Research ArticleAn Efficient Combination among sMRI CSF Cognitive Score andAPOE ε4 Biomarkers for Classification of AD and MCI UsingExtreme Learning Machine

Uttam Khatri and Goo-Rak Kwon

Department of Information and Communication Engineering Chosun University 309 Pilmun-Daero Dong-GuGwangju 61452 Republic of Korea

Correspondence should be addressed to Goo-Rak Kwon grkwonchosunackr

Received 9 November 2019 Revised 13 January 2020 Accepted 17 February 2020 Published 4 June 2020

Guest Editor Horacio Rostro-Gonzalez

Copyright copy 2020 Uttam Khatri and Goo-Rak Kwon is is an open access article distributed under the Creative CommonsAttribution License which permits unrestricted use distribution and reproduction in anymedium provided the original work isproperly cited

Alzheimerrsquos disease (AD) is the most common cause of dementia and a progressive neurodegenerative condition characterized bya decline in cognitive function Symptoms usually appear gradually and worsen over time becoming severe enough to interferewith individual daily tasks us the accurate diagnosis of both AD and the prodromal stage (ie mild cognitive impairment(MCI)) is crucial for timely treatment As AD is inherently dynamic the relationship between AD indicators is unclear and variesover time To address this issue we first aimed at investigating differences in atrophic patterns between individuals with AD andMCI and healthy controls (HCs)en we utilized multiple biomarkers along with filter- and wrapper-based feature selection andan extreme learning machine- (ELM-) based approach with 10-fold cross-validation for classification Increasing efforts arefocusing on the use of multiple biomarkers which can be useful for the diagnosis of AD and MCI However optimumcombinations have yet to be identified and most multimodal analyses use only volumetric measures obtained from magneticresonance imaging (MRI) Anatomical structural MRI (sMRI) measures have also so far mostly been used separately e fullpossibilities of using anatomical MRI for AD detection have thus yet to be explored In this study three measures (corticalthickness surface area and gray matter volume) obtained from sMRI through preprocessing for brain atrophy measurementscerebrospinal fluid (CSF) for quantification of specific proteins cognitive score as a measure of cognitive performance andAPOE ε4 allele status were utilized Our results show that a combination of specific biomarkers performs well with accuracies of9731 for classifying AD vs HC 9172 for MCI vs HC 8791 for MCI vs AD and 8338 for MCIs vs MCIc respectivelywhen evaluated using the proposed algorithm Meanwhile the areas under the curve (AUC) from the receiver operatingcharacteristic (ROC) curves combining multiple biomarkers provided better classification performance e proposed featurescombination and selection algorithm effectively classified AD and MCI and MCIs vs MCIc the most challenging classificationtask and therefore could increase the accuracy of AD classification in clinical practice Furthermore we compared the per-formance of the proposed method with SVM classifiers using a cross-validation method with Alzheimerrsquos Disease NeuroimagingInitiative (ADNI) datasets

1 Introduction

Alzheimerrsquos disease (AD) is a progressive and irreversibleneurodegenerative disorder of the central nervous systemcharacterized by abnormal accumulation of neurofibrillarytangles and amyloid plaques in the brain affecting the be-havior thinking and memory of an individual [1]

Alzheimerrsquos disease occurs in late life and is the mostcommon form of dementia for which there is no cure Anestimated 57 million Americans are living with AD in 2018By 2050 this figure is projected to rise to nearly 14 million[2] Although some currently available treatments maytemporarily decelerate the progression none have demon-strable effectiveness in treating patients with AD A

HindawiComputational Intelligence and NeuroscienceVolume 2020 Article ID 8015156 18 pageshttpsdoiorg10115520208015156

promising amount of ongoing research [3ndash6] is focused ondifferent biomarker-based techniques in an effort to detectearly AD-related changes and characterize prominent at-rophy patterns during the prodromal stages when mildsymptoms are the only evidence of the disease us it isimportant to develop strategies to enable timely treatmentand delay progression during the early stages of AD beforethe onset of clinical symptoms As a result the concept ofmild cognitive impairment (MCI) was introduced MCI atransitional stage between healthy (normal) controls (HC)and AD patients is defined to describe people who havemildsymptoms of brain dysfunction but can still perform ev-eryday tasks Identifying potentially highly sensitive diag-nostic biomarkers that change with disease progression maysupport the physician in making a correct diagnosis If AD isdetected during the early stage of MCI the number ofpatients could be reduced by nearly one-third throughrehabilitation exercises and appropriate medication [7]Patients in the MCI stage have a high risk of progressing todementia [8ndash10] and can be categorized as stable MCI(MCIs) or convertible MCI (MCIc) which is also known asprogressive MCI (pMCI) Some MCI patients progress toAD within a specific time frame while others remain stableReports have shown that 10ndash15 of MCI patients progressto AD each year and 80 of these will have converted to ADafter approximately 5-6 years of follow-up [9 11] It iscrucial to find biomarkers that distinguish patients who haveMCI and later converted to AD (MCIc) from those who donot convert to AD andHCus early identification ofMCIindividuals is increasingly clinically important in potentiallydelaying or preventing the conversion from MCI to AD Toidentify biomarkers for MCI and AD various machinelearning methods have been applied which have improvedthe prediction and performance and more importantly thediscrimination of patients with MCIs from those MCI pa-tients who will progress to develop AD (MCIc) [12] Variousbiomarkers have been identified for the diagnosis of MCIand AD including functional and structural neuroimagingmeasures as well as cognitive score APOE ε4 allele statusand cerebrospinal fluid (CSF) markers e most recentcriteria for AD diagnosis [13] suggest that neuroimaging andbiological measures may play a vital role in the early de-tection of AD and the monitoring of the prodromal stage

Imaging biomarkers are considered important indicatorsin the diagnosis of AD and MCI With the development ofneuroimaging technology structural magnetic resonanceimaging (sMRI) techniques have become widely popular andcan be used to locate more subtle morphological changes inbrain disorders [14 15] For dementia patients MRI is usedin the standard clinical assessment ere have been a largenumber of studies aiming to identify imaging biomarkers forthe diagnosis of AD and the prediction of MCI progressionA majority of the well-established structural MRI bio-markers are mainly based on cerebral atrophy measure-ments or ventricular expansion Imaging biomarkers suchas cortical thickness [16ndash19] voxel-wise tissue probability[20ndash22] and volume [23ndash25] can show AD-associated at-rophy patterns and serve as effective biomarkers to classifyAD and MCI So far to our knowledge gradual cerebral

atrophy is one of the obvious and major changes in AD andthe pattern of atrophy can be analyzed via high-resolutionMRI technology Morphology-related cortical volumecortical thickness and cortical area measurements have beenutilized to better understand the fundamental pathophysi-ology of AD diagnosis However the majority of thesestudies [26ndash28] are mainly focused on differences in corticaland gray matter volumes Reference to surface area andother biomarkers are still lacking in this regard e com-bination of different cortical metrics (ie volume area andthickness) across multiple brain regions may better distin-guish between AD patients and HC erefore advancedmachine learning with multivariate approaches which canestablish the subtle relationship between multiple regionsand metrics [29 30] is potentially useful in assisting withprediction and diagnosis of AD Besides structural changesidentified by MRI other interesting biomarkers for ADdetection include CSF components cognitive score and thepresence of the APOE ε4 allele Numerous CSF cognitiveand APOE ε4 allele studies [12 31 32] have been carried outfor the classification of AD and MCI CSF biomarkers thathave been utilized in several studies include hyper-phosphorylated tau (P-tau) total tau (T-tau) and the Aβ42amino acid ese three CSF components provide valuableinformation for the identification of AD as patients haveabnormally low levels of Aβ42 and high levels of P-tau andT-tau [33] It has been shown that a combination of T-tauand CSF component measures provides outstanding clas-sification accuracy for separating HC fromAD patients withhigh sensitivity and specificity [34] Furthermore geneticrisk factors also impact the imaging and biological markersof AD classification Several previous studies [35] haveshown that the presence of a specific variant of the apoli-poprotein E gene (APOE) is a crucial risk factor associatedwith late-onset AD APOE has three majorsrsquo alleles ε2 ε3and ε4 In comparison with noncarriers AD patient carriersof the ε4 allele typically have low CSF Aβ42 and elevated CSFlevels of P-tau and T-tau along with accelerated atrophypatterns on MRI Various aspects of pathological patternsassociated with AD can be revealed by diverse biomarkersthus complementary biomarkers might assist diagnosisIt has been shown that a combination of different modal-ities of biomarkers can enhance diagnostic performance[25 36ndash40] Some recent papers of note [41ndash43] havedemonstrated the feasibility of machine learning ap-proaches One of the frequently usedmethods for solving theclassification problem is the support vector machine (SVM)A number of studies have applied the SVM for AD pre-diction and classification [24 39 44 45] In the field ofmachine learning deep learning has gained popularity andbecome a promising technology Deep learning relates tomultilevel representation learning and abstraction and hasresulted in a significant improvement in performance in thefield of data analysis and image classification In recent yearsuse of deep learning techniques for multimodal data analysisand classification has greatly increased For exampleseamless information is obtained using stacked autoen-coders from various types of media [46] To obtain jointrepresentation of text and images a multimodal deep belief

2 Computational Intelligence and Neuroscience

network was developed [47] Another study [48] proposed amultisource deep learning method to analyze human poseestimation A further study [49] developed a modal ofcombining MRI positron emission tomography (PET) andCSF modalities using stacked autoencoders to obtain au-tomatic classification of AD Backpropagation algorithmsare used to learn by most deep learning architectures whichiteratively adjust the parameters For this reason to reachgood generalization performance conventional neuralnetworks use many iterations [50] To overcome this situ-ation Huang et al [51] proposed an extreme learningmachine (ELM) in contrast to traditional methods withgood computational efficiency by randomly assigning weightin the input layer and analytically calculating hidden layerweights In another study [52] the authors used the dual-treecomplex wavelet transform (DTCWT) combined with ELMclassifiers and achieved good accuracy in AD classificationSimilarly in a further study [53] the authors used ELMclassifiers with multivariate pattern analysis to classify ADusing functional MRI (fMRI) data and achieved outstandingperformance e majority of the current literature regardsELM to be a good machine learning tool [54 55] ELMrsquosmajor strength is that the hidden layerrsquos learning parametersincluding the input weights and biases do not have to beiteratively tuned as in single hidden layer feedforward neural(SLFN) networks Because of this ELM costs less and iscapable of achieving faster speeds [54] Furthermore it is themost favored of in machine learning methods compared toits predecessors Some of the other commendable attributesof ELM include good generalization accuracy and perfor-mance a simple learning algorithm improved efficiencynonlinear transformation during the training phase pos-session of a unified solution to different practical applica-tions lack of local minimal and overfitting and the need forfewer optimizations as compared to SVM [55]erefore inthis study we were motivated to use an extreme learningmachine to achieve optimum classification accuracy in theidentification of AD

Our results using the Alzheimerrsquos Disease Neuro-imaging Initiative (ADNI) dataset (including both MCIpatients who did not convert to AD and MCI patients whoconverted to AD within 36 months) demonstrate the utilityof the suggested method Structural MRI data were firstlypreprocessed by FreeSurfer (version 600) to obtain threetypes of measures and statistical analysis was performedusing query design estimate contrast (QDEC) As well ascortical thickness and gray matter volume and surface areawe also utilized CSF markers APOE ε4 allele status andcognitive score To validate the effectiveness of our methodwe compared the classification performance with linear-SVM and RBF-SVM e general block diagram in Figure 1shows the workflow of the proposed method

2 Material and Methods

21 Data All data used in this analysis were obtained fromthe ADNI database e ADNI was initiated in 2003 aspublic-private partnership under the Principal InvestigatorMichael W Weiner MD e primary objective of ADNI is

to investigate whether imaging modalities such as MRIPET other neuropsychological assessments and clinical andbiological markers can be combined to for the early de-tection of AD and progression of the prodromal state (ieMCI) Demographic information raw neuroimaging dataCSF components APOE genotype diagnostic informationand neuropsychological test scores are publically available atthe ADNI data repository (httpadniloniuscedu) In-formed consent was obtained from all participants and thestudy was approved by the Institutional Review Board ofeach data site (for more information see httpadniloniusceduwp-contentthemesfreshnews-dev-v2documentspolicyADNI_Acknowledgement_List205-29-18pdf)

For this study we utilized MRI CSF and APOE ge-notype data e resulting study cohort included patientsaffected by AD patients with MCI and healthy controlsSociodemographical and clinical information of the par-ticipants is shown in Table 1

22 Data Acquisition Structural MRI data were acquiredusing either Siemens GE or Philips scanners at ADNIparticipating sites Since the image acquisition protocolsvaried for each scanner the image normalization steps wereprovided by ADNI Corrections included calibration ge-ometry distortion and intensity nonuniformity reductionDetailed information is available at the ADNI website(httpadniloniuscedu) ese corrections were appliedon each MPRAGE image following the image preprocessingsteps In this study we utilized T1-weighted images whichwere collected and reviewed for quality and correction interms of data format and alignment Finally images with256times 256times176 resolution and 1times 1times 1mm voxel size werecollected

CSF data were collected in the morning after overnightfasting with the use of a 20- or 24-G spinal needle Within 1hour of acquisition CSF was frozen and transported to theADNI core laboratory at the Medical Center of PennsylvaniaUniversity

e ADNI biomarker core laboratory also providedgenotype and gene expression data for each participant inthis study which were obtained from peripheral bloodsamples e genetic feature was a single categorical variablefor each participant taking one of five possible values (ε2ε3) (ε2 ε4) (ε3 ε3) (ε3 ε4) or (ε4 ε4) In this study wespecifically analyzed APOE ε4 allele status (carrier vsnoncarrier)

Cognitive score obtained from the Mini-Mental StateExamination (MMSE) at baseline was used as the measureof the patientrsquos cognitive performance

23 FreeSurfer Analysis of MRI We applied the recon-allFreeSurfer pipeline (version 600) which is freely accessibleat httpsurfernmrmghharvardedu to sMRI images forcortical reconstruction and volumetric segmentation [56]is pipeline automatically generated reliable volume andthickness segmentation of white matter gray matter andsubcortical volume Cortical reconstruction and subcorticalvolumetric segmentation include removal of nonbrain

