have used before
1 1 2 2ˆ ˆˆ( ) ( ) ( ) ( ) ( ) ( )y t G q u t H q e t E e t λminus minus= + =θ θ
bull The goal is to fine the parameter vector that minimizes the prediction error eg
[ ]22
1 1
1 1 ˆ( ) ( ) ( ) ( | )N N
Nt t
V t y t y tN N
ε= =
= = minussum sumθ θ θ 1 1 1ˆˆ( ) ( )[ ( ) ( ) ( )]t H q y t G q u tε minus minus minus= minusθ
bull The question is whether for this parametrization the closed loop system is identifiable ie whether there exists θ such that
ˆ ˆs sG G H H= =
bull This depends on the parametrization and the true system form (eg feedback law)
bull Example Let the system
Model
2 2( ) ( 1) ( 1) ( ) ( )( ) ( )y t ay t bu t e t E e tu t fy t
λ+ minus = minus + == minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +
( ) ( ) ( 1) ( )y t a fb y t e t+ + minus =
If we use eg least squares we can estimate only the linear combination a+fb
Closed‐loop systemsbull Assume we have two different regulators f1 and f2 which are active for fractions γ1 and γ2 of the total time ie
bull As before the system is2 2( ) ( 1) ( 1) ( ) ( )
( ) ( )y t ay t bu t e t E e t
fλ+ minus = minus + =
and the model( ) ( )iu t f y t= minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +( ) ( ) ( 1) ( )
ˆˆ( ) ( ) ( ) ( 1)i iy t a bf y t e t
t y t a bf y tε
+ + minus =
= + + minus
therefore
bull Cost functionN
( ) ( ) ( ) ( 1)i i it y t a bf y tε + +2
2 22
ˆˆ( ) ( ) 11 ( )
i ii
i
a bf a bfE ta bf
ε λ⎡ ⎤+ minus minus
= +⎢ ⎥minus +⎢ ⎥⎣ ⎦
2
1
2 22 2 21 1 2 2
1 22 2
1ˆˆlim ( ) ( )
ˆ ˆˆ ˆ( ) ( )1 ( ) 1 ( )
N
N Nt
V a b tN
a bf a bf a bf a bfa bf a bf
ε
λ γ λ γ λ
rarrinfin=
=
+ minus minus + minus minus= + +
minus + minus +
sum θ
bull We can see that ‐minimum
1 21 ( ) 1 ( )a bf a bf+ +
2ˆˆ( ) ( )N NV a b V a b λge =
Closed‐loop systemsbull Global minimum Solution of
bull unique solution for f1nef2
bull Indirect identificationbull Knowledge of v(t) and F(q‐1) L(q-1) is requiredbull Using one of the examined methods (LS PEM) we identify the total system between v ybull Using knowledge of L F we can get estimates of Gs HsUsing knowledge of L F we can get estimates of Gs Hs
Hs(q-1)
e(t)1( ) ( ( ) ( ))
1
( ) 1 ( ) ( )
s ss
s
y t G Lv t H e tFG
FG Fu t Lv t H e t
= ++
⎡ ⎤⎢ ⎥
bullComparing this to the general LTI structure
we can get estimates of
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
( ) 1 ( ) ( )1 1
ss
s s
u t Lv t H e tFG FG
= minus minus⎢ ⎥+ +⎣ ⎦
1 1ˆ ˆ( ) ( ) ( ) ( ) ( )y t G q v t H q e tminus minus= +θ θ1we can get estimates of
and use knowledge of LF to obtain GsHs
F(q-1)
11
11
ss
ss
G G LFG
H HFG
=+
=+
Nonlinear systems
S(τ) y(t)x(t) S(Ax1+Bx2)neAS(x1)+BS(x2)
bull Most real‐life systems are nonlinear linearity is an approximation that may or may not yield satisfactory resultsbull Nonlinear models identification methods
N t i (V lt Wi i )bull Nonparametric (VolterraWiener series)bull Parametric (nonlinear ARX ARMAX models nonlinear differential equations)bull Block diagrams Neural network models
yu z
A b f th l i t ti t th t d l f IO d t
Gyu z
21 2
( ) ( ) ( )( ) ( ) ( )z t g t u ty t c z t c z t
=
= +
bull As before the goal is to estimate the system model from IO databull More complicated problem ndash number of free parameters much higherbull No simple relations exist in the frequency domain as for linear systems (the output of a nonlinear system to a sinusoidal signal is not a sinusoidal signal with the same frequency but hibit h i t l )exhibits harmonic components also)
bull Generally more difficult to obtain theoretical results for convergence consistency etc compared to linear systems
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
Volterra‐Wiener seriesbull Generalization of convolution sum bull Goal estimate Volterra kernels kn from IO measurements
system memory
k 1(m
)k 2
(m1m
2)k
system memory
k
131313m
m1m2 m1 m2
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Estimationbull Least‐squares we can write the above relation as a linear regression problems ‐ very large number of free parameters (order ΜQ) bull Cross‐correlation method Estimate high‐order cross‐correlation functions between input output eg
As we saw for linear systems also long data records are typically needed to obtain accurate estimates for these cross‐correlation functions
1 2 1 2( ) ( ) ( ) ( )xxy E y t x t x tϕ τ τ τ τ= + +
Nonlinear systems⎫⎧
sum sum sum=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Efficient Volterra kernel estimationF ti i R d ti f f tbull Function expansions Reduction of free parameters
bull One of the most widely used basis set is the Laguerre basis( ) 2 1 2
0( ) (1 ) ( 1) (1 ) 0 1
jj m j m m
jm
jb a
m mτ τ
τ α α α αminus minus
=
⎛ ⎞⎛ ⎞= minus minus minus lt lt⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠sum
03
04
05α =01α =02α =04
1 1 1
1
1 10 0
( ) ( ) ( )q
q
L L
q q j j j j qj j
k m m c b m b m= =
= sum sum
V + b
0
01
02y=Vc+ε vj=xbj
cest=[VTV]-1VTy
bull Orthonormal basis in [0infin) ndash well‐
-03
-02
-01Orthonormal basis in [0 infin) wellconditioned estimation of causal systems
bull Exponential decay suitable for finite memory systems
0 5 10 15 20 25 30 35 40 45 50-05
-04
Time lags
y ybull α critical for model accuracy
parsimony
Nonlinear systems
bull Polynomial activation ffunctionsbull Free parameters ~ QMbull Equivalent to Volterra model
M
sum=
minus=j
jii jnxwnv0
)()(miiiiiii zazaazf 110 )( +++=
H
sum=
=H
iijijijmiimm m
wwwacjjjk1
21 )(11
Nonlinear systems
bull Combination of neural networks with functionexpansions (Laguerre‐Volterra networks)
x(n)
expansions (Laguerre Volterra networks)
sum=
=Q
m
mkkmkk nucnuf
1 ))(())((
m 1
bull Drastic reduction of free parameters
vL(n)b0 bj bL
v0(n) vj(n)
hellip hellip
wwK0
w1jw1L
Drastic reduction of free parameters (Q+L+1)∙K+2 bull K number of hidden units f1 fK
hellip
w10wKL
K0 wKj
+
y(n)
y0
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
bull Cross‐validationbull Sufficient data pointsbull System (almost) time‐invariant
M d l ibull Model comparison
( )
St ti ti l it i
1 2 2 1
2 2
( ) ( ) ( )
MSE MSE p pFMSE N pminus minus
=minus 2 1 2p p N pF minus minus 2 1
2p pχ minus
2ˆ pbull Statistical criteria 2ˆ( ) ( )1 N NpAIC p VN
= +θ
1 ˆ( ) ( )1 N Np NFPE p Vp N
+=
minusθ
bull Recursive identificationU d t th ti t t h ti i t i t f
lnˆ( ) ( )1 N Np NMDL p VN
= +θ
ˆ( 1)θˆ( )θbull Update the estimates at each time point ie get frombull Recursive least‐squares
( 1)t minusθ( )tθ
ˆ ˆ( ) ( 1) ( ) ( )t t t tε= minus +θ θ K( 1) ( ) ( ) ( 1)( ) ( 1)1 ( ) ( 1) ( )
T
T
t t t tt tt t t
minus minus= minus minus
+ minusP φ φ PP P
φ P φˆ( ) ( ) ( ) ( 1)Tt y t t tε = minus minusφ θ
1 ( ) ( 1) ( )t t t+φ P φ( 1) ( )( ) ( ) ( )
1 ( ) ( 1) ( )T
t tt t tt t tminus
= =+ minus
P φK P φφ P φ
bull Time‐varying systemsbull Modified cost function 2
1( ) ( )
tt s
ts
V sλ εminus
=
=sumθ
λ forgetting factor (between 0 and 1) Smaller λ faster adaptation but more sensitivity to noise
ˆ ˆ( ) ( 1) ( ) ( )t t t tε= minus +θ θ K( 1) ( ) ( ) ( 1)1( ) ( 1)
( ) ( 1) ( )
T
T
t t t tt tt t tλ λ
⎡ ⎤minus minus= minus minus⎢ ⎥+ minus⎣ ⎦
P φ φ PP Pφ P φ
ˆ( ) ( ) ( ) ( 1)Tt y t t tε = minus minusφ θ( ) ( ) ( )⎣ ⎦φ φ
( 1) ( )( ) ( ) ( )( ) ( 1) ( )T
t tt t tt t tλminus
= =+ minus
P φK P φφ P φ
ˆ (0) 0=θ(0) 0(0) ρ
==
θP Ι
Closed‐loop systemsbull Many true systems (industrial physiological) operate with feedback
bull Often the open‐loop system may be unstable so experiments are performed in closed‐loop fi ti ( i i d t i l t )configurations (eg in industrial systems)
bull We have already seen (nonparametric identification ndash HW 2) that the presence of feedback may affect our estimates considerablybull Eg the estimation of G from measurements of uy is significantly affected by the fact that the noise d i t i l l t dand input signals are correlated
(t) f l t i t i t i th l tHs(q-1)
e(t)1 1 2 2
1 1
( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( )s sy t G q u t H q e t E e t
u t F q y t L q v t
λminus minus
minus minus
= + =
= minus +
v(t) reference valuesetpoint or noise entering the regulatorbull Goal Identification of Gs(q-1) Hs(q-1) bull Cases
bull v(t) observable non‐observableF( 1) k k
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
bull F(q-1) knownunknownbull Assumptions
bull Gs(0)=0 (no instantaneous effects of u on y)bull The subsystems between ve and y are stable ie LHs and[1+GF] 1 t bl
F(q-1)
-
[1+GF]‐1 are stablebull v(t) is stationary and persistently exciting of sufficient orderbull v(t) and e(s) are independent for any ts
Closed‐loop systemsbull The general relations that describe this system are
1( ) ( ( ) ( ))s sy t G Lv t H e t= +( ) ( ( ) ( ))1
( ) 1 ( ) ( )1 1
s ss
ss
s s
yFG
FG Fu t Lv t H e tFG FG
+
⎡ ⎤= minus minus⎢ ⎥+ +⎣ ⎦
Hs(q-1)
e(t)
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
F(q-1)
-
Closed‐loop systemsbull Often control systems applications aim to minimize the difference between a reference signal v(t) and y(t) which implies small u(t) values ndash not ideal for identification purposesbull Spectral analysis
W h b f (HW2) th t th t i ti t f G ( 1) i th f d ibull We have seen before (HW2) that the nonparametric estimate of Gs(q‐1) in the frequency domain (assuming L=1) ie
i ld bi d ti t I th l th bi i
ˆ ( ) ( ) ( ) ( ) ( )ˆ ( ) ˆ ( ) ( )( )yu s vv zz
vv zzuu
G FGω ω ω ω ωω
ω ωωΦ Φ minusΦ
= =Φ +ΦΦ
yields biased estimates In the general case the bias is
where
1 ( ) ( ) ( )ˆ ( ) ( )( ) ( ) ( )
s zzs
vv zz
G FG GFω ω ωω ωω ω ω
+ Φminus = minus
Φ +ΦHs(q-1)
e(t)
For e(t)=0 no bias in the estimateFor v(t)=0 we have
1 1( ) ( ) ( ) ( )sz t F q H q e tminus minus=
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
1ˆ ( )( )
GF
ωω
minusF(q-1)
-
Closed‐loop systemsbull Modified spectral analysis
bull Assuming that the external input signal ν(t) is measurable ( ) ( ) ( ) ( ) ( )vy s vu s veG Hω ω ω ω ωΦ = Φ + Φ =
Therefore we can obtain the following (unbiased) estimate
Ti d i t l h f id tifi ti f l d l t
( ) ( )s vuG ω ω= Φ
ˆ ( )ˆ ( ) ˆ ( )vy
vu
Gω
ωω
Φ=Φ
bull Time domain ndash two general approaches for identification of closed‐loop systems arebull Direct identification We treat the system as an open‐loop systembull Indirect identification We assume that the external input reference signal v(t) is measurable and that the feedback law F is known
Hs(q-1)
e(t)
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
F(q-1)
Closed‐loop systemsbull Direct identification We use the measurements of u(t) and y(t) and the general model structure we have used before
1 1 2 2ˆ ˆˆ( ) ( ) ( ) ( ) ( ) ( )y t G q u t H q e t E e t λminus minus= + =θ θ
bull The goal is to fine the parameter vector that minimizes the prediction error eg
[ ]22
1 1
1 1 ˆ( ) ( ) ( ) ( | )N N
Nt t
V t y t y tN N
ε= =
= = minussum sumθ θ θ 1 1 1ˆˆ( ) ( )[ ( ) ( ) ( )]t H q y t G q u tε minus minus minus= minusθ
bull The question is whether for this parametrization the closed loop system is identifiable ie whether there exists θ such that
ˆ ˆs sG G H H= =
bull This depends on the parametrization and the true system form (eg feedback law)
bull Example Let the system
Model
2 2( ) ( 1) ( 1) ( ) ( )( ) ( )y t ay t bu t e t E e tu t fy t
λ+ minus = minus + == minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +
( ) ( ) ( 1) ( )y t a fb y t e t+ + minus =
If we use eg least squares we can estimate only the linear combination a+fb
Closed‐loop systemsbull Assume we have two different regulators f1 and f2 which are active for fractions γ1 and γ2 of the total time ie