Computational Intelligence and Neuroscience 3

network was developed [47] Another study [48] proposed amultisource deep learning method to analyze human poseestimation A further study [49] developed a modal ofcombining MRI positron emission tomography (PET) andCSF modalities using stacked autoencoders to obtain au-tomatic classification of AD Backpropagation algorithmsare used to learn by most deep learning architectures whichiteratively adjust the parameters For this reason to reachgood generalization performance conventional neuralnetworks use many iterations [50] To overcome this situ-ation Huang et al [51] proposed an extreme learningmachine (ELM) in contrast to traditional methods withgood computational efficiency by randomly assigning weightin the input layer and analytically calculating hidden layerweights In another study [52] the authors used the dual-treecomplex wavelet transform (DTCWT) combined with ELMclassifiers and achieved good accuracy in AD classificationSimilarly in a further study [53] the authors used ELMclassifiers with multivariate pattern analysis to classify ADusing functional MRI (fMRI) data and achieved outstandingperformance e majority of the current literature regardsELM to be a good machine learning tool [54 55] ELMrsquosmajor strength is that the hidden layerrsquos learning parametersincluding the input weights and biases do not have to beiteratively tuned as in single hidden layer feedforward neural(SLFN) networks Because of this ELM costs less and iscapable of achieving faster speeds [54] Furthermore it is themost favored of in machine learning methods compared toits predecessors Some of the other commendable attributesof ELM include good generalization accuracy and perfor-mance a simple learning algorithm improved efficiencynonlinear transformation during the training phase pos-session of a unified solution to different practical applica-tions lack of local minimal and overfitting and the need forfewer optimizations as compared to SVM [55]erefore inthis study we were motivated to use an extreme learningmachine to achieve optimum classification accuracy in theidentification of AD

Our results using the Alzheimerrsquos Disease Neuro-imaging Initiative (ADNI) dataset (including both MCIpatients who did not convert to AD and MCI patients whoconverted to AD within 36 months) demonstrate the utilityof the suggested method Structural MRI data were firstlypreprocessed by FreeSurfer (version 600) to obtain threetypes of measures and statistical analysis was performedusing query design estimate contrast (QDEC) As well ascortical thickness and gray matter volume and surface areawe also utilized CSF markers APOE ε4 allele status andcognitive score To validate the effectiveness of our methodwe compared the classification performance with linear-SVM and RBF-SVM e general block diagram in Figure 1shows the workflow of the proposed method

2 Material and Methods

21 Data All data used in this analysis were obtained fromthe ADNI database e ADNI was initiated in 2003 aspublic-private partnership under the Principal InvestigatorMichael W Weiner MD e primary objective of ADNI is

to investigate whether imaging modalities such as MRIPET other neuropsychological assessments and clinical andbiological markers can be combined to for the early de-tection of AD and progression of the prodromal state (ieMCI) Demographic information raw neuroimaging dataCSF components APOE genotype diagnostic informationand neuropsychological test scores are publically available atthe ADNI data repository (httpadniloniuscedu) In-formed consent was obtained from all participants and thestudy was approved by the Institutional Review Board ofeach data site (for more information see httpadniloniusceduwp-contentthemesfreshnews-dev-v2documentspolicyADNI_Acknowledgement_List205-29-18pdf)

For this study we utilized MRI CSF and APOE ge-notype data e resulting study cohort included patientsaffected by AD patients with MCI and healthy controlsSociodemographical and clinical information of the par-ticipants is shown in Table 1

22 Data Acquisition Structural MRI data were acquiredusing either Siemens GE or Philips scanners at ADNIparticipating sites Since the image acquisition protocolsvaried for each scanner the image normalization steps wereprovided by ADNI Corrections included calibration ge-ometry distortion and intensity nonuniformity reductionDetailed information is available at the ADNI website(httpadniloniuscedu) ese corrections were appliedon each MPRAGE image following the image preprocessingsteps In this study we utilized T1-weighted images whichwere collected and reviewed for quality and correction interms of data format and alignment Finally images with256times 256times176 resolution and 1times 1times 1mm voxel size werecollected

CSF data were collected in the morning after overnightfasting with the use of a 20- or 24-G spinal needle Within 1hour of acquisition CSF was frozen and transported to theADNI core laboratory at the Medical Center of PennsylvaniaUniversity

e ADNI biomarker core laboratory also providedgenotype and gene expression data for each participant inthis study which were obtained from peripheral bloodsamples e genetic feature was a single categorical variablefor each participant taking one of five possible values (ε2ε3) (ε2 ε4) (ε3 ε3) (ε3 ε4) or (ε4 ε4) In this study wespecifically analyzed APOE ε4 allele status (carrier vsnoncarrier)

Cognitive score obtained from the Mini-Mental StateExamination (MMSE) at baseline was used as the measureof the patientrsquos cognitive performance

23 FreeSurfer Analysis of MRI We applied the recon-allFreeSurfer pipeline (version 600) which is freely accessibleat httpsurfernmrmghharvardedu to sMRI images forcortical reconstruction and volumetric segmentation [56]is pipeline automatically generated reliable volume andthickness segmentation of white matter gray matter andsubcortical volume Cortical reconstruction and subcorticalvolumetric segmentation include removal of nonbrain

Computational Intelligence and Neuroscience 3

tissues Talairach transformations segmentation of sub-cortical gray and white matter regions intensity standard-ization and Atlas registration After these steps a corticalsurface mesh model was generated and finally the 34cortical regions were obtained from cortical surface par-cellation based on sulcal and gyral landmarks for bothhemispheres corresponding to Desikan et al [57] For sta-tistical analysis purposes smoothing was carried out usingrecon-all with the qcache option in FreeSurfer e QDECtool within FreeSurfer was utilized to analyze differences in

cortical thickness surface area and gray matter volumebetween HC MCI and AD individuals Statistical signifi-cance levels were corrected for both hemispheres using thefalse discovery rate (FDR) plt 005 to control for multiplecomparisons [58]

24 Machine Learning-Based Prediction and Analysis Anoverview of the prediction framework developed for thisstudy is shown in Figure 1 e framework consists of four

FreeSurfer-basedpreprocessing of structural

MRI images

-Parcellation into ROIs-Cortical segmentation

-Subcortical segmentaion-White matter segmentaion

Features combination

Features normalization

Statistical analysis Qdec surface group analysis

Cognitive score

Cortical thickness surface area and gray

matter volume

CSF

91 230 109103 49 127

Genetic

APOE allele Aβ42 T-tau and P-tau MMSE

Cognition

sMRI

Performance evaluation(ACC SEN SPE AUC)

9 subsets for training

1 subset for testing

10-fold cross-validation

SVM ELM

Filter + wrapper

Figure 1 Block diagram of the proposed framework

4 Computational Intelligence and Neuroscience

major steps feature extraction feature combination nor-malization and feature selection and classification We usedtwo machine learning classification algorithms SVM andELM

25 Feature Selection e feature selection algorithm is anessential part of a machine learning approach facilitatingdata understanding reducing storage requirements andtraining-testing times and improving the accuracy ofclassification Importantly feature selection was performedusing only the training dataset and then applied on the testset Before feature selection we performed feature nor-malization and all feature sets were normalized to unitvariance and zero mean to reduce redundancy and improvedata integrity between the feature sets For a given datamatrix X columns represent features and rows represent theparticipantsrsquo normalized matrix Xnorm with an element (i j)which was calculated as shown in equation (1) After featurenormalization we used a combination of filter and wrapperalgorithms for feature selection We used a filter method andsorted features based on their minimum redundancymaximum relevance (MRMR) scores MRMR has beenpreviously described [59] e MRMR score for a feature setS is defined in equation (2)

Xnorm x(ij) minus mean Xj1113872 1113873

std Xj1113872 1113873 (1)

MRMR MAXs

1|s|

1113944fiisinS

I fi c( 1113857 minus1

|S|21113944 I fi fj1113872 1113873

⎧⎪⎨

⎪⎩

⎫⎪⎬

⎪⎭

(2)

where the relevance of a feature set S for k classes C

c1 c2 ck1113864 1113865 is defined by the average value of mutualinformation between the individual feature fi andC and theredundancy of all features in the feature set S is the averagevalue of mutual information between features fi and fj etop 60 features identified by the MRMR algorithm were usedin a wrapper algorithm to find an optimal subset of featuresWe developed and validated a sequential feature selection(SFS) algorithm as a wrapper feature selection method forthis study e SFS algorithm has been previously described

in detail [60] Briefly different subsets of features wereselected from the top 60 features identified by the MRMRalgorithm and then the accuracy of the ELM classifiers basedon these subsets of features was calculated

26 SVM Classifier Generally the SVM [61] is a binaryclassifier which is applicable to both separable and non-separable datasets It has been successfully utilized in theneuroimaging field and has become the most popular ma-chine learning algorithm in the field of neuroscience duringthe past decadee SVM is a supervised classifier that uses atraining dataset to find an optimal separating hyperplane inan n-dimensional space e optimal hyperplane is one thatbest separates the two target participant groups In ourstudy we utilized both linear SVM and nonlinear SVMbased on radial basis function (RBF) kernels An RBF kernelperforms better than a linear kernel for a small number offeature sets A regularization constant C and a set of kernelhyperparameters c (gamma) need to be tuned in SVMsese parameters were optimized using a cross-validation(CV) method is procedure was repeated 1000 times eachtime randomly selecting a new set of 10 held-out participantsto obtain optimum hyperparameters optimization In thismethod the search scales for regularization constant andgamma values were set to C (0001 001 01 1 10 1001000) and c (0001 001 01 1 10 100 1000) respectivelyemaximum validation accuracy was obtained at C 1 andc 01 e tuned parameters were used to predict the ac-curacy value on the test dataset

27 ELM Classifier e extreme learning machine iscomposed of a hidden layer in between the input and outputlayers [51] Whereas weights and biases are required to foradjustment by gradient-based learning algorithms on tra-ditional feedforward neural networks for all layers in theELM hidden layer biases and input weights are arbitrarilyassigned without iterative processes and output weights arecomputed by solving a single hidden layer system [50]uscompared to traditional neural networks the ELM learnsmuch faster and it is widely used in various regression andclassification tasks being an efficient and reliable learningalgorithm [62ndash65] Particularly for N training samples(X(j) I(j))|X(j) isinRp and I(j) isinRq and j 1 2 Nthe output in ELM oj with nh hidden neurons can berepresented as shown in the following equationfd3

oj 1113944

nh

i1βT

i a wTi X

(j)+ bi1113872 1113873 1113944

nh

i1βT

i hi X(j)

1113872 1113873 h X(j)

1113872 1113873Tβ

(3)

where X(j) and I(j) are the jth input and target vectorsrespectively e parameters p and q are the input and targetvector dimensions respectively Additionally oj isinR

q sig-nifies the output of the ELM for the jth training samplewi isinR

p indicates the input weight that links the inputnodes to the ith hidden node bi represents the bias of the ith

hidden node and a(middot) signifies the activation function forthe given hidden layer β [β1 βnh

]T are the values of

Table 1 Baseline clinical and sociodemographical information ofthe participants of this study

Group AD MCI HC MCIs MCIcNo ofparticipants 53 77 57 35 42

Femalemale 2033 3443 3225 1322 2121

Age 744plusmn 78 741plusmn 72 756plusmn 52 739plusmn 72 743plusmn 72Education 151plusmn 32 159plusmn 29 157plusmn 28 161plusmn 29 158plusmn 29MMSE 235plusmn 18 269plusmn 18 291plusmn 09 272plusmn 17 266plusmn 18CDR 07plusmn 02 05 0 05 05e entries for age gender education and MMSE denote mean andstandard deviation for each group MMSE Mini-Mental State Exam CDRclinical dementia ratio

Computational Intelligence and Neuroscience 5

the output weights between the output neurons and thehidden layer h(X(j)) [h1(X(j) hnh

(X(j)))]T is theoutput vector of the hidden layer with respect to the jth

training sample X(j) middot hi(X(j)) is the output of the ith hiddenlayer for the jth training sample To obtain the optimalhidden layer weights 1113954β with respect to N training samplescan be considered to solve the following optimizationproblem

minβ

λ Hβ minus L2

+β2 (4)

where H [h(X(1) h(X(N)))]T and L [I(1)

I(N)]TEquation (4) represents the optimization problem and

its optimal solution 1113954β can be analytically obtained as follows

1113954β HT 1

λI + HH

T1113874 1113875

minus1L (5)

where λ is a regularization parameter and I represents theidentity matrix After finding the optimal solution 1113954β theoutput of the ELM on test data Xtest is determined asfollows

otest h Xtest( 1113857TH

T 1λ

I + HHT

1113874 1113875minus1

L (6)

In this proposed method the hidden node number wasset between 1 and 500 and we selected a sigmoid as anactivation function Further we used a grid searchmethod totune the ELM parameter on the training dataset in order toachieve optimum cross-validated validation accuracySimilarly to minimize the random effects during the weightinitializations each parameter of the number of hiddennodes was used 100 times and the average performance wascalculated

28 Cross-Validation and Performance Evaluation We usedthe k-fold cross-validation (KCV k 10) method for cross-validation All participants were randomly divided into 10equally sized subsets using the KCV (k 10) cross-validationapproach In each fold of the KCV 90 of the data were usedto train the model based on a subset of features and then10 of the data (cross-validation set) were used to calculatethe accuracy of the ELM classifier Accuracies of the ELMclassifiers corresponding to all subsets of features werecalculated and classified with maximum accuracy and thecorresponding optimal subset of features was identifiedSimilarly classification performance was evaluated on ac-curacy (ACC) specificity (SPE) and sensitivity (SEN) TPFP FN and TN represent the number of true positives falsepositives false negatives and true negatives respectively Interms of numerical values ACC SPE and SEN can becalculated as follows

accuracy(ACC) (TN + TN)

(TP + TN + FP + FN) (7)

sensitivity(SEN) (TP)

(TP + FN) (8)

specificity(SPE) (TN)

(TN + FP) (9)

e other effective way to evaluate results for a classifieris the receiver operating characteristic (ROC) curve eROC curve is the plot of true-positive rate against false-positive rate by changing the discrimination threshold andtherefore summarizing the classifierrsquos performance eROC curve is usually represented by the area under the curve(AUC) which is denoted by a number between 0 and 100

3 Results and Analysis

31 Statistical Analysis Cortical thickness gray matter(GM) volume and surface area were analyzed using asurface-based group analysis in FreeSurferrsquos QDEC (version15) First the spatial cortical thickness GM volume andsurface area of both hemispheres were smoothed with acircularly symmetric Gaussian kernel of 10mm full-widthhalf-maximum to normally distribute the results en weemployed a general linear model (GLM) analysis with agesex and education as the nuisance factors in the designmatrix to directly compare the three parameters in bothhemispheres of the AD vs HC HC vs MCI AD vs MCIand MCIs vs MCIc groups Statistical analysis results re-garding cortical thickness surface area and gray mattervolume are shown in Figure 2 e DesikanndashKilliany Atlasdivides the human cerebral cortex into 34 cortical regions ineach hemisphere As there were a high number of atrophicregions we present only the top-ranked regions with sig-nificant differences e atrophic regions for three param-eters are listed below

Table 2 presents the atrophy position and range ofclusters for the differences in gray matter volume corticalthickness and surface area at each vertex between HC MCIand AD groups by QDEC analysis In this table only the topfeatures which have significant cluster differences for eachkind of parameter are provided From the statistically sig-nificant brain regions shown in Figure 2 and Table 2 weobserve the following