bull As before the system is2 2( ) ( 1) ( 1) ( ) ( )
( ) ( )y t ay t bu t e t E e t
fλ+ minus = minus + =
and the model( ) ( )iu t f y t= minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +( ) ( ) ( 1) ( )
ˆˆ( ) ( ) ( ) ( 1)i iy t a bf y t e t
t y t a bf y tε
+ + minus =
= + + minus
therefore
bull Cost functionN
( ) ( ) ( ) ( 1)i i it y t a bf y tε + +2
2 22
ˆˆ( ) ( ) 11 ( )
i ii
i
a bf a bfE ta bf
ε λ⎡ ⎤+ minus minus
= +⎢ ⎥minus +⎢ ⎥⎣ ⎦
2
1
2 22 2 21 1 2 2
1 22 2
1ˆˆlim ( ) ( )
ˆ ˆˆ ˆ( ) ( )1 ( ) 1 ( )
N
N Nt
V a b tN
a bf a bf a bf a bfa bf a bf
ε
λ γ λ γ λ
rarrinfin=
=
+ minus minus + minus minus= + +
minus + minus +
sum θ
bull We can see that ‐minimum
1 21 ( ) 1 ( )a bf a bf+ +
2ˆˆ( ) ( )N NV a b V a b λge =
Closed‐loop systemsbull Global minimum Solution of
bull unique solution for f1nef2
bull Indirect identificationbull Knowledge of v(t) and F(q‐1) L(q-1) is requiredbull Using one of the examined methods (LS PEM) we identify the total system between v ybull Using knowledge of L F we can get estimates of Gs HsUsing knowledge of L F we can get estimates of Gs Hs
Hs(q-1)
e(t)1( ) ( ( ) ( ))
1
( ) 1 ( ) ( )
s ss
s
y t G Lv t H e tFG
FG Fu t Lv t H e t
= ++
⎡ ⎤⎢ ⎥
bullComparing this to the general LTI structure
we can get estimates of
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
( ) 1 ( ) ( )1 1
ss
s s
u t Lv t H e tFG FG
= minus minus⎢ ⎥+ +⎣ ⎦
1 1ˆ ˆ( ) ( ) ( ) ( ) ( )y t G q v t H q e tminus minus= +θ θ1we can get estimates of
and use knowledge of LF to obtain GsHs
F(q-1)
11
11
ss
ss
G G LFG
H HFG
=+
=+
Nonlinear systems
S(τ) y(t)x(t) S(Ax1+Bx2)neAS(x1)+BS(x2)
bull Most real‐life systems are nonlinear linearity is an approximation that may or may not yield satisfactory resultsbull Nonlinear models identification methods
N t i (V lt Wi i )bull Nonparametric (VolterraWiener series)bull Parametric (nonlinear ARX ARMAX models nonlinear differential equations)bull Block diagrams Neural network models
yu z
A b f th l i t ti t th t d l f IO d t
Gyu z
21 2
( ) ( ) ( )( ) ( ) ( )z t g t u ty t c z t c z t
=
= +
bull As before the goal is to estimate the system model from IO databull More complicated problem ndash number of free parameters much higherbull No simple relations exist in the frequency domain as for linear systems (the output of a nonlinear system to a sinusoidal signal is not a sinusoidal signal with the same frequency but hibit h i t l )exhibits harmonic components also)
bull Generally more difficult to obtain theoretical results for convergence consistency etc compared to linear systems
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
Volterra‐Wiener seriesbull Generalization of convolution sum bull Goal estimate Volterra kernels kn from IO measurements
system memory
k 1(m
)k 2
(m1m
2)k
system memory
k
131313m
m1m2 m1 m2
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Estimationbull Least‐squares we can write the above relation as a linear regression problems ‐ very large number of free parameters (order ΜQ) bull Cross‐correlation method Estimate high‐order cross‐correlation functions between input output eg
As we saw for linear systems also long data records are typically needed to obtain accurate estimates for these cross‐correlation functions
1 2 1 2( ) ( ) ( ) ( )xxy E y t x t x tϕ τ τ τ τ= + +
Nonlinear systems⎫⎧
sum sum sum=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Efficient Volterra kernel estimationF ti i R d ti f f tbull Function expansions Reduction of free parameters
bull One of the most widely used basis set is the Laguerre basis( ) 2 1 2
0( ) (1 ) ( 1) (1 ) 0 1
jj m j m m
jm
jb a
m mτ τ
τ α α α αminus minus
=
⎛ ⎞⎛ ⎞= minus minus minus lt lt⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠sum
03
04
05α =01α =02α =04
1 1 1
1
1 10 0
( ) ( ) ( )q
q
L L
q q j j j j qj j
k m m c b m b m= =
= sum sum
V + b
0
01
02y=Vc+ε vj=xbj
cest=[VTV]-1VTy
bull Orthonormal basis in [0infin) ndash well‐
-03
-02
-01Orthonormal basis in [0 infin) wellconditioned estimation of causal systems
bull Exponential decay suitable for finite memory systems
0 5 10 15 20 25 30 35 40 45 50-05
-04
Time lags
y ybull α critical for model accuracy
parsimony
Nonlinear systems
bull Polynomial activation ffunctionsbull Free parameters ~ QMbull Equivalent to Volterra model
M
sum=
minus=j
jii jnxwnv0
)()(miiiiiii zazaazf 110 )( +++=
H
sum=
=H
iijijijmiimm m
wwwacjjjk1
21 )(11
Nonlinear systems
bull Combination of neural networks with functionexpansions (Laguerre‐Volterra networks)
x(n)
expansions (Laguerre Volterra networks)
sum=
=Q
m
mkkmkk nucnuf
1 ))(())((
m 1
bull Drastic reduction of free parameters
vL(n)b0 bj bL
v0(n) vj(n)
hellip hellip
wwK0
w1jw1L
Drastic reduction of free parameters (Q+L+1)∙K+2 bull K number of hidden units f1 fK
hellip
w10wKL
K0 wKj
+
y(n)
y0
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
bull Time‐varying systemsbull Modified cost function 2
1( ) ( )
tt s
ts
V sλ εminus
=
=sumθ
λ forgetting factor (between 0 and 1) Smaller λ faster adaptation but more sensitivity to noise
ˆ ˆ( ) ( 1) ( ) ( )t t t tε= minus +θ θ K( 1) ( ) ( ) ( 1)1( ) ( 1)
( ) ( 1) ( )
T
T
t t t tt tt t tλ λ
⎡ ⎤minus minus= minus minus⎢ ⎥+ minus⎣ ⎦
P φ φ PP Pφ P φ
ˆ( ) ( ) ( ) ( 1)Tt y t t tε = minus minusφ θ( ) ( ) ( )⎣ ⎦φ φ
( 1) ( )( ) ( ) ( )( ) ( 1) ( )T
t tt t tt t tλminus
= =+ minus
P φK P φφ P φ
ˆ (0) 0=θ(0) 0(0) ρ
==
θP Ι
Closed‐loop systemsbull Many true systems (industrial physiological) operate with feedback
bull Often the open‐loop system may be unstable so experiments are performed in closed‐loop fi ti ( i i d t i l t )configurations (eg in industrial systems)
bull We have already seen (nonparametric identification ndash HW 2) that the presence of feedback may affect our estimates considerablybull Eg the estimation of G from measurements of uy is significantly affected by the fact that the noise d i t i l l t dand input signals are correlated
(t) f l t i t i t i th l tHs(q-1)
e(t)1 1 2 2
1 1
( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( )s sy t G q u t H q e t E e t
u t F q y t L q v t
λminus minus
minus minus
= + =
= minus +
v(t) reference valuesetpoint or noise entering the regulatorbull Goal Identification of Gs(q-1) Hs(q-1) bull Cases
bull v(t) observable non‐observableF( 1) k k
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
bull F(q-1) knownunknownbull Assumptions
bull Gs(0)=0 (no instantaneous effects of u on y)bull The subsystems between ve and y are stable ie LHs and[1+GF] 1 t bl
F(q-1)
-
[1+GF]‐1 are stablebull v(t) is stationary and persistently exciting of sufficient orderbull v(t) and e(s) are independent for any ts
Closed‐loop systemsbull The general relations that describe this system are
1( ) ( ( ) ( ))s sy t G Lv t H e t= +( ) ( ( ) ( ))1
( ) 1 ( ) ( )1 1
s ss
ss
s s
yFG
FG Fu t Lv t H e tFG FG
+
⎡ ⎤= minus minus⎢ ⎥+ +⎣ ⎦
Hs(q-1)
e(t)
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
F(q-1)
-
Closed‐loop systemsbull Often control systems applications aim to minimize the difference between a reference signal v(t) and y(t) which implies small u(t) values ndash not ideal for identification purposesbull Spectral analysis
W h b f (HW2) th t th t i ti t f G ( 1) i th f d ibull We have seen before (HW2) that the nonparametric estimate of Gs(q‐1) in the frequency domain (assuming L=1) ie
i ld bi d ti t I th l th bi i
ˆ ( ) ( ) ( ) ( ) ( )ˆ ( ) ˆ ( ) ( )( )yu s vv zz
vv zzuu
G FGω ω ω ω ωω
ω ωωΦ Φ minusΦ
= =Φ +ΦΦ
yields biased estimates In the general case the bias is
where
1 ( ) ( ) ( )ˆ ( ) ( )( ) ( ) ( )
s zzs
vv zz
G FG GFω ω ωω ωω ω ω
+ Φminus = minus
Φ +ΦHs(q-1)
e(t)
For e(t)=0 no bias in the estimateFor v(t)=0 we have
1 1( ) ( ) ( ) ( )sz t F q H q e tminus minus=
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
1ˆ ( )( )
GF
ωω
minusF(q-1)
-
Closed‐loop systemsbull Modified spectral analysis
bull Assuming that the external input signal ν(t) is measurable ( ) ( ) ( ) ( ) ( )vy s vu s veG Hω ω ω ω ωΦ = Φ + Φ =
Therefore we can obtain the following (unbiased) estimate
Ti d i t l h f id tifi ti f l d l t
( ) ( )s vuG ω ω= Φ
ˆ ( )ˆ ( ) ˆ ( )vy
vu
Gω
ωω
Φ=Φ
bull Time domain ndash two general approaches for identification of closed‐loop systems arebull Direct identification We treat the system as an open‐loop systembull Indirect identification We assume that the external input reference signal v(t) is measurable and that the feedback law F is known
Hs(q-1)
e(t)
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
F(q-1)
Closed‐loop systemsbull Direct identification We use the measurements of u(t) and y(t) and the general model structure we have used before
1 1 2 2ˆ ˆˆ( ) ( ) ( ) ( ) ( ) ( )y t G q u t H q e t E e t λminus minus= + =θ θ
bull The goal is to fine the parameter vector that minimizes the prediction error eg
[ ]22
1 1
1 1 ˆ( ) ( ) ( ) ( | )N N
Nt t
V t y t y tN N
ε= =
= = minussum sumθ θ θ 1 1 1ˆˆ( ) ( )[ ( ) ( ) ( )]t H q y t G q u tε minus minus minus= minusθ
bull The question is whether for this parametrization the closed loop system is identifiable ie whether there exists θ such that
ˆ ˆs sG G H H= =
bull This depends on the parametrization and the true system form (eg feedback law)
bull Example Let the system
Model
2 2( ) ( 1) ( 1) ( ) ( )( ) ( )y t ay t bu t e t E e tu t fy t
λ+ minus = minus + == minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +
( ) ( ) ( 1) ( )y t a fb y t e t+ + minus =
If we use eg least squares we can estimate only the linear combination a+fb
Closed‐loop systemsbull Assume we have two different regulators f1 and f2 which are active for fractions γ1 and γ2 of the total time ie
bull As before the system is2 2( ) ( 1) ( 1) ( ) ( )
( ) ( )y t ay t bu t e t E e t
fλ+ minus = minus + =
and the model( ) ( )iu t f y t= minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +( ) ( ) ( 1) ( )
ˆˆ( ) ( ) ( ) ( 1)i iy t a bf y t e t
t y t a bf y tε
+ + minus =
= + + minus
therefore
bull Cost functionN
( ) ( ) ( ) ( 1)i i it y t a bf y tε + +2
2 22
ˆˆ( ) ( ) 11 ( )
i ii
i
a bf a bfE ta bf
ε λ⎡ ⎤+ minus minus
= +⎢ ⎥minus +⎢ ⎥⎣ ⎦
2
1
2 22 2 21 1 2 2
1 22 2
1ˆˆlim ( ) ( )
ˆ ˆˆ ˆ( ) ( )1 ( ) 1 ( )
N
N Nt
V a b tN
a bf a bf a bf a bfa bf a bf
ε
λ γ λ γ λ
rarrinfin=
=
+ minus minus + minus minus= + +
minus + minus +
sum θ
bull We can see that ‐minimum
1 21 ( ) 1 ( )a bf a bf+ +
2ˆˆ( ) ( )N NV a b V a b λge =
Closed‐loop systemsbull Global minimum Solution of
bull unique solution for f1nef2
bull Indirect identificationbull Knowledge of v(t) and F(q‐1) L(q-1) is requiredbull Using one of the examined methods (LS PEM) we identify the total system between v ybull Using knowledge of L F we can get estimates of Gs HsUsing knowledge of L F we can get estimates of Gs Hs
Hs(q-1)
e(t)1( ) ( ( ) ( ))
1
( ) 1 ( ) ( )
s ss
s
y t G Lv t H e tFG
FG Fu t Lv t H e t
= ++
⎡ ⎤⎢ ⎥
bullComparing this to the general LTI structure
we can get estimates of
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
( ) 1 ( ) ( )1 1
ss
s s
u t Lv t H e tFG FG
= minus minus⎢ ⎥+ +⎣ ⎦
1 1ˆ ˆ( ) ( ) ( ) ( ) ( )y t G q v t H q e tminus minus= +θ θ1we can get estimates of
and use knowledge of LF to obtain GsHs
F(q-1)
11
11
ss
ss
G G LFG
H HFG
=+
=+
Nonlinear systems
S(τ) y(t)x(t) S(Ax1+Bx2)neAS(x1)+BS(x2)
bull Most real‐life systems are nonlinear linearity is an approximation that may or may not yield satisfactory resultsbull Nonlinear models identification methods