(1) e cortical thicknesses of the left insula left cuneusparacentral right rostral middle frontal and rightpars opercularis areas were thinner in the AD groupcompared with the HC group For HC vs MCI thecortical thicknesses of the left precuneus left lingualleft and right insula right pars triangularis and rightinferior parietal areas were thinner Similarly for ADvs MCI the cortical thicknesses of the left inferiorparietal right lateral occipital right inferior parietaland left and right superior temporal areas showedthe most atrophy For MCIs vs MCIc the corticalthicknesses of the left inferior parietal left para-hippocampal right temporal pole and right superiortemporal areas showed the greatest differences

(2) Regarding surface area the AD group had smallervalues than the HC group in the left and rightparacentral left lateral orbitofrontal right inferiorparietal right posterior cingulate and right inferior

6 Computational Intelligence and Neuroscience

parietal areas For HC vs MCI the left superiorfrontal left postcentral left superior parietal rightsupramarginal right fusiform right precuneus andright precentral areas showed the lowest valuesSimilarly for AD vs MCI the left superior parietalleft and right precentral left paracentral right in-ferior parietal and right superior frontal areas

showed the largest decreases in surface area ForMCIs vs MCIc the left fusiform left lateral occipitalleft parahippocampal right superior parietal andright postcentral areas showed the most differences

(3) In comparison with the HC group the volume ofgray matter of the left caudal anterior cingulate leftorbitalis left lateral orbitofrontal right lateral

icknessLe hemisphere Right hemisphere

VolumeRight hemisphereLe hemisphere

text

AreaLe hemisphere Right hemisphere

ndash500 500ndash250 250000

AD

vs

MCI

MCI

s vs

MCI

cA

D v

s H

CH

C vs

MCI

Figure 2 Differences in cortical thickness area and volume in patients at different stages of Alzheimerrsquos diseasee colored bar representsthe significance level of clusters e significance threshold was set at plt 005

Computational Intelligence and Neuroscience 7

Table 2 Cluster differences in cortical thickness area and volume in AD MCI and MCIc patients

Feature RegionCoordinates

Vertex Value Size (mm2)x y z

AD vs HC

ickness

Left insula minus295 179 119 3165 minus23512 361947Left parahippocampal minus313 minus417 minus87 557 minus22303 1073627

Left cuneus minus479 minus209 428 1039 minus23188 282428Right rostral middle frontal 385 43 7 1224 minus21737 148Right superior temporal 533 minus52 minus52 468 minus28676 25628Right pars opercularis 503 102 88 1242 minus36071 5815525

Area

Left paracentral minus158 minus355 494 1197 minus22767 701872Left lateral orbitofrontal minus533 minus2 75 489 minus21962 321823

Right paracentral 125 minus371 554 1233 minus22555 4255Right inferior parietal 34 minus519 374 1133 minus21026 721305

Right posterior cingulate 14 minus304 366 1201 minus20979 141667Right inferior parietal 395 minus447 346 7 minus20263 5076

Volume

Left caudal anterior cingulate minus5 184 264 2073 minus32264 83461Left orbitalis minus287 minus602 417 1103 23437 120255

Left lateral orbitofrontal minus594 minus69 94 412 minus20146 63567Right lateral orbitofrontal 304 229 minus203 631 minus27411 20256Right rostral middle frontal 348 51 47 198 minus2717 109882

Right pars opercularis 468 152 87 136 minus2717 80771AD vs MCI

ickness

Left inferior parietal minus436 minus615 33 939 27022 43937Left fusiform minus40 minus537 minus203 254 2459 14984

Left superior temporal minus477 minus25 minus91 210 minus25577 32803Right lateral occipital 261 minus939 37 363 34944 363Right inferior parietal 377 minus717 428 1129 33654 64624Right superior temporal 507 minus143 minus24 722 minus29715 34228

Area

Left superior parietal 284 minus51 427 1702 minus25085 64276Left precentral 322 minus218 608 109 18969 5154Left paracentral 141 minus367 524 238 minus18959 7968

Right inferior parietal minus352 minus872 137 50 minus18099 3453Right precentral minus542 minus11 71 51 minus17244 2134

Right superior frontal minus74 4 665 34 16541 1786

Volume

Left superior parietal minus285 minus588 40 540 35868 2161Left superior frontal minus85 412 303 48 minus26334 3398

Left caudal anterior cingulate minus49 113 321 314 minus2549 15007Right entorhinal 215 minus89 minus295 207 minus31433 5569

Right lateral orbitofrontal 428 278 minus137 198 minus31034 12501Right superior parietal 321 minus442 412 115 minus25309 4029

HC vs MCI

ickness

Left insula minus364 minus94 minus117 811 minus35988 30994Left precuneus minus6 minus689 417 628 23682 32202Left lingual minus293 minus445 minus66 612 minus27175 27916Right insula 358 minus106 minus74 1154 minus31385 43662

Right pars triangularis 389 317 1 372 minus21264 20487Right inferior parietal 351 minus72 409 285 minus27121 1522

Area

Left superior frontal minus178 315 496 237 17293 15688Left postcentral minus521 minus227 507 264 17286 12275

Left superior parietal minus32 minus493 473 141 minus17572 6604Right supramarginal 534 minus474 35 1336 27163 67685

Right fusiform 347 minus12 minus341 551 20892 32284Right precuneus 182 minus773 277 482 17491 31336Right precentral 207 minus306 539 328 minus19165 11545

Volume

Left supramarginal 521 minus463 226 441 27521 21023Left cuneus 171 minus695 168 628 2611 48174

Left precentral 212 minus307 538 376 minus21546 12702Right parahippocampal minus238 minus361 minus156 278 minus18212 12768Right superior parietal minus91 minus743 465 598 25199 30367Right superior frontal minus181 333 396 210 26035 11472

8 Computational Intelligence and Neuroscience

orbitofrontal right rostral middle frontal and rightpars opercularis areas was lower in the AD groupFor AD vs MCI the left superior parietal left su-perior frontal left caudal anterior cingulate rightentorhinal and right superior parietal areas showedthe largest decreases in volume For HC vs MCI theleft supramarginal left cuneus left precentral rightparahippocampal and right superior parietal areasshowed the most volume atrophy Similarly forMCIs vs MCIc the left bankssts left precentral leftrostral middle frontal right precentral and rightfusiform areas had the largest decreases in volume

Moreover from this analysis we observe that areathickness and volume in the AD group were significantlydecreased in comparison with the HC group Similarly therewas significant atrophy in the MCIc group in comparisonwith the MCIs group and there was little difference in theatrophy pattern among AD and MCI patients

32 Feature Selection and Classification e individualfeature set was selected using an MRMR (filter) and a se-quential feature selection (wrapper) method to identify theoptimal feature set for different groups and improve theclassifier accuracy Similarly we combined all the feature setsfrom different measures and applied the MRMR and SFSmethod to select the optimal features Multiple measuresfeature sets were created by combining cortical thicknessarea and volume from sMRI three component measure-ments from CSF APOE ε4 status andMMSE score First weselected the top 60 features from theMRMR algorithm basedon the features score and then we applied a sequentialfeature selection algorithm on these top 60 ranked featureswhich gave the optimal feature set to achieve the maximumclassification accuracy on ELM classifiers Figure 3 shows the

number of optimal features sequence after sequential featureselection on the top 60 ranked features obtained from theMRMR algorithm using ELM classifier Figure 3 shows onlythe selected features for the combined feature set From thiswe can assume that the proposed feature selection withcross-validation provides the optimal feature vector forinput to the classifiers In this proposed method classifi-cation performance was quantified by the number of featuresselected versus accuracy and area under the ROC curve

To further analyze the effectiveness of the purposeclassification method combining different measures wecalculated the AUCs for the concatenation of all featuresFigure 4shows the receiver operating curves for individualfeatures and all feature (imaging and nonimaging bio-markers ie multiple features) combinations for eachclassification group using the ELM classifier

4 Discussion

41 Performance Analysis In this study we first performedstatistical analysis and pattern classification to differentiateand identify atrophy patterns for the four groups (AD HCMCIs and MCIc) Individual sMRI was preprocessed usingthe FreeSurfer tool After preprocessing the statisticalanalysis results of sMRI QDEC was applied and finally weperformed the classification task by the proposed featureselection and classification method respectively For brainatrophy analysis we used three types of sMRI corticalmetrics (cortical thickness gray matter volume and surfacearea) In comparing the AD group with the HC group theinsula pars opercularis parahippocampal and superiortemporal areas were severely affected in terms of corticalthickness gray matter volume and surface area e corticalthickness of the left hemisphere was thinner than that of theleft hemisphere Similarly for AD vs MCI the most atrophy

Table 2 Continued

Feature RegionCoordinates

Vertex Value Size (mm2)x y z

MCIs vs MCIc

ickness

Left inferior parietal minus463 minus60 111 2770 minus37613 142748Left superior temporal minus455 minus04 minus208 1067 minus26742 46624Left parahippocampal minus317 minus403 minus101 550 minus23925 23985Right temporal pole 292 9 minus38 1449 minus55983 82282

Right superior temporal 623 minus347 152 1292 minus31646 54997Right inferior temporal 516 minus566 minus37 882 minus30694 50791

Right precentral 488 minus65 405 858 minus28315 35193

Area

Left fusiform minus364 minus295 minus22 659 minus30958 31221Left lateral occipital minus194 minus99 minus151 326 26022 25061Left parahippocampal minus188 minus335 minus14 224 minus20347 9759Right superior temporal 481 minus329 22 1515 minus25815 5798Right superior parietal 106 minus527 651 1697 minus22367 64152

Right postcentral 338 minus29 519 799 minus20859 36723

Volume

Left bankssts minus579 minus469 minus1 2196 minus26263 116258Left precentral minus361 minus22 517 2736 minus25339 10731

Left rostral middle frontal minus224 279 328 582 minus23081 32686Right superior temporal 623 minus347 152 3392 minus44677 144996

Right precentral 495 minus62 411 4103 minus32472 181394Right fusiform 339 minus378 minus228 782 minus31822 42505

Computational Intelligence and Neuroscience 9

was seen in the left inferior parietal and right lateral occipitalareas For HC vs MCI the supramarginal and cuneus areasshowed the most atrophy For MCIs vs MCIc the superiortemporal and precentral areas showed the largest decreasesin thickness volume and area Based on these differencesthe majority of the decrease in cortical thickness gray mattervolume and surface area appears mostly in the frontal lobetemporal lobe occipital lobe cingulate gyrus and parietallobe is phenomenon strongly agrees with findings relatedto atrophy patterns seen in previous studies [14 66] eseregions are mainly involved in the expression of personalitymotor execution complex cognitive behavior and decisionmaking [50] In addition we present the less commonanalysis of MCIs vs MCIc For MCIc the most atrophy wasseen in the superior temporal bankssts precentral inferior

parietal and insula areas which show potential for the earlyrecognition of progression to Alzheimerrsquos disease

To test the effectiveness of our purposed featurescombination method (ie imaging and nonimaging fea-tures) we adapted the individual feature from sMRI and CSFseparately to conduct the study although regarding geneticand cognitive features we combined them to test the per-formance and then compared the accuracy with the accuracyof the combination of all features For the proposed featureselection we used a combination of filter and wrapper al-gorithms and compared the performance of the classifiers onthe selected feature set In this text SVM-linear SVM-RBFand ELM were used for classification e classificationresults for the compared methods are presented inTables 3ndash6 and also in graphical form in Figure 5 As shown

AD vs HC

Number of selected features = 27

05

101520253035404550556065707580859095

100

Accu

racy

()

5 10 15 20 25 30 35 40 45 50 55 600

Number of features sequence

(a)

AD vs MCI

Number of selected features = 17

05

101520253035404550556065707580859095

100

Accu

racy

()

5 10 15 20 25 30 35 40 45 50 55 600

Number of feature sequence

(b)HC vs MCI

Number of selected features = 25

5 10 15 20 25 30 35 40 45 50 55 600

Number of feature sequence

05

101520253035404550556065707580859095

100

Accu

racy

()

(c)

MCIs vs MCIc

Number of selected features = 13

05

101520253035404550556065707580859095

100Ac

cura

cy (

)

5 10 15 20 25 30 35 40 45 50 55 600

Number of features sequence

(d)

Figure 3 Number of feature sets selected from the sequential feature selection algorithm on the top 60 ranked features obtained using theMRMR algorithm Only the number of features versus accuracy for the combination of all feature sets using ELM classifiers is shown (a) ADvs HC feature subset (b) AD vs MCI feature subset (c) HC vs MCI feature subset (d) MCIs vs MCIc feature subset

10 Computational Intelligence and Neuroscience

AD vs HC

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(a)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

AD vs MCI

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(b)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

HC vs MCI

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(c)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

MCIs vs MCIc

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(d)

Figure 4 ROC curves for discriminating AD MCIs MCIc and HC (healthy controls) ROC curves are plotted for the five biomarkersseparately and for the concatenation of all features (ie all five biomarkers measures) (a) ROC curves for AD vs HC classification (b) forAD vs MCI classification (c) for HC vs MCI classification and (d) for MCIs vs MCIc classification using two-stage feature selection withELM classifiers

Computational Intelligence and Neuroscience 11

in the tables the performance accuracy of feature fusion isnoticeably improved when compared with that of the in-dividual feature set and there are distinct degrees of ele-vation in other indexes the specificity index is particularlymore noticeable In contrast performance accuracy basedon the nonimaging features is almost the same as the im-aging features for AD diagnosis but far superior for MCIdiagnosis For AD vs HC the accuracy obtained by CSFmeasures genetics and cognitive score showed a lowerincrease although for MCI vs HC AD vs MCI and MCIsvs MCIc the accuracy was higher e result showed thatthe ELM classifier achieves better classification scorescompared to the SVM classifier (linear and RBF-SVM) for

both single modality feature sets as well as all featurescombined (shown in Figure 6) e ELM is good machinelearning tool and the major strength is that the hiddenlayerrsquos learning parameters including input weight andbiases do not tune iteratively as in SLFN e ELM offersmany advantages over other learning algorithms [54] Fromthe results shown in Tables 3ndash6 for AD vs HC the clas-sification result increased by 5ndash7 as compared to SVMclassifiers with 9804 sensitivity and 9628 specificityFor AD vs MCI classification accuracy increased by 5ndash9with 92 sensitivity and 9733 accuracy For HC vs MCIclassification accuracy increased by 6ndash10 with 9123sensitivity and 9913 specificity Similarly for MCIs vs

Table 3 10-fold cross-validated classification performance for AD vs HC

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 8513 9207 8544 086 8068 8568 8381 7717 8317 7381Surface area 8230 8930 9154 085 7616 8616 7514 7867 8767 7262Volume 8790 9387 8475 088 7929 7429 7833 8000 7710 6324CSF 8973 9633 8738 089 7905 8905 7417 7533 8333 6857APOE+MMSE 8645 9513 8700 093 8203 8703 8467 8416 8305 7879Concatenation (all_features_set) 9731 9804 9628 097 9133 9333 8757 9350 955 9058