N t i (V lt Wi i )bull Nonparametric (VolterraWiener series)bull Parametric (nonlinear ARX ARMAX models nonlinear differential equations)bull Block diagrams Neural network models
yu z
A b f th l i t ti t th t d l f IO d t
Gyu z
21 2
( ) ( ) ( )( ) ( ) ( )z t g t u ty t c z t c z t
=
= +
bull As before the goal is to estimate the system model from IO databull More complicated problem ndash number of free parameters much higherbull No simple relations exist in the frequency domain as for linear systems (the output of a nonlinear system to a sinusoidal signal is not a sinusoidal signal with the same frequency but hibit h i t l )exhibits harmonic components also)
bull Generally more difficult to obtain theoretical results for convergence consistency etc compared to linear systems
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
Volterra‐Wiener seriesbull Generalization of convolution sum bull Goal estimate Volterra kernels kn from IO measurements
system memory
k 1(m
)k 2
(m1m
2)k
system memory
k
131313m
m1m2 m1 m2
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Estimationbull Least‐squares we can write the above relation as a linear regression problems ‐ very large number of free parameters (order ΜQ) bull Cross‐correlation method Estimate high‐order cross‐correlation functions between input output eg
As we saw for linear systems also long data records are typically needed to obtain accurate estimates for these cross‐correlation functions
1 2 1 2( ) ( ) ( ) ( )xxy E y t x t x tϕ τ τ τ τ= + +
Nonlinear systems⎫⎧
sum sum sum=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Efficient Volterra kernel estimationF ti i R d ti f f tbull Function expansions Reduction of free parameters
bull One of the most widely used basis set is the Laguerre basis( ) 2 1 2
0( ) (1 ) ( 1) (1 ) 0 1
jj m j m m
jm
jb a
m mτ τ
τ α α α αminus minus
=
⎛ ⎞⎛ ⎞= minus minus minus lt lt⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠sum
03
04
05α =01α =02α =04
1 1 1
1
1 10 0
( ) ( ) ( )q
q
L L
q q j j j j qj j
k m m c b m b m= =
= sum sum
V + b
0
01
02y=Vc+ε vj=xbj
cest=[VTV]-1VTy
bull Orthonormal basis in [0infin) ndash well‐
-03
-02
-01Orthonormal basis in [0 infin) wellconditioned estimation of causal systems
bull Exponential decay suitable for finite memory systems
0 5 10 15 20 25 30 35 40 45 50-05
-04
Time lags
y ybull α critical for model accuracy
parsimony
Nonlinear systems
bull Polynomial activation ffunctionsbull Free parameters ~ QMbull Equivalent to Volterra model
M
sum=
minus=j
jii jnxwnv0
)()(miiiiiii zazaazf 110 )( +++=
H
sum=
=H
iijijijmiimm m
wwwacjjjk1
21 )(11
Nonlinear systems
bull Combination of neural networks with functionexpansions (Laguerre‐Volterra networks)
x(n)
expansions (Laguerre Volterra networks)
sum=
=Q
m
mkkmkk nucnuf
1 ))(())((
m 1
bull Drastic reduction of free parameters
vL(n)b0 bj bL
v0(n) vj(n)
hellip hellip
wwK0
w1jw1L
Drastic reduction of free parameters (Q+L+1)∙K+2 bull K number of hidden units f1 fK
hellip
w10wKL
K0 wKj
+
y(n)
y0
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Closed‐loop systemsbull Many true systems (industrial physiological) operate with feedback
bull Often the open‐loop system may be unstable so experiments are performed in closed‐loop fi ti ( i i d t i l t )configurations (eg in industrial systems)
bull We have already seen (nonparametric identification ndash HW 2) that the presence of feedback may affect our estimates considerablybull Eg the estimation of G from measurements of uy is significantly affected by the fact that the noise d i t i l l t dand input signals are correlated
(t) f l t i t i t i th l tHs(q-1)
e(t)1 1 2 2
1 1
( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( )s sy t G q u t H q e t E e t
u t F q y t L q v t
λminus minus
minus minus
= + =
= minus +
v(t) reference valuesetpoint or noise entering the regulatorbull Goal Identification of Gs(q-1) Hs(q-1) bull Cases
bull v(t) observable non‐observableF( 1) k k
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
bull F(q-1) knownunknownbull Assumptions
bull Gs(0)=0 (no instantaneous effects of u on y)bull The subsystems between ve and y are stable ie LHs and[1+GF] 1 t bl
F(q-1)
-
[1+GF]‐1 are stablebull v(t) is stationary and persistently exciting of sufficient orderbull v(t) and e(s) are independent for any ts
Closed‐loop systemsbull The general relations that describe this system are
1( ) ( ( ) ( ))s sy t G Lv t H e t= +( ) ( ( ) ( ))1
( ) 1 ( ) ( )1 1
s ss
ss
s s
yFG
FG Fu t Lv t H e tFG FG
+
⎡ ⎤= minus minus⎢ ⎥+ +⎣ ⎦
Hs(q-1)
e(t)
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
F(q-1)
-
Closed‐loop systemsbull Often control systems applications aim to minimize the difference between a reference signal v(t) and y(t) which implies small u(t) values ndash not ideal for identification purposesbull Spectral analysis
W h b f (HW2) th t th t i ti t f G ( 1) i th f d ibull We have seen before (HW2) that the nonparametric estimate of Gs(q‐1) in the frequency domain (assuming L=1) ie
i ld bi d ti t I th l th bi i
ˆ ( ) ( ) ( ) ( ) ( )ˆ ( ) ˆ ( ) ( )( )yu s vv zz
vv zzuu
G FGω ω ω ω ωω
ω ωωΦ Φ minusΦ
= =Φ +ΦΦ
yields biased estimates In the general case the bias is
where
1 ( ) ( ) ( )ˆ ( ) ( )( ) ( ) ( )
s zzs
vv zz
G FG GFω ω ωω ωω ω ω
+ Φminus = minus
Φ +ΦHs(q-1)
e(t)
For e(t)=0 no bias in the estimateFor v(t)=0 we have
1 1( ) ( ) ( ) ( )sz t F q H q e tminus minus=
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
1ˆ ( )( )
GF
ωω
minusF(q-1)
-
Closed‐loop systemsbull Modified spectral analysis
bull Assuming that the external input signal ν(t) is measurable ( ) ( ) ( ) ( ) ( )vy s vu s veG Hω ω ω ω ωΦ = Φ + Φ =
Therefore we can obtain the following (unbiased) estimate
Ti d i t l h f id tifi ti f l d l t
( ) ( )s vuG ω ω= Φ
ˆ ( )ˆ ( ) ˆ ( )vy
vu
Gω
ωω
Φ=Φ
bull Time domain ndash two general approaches for identification of closed‐loop systems arebull Direct identification We treat the system as an open‐loop systembull Indirect identification We assume that the external input reference signal v(t) is measurable and that the feedback law F is known
Hs(q-1)
e(t)
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
F(q-1)
Closed‐loop systemsbull Direct identification We use the measurements of u(t) and y(t) and the general model structure we have used before
1 1 2 2ˆ ˆˆ( ) ( ) ( ) ( ) ( ) ( )y t G q u t H q e t E e t λminus minus= + =θ θ
bull The goal is to fine the parameter vector that minimizes the prediction error eg
[ ]22
1 1
1 1 ˆ( ) ( ) ( ) ( | )N N
Nt t
V t y t y tN N
ε= =
= = minussum sumθ θ θ 1 1 1ˆˆ( ) ( )[ ( ) ( ) ( )]t H q y t G q u tε minus minus minus= minusθ
bull The question is whether for this parametrization the closed loop system is identifiable ie whether there exists θ such that
ˆ ˆs sG G H H= =
bull This depends on the parametrization and the true system form (eg feedback law)
bull Example Let the system
Model
2 2( ) ( 1) ( 1) ( ) ( )( ) ( )y t ay t bu t e t E e tu t fy t
λ+ minus = minus + == minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +
( ) ( ) ( 1) ( )y t a fb y t e t+ + minus =
If we use eg least squares we can estimate only the linear combination a+fb
Closed‐loop systemsbull Assume we have two different regulators f1 and f2 which are active for fractions γ1 and γ2 of the total time ie
bull As before the system is2 2( ) ( 1) ( 1) ( ) ( )
( ) ( )y t ay t bu t e t E e t
fλ+ minus = minus + =
and the model( ) ( )iu t f y t= minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +( ) ( ) ( 1) ( )
ˆˆ( ) ( ) ( ) ( 1)i iy t a bf y t e t
t y t a bf y tε
+ + minus =
= + + minus
therefore
bull Cost functionN
( ) ( ) ( ) ( 1)i i it y t a bf y tε + +2
2 22
ˆˆ( ) ( ) 11 ( )
i ii
i
a bf a bfE ta bf
ε λ⎡ ⎤+ minus minus
= +⎢ ⎥minus +⎢ ⎥⎣ ⎦
2
1
2 22 2 21 1 2 2
1 22 2
1ˆˆlim ( ) ( )
ˆ ˆˆ ˆ( ) ( )1 ( ) 1 ( )
N
N Nt
V a b tN
a bf a bf a bf a bfa bf a bf
ε
λ γ λ γ λ
rarrinfin=
=
+ minus minus + minus minus= + +
minus + minus +
sum θ
bull We can see that ‐minimum
1 21 ( ) 1 ( )a bf a bf+ +
2ˆˆ( ) ( )N NV a b V a b λge =
Closed‐loop systemsbull Global minimum Solution of
bull unique solution for f1nef2
bull Indirect identificationbull Knowledge of v(t) and F(q‐1) L(q-1) is requiredbull Using one of the examined methods (LS PEM) we identify the total system between v ybull Using knowledge of L F we can get estimates of Gs HsUsing knowledge of L F we can get estimates of Gs Hs
Hs(q-1)
e(t)1( ) ( ( ) ( ))
1
( ) 1 ( ) ( )
s ss
s
y t G Lv t H e tFG
FG Fu t Lv t H e t
= ++
⎡ ⎤⎢ ⎥
bullComparing this to the general LTI structure
we can get estimates of
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
( ) 1 ( ) ( )1 1
ss
s s
u t Lv t H e tFG FG
= minus minus⎢ ⎥+ +⎣ ⎦
1 1ˆ ˆ( ) ( ) ( ) ( ) ( )y t G q v t H q e tminus minus= +θ θ1we can get estimates of
and use knowledge of LF to obtain GsHs
F(q-1)
11
11
ss
ss
G G LFG
H HFG
=+
=+
Nonlinear systems
S(τ) y(t)x(t) S(Ax1+Bx2)neAS(x1)+BS(x2)
bull Most real‐life systems are nonlinear linearity is an approximation that may or may not yield satisfactory resultsbull Nonlinear models identification methods
N t i (V lt Wi i )bull Nonparametric (VolterraWiener series)bull Parametric (nonlinear ARX ARMAX models nonlinear differential equations)bull Block diagrams Neural network models
yu z
A b f th l i t ti t th t d l f IO d t
Gyu z
21 2
( ) ( ) ( )( ) ( ) ( )z t g t u ty t c z t c z t
=
= +
bull As before the goal is to estimate the system model from IO databull More complicated problem ndash number of free parameters much higherbull No simple relations exist in the frequency domain as for linear systems (the output of a nonlinear system to a sinusoidal signal is not a sinusoidal signal with the same frequency but hibit h i t l )exhibits harmonic components also)
bull Generally more difficult to obtain theoretical results for convergence consistency etc compared to linear systems
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
Volterra‐Wiener seriesbull Generalization of convolution sum bull Goal estimate Volterra kernels kn from IO measurements
system memory
k 1(m
)k 2
(m1m
2)k
system memory
k
131313m
m1m2 m1 m2
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Estimationbull Least‐squares we can write the above relation as a linear regression problems ‐ very large number of free parameters (order ΜQ) bull Cross‐correlation method Estimate high‐order cross‐correlation functions between input output eg
As we saw for linear systems also long data records are typically needed to obtain accurate estimates for these cross‐correlation functions
1 2 1 2( ) ( ) ( ) ( )xxy E y t x t x tϕ τ τ τ τ= + +
Nonlinear systems⎫⎧
sum sum sum=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Efficient Volterra kernel estimationF ti i R d ti f f tbull Function expansions Reduction of free parameters
bull One of the most widely used basis set is the Laguerre basis( ) 2 1 2
0( ) (1 ) ( 1) (1 ) 0 1
jj m j m m
jm
jb a
m mτ τ
τ α α α αminus minus
=
⎛ ⎞⎛ ⎞= minus minus minus lt lt⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠sum
03
04
05α =01α =02α =04
1 1 1
1
1 10 0
( ) ( ) ( )q
q
L L
q q j j j j qj j
k m m c b m b m= =
= sum sum
V + b
0
01
02y=Vc+ε vj=xbj
cest=[VTV]-1VTy
bull Orthonormal basis in [0infin) ndash well‐
-03
-02
-01Orthonormal basis in [0 infin) wellconditioned estimation of causal systems
bull Exponential decay suitable for finite memory systems
0 5 10 15 20 25 30 35 40 45 50-05
-04
Time lags
y ybull α critical for model accuracy
parsimony
Nonlinear systems
bull Polynomial activation ffunctionsbull Free parameters ~ QMbull Equivalent to Volterra model
M
sum=
minus=j
jii jnxwnv0
)()(miiiiiii zazaazf 110 )( +++=
H
sum=
=H
iijijijmiimm m
wwwacjjjk1
21 )(11
Nonlinear systems
bull Combination of neural networks with functionexpansions (Laguerre‐Volterra networks)
x(n)
expansions (Laguerre Volterra networks)
sum=
=Q
m
mkkmkk nucnuf
1 ))(())((
m 1
bull Drastic reduction of free parameters
vL(n)b0 bj bL
v0(n) vj(n)
hellip hellip
wwK0
w1jw1L
Drastic reduction of free parameters (Q+L+1)∙K+2 bull K number of hidden units f1 fK
hellip
w10wKL
K0 wKj
+
y(n)
y0
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Closed‐loop systemsbull The general relations that describe this system are
1( ) ( ( ) ( ))s sy t G Lv t H e t= +( ) ( ( ) ( ))1
( ) 1 ( ) ( )1 1
s ss
ss
s s
yFG
FG Fu t Lv t H e tFG FG
+
⎡ ⎤= minus minus⎢ ⎥+ +⎣ ⎦
Hs(q-1)
e(t)
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
F(q-1)
-
Closed‐loop systemsbull Often control systems applications aim to minimize the difference between a reference signal v(t) and y(t) which implies small u(t) values ndash not ideal for identification purposesbull Spectral analysis
W h b f (HW2) th t th t i ti t f G ( 1) i th f d ibull We have seen before (HW2) that the nonparametric estimate of Gs(q‐1) in the frequency domain (assuming L=1) ie
i ld bi d ti t I th l th bi i
ˆ ( ) ( ) ( ) ( ) ( )ˆ ( ) ˆ ( ) ( )( )yu s vv zz
vv zzuu
G FGω ω ω ω ωω
ω ωωΦ Φ minusΦ
= =Φ +ΦΦ
yields biased estimates In the general case the bias is
where
1 ( ) ( ) ( )ˆ ( ) ( )( ) ( ) ( )
s zzs
vv zz
G FG GFω ω ωω ωω ω ω
+ Φminus = minus
Φ +ΦHs(q-1)
e(t)
For e(t)=0 no bias in the estimateFor v(t)=0 we have
1 1( ) ( ) ( ) ( )sz t F q H q e tminus minus=
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
1ˆ ( )( )
GF
ωω
minusF(q-1)
-
Closed‐loop systemsbull Modified spectral analysis
bull Assuming that the external input signal ν(t) is measurable ( ) ( ) ( ) ( ) ( )vy s vu s veG Hω ω ω ω ωΦ = Φ + Φ =
Therefore we can obtain the following (unbiased) estimate
Ti d i t l h f id tifi ti f l d l t
( ) ( )s vuG ω ω= Φ
ˆ ( )ˆ ( ) ˆ ( )vy
vu
Gω
ωω
Φ=Φ
bull Time domain ndash two general approaches for identification of closed‐loop systems arebull Direct identification We treat the system as an open‐loop systembull Indirect identification We assume that the external input reference signal v(t) is measurable and that the feedback law F is known
Hs(q-1)
e(t)
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
F(q-1)
Closed‐loop systemsbull Direct identification We use the measurements of u(t) and y(t) and the general model structure we have used before
1 1 2 2ˆ ˆˆ( ) ( ) ( ) ( ) ( ) ( )y t G q u t H q e t E e t λminus minus= + =θ θ
bull The goal is to fine the parameter vector that minimizes the prediction error eg
[ ]22
1 1
1 1 ˆ( ) ( ) ( ) ( | )N N
Nt t
V t y t y tN N
ε= =
= = minussum sumθ θ θ 1 1 1ˆˆ( ) ( )[ ( ) ( ) ( )]t H q y t G q u tε minus minus minus= minusθ
bull The question is whether for this parametrization the closed loop system is identifiable ie whether there exists θ such that
ˆ ˆs sG G H H= =
bull This depends on the parametrization and the true system form (eg feedback law)
bull Example Let the system
Model
2 2( ) ( 1) ( 1) ( ) ( )( ) ( )y t ay t bu t e t E e tu t fy t
λ+ minus = minus + == minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +
( ) ( ) ( 1) ( )y t a fb y t e t+ + minus =
If we use eg least squares we can estimate only the linear combination a+fb
Closed‐loop systemsbull Assume we have two different regulators f1 and f2 which are active for fractions γ1 and γ2 of the total time ie
bull As before the system is2 2( ) ( 1) ( 1) ( ) ( )
( ) ( )y t ay t bu t e t E e t
fλ+ minus = minus + =
and the model( ) ( )iu t f y t= minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +( ) ( ) ( 1) ( )
ˆˆ( ) ( ) ( ) ( 1)i iy t a bf y t e t
t y t a bf y tε
+ + minus =
= + + minus
therefore
bull Cost functionN
( ) ( ) ( ) ( 1)i i it y t a bf y tε + +2
2 22
ˆˆ( ) ( ) 11 ( )
i ii
i
a bf a bfE ta bf
ε λ⎡ ⎤+ minus minus
= +⎢ ⎥minus +⎢ ⎥⎣ ⎦
2
1
2 22 2 21 1 2 2
1 22 2
1ˆˆlim ( ) ( )
ˆ ˆˆ ˆ( ) ( )1 ( ) 1 ( )
N
N Nt
V a b tN
a bf a bf a bf a bfa bf a bf
ε
λ γ λ γ λ
rarrinfin=
=
+ minus minus + minus minus= + +
minus + minus +
sum θ
bull We can see that ‐minimum
1 21 ( ) 1 ( )a bf a bf+ +
2ˆˆ( ) ( )N NV a b V a b λge =
Closed‐loop systemsbull Global minimum Solution of
bull unique solution for f1nef2
bull Indirect identificationbull Knowledge of v(t) and F(q‐1) L(q-1) is requiredbull Using one of the examined methods (LS PEM) we identify the total system between v ybull Using knowledge of L F we can get estimates of Gs HsUsing knowledge of L F we can get estimates of Gs Hs
Hs(q-1)
e(t)1( ) ( ( ) ( ))
1
( ) 1 ( ) ( )
s ss
s
y t G Lv t H e tFG
FG Fu t Lv t H e t
= ++
⎡ ⎤⎢ ⎥
bullComparing this to the general LTI structure
we can get estimates of
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
( ) 1 ( ) ( )1 1
ss
s s
u t Lv t H e tFG FG
= minus minus⎢ ⎥+ +⎣ ⎦
1 1ˆ ˆ( ) ( ) ( ) ( ) ( )y t G q v t H q e tminus minus= +θ θ1we can get estimates of
and use knowledge of LF to obtain GsHs
F(q-1)
11
11
ss
ss
G G LFG
H HFG
=+
=+
Nonlinear systems
S(τ) y(t)x(t) S(Ax1+Bx2)neAS(x1)+BS(x2)
bull Most real‐life systems are nonlinear linearity is an approximation that may or may not yield satisfactory resultsbull Nonlinear models identification methods
N t i (V lt Wi i )bull Nonparametric (VolterraWiener series)bull Parametric (nonlinear ARX ARMAX models nonlinear differential equations)bull Block diagrams Neural network models
yu z
A b f th l i t ti t th t d l f IO d t
Gyu z
21 2
( ) ( ) ( )( ) ( ) ( )z t g t u ty t c z t c z t
=
= +
bull As before the goal is to estimate the system model from IO databull More complicated problem ndash number of free parameters much higherbull No simple relations exist in the frequency domain as for linear systems (the output of a nonlinear system to a sinusoidal signal is not a sinusoidal signal with the same frequency but hibit h i t l )exhibits harmonic components also)
bull Generally more difficult to obtain theoretical results for convergence consistency etc compared to linear systems
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
Volterra‐Wiener seriesbull Generalization of convolution sum bull Goal estimate Volterra kernels kn from IO measurements
system memory
k 1(m
)k 2
(m1m
2)k
system memory
k
131313m
m1m2 m1 m2
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Estimationbull Least‐squares we can write the above relation as a linear regression problems ‐ very large number of free parameters (order ΜQ) bull Cross‐correlation method Estimate high‐order cross‐correlation functions between input output eg
As we saw for linear systems also long data records are typically needed to obtain accurate estimates for these cross‐correlation functions
1 2 1 2( ) ( ) ( ) ( )xxy E y t x t x tϕ τ τ τ τ= + +
Nonlinear systems⎫⎧
sum sum sum=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Efficient Volterra kernel estimationF ti i R d ti f f tbull Function expansions Reduction of free parameters
bull One of the most widely used basis set is the Laguerre basis( ) 2 1 2
0( ) (1 ) ( 1) (1 ) 0 1
jj m j m m
jm
jb a
m mτ τ
τ α α α αminus minus
=
⎛ ⎞⎛ ⎞= minus minus minus lt lt⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠sum
03
04
05α =01α =02α =04
1 1 1
1
1 10 0
( ) ( ) ( )q
q
L L
q q j j j j qj j
k m m c b m b m= =
= sum sum
V + b
0
01
02y=Vc+ε vj=xbj
cest=[VTV]-1VTy
bull Orthonormal basis in [0infin) ndash well‐
-03
-02
-01Orthonormal basis in [0 infin) wellconditioned estimation of causal systems
bull Exponential decay suitable for finite memory systems
0 5 10 15 20 25 30 35 40 45 50-05
-04
Time lags
y ybull α critical for model accuracy
parsimony
Nonlinear systems
bull Polynomial activation ffunctionsbull Free parameters ~ QMbull Equivalent to Volterra model
M
sum=
minus=j
jii jnxwnv0
)()(miiiiiii zazaazf 110 )( +++=
H
sum=
=H
iijijijmiimm m
wwwacjjjk1
21 )(11
Nonlinear systems
bull Combination of neural networks with functionexpansions (Laguerre‐Volterra networks)
x(n)
expansions (Laguerre Volterra networks)
sum=
=Q
m
mkkmkk nucnuf
1 ))(())((
m 1
bull Drastic reduction of free parameters
vL(n)b0 bj bL
v0(n) vj(n)
hellip hellip
wwK0
w1jw1L
Drastic reduction of free parameters (Q+L+1)∙K+2 bull K number of hidden units f1 fK
hellip
w10wKL
K0 wKj
+
y(n)
y0
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Closed‐loop systemsbull Often control systems applications aim to minimize the difference between a reference signal v(t) and y(t) which implies small u(t) values ndash not ideal for identification purposesbull Spectral analysis
W h b f (HW2) th t th t i ti t f G ( 1) i th f d ibull We have seen before (HW2) that the nonparametric estimate of Gs(q‐1) in the frequency domain (assuming L=1) ie
i ld bi d ti t I th l th bi i
ˆ ( ) ( ) ( ) ( ) ( )ˆ ( ) ˆ ( ) ( )( )yu s vv zz
vv zzuu
G FGω ω ω ω ωω
ω ωωΦ Φ minusΦ
= =Φ +ΦΦ
yields biased estimates In the general case the bias is
where
1 ( ) ( ) ( )ˆ ( ) ( )( ) ( ) ( )
s zzs
vv zz
G FG GFω ω ωω ωω ω ω
+ Φminus = minus
Φ +ΦHs(q-1)
e(t)
For e(t)=0 no bias in the estimateFor v(t)=0 we have
1 1( ) ( ) ( ) ( )sz t F q H q e tminus minus=
Gs(q-1) y(t)u(t)+
v(t) z(t)
+
s(q )
L(q-1)
1ˆ ( )( )
GF
ωω
minusF(q-1)
-
Closed‐loop systemsbull Modified spectral analysis
bull Assuming that the external input signal ν(t) is measurable ( ) ( ) ( ) ( ) ( )vy s vu s veG Hω ω ω ω ωΦ = Φ + Φ =
Therefore we can obtain the following (unbiased) estimate
Ti d i t l h f id tifi ti f l d l t
( ) ( )s vuG ω ω= Φ
ˆ ( )ˆ ( ) ˆ ( )vy
vu
Gω
ωω
Φ=Φ
bull Time domain ndash two general approaches for identification of closed‐loop systems arebull Direct identification We treat the system as an open‐loop systembull Indirect identification We assume that the external input reference signal v(t) is measurable and that the feedback law F is known
Hs(q-1)
e(t)
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
F(q-1)
Closed‐loop systemsbull Direct identification We use the measurements of u(t) and y(t) and the general model structure we have used before
1 1 2 2ˆ ˆˆ( ) ( ) ( ) ( ) ( ) ( )y t G q u t H q e t E e t λminus minus= + =θ θ
bull The goal is to fine the parameter vector that minimizes the prediction error eg
[ ]22
1 1
1 1 ˆ( ) ( ) ( ) ( | )N N
Nt t
V t y t y tN N
ε= =
= = minussum sumθ θ θ 1 1 1ˆˆ( ) ( )[ ( ) ( ) ( )]t H q y t G q u tε minus minus minus= minusθ
bull The question is whether for this parametrization the closed loop system is identifiable ie whether there exists θ such that
ˆ ˆs sG G H H= =
bull This depends on the parametrization and the true system form (eg feedback law)
bull Example Let the system
Model
2 2( ) ( 1) ( 1) ( ) ( )( ) ( )y t ay t bu t e t E e tu t fy t
λ+ minus = minus + == minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +
( ) ( ) ( 1) ( )y t a fb y t e t+ + minus =
If we use eg least squares we can estimate only the linear combination a+fb
Closed‐loop systemsbull Assume we have two different regulators f1 and f2 which are active for fractions γ1 and γ2 of the total time ie
bull As before the system is2 2( ) ( 1) ( 1) ( ) ( )
( ) ( )y t ay t bu t e t E e t
fλ+ minus = minus + =