Table 4 10-fold cross-validated classification performance for AD vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7010 8130 6610 071 6401 8383 6233 6503 7880 7330Surface area 6160 6670 8430 063 6345 6810 8011 6067 7123 6285Volume 6780 8210 6710 069 5820 7792 6856 6005 7337 6373CSF 6470 6600 8140 073 5607 6275 7573 6123 6871 7937APOE+MMSE 8145 8440 8770 082 7681 6781 8548 7620 6694 8748Concatenation (all_features_set) 8791 9200 9733 089 8350 8864 7672 7831 7445 8262

Table 5 10-fold cross-validated classification performance for HC vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7540 8407 6790 083 7417 8081 7123 6734 7317 7793Surface area 7415 6520 8385 084 6354 6767 6647 6872 7443 6735Volume 7784 8841 7280 084 6547 7473 6875 7283 7827 6570CSF 8330 7393 8402 085 7339 7525 6683 7580 7338 7814APOE+MMSE 7937 9325 7297 089 6829 8173 7668 7023 8124 7805Concatenation (all_features_set) 9172 9123 9913 093 8170 8368 8749 8565 7854 8926

Table 6 10-fold cross-validated classification performance for MCIs vs MCIc

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 6371 7110 3837 064 5718 6553 6038 5005 7118 6813Surface area 6143 7228 7480 069 5425 6322 7617 5827 7223 6485Volume 6785 8440 8773 074 5917 7828 6755 6214 7445 5773CSF 7137 7884 6857 072 6473 6009 7278 6733 7882 6878APOE+MMSE 7609 8713 7209 079 6553 6881 6743 6808 7221 6633Concatenation (all_features_set) 8338 9301 7577 085 7137 8116 7675 7571 7838 8467

12 Computational Intelligence and Neuroscience

MCIc classification accuracy increased by 7ndash12 with9301 sensitivity and 7577 specificity but 8467 and7667 specificity was obtained for linear and RBF-SVMclassifiers respectively which is more than obtained by theELM classifier From our analyses we believe that the ELM

gives better performance as compared to SVM from alearning efficiency standpoint because the ELMrsquos originaldesign has high learning accuracy fast learning speedscalability and the least human intervention [54 55] Fromthe results shown in Table 7 we can see that our suggested

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(a)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(b)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(c)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(d)

Figure 5 Classification results for different Alzheimerrsquos groups using ELM classifiers (a) Classification results for AD vs HC groups (b) forAD vs MCI groups (c) for HC vs MCI groups and (d) for MCIs vs MCIc groups

Computational Intelligence and Neuroscience 13

method achieved better accuracy than other existingmethodsFor classifying AD and HC our method achieved a classi-fication accuracy of 9731 with a sensitivity of 9804 aspecificity of 9628 and an AUC value of 097 Forclassifying MCI and HC our method achieves a classificationaccuracy of 9172 with a sensitivity of 9123 a specificity of9913 and an AUC value of 093 For distinguishing ADfrom MCI our proposed method obtained a classificationaccuracy of 8791 with a sensitivity of 9200 a specificityof 9733 and an AUC value of 089 Similarly for MCIs vsMCIc we achieved outstanding performance as compared toprevious methods by combining multiple features with anaccuracy of 8338 a sensitivity of 9301 a specificity of7577 and an AUC value of 085

42ComparisonwithOtherMethods In the previous sectionwe discussed in detail the findings of the proposed featureselection and classification method In this section we

compare and discuss the findings of our method in com-parison with existing state-of-the-art methods Tables 7 and 8show a comparison of the classification performance of theproposed method with recently published studies which usedmultimodality data to distinguish individuals with AD andMCI from HC Westman et al [67] used MRI and CSFbiomarkers and obtained 918 accuracy for AD vs HC776 for HC vs MCI and 685 for pMCI vs MCIsclassification using the orthogonal partial least squares tolatent structures (OPLS) method Zhang and Shen [39] used amultimodal (MRI PET and CSF) classification method withmultitask feature selection and an SVM classifier for AD andMCI classification By combining MRI PET and CSF datathey achieved a higher accuracy of 933 for AD vs HC and832 accuracy for HC vs MCI classification Similarly theauthors achieved an accuracy of 739 on pMCI vs MCIssamples Hinrichs et al [69] obtained an accuracy of 924 forAD vs HC classification using MRI PET CSF APOE and

ELMSVM_RBFSVM-linear

0

20

40

60

80

100

ACC

()

AD vs HC HC vs MCI MCIs vs MCIcAD vs MCI

Figure 6 Performance comparison of different classifiers

Table 7 Comparison of classification of AD vs HC and HC vs MCI between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection AD vsHC ()

HC vsMCI

Westman et al[67] MRI +CSF OPLS mdash 918 776

Johnson et al [68] MRI + PET+CSF+ cognitive scores Stackedautoencoder

Sparse representationlearning 959 85

Hinrichs et al [69] MRI + PET+CSF+APOE+ cognitive scores MKL mdash 924 naZhang and Shenet al [39] MRI + PET+CSF SVM Multitask feature selection 933 832

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 9301 mdash

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 995 mdash

Maqsood et al[72] sMRI CNN mdash 9285 mdash

Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper(SFS) 9731 9172

14 Computational Intelligence and Neuroscience

cognitive score in multiple kernel learning Similarly Johnsonet al [68] proposed a sparse representation learning featureselection and stacked autoencoder for classifying AD usingMRI PET CSF and cognitive scores as features and obtained959 accuracy for AD vs HC classification and 85 accuracyfor HC vs MCI classification Maqsood et al [72] proposedthe transfer learning classification model of AD for bothbinary and multiclass problems e algorithm utilizes apretrained AlexNet network and fine-tuned the convolutionalneural network (CNN) for classificationis model was fine-tuned over both segmented and unsegmented 3D views of thehuman brain e MRI scans were segmented into thecharacteristic components of GM white matter and CSFeretrained CNNwas then validated using the test data to obtainaccuracies of 896 and 928 for binary and multiclassproblems respectively using the OASIS dataset Beheshtiet al [70] developed a computer-aided diagnosis (CAD)system composed of four systematic stages for analyzingglobal and local differences in the GM of AD patientscompared with HC using the voxel-based morphometry(VBM) methodey used feature ranking based on the t-testand a genetic algorithm with Fisherrsquos criteria as part of theobjective function in the genetic algorithm e authorsutilized the SVM classifier with 10-fold cross-validation toobtain accuracies of 9301 and 75 for AD vs HC andMCIsvs pMCI classification respectively Spasov et al [71] de-veloped a deep learning architecture based on dual learningand an ad hoc layer for 3D separable convolution to identifyMCI patients eir deep learning procedure combined MRIneuropsychological demographic and APOE ε4 data toachieve an accuracy of 86 ey achieved an accuracy of995 for AD vs HC classification In another study Moradiet al used VBM analysis of GM as a feature combined withage and cognitive measures ey achieved an accuracy of82 for pMCI vs MCIs classification From this comparisonwe can infer that the proposed feature combination (MRICSF APOE and MMSE data) is robust or comparable to theother multimodal biomarker methods reported in the liter-ature for both AD vs HC and MCIs vs MCIc classification

5 Conclusion

In conclusion we demonstrated that a combination of threesMRImeasures cortical thickness cortical area cortical volumeand three nonimaging measures CSF components APOE ε4status and MMSE score improves AD diagnosis Furthermore

this combination shows great potential for the early identifi-cation of mild cognitive impairment (the prodromal stage ofAlzheimerrsquos disease) In this method we proposed filter andwrapper feature selection with an ELM classifier for multiplebiomarker-based AD diagnosis which significantly improvedthe classifierrsquos performance Moreover the results were betterthan or comparable with those previously reported particularlyfor the most challenging classification such as HC vs MCI andMCIs vs MCIc e added value of combining different ana-tomical MRI measures should be considered in AD scanningprotocols Only using specific aspects or a single measure ofwhole brain atrophy for AD diagnosis is still common practiceOur results show that clinical AD diagnosis could benefit fromthe combination of multiple measures from an anatomical MRIscan and other nonimaging biomarkers incorporated into anautomated machine learning system Our suggested methodeffectively enhances the diagnostic accuracy ofADandMCI butstill has some drawbacks Future work will include variousimprovements First we will optimize the parameter obtainingprocess Second in order to enhance the effectiveness of thesuggested method the dataset will be increased in terms of thefollowing extension of the longitudinal dataset for better un-derstanding of the progression from MCI and inclusion ofmultimodal data such as PET and fMRI data which providesdifferent insights into the characteristics of AD

Data Availability

e Alzheimerrsquos Disease Neuroimaging Initiative (ADNI)dataset (httpadniloniuscedu) was used in this studyComplete information regarding ADNI investigators can befound at httpadniloniusceduwp_contentuploadshow_to_applyADNI_Acknowledgement_istpd

Conflicts of Interest

e authors declare no conflicts of interest relating to thiswork

Acknowledgments

is work was supported by the National Research Foun-dation of Korea Grant funded by the Korean Government(NRF-2019R1F1A1060166 and NRF-2019R1A4A1029769)Data collection and sharing for this project was funded bythe ADNI (National Institutes of Health grant number U01

Table 8 Comparison of classification of MCIs vs MCIc between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection MCIs vs MCIc()

Westman et al [67] MRI +CSF OPLS mdash 685Zhang and Shen et al[39] MRI + PET+CSF SVM Multitask feature selection 7390

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 7500

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 86

Moradi et al [73] MRI + age and cognitive mdash 82Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper (SFS) 8338

Computational Intelligence and Neuroscience 15

major steps feature extraction feature combination nor-malization and feature selection and classification We usedtwo machine learning classification algorithms SVM andELM

25 Feature Selection e feature selection algorithm is anessential part of a machine learning approach facilitatingdata understanding reducing storage requirements andtraining-testing times and improving the accuracy ofclassification Importantly feature selection was performedusing only the training dataset and then applied on the testset Before feature selection we performed feature nor-malization and all feature sets were normalized to unitvariance and zero mean to reduce redundancy and improvedata integrity between the feature sets For a given datamatrix X columns represent features and rows represent theparticipantsrsquo normalized matrix Xnorm with an element (i j)which was calculated as shown in equation (1) After featurenormalization we used a combination of filter and wrapperalgorithms for feature selection We used a filter method andsorted features based on their minimum redundancymaximum relevance (MRMR) scores MRMR has beenpreviously described [59] e MRMR score for a feature setS is defined in equation (2)

Xnorm x(ij) minus mean Xj1113872 1113873

std Xj1113872 1113873 (1)

MRMR MAXs

1|s|

1113944fiisinS

I fi c( 1113857 minus1

|S|21113944 I fi fj1113872 1113873

⎧⎪⎨

⎪⎩

⎫⎪⎬

⎪⎭

(2)

where the relevance of a feature set S for k classes C

c1 c2 ck1113864 1113865 is defined by the average value of mutualinformation between the individual feature fi andC and theredundancy of all features in the feature set S is the averagevalue of mutual information between features fi and fj etop 60 features identified by the MRMR algorithm were usedin a wrapper algorithm to find an optimal subset of featuresWe developed and validated a sequential feature selection(SFS) algorithm as a wrapper feature selection method forthis study e SFS algorithm has been previously described

in detail [60] Briefly different subsets of features wereselected from the top 60 features identified by the MRMRalgorithm and then the accuracy of the ELM classifiers basedon these subsets of features was calculated

26 SVM Classifier Generally the SVM [61] is a binaryclassifier which is applicable to both separable and non-separable datasets It has been successfully utilized in theneuroimaging field and has become the most popular ma-chine learning algorithm in the field of neuroscience duringthe past decadee SVM is a supervised classifier that uses atraining dataset to find an optimal separating hyperplane inan n-dimensional space e optimal hyperplane is one thatbest separates the two target participant groups In ourstudy we utilized both linear SVM and nonlinear SVMbased on radial basis function (RBF) kernels An RBF kernelperforms better than a linear kernel for a small number offeature sets A regularization constant C and a set of kernelhyperparameters c (gamma) need to be tuned in SVMsese parameters were optimized using a cross-validation(CV) method is procedure was repeated 1000 times eachtime randomly selecting a new set of 10 held-out participantsto obtain optimum hyperparameters optimization In thismethod the search scales for regularization constant andgamma values were set to C (0001 001 01 1 10 1001000) and c (0001 001 01 1 10 100 1000) respectivelyemaximum validation accuracy was obtained at C 1 andc 01 e tuned parameters were used to predict the ac-curacy value on the test dataset

27 ELM Classifier e extreme learning machine iscomposed of a hidden layer in between the input and outputlayers [51] Whereas weights and biases are required to foradjustment by gradient-based learning algorithms on tra-ditional feedforward neural networks for all layers in theELM hidden layer biases and input weights are arbitrarilyassigned without iterative processes and output weights arecomputed by solving a single hidden layer system [50]uscompared to traditional neural networks the ELM learnsmuch faster and it is widely used in various regression andclassification tasks being an efficient and reliable learningalgorithm [62ndash65] Particularly for N training samples(X(j) I(j))|X(j) isinRp and I(j) isinRq and j 1 2 Nthe output in ELM oj with nh hidden neurons can berepresented as shown in the following equationfd3

oj 1113944

nh

i1βT

i a wTi X

(j)+ bi1113872 1113873 1113944

nh

i1βT

i hi X(j)

1113872 1113873 h X(j)

1113872 1113873Tβ

(3)

where X(j) and I(j) are the jth input and target vectorsrespectively e parameters p and q are the input and targetvector dimensions respectively Additionally oj isinR

q sig-nifies the output of the ELM for the jth training samplewi isinR

p indicates the input weight that links the inputnodes to the ith hidden node bi represents the bias of the ith

hidden node and a(middot) signifies the activation function forthe given hidden layer β [β1 βnh

]T are the values of

Table 1 Baseline clinical and sociodemographical information ofthe participants of this study

Group AD MCI HC MCIs MCIcNo ofparticipants 53 77 57 35 42

Femalemale 2033 3443 3225 1322 2121

Age 744plusmn 78 741plusmn 72 756plusmn 52 739plusmn 72 743plusmn 72Education 151plusmn 32 159plusmn 29 157plusmn 28 161plusmn 29 158plusmn 29MMSE 235plusmn 18 269plusmn 18 291plusmn 09 272plusmn 17 266plusmn 18CDR 07plusmn 02 05 0 05 05e entries for age gender education and MMSE denote mean andstandard deviation for each group MMSE Mini-Mental State Exam CDRclinical dementia ratio

Computational Intelligence and Neuroscience 5

the output weights between the output neurons and thehidden layer h(X(j)) [h1(X(j) hnh