and the model( ) ( )iu t f y t= minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +( ) ( ) ( 1) ( )
ˆˆ( ) ( ) ( ) ( 1)i iy t a bf y t e t
t y t a bf y tε
+ + minus =
= + + minus
therefore
bull Cost functionN
( ) ( ) ( ) ( 1)i i it y t a bf y tε + +2
2 22
ˆˆ( ) ( ) 11 ( )
i ii
i
a bf a bfE ta bf
ε λ⎡ ⎤+ minus minus
= +⎢ ⎥minus +⎢ ⎥⎣ ⎦
2
1
2 22 2 21 1 2 2
1 22 2
1ˆˆlim ( ) ( )
ˆ ˆˆ ˆ( ) ( )1 ( ) 1 ( )
N
N Nt
V a b tN
a bf a bf a bf a bfa bf a bf
ε
λ γ λ γ λ
rarrinfin=
=
+ minus minus + minus minus= + +
minus + minus +
sum θ
bull We can see that ‐minimum
1 21 ( ) 1 ( )a bf a bf+ +
2ˆˆ( ) ( )N NV a b V a b λge =
Closed‐loop systemsbull Global minimum Solution of
bull unique solution for f1nef2
bull Indirect identificationbull Knowledge of v(t) and F(q‐1) L(q-1) is requiredbull Using one of the examined methods (LS PEM) we identify the total system between v ybull Using knowledge of L F we can get estimates of Gs HsUsing knowledge of L F we can get estimates of Gs Hs
Hs(q-1)
e(t)1( ) ( ( ) ( ))
1
( ) 1 ( ) ( )
s ss
s
y t G Lv t H e tFG
FG Fu t Lv t H e t
= ++
⎡ ⎤⎢ ⎥
bullComparing this to the general LTI structure
we can get estimates of
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
( ) 1 ( ) ( )1 1
ss
s s
u t Lv t H e tFG FG
= minus minus⎢ ⎥+ +⎣ ⎦
1 1ˆ ˆ( ) ( ) ( ) ( ) ( )y t G q v t H q e tminus minus= +θ θ1we can get estimates of
and use knowledge of LF to obtain GsHs
F(q-1)
11
11
ss
ss
G G LFG
H HFG
=+
=+
Nonlinear systems
S(τ) y(t)x(t) S(Ax1+Bx2)neAS(x1)+BS(x2)
bull Most real‐life systems are nonlinear linearity is an approximation that may or may not yield satisfactory resultsbull Nonlinear models identification methods
N t i (V lt Wi i )bull Nonparametric (VolterraWiener series)bull Parametric (nonlinear ARX ARMAX models nonlinear differential equations)bull Block diagrams Neural network models
yu z
A b f th l i t ti t th t d l f IO d t
Gyu z
21 2
( ) ( ) ( )( ) ( ) ( )z t g t u ty t c z t c z t
=
= +
bull As before the goal is to estimate the system model from IO databull More complicated problem ndash number of free parameters much higherbull No simple relations exist in the frequency domain as for linear systems (the output of a nonlinear system to a sinusoidal signal is not a sinusoidal signal with the same frequency but hibit h i t l )exhibits harmonic components also)
bull Generally more difficult to obtain theoretical results for convergence consistency etc compared to linear systems
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
Volterra‐Wiener seriesbull Generalization of convolution sum bull Goal estimate Volterra kernels kn from IO measurements
system memory
k 1(m
)k 2
(m1m
2)k
system memory
k
131313m
m1m2 m1 m2
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Estimationbull Least‐squares we can write the above relation as a linear regression problems ‐ very large number of free parameters (order ΜQ) bull Cross‐correlation method Estimate high‐order cross‐correlation functions between input output eg
As we saw for linear systems also long data records are typically needed to obtain accurate estimates for these cross‐correlation functions
1 2 1 2( ) ( ) ( ) ( )xxy E y t x t x tϕ τ τ τ τ= + +
Nonlinear systems⎫⎧
sum sum sum=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Efficient Volterra kernel estimationF ti i R d ti f f tbull Function expansions Reduction of free parameters
bull One of the most widely used basis set is the Laguerre basis( ) 2 1 2
0( ) (1 ) ( 1) (1 ) 0 1
jj m j m m
jm
jb a
m mτ τ
τ α α α αminus minus
=
⎛ ⎞⎛ ⎞= minus minus minus lt lt⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠sum
03
04
05α =01α =02α =04
1 1 1
1
1 10 0
( ) ( ) ( )q
q
L L
q q j j j j qj j
k m m c b m b m= =
= sum sum
V + b
0
01
02y=Vc+ε vj=xbj
cest=[VTV]-1VTy
bull Orthonormal basis in [0infin) ndash well‐
-03
-02
-01Orthonormal basis in [0 infin) wellconditioned estimation of causal systems
bull Exponential decay suitable for finite memory systems
0 5 10 15 20 25 30 35 40 45 50-05
-04
Time lags
y ybull α critical for model accuracy
parsimony
Nonlinear systems
bull Polynomial activation ffunctionsbull Free parameters ~ QMbull Equivalent to Volterra model
M
sum=
minus=j
jii jnxwnv0
)()(miiiiiii zazaazf 110 )( +++=
H
sum=
=H
iijijijmiimm m
wwwacjjjk1
21 )(11
Nonlinear systems
bull Combination of neural networks with functionexpansions (Laguerre‐Volterra networks)
x(n)
expansions (Laguerre Volterra networks)
sum=
=Q
m
mkkmkk nucnuf
1 ))(())((
m 1
bull Drastic reduction of free parameters
vL(n)b0 bj bL
v0(n) vj(n)
hellip hellip
wwK0
w1jw1L
Drastic reduction of free parameters (Q+L+1)∙K+2 bull K number of hidden units f1 fK
hellip
w10wKL
K0 wKj
+
y(n)
y0
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Closed‐loop systemsbull Modified spectral analysis
bull Assuming that the external input signal ν(t) is measurable ( ) ( ) ( ) ( ) ( )vy s vu s veG Hω ω ω ω ωΦ = Φ + Φ =
Therefore we can obtain the following (unbiased) estimate
Ti d i t l h f id tifi ti f l d l t
( ) ( )s vuG ω ω= Φ
ˆ ( )ˆ ( ) ˆ ( )vy
vu
Gω
ωω
Φ=Φ
bull Time domain ndash two general approaches for identification of closed‐loop systems arebull Direct identification We treat the system as an open‐loop systembull Indirect identification We assume that the external input reference signal v(t) is measurable and that the feedback law F is known
Hs(q-1)
e(t)
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
F(q-1)
Closed‐loop systemsbull Direct identification We use the measurements of u(t) and y(t) and the general model structure we have used before
1 1 2 2ˆ ˆˆ( ) ( ) ( ) ( ) ( ) ( )y t G q u t H q e t E e t λminus minus= + =θ θ
bull The goal is to fine the parameter vector that minimizes the prediction error eg
[ ]22
1 1
1 1 ˆ( ) ( ) ( ) ( | )N N
Nt t
V t y t y tN N
ε= =
= = minussum sumθ θ θ 1 1 1ˆˆ( ) ( )[ ( ) ( ) ( )]t H q y t G q u tε minus minus minus= minusθ
bull The question is whether for this parametrization the closed loop system is identifiable ie whether there exists θ such that
ˆ ˆs sG G H H= =
bull This depends on the parametrization and the true system form (eg feedback law)
bull Example Let the system
Model
2 2( ) ( 1) ( 1) ( ) ( )( ) ( )y t ay t bu t e t E e tu t fy t
λ+ minus = minus + == minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +
( ) ( ) ( 1) ( )y t a fb y t e t+ + minus =
If we use eg least squares we can estimate only the linear combination a+fb
Closed‐loop systemsbull Assume we have two different regulators f1 and f2 which are active for fractions γ1 and γ2 of the total time ie
bull As before the system is2 2( ) ( 1) ( 1) ( ) ( )
( ) ( )y t ay t bu t e t E e t
fλ+ minus = minus + =
and the model( ) ( )iu t f y t= minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +( ) ( ) ( 1) ( )
ˆˆ( ) ( ) ( ) ( 1)i iy t a bf y t e t
t y t a bf y tε
+ + minus =
= + + minus
therefore
bull Cost functionN
( ) ( ) ( ) ( 1)i i it y t a bf y tε + +2
2 22
ˆˆ( ) ( ) 11 ( )
i ii
i
a bf a bfE ta bf
ε λ⎡ ⎤+ minus minus
= +⎢ ⎥minus +⎢ ⎥⎣ ⎦
2
1
2 22 2 21 1 2 2
1 22 2
1ˆˆlim ( ) ( )
ˆ ˆˆ ˆ( ) ( )1 ( ) 1 ( )
N
N Nt
V a b tN
a bf a bf a bf a bfa bf a bf
ε
λ γ λ γ λ
rarrinfin=
=
+ minus minus + minus minus= + +
minus + minus +
sum θ
bull We can see that ‐minimum
1 21 ( ) 1 ( )a bf a bf+ +
2ˆˆ( ) ( )N NV a b V a b λge =
Closed‐loop systemsbull Global minimum Solution of
bull unique solution for f1nef2
bull Indirect identificationbull Knowledge of v(t) and F(q‐1) L(q-1) is requiredbull Using one of the examined methods (LS PEM) we identify the total system between v ybull Using knowledge of L F we can get estimates of Gs HsUsing knowledge of L F we can get estimates of Gs Hs
Hs(q-1)
e(t)1( ) ( ( ) ( ))
1
( ) 1 ( ) ( )
s ss
s
y t G Lv t H e tFG
FG Fu t Lv t H e t
= ++
⎡ ⎤⎢ ⎥
bullComparing this to the general LTI structure
we can get estimates of
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
( ) 1 ( ) ( )1 1
ss
s s
u t Lv t H e tFG FG
= minus minus⎢ ⎥+ +⎣ ⎦
1 1ˆ ˆ( ) ( ) ( ) ( ) ( )y t G q v t H q e tminus minus= +θ θ1we can get estimates of
and use knowledge of LF to obtain GsHs
F(q-1)
11
11
ss
ss
G G LFG
H HFG
=+
=+
Nonlinear systems
S(τ) y(t)x(t) S(Ax1+Bx2)neAS(x1)+BS(x2)
bull Most real‐life systems are nonlinear linearity is an approximation that may or may not yield satisfactory resultsbull Nonlinear models identification methods
N t i (V lt Wi i )bull Nonparametric (VolterraWiener series)bull Parametric (nonlinear ARX ARMAX models nonlinear differential equations)bull Block diagrams Neural network models
yu z
A b f th l i t ti t th t d l f IO d t
Gyu z
21 2
( ) ( ) ( )( ) ( ) ( )z t g t u ty t c z t c z t
=
= +
bull As before the goal is to estimate the system model from IO databull More complicated problem ndash number of free parameters much higherbull No simple relations exist in the frequency domain as for linear systems (the output of a nonlinear system to a sinusoidal signal is not a sinusoidal signal with the same frequency but hibit h i t l )exhibits harmonic components also)
bull Generally more difficult to obtain theoretical results for convergence consistency etc compared to linear systems
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
Volterra‐Wiener seriesbull Generalization of convolution sum bull Goal estimate Volterra kernels kn from IO measurements
system memory
k 1(m
)k 2
(m1m
2)k
system memory
k
131313m
m1m2 m1 m2
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Estimationbull Least‐squares we can write the above relation as a linear regression problems ‐ very large number of free parameters (order ΜQ) bull Cross‐correlation method Estimate high‐order cross‐correlation functions between input output eg
As we saw for linear systems also long data records are typically needed to obtain accurate estimates for these cross‐correlation functions
1 2 1 2( ) ( ) ( ) ( )xxy E y t x t x tϕ τ τ τ τ= + +
Nonlinear systems⎫⎧
sum sum sum=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Efficient Volterra kernel estimationF ti i R d ti f f tbull Function expansions Reduction of free parameters
bull One of the most widely used basis set is the Laguerre basis( ) 2 1 2
0( ) (1 ) ( 1) (1 ) 0 1
jj m j m m
jm
jb a
m mτ τ
τ α α α αminus minus
=
⎛ ⎞⎛ ⎞= minus minus minus lt lt⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠sum
03
04
05α =01α =02α =04
1 1 1
1
1 10 0
( ) ( ) ( )q
q
L L
q q j j j j qj j
k m m c b m b m= =
= sum sum
V + b
0
01
02y=Vc+ε vj=xbj
cest=[VTV]-1VTy
bull Orthonormal basis in [0infin) ndash well‐
-03
-02
-01Orthonormal basis in [0 infin) wellconditioned estimation of causal systems
bull Exponential decay suitable for finite memory systems
0 5 10 15 20 25 30 35 40 45 50-05
-04
Time lags
y ybull α critical for model accuracy
parsimony
Nonlinear systems
bull Polynomial activation ffunctionsbull Free parameters ~ QMbull Equivalent to Volterra model
M
sum=
minus=j
jii jnxwnv0
)()(miiiiiii zazaazf 110 )( +++=
H
sum=
=H
iijijijmiimm m
wwwacjjjk1
21 )(11
Nonlinear systems
bull Combination of neural networks with functionexpansions (Laguerre‐Volterra networks)
x(n)
expansions (Laguerre Volterra networks)
sum=
=Q
m
mkkmkk nucnuf
1 ))(())((
m 1
bull Drastic reduction of free parameters
vL(n)b0 bj bL
v0(n) vj(n)
hellip hellip
wwK0
w1jw1L
Drastic reduction of free parameters (Q+L+1)∙K+2 bull K number of hidden units f1 fK
hellip
w10wKL
K0 wKj
+
y(n)
y0
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Closed‐loop systemsbull Direct identification We use the measurements of u(t) and y(t) and the general model structure we have used before
1 1 2 2ˆ ˆˆ( ) ( ) ( ) ( ) ( ) ( )y t G q u t H q e t E e t λminus minus= + =θ θ
bull The goal is to fine the parameter vector that minimizes the prediction error eg
[ ]22
1 1
1 1 ˆ( ) ( ) ( ) ( | )N N
Nt t
V t y t y tN N
ε= =
= = minussum sumθ θ θ 1 1 1ˆˆ( ) ( )[ ( ) ( ) ( )]t H q y t G q u tε minus minus minus= minusθ
bull The question is whether for this parametrization the closed loop system is identifiable ie whether there exists θ such that
ˆ ˆs sG G H H= =
bull This depends on the parametrization and the true system form (eg feedback law)
bull Example Let the system
Model
2 2( ) ( 1) ( 1) ( ) ( )( ) ( )y t ay t bu t e t E e tu t fy t
λ+ minus = minus + == minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +
( ) ( ) ( 1) ( )y t a fb y t e t+ + minus =
If we use eg least squares we can estimate only the linear combination a+fb
Closed‐loop systemsbull Assume we have two different regulators f1 and f2 which are active for fractions γ1 and γ2 of the total time ie
bull As before the system is2 2( ) ( 1) ( 1) ( ) ( )
( ) ( )y t ay t bu t e t E e t
fλ+ minus = minus + =
and the model( ) ( )iu t f y t= minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +( ) ( ) ( 1) ( )
ˆˆ( ) ( ) ( ) ( 1)i iy t a bf y t e t
t y t a bf y tε
+ + minus =
= + + minus
therefore
bull Cost functionN
( ) ( ) ( ) ( 1)i i it y t a bf y tε + +2
2 22
ˆˆ( ) ( ) 11 ( )
i ii
i
a bf a bfE ta bf
ε λ⎡ ⎤+ minus minus
= +⎢ ⎥minus +⎢ ⎥⎣ ⎦
2
1
2 22 2 21 1 2 2
1 22 2
1ˆˆlim ( ) ( )
ˆ ˆˆ ˆ( ) ( )1 ( ) 1 ( )
N
N Nt
V a b tN
a bf a bf a bf a bfa bf a bf
ε
λ γ λ γ λ
rarrinfin=
=
+ minus minus + minus minus= + +
minus + minus +
sum θ
bull We can see that ‐minimum
1 21 ( ) 1 ( )a bf a bf+ +
2ˆˆ( ) ( )N NV a b V a b λge =
Closed‐loop systemsbull Global minimum Solution of
bull unique solution for f1nef2
bull Indirect identificationbull Knowledge of v(t) and F(q‐1) L(q-1) is requiredbull Using one of the examined methods (LS PEM) we identify the total system between v ybull Using knowledge of L F we can get estimates of Gs HsUsing knowledge of L F we can get estimates of Gs Hs
Hs(q-1)
e(t)1( ) ( ( ) ( ))
1
( ) 1 ( ) ( )
s ss
s
y t G Lv t H e tFG
FG Fu t Lv t H e t
= ++
⎡ ⎤⎢ ⎥
bullComparing this to the general LTI structure
we can get estimates of
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
( ) 1 ( ) ( )1 1
ss
s s
u t Lv t H e tFG FG
= minus minus⎢ ⎥+ +⎣ ⎦
1 1ˆ ˆ( ) ( ) ( ) ( ) ( )y t G q v t H q e tminus minus= +θ θ1we can get estimates of
and use knowledge of LF to obtain GsHs
F(q-1)
11
11
ss
ss
G G LFG
H HFG
=+
=+
Nonlinear systems
S(τ) y(t)x(t) S(Ax1+Bx2)neAS(x1)+BS(x2)
bull Most real‐life systems are nonlinear linearity is an approximation that may or may not yield satisfactory resultsbull Nonlinear models identification methods
N t i (V lt Wi i )bull Nonparametric (VolterraWiener series)bull Parametric (nonlinear ARX ARMAX models nonlinear differential equations)bull Block diagrams Neural network models
yu z
A b f th l i t ti t th t d l f IO d t
Gyu z
21 2
( ) ( ) ( )( ) ( ) ( )z t g t u ty t c z t c z t
=
= +
bull As before the goal is to estimate the system model from IO databull More complicated problem ndash number of free parameters much higherbull No simple relations exist in the frequency domain as for linear systems (the output of a nonlinear system to a sinusoidal signal is not a sinusoidal signal with the same frequency but hibit h i t l )exhibits harmonic components also)
bull Generally more difficult to obtain theoretical results for convergence consistency etc compared to linear systems
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
Volterra‐Wiener seriesbull Generalization of convolution sum bull Goal estimate Volterra kernels kn from IO measurements
system memory
k 1(m
)k 2
(m1m
2)k
system memory
k
131313m
m1m2 m1 m2
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Estimationbull Least‐squares we can write the above relation as a linear regression problems ‐ very large number of free parameters (order ΜQ) bull Cross‐correlation method Estimate high‐order cross‐correlation functions between input output eg
As we saw for linear systems also long data records are typically needed to obtain accurate estimates for these cross‐correlation functions
1 2 1 2( ) ( ) ( ) ( )xxy E y t x t x tϕ τ τ τ τ= + +
Nonlinear systems⎫⎧
sum sum sum=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Efficient Volterra kernel estimationF ti i R d ti f f tbull Function expansions Reduction of free parameters
bull One of the most widely used basis set is the Laguerre basis( ) 2 1 2
0( ) (1 ) ( 1) (1 ) 0 1
jj m j m m
jm
jb a
m mτ τ
τ α α α αminus minus
=
⎛ ⎞⎛ ⎞= minus minus minus lt lt⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠sum
03
04
05α =01α =02α =04
1 1 1
1
1 10 0
( ) ( ) ( )q
q
L L
q q j j j j qj j
k m m c b m b m= =
= sum sum
V + b
0
01
02y=Vc+ε vj=xbj
cest=[VTV]-1VTy
bull Orthonormal basis in [0infin) ndash well‐
-03
-02
-01Orthonormal basis in [0 infin) wellconditioned estimation of causal systems
bull Exponential decay suitable for finite memory systems
0 5 10 15 20 25 30 35 40 45 50-05
-04
Time lags
y ybull α critical for model accuracy
parsimony
Nonlinear systems
bull Polynomial activation ffunctionsbull Free parameters ~ QMbull Equivalent to Volterra model
M
sum=
minus=j
jii jnxwnv0
)()(miiiiiii zazaazf 110 )( +++=
H
sum=
=H
iijijijmiimm m
wwwacjjjk1
21 )(11
Nonlinear systems
bull Combination of neural networks with functionexpansions (Laguerre‐Volterra networks)
x(n)
expansions (Laguerre Volterra networks)
sum=
=Q
m
mkkmkk nucnuf
1 ))(())((
m 1
bull Drastic reduction of free parameters
vL(n)b0 bj bL
v0(n) vj(n)
hellip hellip
wwK0
w1jw1L
Drastic reduction of free parameters (Q+L+1)∙K+2 bull K number of hidden units f1 fK
hellip
w10wKL
K0 wKj
+
y(n)
y0
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Closed‐loop systemsbull Assume we have two different regulators f1 and f2 which are active for fractions γ1 and γ2 of the total time ie
bull As before the system is2 2( ) ( 1) ( 1) ( ) ( )
( ) ( )y t ay t bu t e t E e t
fλ+ minus = minus + =
and the model( ) ( )iu t f y t= minus
ˆˆ( ) ( 1) ( 1) ( )y t ay t bu t tε+ minus = minus +( ) ( ) ( 1) ( )
ˆˆ( ) ( ) ( ) ( 1)i iy t a bf y t e t
t y t a bf y tε
+ + minus =
= + + minus
therefore
bull Cost functionN
( ) ( ) ( ) ( 1)i i it y t a bf y tε + +2
2 22
ˆˆ( ) ( ) 11 ( )
i ii
i
a bf a bfE ta bf
ε λ⎡ ⎤+ minus minus
= +⎢ ⎥minus +⎢ ⎥⎣ ⎦
2
1
2 22 2 21 1 2 2
1 22 2
1ˆˆlim ( ) ( )
ˆ ˆˆ ˆ( ) ( )1 ( ) 1 ( )
N
N Nt
V a b tN
a bf a bf a bf a bfa bf a bf
ε
λ γ λ γ λ
rarrinfin=
=
+ minus minus + minus minus= + +
minus + minus +
sum θ
bull We can see that ‐minimum
1 21 ( ) 1 ( )a bf a bf+ +
2ˆˆ( ) ( )N NV a b V a b λge =
Closed‐loop systemsbull Global minimum Solution of
bull unique solution for f1nef2
bull Indirect identificationbull Knowledge of v(t) and F(q‐1) L(q-1) is requiredbull Using one of the examined methods (LS PEM) we identify the total system between v ybull Using knowledge of L F we can get estimates of Gs HsUsing knowledge of L F we can get estimates of Gs Hs
Hs(q-1)
e(t)1( ) ( ( ) ( ))
1
( ) 1 ( ) ( )
s ss
s
y t G Lv t H e tFG
FG Fu t Lv t H e t
= ++
⎡ ⎤⎢ ⎥
bullComparing this to the general LTI structure
we can get estimates of
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
( ) 1 ( ) ( )1 1
ss
s s
u t Lv t H e tFG FG
= minus minus⎢ ⎥+ +⎣ ⎦
1 1ˆ ˆ( ) ( ) ( ) ( ) ( )y t G q v t H q e tminus minus= +θ θ1we can get estimates of
and use knowledge of LF to obtain GsHs
F(q-1)
11
11
ss
ss
G G LFG
H HFG
=+
=+
Nonlinear systems
S(τ) y(t)x(t) S(Ax1+Bx2)neAS(x1)+BS(x2)
bull Most real‐life systems are nonlinear linearity is an approximation that may or may not yield satisfactory resultsbull Nonlinear models identification methods
N t i (V lt Wi i )bull Nonparametric (VolterraWiener series)bull Parametric (nonlinear ARX ARMAX models nonlinear differential equations)bull Block diagrams Neural network models
yu z
A b f th l i t ti t th t d l f IO d t
Gyu z
21 2
( ) ( ) ( )( ) ( ) ( )z t g t u ty t c z t c z t
=
= +
bull As before the goal is to estimate the system model from IO databull More complicated problem ndash number of free parameters much higherbull No simple relations exist in the frequency domain as for linear systems (the output of a nonlinear system to a sinusoidal signal is not a sinusoidal signal with the same frequency but hibit h i t l )exhibits harmonic components also)
bull Generally more difficult to obtain theoretical results for convergence consistency etc compared to linear systems
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
Volterra‐Wiener seriesbull Generalization of convolution sum bull Goal estimate Volterra kernels kn from IO measurements
system memory
k 1(m
)k 2
(m1m
2)k
system memory
k
131313m
m1m2 m1 m2
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Estimationbull Least‐squares we can write the above relation as a linear regression problems ‐ very large number of free parameters (order ΜQ) bull Cross‐correlation method Estimate high‐order cross‐correlation functions between input output eg
As we saw for linear systems also long data records are typically needed to obtain accurate estimates for these cross‐correlation functions
1 2 1 2( ) ( ) ( ) ( )xxy E y t x t x tϕ τ τ τ τ= + +
Nonlinear systems⎫⎧
sum sum sum=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Efficient Volterra kernel estimationF ti i R d ti f f tbull Function expansions Reduction of free parameters
bull One of the most widely used basis set is the Laguerre basis( ) 2 1 2
0( ) (1 ) ( 1) (1 ) 0 1
jj m j m m
jm
jb a
m mτ τ
τ α α α αminus minus
=
⎛ ⎞⎛ ⎞= minus minus minus lt lt⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠sum
03
04
05α =01α =02α =04
1 1 1
1
1 10 0
( ) ( ) ( )q
q
L L
q q j j j j qj j
k m m c b m b m= =
= sum sum
V + b
0
01
02y=Vc+ε vj=xbj
cest=[VTV]-1VTy
bull Orthonormal basis in [0infin) ndash well‐
-03
-02
-01Orthonormal basis in [0 infin) wellconditioned estimation of causal systems
bull Exponential decay suitable for finite memory systems
0 5 10 15 20 25 30 35 40 45 50-05
-04
Time lags
y ybull α critical for model accuracy
parsimony
Nonlinear systems
bull Polynomial activation ffunctionsbull Free parameters ~ QMbull Equivalent to Volterra model
M
sum=
minus=j
jii jnxwnv0
)()(miiiiiii zazaazf 110 )( +++=
H
sum=
=H
iijijijmiimm m
wwwacjjjk1
21 )(11
Nonlinear systems
bull Combination of neural networks with functionexpansions (Laguerre‐Volterra networks)
x(n)
expansions (Laguerre Volterra networks)
sum=
=Q
m
mkkmkk nucnuf
1 ))(())((
m 1
bull Drastic reduction of free parameters
vL(n)b0 bj bL
v0(n) vj(n)
hellip hellip
wwK0
w1jw1L
Drastic reduction of free parameters (Q+L+1)∙K+2 bull K number of hidden units f1 fK
hellip
w10wKL
K0 wKj
+
y(n)
y0
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Closed‐loop systemsbull Global minimum Solution of
bull unique solution for f1nef2
bull Indirect identificationbull Knowledge of v(t) and F(q‐1) L(q-1) is requiredbull Using one of the examined methods (LS PEM) we identify the total system between v ybull Using knowledge of L F we can get estimates of Gs HsUsing knowledge of L F we can get estimates of Gs Hs
Hs(q-1)
e(t)1( ) ( ( ) ( ))
1
( ) 1 ( ) ( )
s ss
s
y t G Lv t H e tFG
FG Fu t Lv t H e t
= ++