(X(j)))]T is theoutput vector of the hidden layer with respect to the jth

training sample X(j) middot hi(X(j)) is the output of the ith hiddenlayer for the jth training sample To obtain the optimalhidden layer weights 1113954β with respect to N training samplescan be considered to solve the following optimizationproblem

minβ

λ Hβ minus L2

+β2 (4)

where H [h(X(1) h(X(N)))]T and L [I(1)

I(N)]TEquation (4) represents the optimization problem and

its optimal solution 1113954β can be analytically obtained as follows

1113954β HT 1

λI + HH

T1113874 1113875

minus1L (5)

where λ is a regularization parameter and I represents theidentity matrix After finding the optimal solution 1113954β theoutput of the ELM on test data Xtest is determined asfollows

otest h Xtest( 1113857TH

T 1λ

I + HHT

1113874 1113875minus1

L (6)

In this proposed method the hidden node number wasset between 1 and 500 and we selected a sigmoid as anactivation function Further we used a grid searchmethod totune the ELM parameter on the training dataset in order toachieve optimum cross-validated validation accuracySimilarly to minimize the random effects during the weightinitializations each parameter of the number of hiddennodes was used 100 times and the average performance wascalculated

28 Cross-Validation and Performance Evaluation We usedthe k-fold cross-validation (KCV k 10) method for cross-validation All participants were randomly divided into 10equally sized subsets using the KCV (k 10) cross-validationapproach In each fold of the KCV 90 of the data were usedto train the model based on a subset of features and then10 of the data (cross-validation set) were used to calculatethe accuracy of the ELM classifier Accuracies of the ELMclassifiers corresponding to all subsets of features werecalculated and classified with maximum accuracy and thecorresponding optimal subset of features was identifiedSimilarly classification performance was evaluated on ac-curacy (ACC) specificity (SPE) and sensitivity (SEN) TPFP FN and TN represent the number of true positives falsepositives false negatives and true negatives respectively Interms of numerical values ACC SPE and SEN can becalculated as follows

accuracy(ACC) (TN + TN)

(TP + TN + FP + FN) (7)

sensitivity(SEN) (TP)

(TP + FN) (8)

specificity(SPE) (TN)

(TN + FP) (9)

e other effective way to evaluate results for a classifieris the receiver operating characteristic (ROC) curve eROC curve is the plot of true-positive rate against false-positive rate by changing the discrimination threshold andtherefore summarizing the classifierrsquos performance eROC curve is usually represented by the area under the curve(AUC) which is denoted by a number between 0 and 100

3 Results and Analysis

31 Statistical Analysis Cortical thickness gray matter(GM) volume and surface area were analyzed using asurface-based group analysis in FreeSurferrsquos QDEC (version15) First the spatial cortical thickness GM volume andsurface area of both hemispheres were smoothed with acircularly symmetric Gaussian kernel of 10mm full-widthhalf-maximum to normally distribute the results en weemployed a general linear model (GLM) analysis with agesex and education as the nuisance factors in the designmatrix to directly compare the three parameters in bothhemispheres of the AD vs HC HC vs MCI AD vs MCIand MCIs vs MCIc groups Statistical analysis results re-garding cortical thickness surface area and gray mattervolume are shown in Figure 2 e DesikanndashKilliany Atlasdivides the human cerebral cortex into 34 cortical regions ineach hemisphere As there were a high number of atrophicregions we present only the top-ranked regions with sig-nificant differences e atrophic regions for three param-eters are listed below

Table 2 presents the atrophy position and range ofclusters for the differences in gray matter volume corticalthickness and surface area at each vertex between HC MCIand AD groups by QDEC analysis In this table only the topfeatures which have significant cluster differences for eachkind of parameter are provided From the statistically sig-nificant brain regions shown in Figure 2 and Table 2 weobserve the following

(1) e cortical thicknesses of the left insula left cuneusparacentral right rostral middle frontal and rightpars opercularis areas were thinner in the AD groupcompared with the HC group For HC vs MCI thecortical thicknesses of the left precuneus left lingualleft and right insula right pars triangularis and rightinferior parietal areas were thinner Similarly for ADvs MCI the cortical thicknesses of the left inferiorparietal right lateral occipital right inferior parietaland left and right superior temporal areas showedthe most atrophy For MCIs vs MCIc the corticalthicknesses of the left inferior parietal left para-hippocampal right temporal pole and right superiortemporal areas showed the greatest differences

(2) Regarding surface area the AD group had smallervalues than the HC group in the left and rightparacentral left lateral orbitofrontal right inferiorparietal right posterior cingulate and right inferior

6 Computational Intelligence and Neuroscience

parietal areas For HC vs MCI the left superiorfrontal left postcentral left superior parietal rightsupramarginal right fusiform right precuneus andright precentral areas showed the lowest valuesSimilarly for AD vs MCI the left superior parietalleft and right precentral left paracentral right in-ferior parietal and right superior frontal areas

showed the largest decreases in surface area ForMCIs vs MCIc the left fusiform left lateral occipitalleft parahippocampal right superior parietal andright postcentral areas showed the most differences

(3) In comparison with the HC group the volume ofgray matter of the left caudal anterior cingulate leftorbitalis left lateral orbitofrontal right lateral

icknessLe hemisphere Right hemisphere

VolumeRight hemisphereLe hemisphere

text

AreaLe hemisphere Right hemisphere

ndash500 500ndash250 250000

AD

vs

MCI

MCI

s vs

MCI

cA

D v

s H

CH

C vs

MCI

Figure 2 Differences in cortical thickness area and volume in patients at different stages of Alzheimerrsquos diseasee colored bar representsthe significance level of clusters e significance threshold was set at plt 005

Computational Intelligence and Neuroscience 7

Table 2 Cluster differences in cortical thickness area and volume in AD MCI and MCIc patients

Feature RegionCoordinates

Vertex Value Size (mm2)x y z

AD vs HC

ickness

Left insula minus295 179 119 3165 minus23512 361947Left parahippocampal minus313 minus417 minus87 557 minus22303 1073627

Left cuneus minus479 minus209 428 1039 minus23188 282428Right rostral middle frontal 385 43 7 1224 minus21737 148Right superior temporal 533 minus52 minus52 468 minus28676 25628Right pars opercularis 503 102 88 1242 minus36071 5815525

Area

Left paracentral minus158 minus355 494 1197 minus22767 701872Left lateral orbitofrontal minus533 minus2 75 489 minus21962 321823

Right paracentral 125 minus371 554 1233 minus22555 4255Right inferior parietal 34 minus519 374 1133 minus21026 721305

Right posterior cingulate 14 minus304 366 1201 minus20979 141667Right inferior parietal 395 minus447 346 7 minus20263 5076

Volume

Left caudal anterior cingulate minus5 184 264 2073 minus32264 83461Left orbitalis minus287 minus602 417 1103 23437 120255

Left lateral orbitofrontal minus594 minus69 94 412 minus20146 63567Right lateral orbitofrontal 304 229 minus203 631 minus27411 20256Right rostral middle frontal 348 51 47 198 minus2717 109882

Right pars opercularis 468 152 87 136 minus2717 80771AD vs MCI

ickness

Left inferior parietal minus436 minus615 33 939 27022 43937Left fusiform minus40 minus537 minus203 254 2459 14984

Left superior temporal minus477 minus25 minus91 210 minus25577 32803Right lateral occipital 261 minus939 37 363 34944 363Right inferior parietal 377 minus717 428 1129 33654 64624Right superior temporal 507 minus143 minus24 722 minus29715 34228

Area

Left superior parietal 284 minus51 427 1702 minus25085 64276Left precentral 322 minus218 608 109 18969 5154Left paracentral 141 minus367 524 238 minus18959 7968

Right inferior parietal minus352 minus872 137 50 minus18099 3453Right precentral minus542 minus11 71 51 minus17244 2134

Right superior frontal minus74 4 665 34 16541 1786

Volume

Left superior parietal minus285 minus588 40 540 35868 2161Left superior frontal minus85 412 303 48 minus26334 3398

Left caudal anterior cingulate minus49 113 321 314 minus2549 15007Right entorhinal 215 minus89 minus295 207 minus31433 5569

Right lateral orbitofrontal 428 278 minus137 198 minus31034 12501Right superior parietal 321 minus442 412 115 minus25309 4029

HC vs MCI

ickness

Left insula minus364 minus94 minus117 811 minus35988 30994Left precuneus minus6 minus689 417 628 23682 32202Left lingual minus293 minus445 minus66 612 minus27175 27916Right insula 358 minus106 minus74 1154 minus31385 43662

Right pars triangularis 389 317 1 372 minus21264 20487Right inferior parietal 351 minus72 409 285 minus27121 1522

Area

Left superior frontal minus178 315 496 237 17293 15688Left postcentral minus521 minus227 507 264 17286 12275

Left superior parietal minus32 minus493 473 141 minus17572 6604Right supramarginal 534 minus474 35 1336 27163 67685

Right fusiform 347 minus12 minus341 551 20892 32284Right precuneus 182 minus773 277 482 17491 31336Right precentral 207 minus306 539 328 minus19165 11545

Volume

Left supramarginal 521 minus463 226 441 27521 21023Left cuneus 171 minus695 168 628 2611 48174

Left precentral 212 minus307 538 376 minus21546 12702Right parahippocampal minus238 minus361 minus156 278 minus18212 12768Right superior parietal minus91 minus743 465 598 25199 30367Right superior frontal minus181 333 396 210 26035 11472

8 Computational Intelligence and Neuroscience

orbitofrontal right rostral middle frontal and rightpars opercularis areas was lower in the AD groupFor AD vs MCI the left superior parietal left su-perior frontal left caudal anterior cingulate rightentorhinal and right superior parietal areas showedthe largest decreases in volume For HC vs MCI theleft supramarginal left cuneus left precentral rightparahippocampal and right superior parietal areasshowed the most volume atrophy Similarly forMCIs vs MCIc the left bankssts left precentral leftrostral middle frontal right precentral and rightfusiform areas had the largest decreases in volume

Moreover from this analysis we observe that areathickness and volume in the AD group were significantlydecreased in comparison with the HC group Similarly therewas significant atrophy in the MCIc group in comparisonwith the MCIs group and there was little difference in theatrophy pattern among AD and MCI patients

32 Feature Selection and Classification e individualfeature set was selected using an MRMR (filter) and a se-quential feature selection (wrapper) method to identify theoptimal feature set for different groups and improve theclassifier accuracy Similarly we combined all the feature setsfrom different measures and applied the MRMR and SFSmethod to select the optimal features Multiple measuresfeature sets were created by combining cortical thicknessarea and volume from sMRI three component measure-ments from CSF APOE ε4 status andMMSE score First weselected the top 60 features from theMRMR algorithm basedon the features score and then we applied a sequentialfeature selection algorithm on these top 60 ranked featureswhich gave the optimal feature set to achieve the maximumclassification accuracy on ELM classifiers Figure 3 shows the

number of optimal features sequence after sequential featureselection on the top 60 ranked features obtained from theMRMR algorithm using ELM classifier Figure 3 shows onlythe selected features for the combined feature set From thiswe can assume that the proposed feature selection withcross-validation provides the optimal feature vector forinput to the classifiers In this proposed method classifi-cation performance was quantified by the number of featuresselected versus accuracy and area under the ROC curve

To further analyze the effectiveness of the purposeclassification method combining different measures wecalculated the AUCs for the concatenation of all featuresFigure 4shows the receiver operating curves for individualfeatures and all feature (imaging and nonimaging bio-markers ie multiple features) combinations for eachclassification group using the ELM classifier

4 Discussion

41 Performance Analysis In this study we first performedstatistical analysis and pattern classification to differentiateand identify atrophy patterns for the four groups (AD HCMCIs and MCIc) Individual sMRI was preprocessed usingthe FreeSurfer tool After preprocessing the statisticalanalysis results of sMRI QDEC was applied and finally weperformed the classification task by the proposed featureselection and classification method respectively For brainatrophy analysis we used three types of sMRI corticalmetrics (cortical thickness gray matter volume and surfacearea) In comparing the AD group with the HC group theinsula pars opercularis parahippocampal and superiortemporal areas were severely affected in terms of corticalthickness gray matter volume and surface area e corticalthickness of the left hemisphere was thinner than that of theleft hemisphere Similarly for AD vs MCI the most atrophy

Table 2 Continued

Feature RegionCoordinates

Vertex Value Size (mm2)x y z

MCIs vs MCIc

ickness

Left inferior parietal minus463 minus60 111 2770 minus37613 142748Left superior temporal minus455 minus04 minus208 1067 minus26742 46624Left parahippocampal minus317 minus403 minus101 550 minus23925 23985Right temporal pole 292 9 minus38 1449 minus55983 82282

Right superior temporal 623 minus347 152 1292 minus31646 54997Right inferior temporal 516 minus566 minus37 882 minus30694 50791

Right precentral 488 minus65 405 858 minus28315 35193

Area

Left fusiform minus364 minus295 minus22 659 minus30958 31221Left lateral occipital minus194 minus99 minus151 326 26022 25061Left parahippocampal minus188 minus335 minus14 224 minus20347 9759Right superior temporal 481 minus329 22 1515 minus25815 5798Right superior parietal 106 minus527 651 1697 minus22367 64152

Right postcentral 338 minus29 519 799 minus20859 36723

Volume

Left bankssts minus579 minus469 minus1 2196 minus26263 116258Left precentral minus361 minus22 517 2736 minus25339 10731

Left rostral middle frontal minus224 279 328 582 minus23081 32686Right superior temporal 623 minus347 152 3392 minus44677 144996

Right precentral 495 minus62 411 4103 minus32472 181394Right fusiform 339 minus378 minus228 782 minus31822 42505

Computational Intelligence and Neuroscience 9

was seen in the left inferior parietal and right lateral occipitalareas For HC vs MCI the supramarginal and cuneus areasshowed the most atrophy For MCIs vs MCIc the superiortemporal and precentral areas showed the largest decreasesin thickness volume and area Based on these differencesthe majority of the decrease in cortical thickness gray mattervolume and surface area appears mostly in the frontal lobetemporal lobe occipital lobe cingulate gyrus and parietallobe is phenomenon strongly agrees with findings relatedto atrophy patterns seen in previous studies [14 66] eseregions are mainly involved in the expression of personalitymotor execution complex cognitive behavior and decisionmaking [50] In addition we present the less commonanalysis of MCIs vs MCIc For MCIc the most atrophy wasseen in the superior temporal bankssts precentral inferior

parietal and insula areas which show potential for the earlyrecognition of progression to Alzheimerrsquos disease

To test the effectiveness of our purposed featurescombination method (ie imaging and nonimaging fea-tures) we adapted the individual feature from sMRI and CSFseparately to conduct the study although regarding geneticand cognitive features we combined them to test the per-formance and then compared the accuracy with the accuracyof the combination of all features For the proposed featureselection we used a combination of filter and wrapper al-gorithms and compared the performance of the classifiers onthe selected feature set In this text SVM-linear SVM-RBFand ELM were used for classification e classificationresults for the compared methods are presented inTables 3ndash6 and also in graphical form in Figure 5 As shown