⎡ ⎤⎢ ⎥
bullComparing this to the general LTI structure
we can get estimates of
Gs(q-1) y(t)u(t)+
v(t) z(t)
-+
L(q-1)
( ) 1 ( ) ( )1 1
ss
s s
u t Lv t H e tFG FG
= minus minus⎢ ⎥+ +⎣ ⎦
1 1ˆ ˆ( ) ( ) ( ) ( ) ( )y t G q v t H q e tminus minus= +θ θ1we can get estimates of
and use knowledge of LF to obtain GsHs
F(q-1)
11
11
ss
ss
G G LFG
H HFG
=+
=+
Nonlinear systems
S(τ) y(t)x(t) S(Ax1+Bx2)neAS(x1)+BS(x2)
bull Most real‐life systems are nonlinear linearity is an approximation that may or may not yield satisfactory resultsbull Nonlinear models identification methods
N t i (V lt Wi i )bull Nonparametric (VolterraWiener series)bull Parametric (nonlinear ARX ARMAX models nonlinear differential equations)bull Block diagrams Neural network models
yu z
A b f th l i t ti t th t d l f IO d t
Gyu z
21 2
( ) ( ) ( )( ) ( ) ( )z t g t u ty t c z t c z t
=
= +
bull As before the goal is to estimate the system model from IO databull More complicated problem ndash number of free parameters much higherbull No simple relations exist in the frequency domain as for linear systems (the output of a nonlinear system to a sinusoidal signal is not a sinusoidal signal with the same frequency but hibit h i t l )exhibits harmonic components also)
bull Generally more difficult to obtain theoretical results for convergence consistency etc compared to linear systems
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
Volterra‐Wiener seriesbull Generalization of convolution sum bull Goal estimate Volterra kernels kn from IO measurements
system memory
k 1(m
)k 2
(m1m
2)k
system memory
k
131313m
m1m2 m1 m2
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Estimationbull Least‐squares we can write the above relation as a linear regression problems ‐ very large number of free parameters (order ΜQ) bull Cross‐correlation method Estimate high‐order cross‐correlation functions between input output eg
As we saw for linear systems also long data records are typically needed to obtain accurate estimates for these cross‐correlation functions
1 2 1 2( ) ( ) ( ) ( )xxy E y t x t x tϕ τ τ τ τ= + +
Nonlinear systems⎫⎧
sum sum sum=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Efficient Volterra kernel estimationF ti i R d ti f f tbull Function expansions Reduction of free parameters
bull One of the most widely used basis set is the Laguerre basis( ) 2 1 2
0( ) (1 ) ( 1) (1 ) 0 1
jj m j m m
jm
jb a
m mτ τ
τ α α α αminus minus
=
⎛ ⎞⎛ ⎞= minus minus minus lt lt⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠sum
03
04
05α =01α =02α =04
1 1 1
1
1 10 0
( ) ( ) ( )q
q
L L
q q j j j j qj j
k m m c b m b m= =
= sum sum
V + b
0
01
02y=Vc+ε vj=xbj
cest=[VTV]-1VTy
bull Orthonormal basis in [0infin) ndash well‐
-03
-02
-01Orthonormal basis in [0 infin) wellconditioned estimation of causal systems
bull Exponential decay suitable for finite memory systems
0 5 10 15 20 25 30 35 40 45 50-05
-04
Time lags
y ybull α critical for model accuracy
parsimony
Nonlinear systems
bull Polynomial activation ffunctionsbull Free parameters ~ QMbull Equivalent to Volterra model
M
sum=
minus=j
jii jnxwnv0
)()(miiiiiii zazaazf 110 )( +++=
H
sum=
=H
iijijijmiimm m
wwwacjjjk1
21 )(11
Nonlinear systems
bull Combination of neural networks with functionexpansions (Laguerre‐Volterra networks)
x(n)
expansions (Laguerre Volterra networks)
sum=
=Q
m
mkkmkk nucnuf
1 ))(())((
m 1
bull Drastic reduction of free parameters
vL(n)b0 bj bL
v0(n) vj(n)
hellip hellip
wwK0
w1jw1L
Drastic reduction of free parameters (Q+L+1)∙K+2 bull K number of hidden units f1 fK
hellip
w10wKL
K0 wKj
+
y(n)
y0
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Nonlinear systems
S(τ) y(t)x(t) S(Ax1+Bx2)neAS(x1)+BS(x2)
bull Most real‐life systems are nonlinear linearity is an approximation that may or may not yield satisfactory resultsbull Nonlinear models identification methods
N t i (V lt Wi i )bull Nonparametric (VolterraWiener series)bull Parametric (nonlinear ARX ARMAX models nonlinear differential equations)bull Block diagrams Neural network models
yu z
A b f th l i t ti t th t d l f IO d t
Gyu z
21 2
( ) ( ) ( )( ) ( ) ( )z t g t u ty t c z t c z t
=
= +
bull As before the goal is to estimate the system model from IO databull More complicated problem ndash number of free parameters much higherbull No simple relations exist in the frequency domain as for linear systems (the output of a nonlinear system to a sinusoidal signal is not a sinusoidal signal with the same frequency but hibit h i t l )exhibits harmonic components also)
bull Generally more difficult to obtain theoretical results for convergence consistency etc compared to linear systems
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
Volterra‐Wiener seriesbull Generalization of convolution sum bull Goal estimate Volterra kernels kn from IO measurements
system memory
k 1(m
)k 2
(m1m
2)k
system memory
k
131313m
m1m2 m1 m2
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Estimationbull Least‐squares we can write the above relation as a linear regression problems ‐ very large number of free parameters (order ΜQ) bull Cross‐correlation method Estimate high‐order cross‐correlation functions between input output eg
As we saw for linear systems also long data records are typically needed to obtain accurate estimates for these cross‐correlation functions
1 2 1 2( ) ( ) ( ) ( )xxy E y t x t x tϕ τ τ τ τ= + +
Nonlinear systems⎫⎧
sum sum sum=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Efficient Volterra kernel estimationF ti i R d ti f f tbull Function expansions Reduction of free parameters
bull One of the most widely used basis set is the Laguerre basis( ) 2 1 2
0( ) (1 ) ( 1) (1 ) 0 1
jj m j m m
jm
jb a
m mτ τ
τ α α α αminus minus
=
⎛ ⎞⎛ ⎞= minus minus minus lt lt⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠sum
03
04
05α =01α =02α =04
1 1 1
1
1 10 0
( ) ( ) ( )q
q
L L
q q j j j j qj j
k m m c b m b m= =
= sum sum
V + b
0
01
02y=Vc+ε vj=xbj
cest=[VTV]-1VTy
bull Orthonormal basis in [0infin) ndash well‐
-03
-02
-01Orthonormal basis in [0 infin) wellconditioned estimation of causal systems
bull Exponential decay suitable for finite memory systems
0 5 10 15 20 25 30 35 40 45 50-05
-04
Time lags
y ybull α critical for model accuracy
parsimony
Nonlinear systems
bull Polynomial activation ffunctionsbull Free parameters ~ QMbull Equivalent to Volterra model
M
sum=
minus=j
jii jnxwnv0
)()(miiiiiii zazaazf 110 )( +++=
H
sum=
=H
iijijijmiimm m
wwwacjjjk1
21 )(11
Nonlinear systems
bull Combination of neural networks with functionexpansions (Laguerre‐Volterra networks)
x(n)
expansions (Laguerre Volterra networks)
sum=
=Q
m
mkkmkk nucnuf
1 ))(())((
m 1
bull Drastic reduction of free parameters
vL(n)b0 bj bL
v0(n) vj(n)
hellip hellip
wwK0
w1jw1L
Drastic reduction of free parameters (Q+L+1)∙K+2 bull K number of hidden units f1 fK
hellip
w10wKL
K0 wKj
+
y(n)
y0
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
Volterra‐Wiener seriesbull Generalization of convolution sum bull Goal estimate Volterra kernels kn from IO measurements
system memory
k 1(m
)k 2
(m1m
2)k
system memory
k
131313m
m1m2 m1 m2
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Estimationbull Least‐squares we can write the above relation as a linear regression problems ‐ very large number of free parameters (order ΜQ) bull Cross‐correlation method Estimate high‐order cross‐correlation functions between input output eg
As we saw for linear systems also long data records are typically needed to obtain accurate estimates for these cross‐correlation functions
1 2 1 2( ) ( ) ( ) ( )xxy E y t x t x tϕ τ τ τ τ= + +
Nonlinear systems⎫⎧
sum sum sum=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Efficient Volterra kernel estimationF ti i R d ti f f tbull Function expansions Reduction of free parameters
bull One of the most widely used basis set is the Laguerre basis( ) 2 1 2
0( ) (1 ) ( 1) (1 ) 0 1
jj m j m m
jm
jb a
m mτ τ
τ α α α αminus minus
=
⎛ ⎞⎛ ⎞= minus minus minus lt lt⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠sum
03
04
05α =01α =02α =04
1 1 1
1
1 10 0
( ) ( ) ( )q
q
L L
q q j j j j qj j
k m m c b m b m= =
= sum sum
V + b
0
01
02y=Vc+ε vj=xbj
cest=[VTV]-1VTy
bull Orthonormal basis in [0infin) ndash well‐
-03
-02
-01Orthonormal basis in [0 infin) wellconditioned estimation of causal systems
bull Exponential decay suitable for finite memory systems
0 5 10 15 20 25 30 35 40 45 50-05
-04
Time lags
y ybull α critical for model accuracy
parsimony
Nonlinear systems
bull Polynomial activation ffunctionsbull Free parameters ~ QMbull Equivalent to Volterra model
M
sum=
minus=j
jii jnxwnv0
)()(miiiiiii zazaazf 110 )( +++=
H
sum=
=H
iijijijmiimm m
wwwacjjjk1
21 )(11
Nonlinear systems
bull Combination of neural networks with functionexpansions (Laguerre‐Volterra networks)
x(n)
expansions (Laguerre Volterra networks)
sum=
=Q
m
mkkmkk nucnuf
1 ))(())((
m 1
bull Drastic reduction of free parameters
vL(n)b0 bj bL
v0(n) vj(n)
hellip hellip
wwK0
w1jw1L
Drastic reduction of free parameters (Q+L+1)∙K+2 bull K number of hidden units f1 fK
hellip
w10wKL
K0 wKj
+
y(n)
y0
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Nonlinear systemsM MemoryQ Ordersum sum sum
=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Estimationbull Least‐squares we can write the above relation as a linear regression problems ‐ very large number of free parameters (order ΜQ) bull Cross‐correlation method Estimate high‐order cross‐correlation functions between input output eg
As we saw for linear systems also long data records are typically needed to obtain accurate estimates for these cross‐correlation functions
1 2 1 2( ) ( ) ( ) ( )xxy E y t x t x tϕ τ τ τ τ= + +
Nonlinear systems⎫⎧
sum sum sum=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Efficient Volterra kernel estimationF ti i R d ti f f tbull Function expansions Reduction of free parameters
bull One of the most widely used basis set is the Laguerre basis( ) 2 1 2
0( ) (1 ) ( 1) (1 ) 0 1
jj m j m m
jm
jb a
m mτ τ
τ α α α αminus minus
=
⎛ ⎞⎛ ⎞= minus minus minus lt lt⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠sum
03
04
05α =01α =02α =04
1 1 1
1
1 10 0
( ) ( ) ( )q
q
L L
q q j j j j qj j
k m m c b m b m= =
= sum sum
V + b
0
01
02y=Vc+ε vj=xbj
cest=[VTV]-1VTy
bull Orthonormal basis in [0infin) ndash well‐
-03
-02
-01Orthonormal basis in [0 infin) wellconditioned estimation of causal systems
bull Exponential decay suitable for finite memory systems
0 5 10 15 20 25 30 35 40 45 50-05
-04
Time lags
y ybull α critical for model accuracy
parsimony
Nonlinear systems
bull Polynomial activation ffunctionsbull Free parameters ~ QMbull Equivalent to Volterra model
M
sum=
minus=j
jii jnxwnv0
)()(miiiiiii zazaazf 110 )( +++=
H
sum=
=H
iijijijmiimm m
wwwacjjjk1
21 )(11
Nonlinear systems
bull Combination of neural networks with functionexpansions (Laguerre‐Volterra networks)
x(n)
expansions (Laguerre Volterra networks)
sum=
=Q
m
mkkmkk nucnuf
1 ))(())((
m 1
bull Drastic reduction of free parameters
vL(n)b0 bj bL
v0(n) vj(n)
hellip hellip
wwK0
w1jw1L
Drastic reduction of free parameters (Q+L+1)∙K+2 bull K number of hidden units f1 fK
hellip
w10wKL
K0 wKj
+
y(n)
y0
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Nonlinear systems⎫⎧
sum sum sum=
minus
=
minus
= ⎭⎬⎫
⎩⎨⎧
minusminus=Q
n
M
m
M
mnnn
n
mnxmnxmmkny0
1
0
1
011
1
)()()()(