AD vs HC

Number of selected features = 27

05

101520253035404550556065707580859095

100

Accu

racy

()

5 10 15 20 25 30 35 40 45 50 55 600

Number of features sequence

(a)

AD vs MCI

Number of selected features = 17

05

101520253035404550556065707580859095

100

Accu

racy

()

5 10 15 20 25 30 35 40 45 50 55 600

Number of feature sequence

(b)HC vs MCI

Number of selected features = 25

5 10 15 20 25 30 35 40 45 50 55 600

Number of feature sequence

05

101520253035404550556065707580859095

100

Accu

racy

()

(c)

MCIs vs MCIc

Number of selected features = 13

05

101520253035404550556065707580859095

100Ac

cura

cy (

)

5 10 15 20 25 30 35 40 45 50 55 600

Number of features sequence

(d)

Figure 3 Number of feature sets selected from the sequential feature selection algorithm on the top 60 ranked features obtained using theMRMR algorithm Only the number of features versus accuracy for the combination of all feature sets using ELM classifiers is shown (a) ADvs HC feature subset (b) AD vs MCI feature subset (c) HC vs MCI feature subset (d) MCIs vs MCIc feature subset

10 Computational Intelligence and Neuroscience

AD vs HC

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(a)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

AD vs MCI

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(b)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

HC vs MCI

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(c)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

MCIs vs MCIc

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(d)

Figure 4 ROC curves for discriminating AD MCIs MCIc and HC (healthy controls) ROC curves are plotted for the five biomarkersseparately and for the concatenation of all features (ie all five biomarkers measures) (a) ROC curves for AD vs HC classification (b) forAD vs MCI classification (c) for HC vs MCI classification and (d) for MCIs vs MCIc classification using two-stage feature selection withELM classifiers

Computational Intelligence and Neuroscience 11

in the tables the performance accuracy of feature fusion isnoticeably improved when compared with that of the in-dividual feature set and there are distinct degrees of ele-vation in other indexes the specificity index is particularlymore noticeable In contrast performance accuracy basedon the nonimaging features is almost the same as the im-aging features for AD diagnosis but far superior for MCIdiagnosis For AD vs HC the accuracy obtained by CSFmeasures genetics and cognitive score showed a lowerincrease although for MCI vs HC AD vs MCI and MCIsvs MCIc the accuracy was higher e result showed thatthe ELM classifier achieves better classification scorescompared to the SVM classifier (linear and RBF-SVM) for

both single modality feature sets as well as all featurescombined (shown in Figure 6) e ELM is good machinelearning tool and the major strength is that the hiddenlayerrsquos learning parameters including input weight andbiases do not tune iteratively as in SLFN e ELM offersmany advantages over other learning algorithms [54] Fromthe results shown in Tables 3ndash6 for AD vs HC the clas-sification result increased by 5ndash7 as compared to SVMclassifiers with 9804 sensitivity and 9628 specificityFor AD vs MCI classification accuracy increased by 5ndash9with 92 sensitivity and 9733 accuracy For HC vs MCIclassification accuracy increased by 6ndash10 with 9123sensitivity and 9913 specificity Similarly for MCIs vs

Table 3 10-fold cross-validated classification performance for AD vs HC

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 8513 9207 8544 086 8068 8568 8381 7717 8317 7381Surface area 8230 8930 9154 085 7616 8616 7514 7867 8767 7262Volume 8790 9387 8475 088 7929 7429 7833 8000 7710 6324CSF 8973 9633 8738 089 7905 8905 7417 7533 8333 6857APOE+MMSE 8645 9513 8700 093 8203 8703 8467 8416 8305 7879Concatenation (all_features_set) 9731 9804 9628 097 9133 9333 8757 9350 955 9058

Table 4 10-fold cross-validated classification performance for AD vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7010 8130 6610 071 6401 8383 6233 6503 7880 7330Surface area 6160 6670 8430 063 6345 6810 8011 6067 7123 6285Volume 6780 8210 6710 069 5820 7792 6856 6005 7337 6373CSF 6470 6600 8140 073 5607 6275 7573 6123 6871 7937APOE+MMSE 8145 8440 8770 082 7681 6781 8548 7620 6694 8748Concatenation (all_features_set) 8791 9200 9733 089 8350 8864 7672 7831 7445 8262

Table 5 10-fold cross-validated classification performance for HC vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7540 8407 6790 083 7417 8081 7123 6734 7317 7793Surface area 7415 6520 8385 084 6354 6767 6647 6872 7443 6735Volume 7784 8841 7280 084 6547 7473 6875 7283 7827 6570CSF 8330 7393 8402 085 7339 7525 6683 7580 7338 7814APOE+MMSE 7937 9325 7297 089 6829 8173 7668 7023 8124 7805Concatenation (all_features_set) 9172 9123 9913 093 8170 8368 8749 8565 7854 8926

Table 6 10-fold cross-validated classification performance for MCIs vs MCIc

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 6371 7110 3837 064 5718 6553 6038 5005 7118 6813Surface area 6143 7228 7480 069 5425 6322 7617 5827 7223 6485Volume 6785 8440 8773 074 5917 7828 6755 6214 7445 5773CSF 7137 7884 6857 072 6473 6009 7278 6733 7882 6878APOE+MMSE 7609 8713 7209 079 6553 6881 6743 6808 7221 6633Concatenation (all_features_set) 8338 9301 7577 085 7137 8116 7675 7571 7838 8467

12 Computational Intelligence and Neuroscience

MCIc classification accuracy increased by 7ndash12 with9301 sensitivity and 7577 specificity but 8467 and7667 specificity was obtained for linear and RBF-SVMclassifiers respectively which is more than obtained by theELM classifier From our analyses we believe that the ELM

gives better performance as compared to SVM from alearning efficiency standpoint because the ELMrsquos originaldesign has high learning accuracy fast learning speedscalability and the least human intervention [54 55] Fromthe results shown in Table 7 we can see that our suggested

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(a)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(b)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(c)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(d)

Figure 5 Classification results for different Alzheimerrsquos groups using ELM classifiers (a) Classification results for AD vs HC groups (b) forAD vs MCI groups (c) for HC vs MCI groups and (d) for MCIs vs MCIc groups

Computational Intelligence and Neuroscience 13

method achieved better accuracy than other existingmethodsFor classifying AD and HC our method achieved a classi-fication accuracy of 9731 with a sensitivity of 9804 aspecificity of 9628 and an AUC value of 097 Forclassifying MCI and HC our method achieves a classificationaccuracy of 9172 with a sensitivity of 9123 a specificity of9913 and an AUC value of 093 For distinguishing ADfrom MCI our proposed method obtained a classificationaccuracy of 8791 with a sensitivity of 9200 a specificityof 9733 and an AUC value of 089 Similarly for MCIs vsMCIc we achieved outstanding performance as compared toprevious methods by combining multiple features with anaccuracy of 8338 a sensitivity of 9301 a specificity of7577 and an AUC value of 085

42ComparisonwithOtherMethods In the previous sectionwe discussed in detail the findings of the proposed featureselection and classification method In this section we

compare and discuss the findings of our method in com-parison with existing state-of-the-art methods Tables 7 and 8show a comparison of the classification performance of theproposed method with recently published studies which usedmultimodality data to distinguish individuals with AD andMCI from HC Westman et al [67] used MRI and CSFbiomarkers and obtained 918 accuracy for AD vs HC776 for HC vs MCI and 685 for pMCI vs MCIsclassification using the orthogonal partial least squares tolatent structures (OPLS) method Zhang and Shen [39] used amultimodal (MRI PET and CSF) classification method withmultitask feature selection and an SVM classifier for AD andMCI classification By combining MRI PET and CSF datathey achieved a higher accuracy of 933 for AD vs HC and832 accuracy for HC vs MCI classification Similarly theauthors achieved an accuracy of 739 on pMCI vs MCIssamples Hinrichs et al [69] obtained an accuracy of 924 forAD vs HC classification using MRI PET CSF APOE and

ELMSVM_RBFSVM-linear

0

20

40

60

80

100

ACC

()

AD vs HC HC vs MCI MCIs vs MCIcAD vs MCI

Figure 6 Performance comparison of different classifiers

Table 7 Comparison of classification of AD vs HC and HC vs MCI between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection AD vsHC ()

HC vsMCI

Westman et al[67] MRI +CSF OPLS mdash 918 776

Johnson et al [68] MRI + PET+CSF+ cognitive scores Stackedautoencoder

Sparse representationlearning 959 85

Hinrichs et al [69] MRI + PET+CSF+APOE+ cognitive scores MKL mdash 924 naZhang and Shenet al [39] MRI + PET+CSF SVM Multitask feature selection 933 832

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 9301 mdash

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 995 mdash

Maqsood et al[72] sMRI CNN mdash 9285 mdash

Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper(SFS) 9731 9172

14 Computational Intelligence and Neuroscience

cognitive score in multiple kernel learning Similarly Johnsonet al [68] proposed a sparse representation learning featureselection and stacked autoencoder for classifying AD usingMRI PET CSF and cognitive scores as features and obtained959 accuracy for AD vs HC classification and 85 accuracyfor HC vs MCI classification Maqsood et al [72] proposedthe transfer learning classification model of AD for bothbinary and multiclass problems e algorithm utilizes apretrained AlexNet network and fine-tuned the convolutionalneural network (CNN) for classificationis model was fine-tuned over both segmented and unsegmented 3D views of thehuman brain e MRI scans were segmented into thecharacteristic components of GM white matter and CSFeretrained CNNwas then validated using the test data to obtainaccuracies of 896 and 928 for binary and multiclassproblems respectively using the OASIS dataset Beheshtiet al [70] developed a computer-aided diagnosis (CAD)system composed of four systematic stages for analyzingglobal and local differences in the GM of AD patientscompared with HC using the voxel-based morphometry(VBM) methodey used feature ranking based on the t-testand a genetic algorithm with Fisherrsquos criteria as part of theobjective function in the genetic algorithm e authorsutilized the SVM classifier with 10-fold cross-validation toobtain accuracies of 9301 and 75 for AD vs HC andMCIsvs pMCI classification respectively Spasov et al [71] de-veloped a deep learning architecture based on dual learningand an ad hoc layer for 3D separable convolution to identifyMCI patients eir deep learning procedure combined MRIneuropsychological demographic and APOE ε4 data toachieve an accuracy of 86 ey achieved an accuracy of995 for AD vs HC classification In another study Moradiet al used VBM analysis of GM as a feature combined withage and cognitive measures ey achieved an accuracy of82 for pMCI vs MCIs classification From this comparisonwe can infer that the proposed feature combination (MRICSF APOE and MMSE data) is robust or comparable to theother multimodal biomarker methods reported in the liter-ature for both AD vs HC and MCIs vs MCIc classification

5 Conclusion

In conclusion we demonstrated that a combination of threesMRImeasures cortical thickness cortical area cortical volumeand three nonimaging measures CSF components APOE ε4status and MMSE score improves AD diagnosis Furthermore

this combination shows great potential for the early identifi-cation of mild cognitive impairment (the prodromal stage ofAlzheimerrsquos disease) In this method we proposed filter andwrapper feature selection with an ELM classifier for multiplebiomarker-based AD diagnosis which significantly improvedthe classifierrsquos performance Moreover the results were betterthan or comparable with those previously reported particularlyfor the most challenging classification such as HC vs MCI andMCIs vs MCIc e added value of combining different ana-tomical MRI measures should be considered in AD scanningprotocols Only using specific aspects or a single measure ofwhole brain atrophy for AD diagnosis is still common practiceOur results show that clinical AD diagnosis could benefit fromthe combination of multiple measures from an anatomical MRIscan and other nonimaging biomarkers incorporated into anautomated machine learning system Our suggested methodeffectively enhances the diagnostic accuracy ofADandMCI butstill has some drawbacks Future work will include variousimprovements First we will optimize the parameter obtainingprocess Second in order to enhance the effectiveness of thesuggested method the dataset will be increased in terms of thefollowing extension of the longitudinal dataset for better un-derstanding of the progression from MCI and inclusion ofmultimodal data such as PET and fMRI data which providesdifferent insights into the characteristics of AD

Data Availability

e Alzheimerrsquos Disease Neuroimaging Initiative (ADNI)dataset (httpadniloniuscedu) was used in this studyComplete information regarding ADNI investigators can befound at httpadniloniusceduwp_contentuploadshow_to_applyADNI_Acknowledgement_istpd

Conflicts of Interest

e authors declare no conflicts of interest relating to thiswork

Acknowledgments

is work was supported by the National Research Foun-dation of Korea Grant funded by the Korean Government(NRF-2019R1F1A1060166 and NRF-2019R1A4A1029769)Data collection and sharing for this project was funded bythe ADNI (National Institutes of Health grant number U01

Table 8 Comparison of classification of MCIs vs MCIc between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection MCIs vs MCIc()

Westman et al [67] MRI +CSF OPLS mdash 685Zhang and Shen et al[39] MRI + PET+CSF SVM Multitask feature selection 7390

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 7500

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 86

Moradi et al [73] MRI + age and cognitive mdash 82Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper (SFS) 8338

Computational Intelligence and Neuroscience 15

the output weights between the output neurons and thehidden layer h(X(j)) [h1(X(j) hnh

(X(j)))]T is theoutput vector of the hidden layer with respect to the jth

training sample X(j) middot hi(X(j)) is the output of the ith hiddenlayer for the jth training sample To obtain the optimalhidden layer weights 1113954β with respect to N training samplescan be considered to solve the following optimizationproblem

minβ

λ Hβ minus L2

+β2 (4)

where H [h(X(1) h(X(N)))]T and L [I(1)

I(N)]TEquation (4) represents the optimization problem and

its optimal solution 1113954β can be analytically obtained as follows

1113954β HT 1

λI + HH

T1113874 1113875

minus1L (5)

where λ is a regularization parameter and I represents theidentity matrix After finding the optimal solution 1113954β theoutput of the ELM on test data Xtest is determined asfollows

otest h Xtest( 1113857TH

T 1λ

I + HHT

1113874 1113875minus1

L (6)

In this proposed method the hidden node number wasset between 1 and 500 and we selected a sigmoid as anactivation function Further we used a grid searchmethod totune the ELM parameter on the training dataset in order toachieve optimum cross-validated validation accuracySimilarly to minimize the random effects during the weightinitializations each parameter of the number of hiddennodes was used 100 times and the average performance wascalculated

28 Cross-Validation and Performance Evaluation We usedthe k-fold cross-validation (KCV k 10) method for cross-validation All participants were randomly divided into 10equally sized subsets using the KCV (k 10) cross-validationapproach In each fold of the KCV 90 of the data were usedto train the model based on a subset of features and then10 of the data (cross-validation set) were used to calculatethe accuracy of the ELM classifier Accuracies of the ELMclassifiers corresponding to all subsets of features werecalculated and classified with maximum accuracy and thecorresponding optimal subset of features was identifiedSimilarly classification performance was evaluated on ac-curacy (ACC) specificity (SPE) and sensitivity (SEN) TPFP FN and TN represent the number of true positives falsepositives false negatives and true negatives respectively Interms of numerical values ACC SPE and SEN can becalculated as follows

accuracy(ACC) (TN + TN)

(TP + TN + FP + FN) (7)

sensitivity(SEN) (TP)

(TP + FN) (8)

specificity(SPE) (TN)

(TN + FP) (9)

e other effective way to evaluate results for a classifieris the receiver operating characteristic (ROC) curve eROC curve is the plot of true-positive rate against false-positive rate by changing the discrimination threshold andtherefore summarizing the classifierrsquos performance eROC curve is usually represented by the area under the curve(AUC) which is denoted by a number between 0 and 100

3 Results and Analysis

31 Statistical Analysis Cortical thickness gray matter(GM) volume and surface area were analyzed using asurface-based group analysis in FreeSurferrsquos QDEC (version15) First the spatial cortical thickness GM volume andsurface area of both hemispheres were smoothed with acircularly symmetric Gaussian kernel of 10mm full-widthhalf-maximum to normally distribute the results en weemployed a general linear model (GLM) analysis with agesex and education as the nuisance factors in the designmatrix to directly compare the three parameters in bothhemispheres of the AD vs HC HC vs MCI AD vs MCIand MCIs vs MCIc groups Statistical analysis results re-garding cortical thickness surface area and gray mattervolume are shown in Figure 2 e DesikanndashKilliany Atlasdivides the human cerebral cortex into 34 cortical regions ineach hemisphere As there were a high number of atrophicregions we present only the top-ranked regions with sig-nificant differences e atrophic regions for three param-eters are listed below

Table 2 presents the atrophy position and range ofclusters for the differences in gray matter volume corticalthickness and surface area at each vertex between HC MCIand AD groups by QDEC analysis In this table only the topfeatures which have significant cluster differences for eachkind of parameter are provided From the statistically sig-nificant brain regions shown in Figure 2 and Table 2 weobserve the following

(1) e cortical thicknesses of the left insula left cuneusparacentral right rostral middle frontal and rightpars opercularis areas were thinner in the AD groupcompared with the HC group For HC vs MCI thecortical thicknesses of the left precuneus left lingualleft and right insula right pars triangularis and rightinferior parietal areas were thinner Similarly for ADvs MCI the cortical thicknesses of the left inferiorparietal right lateral occipital right inferior parietaland left and right superior temporal areas showedthe most atrophy For MCIs vs MCIc the corticalthicknesses of the left inferior parietal left para-hippocampal right temporal pole and right superiortemporal areas showed the greatest differences

(2) Regarding surface area the AD group had smallervalues than the HC group in the left and rightparacentral left lateral orbitofrontal right inferiorparietal right posterior cingulate and right inferior

6 Computational Intelligence and Neuroscience

parietal areas For HC vs MCI the left superiorfrontal left postcentral left superior parietal rightsupramarginal right fusiform right precuneus andright precentral areas showed the lowest valuesSimilarly for AD vs MCI the left superior parietalleft and right precentral left paracentral right in-ferior parietal and right superior frontal areas

showed the largest decreases in surface area ForMCIs vs MCIc the left fusiform left lateral occipitalleft parahippocampal right superior parietal andright postcentral areas showed the most differences

(3) In comparison with the HC group the volume ofgray matter of the left caudal anterior cingulate leftorbitalis left lateral orbitofrontal right lateral

icknessLe hemisphere Right hemisphere

VolumeRight hemisphereLe hemisphere

text

AreaLe hemisphere Right hemisphere

ndash500 500ndash250 250000

AD

vs

MCI

MCI

s vs

MCI

cA

D v

s H

CH

C vs

MCI

Figure 2 Differences in cortical thickness area and volume in patients at different stages of Alzheimerrsquos diseasee colored bar representsthe significance level of clusters e significance threshold was set at plt 005

Computational Intelligence and Neuroscience 7

Table 2 Cluster differences in cortical thickness area and volume in AD MCI and MCIc patients

Feature RegionCoordinates

Vertex Value Size (mm2)x y z

AD vs HC

ickness

Left insula minus295 179 119 3165 minus23512 361947Left parahippocampal minus313 minus417 minus87 557 minus22303 1073627

Left cuneus minus479 minus209 428 1039 minus23188 282428Right rostral middle frontal 385 43 7 1224 minus21737 148Right superior temporal 533 minus52 minus52 468 minus28676 25628Right pars opercularis 503 102 88 1242 minus36071 5815525

Area

Left paracentral minus158 minus355 494 1197 minus22767 701872Left lateral orbitofrontal minus533 minus2 75 489 minus21962 321823

Right paracentral 125 minus371 554 1233 minus22555 4255Right inferior parietal 34 minus519 374 1133 minus21026 721305

Right posterior cingulate 14 minus304 366 1201 minus20979 141667Right inferior parietal 395 minus447 346 7 minus20263 5076

Volume

Left caudal anterior cingulate minus5 184 264 2073 minus32264 83461Left orbitalis minus287 minus602 417 1103 23437 120255

Left lateral orbitofrontal minus594 minus69 94 412 minus20146 63567Right lateral orbitofrontal 304 229 minus203 631 minus27411 20256Right rostral middle frontal 348 51 47 198 minus2717 109882

Right pars opercularis 468 152 87 136 minus2717 80771AD vs MCI

ickness

Left inferior parietal minus436 minus615 33 939 27022 43937Left fusiform minus40 minus537 minus203 254 2459 14984

Left superior temporal minus477 minus25 minus91 210 minus25577 32803Right lateral occipital 261 minus939 37 363 34944 363Right inferior parietal 377 minus717 428 1129 33654 64624Right superior temporal 507 minus143 minus24 722 minus29715 34228

Area

Left superior parietal 284 minus51 427 1702 minus25085 64276Left precentral 322 minus218 608 109 18969 5154Left paracentral 141 minus367 524 238 minus18959 7968

Right inferior parietal minus352 minus872 137 50 minus18099 3453Right precentral minus542 minus11 71 51 minus17244 2134

Right superior frontal minus74 4 665 34 16541 1786

Volume

Left superior parietal minus285 minus588 40 540 35868 2161Left superior frontal minus85 412 303 48 minus26334 3398

Left caudal anterior cingulate minus49 113 321 314 minus2549 15007Right entorhinal 215 minus89 minus295 207 minus31433 5569

Right lateral orbitofrontal 428 278 minus137 198 minus31034 12501Right superior parietal 321 minus442 412 115 minus25309 4029

HC vs MCI

ickness

Left insula minus364 minus94 minus117 811 minus35988 30994Left precuneus minus6 minus689 417 628 23682 32202Left lingual minus293 minus445 minus66 612 minus27175 27916Right insula 358 minus106 minus74 1154 minus31385 43662

Right pars triangularis 389 317 1 372 minus21264 20487Right inferior parietal 351 minus72 409 285 minus27121 1522

Area

Left superior frontal minus178 315 496 237 17293 15688Left postcentral minus521 minus227 507 264 17286 12275

Left superior parietal minus32 minus493 473 141 minus17572 6604Right supramarginal 534 minus474 35 1336 27163 67685

Right fusiform 347 minus12 minus341 551 20892 32284Right precuneus 182 minus773 277 482 17491 31336Right precentral 207 minus306 539 328 minus19165 11545

Volume

Left supramarginal 521 minus463 226 441 27521 21023Left cuneus 171 minus695 168 628 2611 48174

Left precentral 212 minus307 538 376 minus21546 12702Right parahippocampal minus238 minus361 minus156 278 minus18212 12768Right superior parietal minus91 minus743 465 598 25199 30367Right superior frontal minus181 333 396 210 26035 11472

8 Computational Intelligence and Neuroscience

orbitofrontal right rostral middle frontal and rightpars opercularis areas was lower in the AD groupFor AD vs MCI the left superior parietal left su-perior frontal left caudal anterior cingulate rightentorhinal and right superior parietal areas showedthe largest decreases in volume For HC vs MCI theleft supramarginal left cuneus left precentral rightparahippocampal and right superior parietal areasshowed the most volume atrophy Similarly forMCIs vs MCIc the left bankssts left precentral leftrostral middle frontal right precentral and rightfusiform areas had the largest decreases in volume

Moreover from this analysis we observe that areathickness and volume in the AD group were significantlydecreased in comparison with the HC group Similarly therewas significant atrophy in the MCIc group in comparisonwith the MCIs group and there was little difference in theatrophy pattern among AD and MCI patients

32 Feature Selection and Classification e individualfeature set was selected using an MRMR (filter) and a se-quential feature selection (wrapper) method to identify theoptimal feature set for different groups and improve theclassifier accuracy Similarly we combined all the feature setsfrom different measures and applied the MRMR and SFSmethod to select the optimal features Multiple measuresfeature sets were created by combining cortical thicknessarea and volume from sMRI three component measure-ments from CSF APOE ε4 status andMMSE score First weselected the top 60 features from theMRMR algorithm basedon the features score and then we applied a sequentialfeature selection algorithm on these top 60 ranked featureswhich gave the optimal feature set to achieve the maximumclassification accuracy on ELM classifiers Figure 3 shows the

number of optimal features sequence after sequential featureselection on the top 60 ranked features obtained from theMRMR algorithm using ELM classifier Figure 3 shows onlythe selected features for the combined feature set From thiswe can assume that the proposed feature selection withcross-validation provides the optimal feature vector forinput to the classifiers In this proposed method classifi-cation performance was quantified by the number of featuresselected versus accuracy and area under the ROC curve

To further analyze the effectiveness of the purposeclassification method combining different measures wecalculated the AUCs for the concatenation of all featuresFigure 4shows the receiver operating curves for individualfeatures and all feature (imaging and nonimaging bio-markers ie multiple features) combinations for eachclassification group using the ELM classifier

4 Discussion

41 Performance Analysis In this study we first performedstatistical analysis and pattern classification to differentiateand identify atrophy patterns for the four groups (AD HCMCIs and MCIc) Individual sMRI was preprocessed usingthe FreeSurfer tool After preprocessing the statisticalanalysis results of sMRI QDEC was applied and finally weperformed the classification task by the proposed featureselection and classification method respectively For brainatrophy analysis we used three types of sMRI corticalmetrics (cortical thickness gray matter volume and surfacearea) In comparing the AD group with the HC group theinsula pars opercularis parahippocampal and superiortemporal areas were severely affected in terms of corticalthickness gray matter volume and surface area e corticalthickness of the left hemisphere was thinner than that of theleft hemisphere Similarly for AD vs MCI the most atrophy

Table 2 Continued

Feature RegionCoordinates

Vertex Value Size (mm2)x y z

MCIs vs MCIc

ickness

Left inferior parietal minus463 minus60 111 2770 minus37613 142748Left superior temporal minus455 minus04 minus208 1067 minus26742 46624Left parahippocampal minus317 minus403 minus101 550 minus23925 23985Right temporal pole 292 9 minus38 1449 minus55983 82282

Right superior temporal 623 minus347 152 1292 minus31646 54997Right inferior temporal 516 minus566 minus37 882 minus30694 50791

Right precentral 488 minus65 405 858 minus28315 35193

Area

Left fusiform minus364 minus295 minus22 659 minus30958 31221Left lateral occipital minus194 minus99 minus151 326 26022 25061Left parahippocampal minus188 minus335 minus14 224 minus20347 9759Right superior temporal 481 minus329 22 1515 minus25815 5798Right superior parietal 106 minus527 651 1697 minus22367 64152

Right postcentral 338 minus29 519 799 minus20859 36723

Volume

Left bankssts minus579 minus469 minus1 2196 minus26263 116258Left precentral minus361 minus22 517 2736 minus25339 10731

Left rostral middle frontal minus224 279 328 582 minus23081 32686Right superior temporal 623 minus347 152 3392 minus44677 144996

Right precentral 495 minus62 411 4103 minus32472 181394Right fusiform 339 minus378 minus228 782 minus31822 42505

Computational Intelligence and Neuroscience 9

was seen in the left inferior parietal and right lateral occipitalareas For HC vs MCI the supramarginal and cuneus areasshowed the most atrophy For MCIs vs MCIc the superiortemporal and precentral areas showed the largest decreasesin thickness volume and area Based on these differencesthe majority of the decrease in cortical thickness gray mattervolume and surface area appears mostly in the frontal lobetemporal lobe occipital lobe cingulate gyrus and parietallobe is phenomenon strongly agrees with findings relatedto atrophy patterns seen in previous studies [14 66] eseregions are mainly involved in the expression of personalitymotor execution complex cognitive behavior and decisionmaking [50] In addition we present the less commonanalysis of MCIs vs MCIc For MCIc the most atrophy wasseen in the superior temporal bankssts precentral inferior

parietal and insula areas which show potential for the earlyrecognition of progression to Alzheimerrsquos disease

To test the effectiveness of our purposed featurescombination method (ie imaging and nonimaging fea-tures) we adapted the individual feature from sMRI and CSFseparately to conduct the study although regarding geneticand cognitive features we combined them to test the per-formance and then compared the accuracy with the accuracyof the combination of all features For the proposed featureselection we used a combination of filter and wrapper al-gorithms and compared the performance of the classifiers onthe selected feature set In this text SVM-linear SVM-RBFand ELM were used for classification e classificationresults for the compared methods are presented inTables 3ndash6 and also in graphical form in Figure 5 As shown

AD vs HC

Number of selected features = 27

05

101520253035404550556065707580859095

100

Accu

racy

()

5 10 15 20 25 30 35 40 45 50 55 600

Number of features sequence

(a)

AD vs MCI

Number of selected features = 17

05

101520253035404550556065707580859095

100

Accu

racy

()

5 10 15 20 25 30 35 40 45 50 55 600

Number of feature sequence

(b)HC vs MCI

Number of selected features = 25

5 10 15 20 25 30 35 40 45 50 55 600

Number of feature sequence

05

101520253035404550556065707580859095

100

Accu

racy

()

(c)

MCIs vs MCIc

Number of selected features = 13

05

101520253035404550556065707580859095

100Ac

cura

cy (

)

5 10 15 20 25 30 35 40 45 50 55 600

Number of features sequence

(d)

Figure 3 Number of feature sets selected from the sequential feature selection algorithm on the top 60 ranked features obtained using theMRMR algorithm Only the number of features versus accuracy for the combination of all feature sets using ELM classifiers is shown (a) ADvs HC feature subset (b) AD vs MCI feature subset (c) HC vs MCI feature subset (d) MCIs vs MCIc feature subset