bull Efficient Volterra kernel estimationF ti i R d ti f f tbull Function expansions Reduction of free parameters
bull One of the most widely used basis set is the Laguerre basis( ) 2 1 2
0( ) (1 ) ( 1) (1 ) 0 1
jj m j m m
jm
jb a
m mτ τ
τ α α α αminus minus
=
⎛ ⎞⎛ ⎞= minus minus minus lt lt⎜ ⎟⎜ ⎟
⎝ ⎠⎝ ⎠sum
03
04
05α =01α =02α =04
1 1 1
1
1 10 0
( ) ( ) ( )q
q
L L
q q j j j j qj j
k m m c b m b m= =
= sum sum
V + b
0
01
02y=Vc+ε vj=xbj
cest=[VTV]-1VTy
bull Orthonormal basis in [0infin) ndash well‐
-03
-02
-01Orthonormal basis in [0 infin) wellconditioned estimation of causal systems
bull Exponential decay suitable for finite memory systems
0 5 10 15 20 25 30 35 40 45 50-05
-04
Time lags
y ybull α critical for model accuracy
parsimony
Nonlinear systems
bull Polynomial activation ffunctionsbull Free parameters ~ QMbull Equivalent to Volterra model
M
sum=
minus=j
jii jnxwnv0
)()(miiiiiii zazaazf 110 )( +++=
H
sum=
=H
iijijijmiimm m
wwwacjjjk1
21 )(11
Nonlinear systems
bull Combination of neural networks with functionexpansions (Laguerre‐Volterra networks)
x(n)
expansions (Laguerre Volterra networks)
sum=
=Q
m
mkkmkk nucnuf
1 ))(())((
m 1
bull Drastic reduction of free parameters
vL(n)b0 bj bL
v0(n) vj(n)
hellip hellip
wwK0
w1jw1L
Drastic reduction of free parameters (Q+L+1)∙K+2 bull K number of hidden units f1 fK
hellip
w10wKL
K0 wKj
+
y(n)
y0
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Nonlinear systems
bull Polynomial activation ffunctionsbull Free parameters ~ QMbull Equivalent to Volterra model
M
sum=
minus=j
jii jnxwnv0
)()(miiiiiii zazaazf 110 )( +++=
H
sum=
=H
iijijijmiimm m
wwwacjjjk1
21 )(11
Nonlinear systems
bull Combination of neural networks with functionexpansions (Laguerre‐Volterra networks)
x(n)
expansions (Laguerre Volterra networks)
sum=
=Q
m
mkkmkk nucnuf
1 ))(())((
m 1
bull Drastic reduction of free parameters
vL(n)b0 bj bL
v0(n) vj(n)
hellip hellip
wwK0
w1jw1L
Drastic reduction of free parameters (Q+L+1)∙K+2 bull K number of hidden units f1 fK
hellip
w10wKL
K0 wKj
+
y(n)
y0
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Nonlinear systems
bull Combination of neural networks with functionexpansions (Laguerre‐Volterra networks)
x(n)
expansions (Laguerre Volterra networks)
sum=
=Q
m
mkkmkk nucnuf
1 ))(())((
m 1
bull Drastic reduction of free parameters
vL(n)b0 bj bL
v0(n) vj(n)
hellip hellip
wwK0
w1jw1L
Drastic reduction of free parameters (Q+L+1)∙K+2 bull K number of hidden units f1 fK
hellip
w10wKL
K0 wKj
+
y(n)
y0
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Nonlinear systemsy
bull LVN trainingndash Iterative gradient descent (backpropagation)ndash Even the Laguerre parameter α can be trained
bull Recursive relations for Laguerre filter outputsbull Training
⎟⎞
⎜⎛ partJ )1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rjkw
rjkw
rjk
rjk
rjk
rjk dw
wJwdwww μγ
)1(
)(
)(
)(
)1(
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛
partpart
minus=+= rkmc
kc
rkm
rkm
rkm
rkm dc
cJcdccc μγ
⎠⎝ part rkmc
)1(0
0
)(0
)(0
)(0
)1(0
minus+ +⎟⎟⎠
⎞⎜⎜⎝
⎛partpart
minus=+= ry
ry
rrrr dyyJydyyy μγ
( 1) ( ) ( ) ( ) ( 1) r r r r rJd dβ β β β γ μ β β α+ minus⎛ ⎞part= + = minus + =⎜ ⎟
bull Equivalence to Volterra modelK L L
r
d dβ ββ β β β γ μ β β αβ
= + = minus + =⎜ ⎟part⎝ ⎠
sum sum sum= = =
=K
k
L
j
L
jnjjjkjkknnn
n
nnmbmbwwcmmk
1 0 011
1
11)()()(
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Nonlinear systemsbull Each input convolved
with different filterx1(n) xI(n)
Nonlinear systems
with different filter bankndash Distinct α parameters
diff i d i)((1)
0 nv
hellip hellip (1)1Lb
(1)jb(1)
0b
)((I)0 nv )((I) nv j )((I)
I nvL
hellip hellip (I)ILb
(I)jb(I)
0b)((1)
1 nvL
hellip)((1) nv j
different input dynamics
ndash Nonlinear interactions between inputs modeled
N b f f
(1)01w
(1) 1
w LK(1)
0wK(1)
w jK
(1)1w j
(1)1 1w L
)( )(j )(I(I)
01w
(I)0wK
(I)w jK
(I)1w j
(I)1 Iw L
(I)w LIK
bull Number of free parameters
⎞⎛ I
fKhellipf1
11
++sdot⎟⎠
⎞⎜⎝
⎛++sum
=
NKNQLI
ii
+ y0+
y(n)
I number of inputs
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Some examples ndash biological systems identificationp g y
bull Noninvasive accurate measurement of a wide variety of biological signals and programmable micropumps for pharmacological substance infusion
bull Rich spontaneous variability in physiological signals
bull We can obtain information about the function of biological physiological systemsndash Homeostatic mechanisms eg pressure regulation blood flow in the brain
ndash Functional connectivity in the brain (EEGMEGfMRI measurements)
bull Control of physiological signals based on mathematical models (model‐predictive control))
ndash Glucose control in diabetics with insulin micropumps
bull Obtain biomarkers for the early diagnosis of pathophysiological condition better understanding of the mechanisms that cause these conditions
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Cerebral blood flow regulation
Autoregulation homeostatic regulation of own blood flow Brain extremely effective autoregulationBrain extremely effective autoregulation
2 of body mass15 of total cardiac output 20 O2 consumption
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Assessment of dynamic autoregulationAssessment of dynamic autoregulation
Step response to inducedABP CO2 changesABP CO2 changes
Thigh cuff deflationHypercapnia hypocapnia
Spontaneous variability MABP
CBFV
Spontaneous variabilityNormal conditions all naturally occurring frequenciesCBFV variability correlated toΜΑBP
MABP
y CO2 variabilityLinear methods
High pass ABP‐CBFV characteristic[Zhang et al Amer J Physiol 1998]Low coherence lt 007 Hz Nonlinearities influence of CO2
Nonlinearmethods
22
Nonlinearmethods Larger fraction of CBFV variability explained [Mitsis et al Ann Biomed Eng 2002]
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Nonlinear model of cerebral h d ihemodynamics
ΑBP CO2
Nonlinear two‐input model of cerebral hemodynamics ΑBP
hellip hellip (1)1Lb
(1)jb
(1)0b hellip hellip (2)
ILb(2)jb
(2)0b
CO2of cerebral hemodynamicsInputs ABP PETCO2variations
Dynamic pressureautoregulation
DynamicCO2 reactivity
a at o s
Output CBFV variations
Simultaneous assessment of fKhellipf1dynamic pressure
autoregulation CO2reactivity
+
CBFV
reactivity
Includes MABP‐CO2interactions
23
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Cerebral hemodynamics under resting conditions
Experimental data14 subjects
conditions
Resting conditions 45 minsTraining data 6 min (360 points)Validation data 1 minValidation data 1 min
Systemorder
NMSE []MABP CO2 ΜΑBP amp CO2
2424
orderLinear (Q=1) 328plusmn132 716plusmn121 248plusmn111
Nonlinear (Q=3) 200plusmn92 515plusmn104 145plusmn69
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Cerebral hemodynamics under resting conditions Model predictionsconditions ndash Model predictions
Dynamic range VLF 0005‐004 Hz LF 004‐015 Hz HF 015‐030 HzNonlinearities prominent in VLFCO2 accounts mainly for VLF CBFV variations (lt004Hz) MABP dominates in HF
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Cerebral hemodynamics under resting conditions ndashLinear componentsLinear components
MABP HP characteristicSlow MABP changes regulated more g geffectively
PETCO2 LP characteristicSlow CO2 changes have more effect on CBFon CBF
26
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Cerebral hemodynamics under resting conditions ndashNonlinear components
Second order kernels
Nonlinear components
Second‐order kernelsMost power in VLF LF
Relative contribution of NL terms more significantsignificant
for PETCO2
MABP PETCO2
Nonlinear to linear terms power ratio
031plusmn013 118plusmn045
27
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Cerebral hemodynamics under resting conditions ‐nonstationarity
k1MABP k1 CO2
nonstationarity
1CO2
Tracking 6 min sliding data segments
28
Tracking 6‐min sliding data segments
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
From systemic to regional hemodynamics ‐functional MRIfunctional MRI
Neural activationRegional changes inblood flow oxygenationOxy Deoxygenated hemoglobin Different magnetic properties ‐ BOLD signalPrecise nature of neurovascular
2929
neurovascular coupling not known
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Respiratory network imaging Voxel‐wise l ianalysis
RestingRestingRestingResting
30EndEnd--tidal tidal forcingforcing
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Modeling of regional CO2 reactivityModeling of regional CO2 reactivity
bull Definition of anatomical and functional regions ofand functional regions of interest
bull CO2 reactivity Averaged O i i i hi OBOLD time‐series within ROI
AV thalamus
31
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Regional dynamic CO2 reactivityRegional dynamic CO2 reactivity
bull Significant regional variabilitybull Differential responses to small large CO changes (undershoot
32
bull Differential responses to small large CO2 changes (undershoot during resting only)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying h d h ll d (l l ) h d l
33
their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
bull This problem is also particularly relevant in neurological disorders such as epilepsy (seizure prediction)
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Functional brain connectivityFunctional brain connectivity
bull Brain function relies on multiple complex interconnections between different regionsbull Brain function relies on multiple complex interconnections between different regions that may may not be task specific
bull We can assess functional connectivity from EEGMEGfMRI measurements by quantifying their dynamic interactions with well‐suited (linear or nonlinear) methods eg correlation coherence directed coherence mutual information etc
34
coherence directed coherence mutual information etcbull Combination with graph‐theoretic measures bull This problem is also particularly relevant in neurological disorders such as epilepsy
(seizure prediction)
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Metabolic system ndash Diabetes and glucose b limetabolism
bull Diabetes mellitusndash Type 1 β cells do not produce insulinyp pndash Type 2 Insulin not utilized properlyndash Many long‐term implications
bull Interaction between key variables y(glucose insulin glucagon)ndash Currently ODE models specialized
experimental protocols (IVGTT)experimental protocols (IVGTT)ndash Continuous glucose sensors
insulin micropumpsbull Extraction of richer informationbull Extraction of richer informationbull Detection of subtle changes (Insulin secretion pattern sensitivity) improved early diagnosis
bull Model based glucose control DelaypreventionModel based glucose control Delayprevention of long‐term implications
35
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Glucose metabolism models in diabetic patientsGlucose metabolism models in diabetic patients
Mode 1 11NL 1
-10
0
10
0 1
0
01
02Mode 1
v1
z1
v1
z1NL 1
150
-8 -6 -4 -2 0 2 4-30
-20
0 100 200 300 400 500-03
-02
-01
Time [min]Insulin
v1
0 2 4 Glucose
v1
100Sensor Glucose [mg dl]
20
40
60
[ ]
-0 1
-005
0Mode 2
v2
z2
v2
z2NL 2
0 50 100 150 200 250 300 350 400 450 5000
50
Insulin uptake [microUnitsml]
-6 -4 -2 0-20
0
20
0 100 200 300 400 500-02
-015
01
v2
-6 -4 -2 00Time [min]
v2
0 50 100 150 200 250 300 350 400 450 500Time [min]
Time [min]
36
Glucose controlGlucose control
37
Glucose controlGlucose control
37