10 Computational Intelligence and Neuroscience

AD vs HC

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(a)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

AD vs MCI

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(b)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

HC vs MCI

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(c)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

MCIs vs MCIc

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(d)

Figure 4 ROC curves for discriminating AD MCIs MCIc and HC (healthy controls) ROC curves are plotted for the five biomarkersseparately and for the concatenation of all features (ie all five biomarkers measures) (a) ROC curves for AD vs HC classification (b) forAD vs MCI classification (c) for HC vs MCI classification and (d) for MCIs vs MCIc classification using two-stage feature selection withELM classifiers

Computational Intelligence and Neuroscience 11

in the tables the performance accuracy of feature fusion isnoticeably improved when compared with that of the in-dividual feature set and there are distinct degrees of ele-vation in other indexes the specificity index is particularlymore noticeable In contrast performance accuracy basedon the nonimaging features is almost the same as the im-aging features for AD diagnosis but far superior for MCIdiagnosis For AD vs HC the accuracy obtained by CSFmeasures genetics and cognitive score showed a lowerincrease although for MCI vs HC AD vs MCI and MCIsvs MCIc the accuracy was higher e result showed thatthe ELM classifier achieves better classification scorescompared to the SVM classifier (linear and RBF-SVM) for

both single modality feature sets as well as all featurescombined (shown in Figure 6) e ELM is good machinelearning tool and the major strength is that the hiddenlayerrsquos learning parameters including input weight andbiases do not tune iteratively as in SLFN e ELM offersmany advantages over other learning algorithms [54] Fromthe results shown in Tables 3ndash6 for AD vs HC the clas-sification result increased by 5ndash7 as compared to SVMclassifiers with 9804 sensitivity and 9628 specificityFor AD vs MCI classification accuracy increased by 5ndash9with 92 sensitivity and 9733 accuracy For HC vs MCIclassification accuracy increased by 6ndash10 with 9123sensitivity and 9913 specificity Similarly for MCIs vs

Table 3 10-fold cross-validated classification performance for AD vs HC

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 8513 9207 8544 086 8068 8568 8381 7717 8317 7381Surface area 8230 8930 9154 085 7616 8616 7514 7867 8767 7262Volume 8790 9387 8475 088 7929 7429 7833 8000 7710 6324CSF 8973 9633 8738 089 7905 8905 7417 7533 8333 6857APOE+MMSE 8645 9513 8700 093 8203 8703 8467 8416 8305 7879Concatenation (all_features_set) 9731 9804 9628 097 9133 9333 8757 9350 955 9058

Table 4 10-fold cross-validated classification performance for AD vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7010 8130 6610 071 6401 8383 6233 6503 7880 7330Surface area 6160 6670 8430 063 6345 6810 8011 6067 7123 6285Volume 6780 8210 6710 069 5820 7792 6856 6005 7337 6373CSF 6470 6600 8140 073 5607 6275 7573 6123 6871 7937APOE+MMSE 8145 8440 8770 082 7681 6781 8548 7620 6694 8748Concatenation (all_features_set) 8791 9200 9733 089 8350 8864 7672 7831 7445 8262

Table 5 10-fold cross-validated classification performance for HC vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7540 8407 6790 083 7417 8081 7123 6734 7317 7793Surface area 7415 6520 8385 084 6354 6767 6647 6872 7443 6735Volume 7784 8841 7280 084 6547 7473 6875 7283 7827 6570CSF 8330 7393 8402 085 7339 7525 6683 7580 7338 7814APOE+MMSE 7937 9325 7297 089 6829 8173 7668 7023 8124 7805Concatenation (all_features_set) 9172 9123 9913 093 8170 8368 8749 8565 7854 8926

Table 6 10-fold cross-validated classification performance for MCIs vs MCIc

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 6371 7110 3837 064 5718 6553 6038 5005 7118 6813Surface area 6143 7228 7480 069 5425 6322 7617 5827 7223 6485Volume 6785 8440 8773 074 5917 7828 6755 6214 7445 5773CSF 7137 7884 6857 072 6473 6009 7278 6733 7882 6878APOE+MMSE 7609 8713 7209 079 6553 6881 6743 6808 7221 6633Concatenation (all_features_set) 8338 9301 7577 085 7137 8116 7675 7571 7838 8467

12 Computational Intelligence and Neuroscience

MCIc classification accuracy increased by 7ndash12 with9301 sensitivity and 7577 specificity but 8467 and7667 specificity was obtained for linear and RBF-SVMclassifiers respectively which is more than obtained by theELM classifier From our analyses we believe that the ELM

gives better performance as compared to SVM from alearning efficiency standpoint because the ELMrsquos originaldesign has high learning accuracy fast learning speedscalability and the least human intervention [54 55] Fromthe results shown in Table 7 we can see that our suggested

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(a)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(b)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(c)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(d)

Figure 5 Classification results for different Alzheimerrsquos groups using ELM classifiers (a) Classification results for AD vs HC groups (b) forAD vs MCI groups (c) for HC vs MCI groups and (d) for MCIs vs MCIc groups

Computational Intelligence and Neuroscience 13

method achieved better accuracy than other existingmethodsFor classifying AD and HC our method achieved a classi-fication accuracy of 9731 with a sensitivity of 9804 aspecificity of 9628 and an AUC value of 097 Forclassifying MCI and HC our method achieves a classificationaccuracy of 9172 with a sensitivity of 9123 a specificity of9913 and an AUC value of 093 For distinguishing ADfrom MCI our proposed method obtained a classificationaccuracy of 8791 with a sensitivity of 9200 a specificityof 9733 and an AUC value of 089 Similarly for MCIs vsMCIc we achieved outstanding performance as compared toprevious methods by combining multiple features with anaccuracy of 8338 a sensitivity of 9301 a specificity of7577 and an AUC value of 085

42ComparisonwithOtherMethods In the previous sectionwe discussed in detail the findings of the proposed featureselection and classification method In this section we

compare and discuss the findings of our method in com-parison with existing state-of-the-art methods Tables 7 and 8show a comparison of the classification performance of theproposed method with recently published studies which usedmultimodality data to distinguish individuals with AD andMCI from HC Westman et al [67] used MRI and CSFbiomarkers and obtained 918 accuracy for AD vs HC776 for HC vs MCI and 685 for pMCI vs MCIsclassification using the orthogonal partial least squares tolatent structures (OPLS) method Zhang and Shen [39] used amultimodal (MRI PET and CSF) classification method withmultitask feature selection and an SVM classifier for AD andMCI classification By combining MRI PET and CSF datathey achieved a higher accuracy of 933 for AD vs HC and832 accuracy for HC vs MCI classification Similarly theauthors achieved an accuracy of 739 on pMCI vs MCIssamples Hinrichs et al [69] obtained an accuracy of 924 forAD vs HC classification using MRI PET CSF APOE and

ELMSVM_RBFSVM-linear

0

20

40

60

80

100

ACC

()

AD vs HC HC vs MCI MCIs vs MCIcAD vs MCI

Figure 6 Performance comparison of different classifiers

Table 7 Comparison of classification of AD vs HC and HC vs MCI between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection AD vsHC ()

HC vsMCI

Westman et al[67] MRI +CSF OPLS mdash 918 776

Johnson et al [68] MRI + PET+CSF+ cognitive scores Stackedautoencoder

Sparse representationlearning 959 85

Hinrichs et al [69] MRI + PET+CSF+APOE+ cognitive scores MKL mdash 924 naZhang and Shenet al [39] MRI + PET+CSF SVM Multitask feature selection 933 832

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 9301 mdash

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 995 mdash

Maqsood et al[72] sMRI CNN mdash 9285 mdash

Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper(SFS) 9731 9172

14 Computational Intelligence and Neuroscience

cognitive score in multiple kernel learning Similarly Johnsonet al [68] proposed a sparse representation learning featureselection and stacked autoencoder for classifying AD usingMRI PET CSF and cognitive scores as features and obtained959 accuracy for AD vs HC classification and 85 accuracyfor HC vs MCI classification Maqsood et al [72] proposedthe transfer learning classification model of AD for bothbinary and multiclass problems e algorithm utilizes apretrained AlexNet network and fine-tuned the convolutionalneural network (CNN) for classificationis model was fine-tuned over both segmented and unsegmented 3D views of thehuman brain e MRI scans were segmented into thecharacteristic components of GM white matter and CSFeretrained CNNwas then validated using the test data to obtainaccuracies of 896 and 928 for binary and multiclassproblems respectively using the OASIS dataset Beheshtiet al [70] developed a computer-aided diagnosis (CAD)system composed of four systematic stages for analyzingglobal and local differences in the GM of AD patientscompared with HC using the voxel-based morphometry(VBM) methodey used feature ranking based on the t-testand a genetic algorithm with Fisherrsquos criteria as part of theobjective function in the genetic algorithm e authorsutilized the SVM classifier with 10-fold cross-validation toobtain accuracies of 9301 and 75 for AD vs HC andMCIsvs pMCI classification respectively Spasov et al [71] de-veloped a deep learning architecture based on dual learningand an ad hoc layer for 3D separable convolution to identifyMCI patients eir deep learning procedure combined MRIneuropsychological demographic and APOE ε4 data toachieve an accuracy of 86 ey achieved an accuracy of995 for AD vs HC classification In another study Moradiet al used VBM analysis of GM as a feature combined withage and cognitive measures ey achieved an accuracy of82 for pMCI vs MCIs classification From this comparisonwe can infer that the proposed feature combination (MRICSF APOE and MMSE data) is robust or comparable to theother multimodal biomarker methods reported in the liter-ature for both AD vs HC and MCIs vs MCIc classification

5 Conclusion

In conclusion we demonstrated that a combination of threesMRImeasures cortical thickness cortical area cortical volumeand three nonimaging measures CSF components APOE ε4status and MMSE score improves AD diagnosis Furthermore

this combination shows great potential for the early identifi-cation of mild cognitive impairment (the prodromal stage ofAlzheimerrsquos disease) In this method we proposed filter andwrapper feature selection with an ELM classifier for multiplebiomarker-based AD diagnosis which significantly improvedthe classifierrsquos performance Moreover the results were betterthan or comparable with those previously reported particularlyfor the most challenging classification such as HC vs MCI andMCIs vs MCIc e added value of combining different ana-tomical MRI measures should be considered in AD scanningprotocols Only using specific aspects or a single measure ofwhole brain atrophy for AD diagnosis is still common practiceOur results show that clinical AD diagnosis could benefit fromthe combination of multiple measures from an anatomical MRIscan and other nonimaging biomarkers incorporated into anautomated machine learning system Our suggested methodeffectively enhances the diagnostic accuracy ofADandMCI butstill has some drawbacks Future work will include variousimprovements First we will optimize the parameter obtainingprocess Second in order to enhance the effectiveness of thesuggested method the dataset will be increased in terms of thefollowing extension of the longitudinal dataset for better un-derstanding of the progression from MCI and inclusion ofmultimodal data such as PET and fMRI data which providesdifferent insights into the characteristics of AD

Data Availability

e Alzheimerrsquos Disease Neuroimaging Initiative (ADNI)dataset (httpadniloniuscedu) was used in this studyComplete information regarding ADNI investigators can befound at httpadniloniusceduwp_contentuploadshow_to_applyADNI_Acknowledgement_istpd

Conflicts of Interest

e authors declare no conflicts of interest relating to thiswork

Acknowledgments

is work was supported by the National Research Foun-dation of Korea Grant funded by the Korean Government(NRF-2019R1F1A1060166 and NRF-2019R1A4A1029769)Data collection and sharing for this project was funded bythe ADNI (National Institutes of Health grant number U01

Table 8 Comparison of classification of MCIs vs MCIc between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection MCIs vs MCIc()

Westman et al [67] MRI +CSF OPLS mdash 685Zhang and Shen et al[39] MRI + PET+CSF SVM Multitask feature selection 7390

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 7500

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 86

Moradi et al [73] MRI + age and cognitive mdash 82Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper (SFS) 8338

Computational Intelligence and Neuroscience 15

parietal areas For HC vs MCI the left superiorfrontal left postcentral left superior parietal rightsupramarginal right fusiform right precuneus andright precentral areas showed the lowest valuesSimilarly for AD vs MCI the left superior parietalleft and right precentral left paracentral right in-ferior parietal and right superior frontal areas

showed the largest decreases in surface area ForMCIs vs MCIc the left fusiform left lateral occipitalleft parahippocampal right superior parietal andright postcentral areas showed the most differences

(3) In comparison with the HC group the volume ofgray matter of the left caudal anterior cingulate leftorbitalis left lateral orbitofrontal right lateral

icknessLe hemisphere Right hemisphere

VolumeRight hemisphereLe hemisphere

text

AreaLe hemisphere Right hemisphere

ndash500 500ndash250 250000

AD

vs

MCI

MCI

s vs

MCI

cA

D v

s H

CH

C vs

MCI

Figure 2 Differences in cortical thickness area and volume in patients at different stages of Alzheimerrsquos diseasee colored bar representsthe significance level of clusters e significance threshold was set at plt 005

Computational Intelligence and Neuroscience 7

Table 2 Cluster differences in cortical thickness area and volume in AD MCI and MCIc patients

Feature RegionCoordinates

Vertex Value Size (mm2)x y z

AD vs HC

ickness

Left insula minus295 179 119 3165 minus23512 361947Left parahippocampal minus313 minus417 minus87 557 minus22303 1073627

Left cuneus minus479 minus209 428 1039 minus23188 282428Right rostral middle frontal 385 43 7 1224 minus21737 148Right superior temporal 533 minus52 minus52 468 minus28676 25628Right pars opercularis 503 102 88 1242 minus36071 5815525

Area

Left paracentral minus158 minus355 494 1197 minus22767 701872Left lateral orbitofrontal minus533 minus2 75 489 minus21962 321823

Right paracentral 125 minus371 554 1233 minus22555 4255Right inferior parietal 34 minus519 374 1133 minus21026 721305

Right posterior cingulate 14 minus304 366 1201 minus20979 141667Right inferior parietal 395 minus447 346 7 minus20263 5076

Volume

Left caudal anterior cingulate minus5 184 264 2073 minus32264 83461Left orbitalis minus287 minus602 417 1103 23437 120255

Left lateral orbitofrontal minus594 minus69 94 412 minus20146 63567Right lateral orbitofrontal 304 229 minus203 631 minus27411 20256Right rostral middle frontal 348 51 47 198 minus2717 109882

Right pars opercularis 468 152 87 136 minus2717 80771AD vs MCI

ickness

Left inferior parietal minus436 minus615 33 939 27022 43937Left fusiform minus40 minus537 minus203 254 2459 14984

Left superior temporal minus477 minus25 minus91 210 minus25577 32803Right lateral occipital 261 minus939 37 363 34944 363Right inferior parietal 377 minus717 428 1129 33654 64624Right superior temporal 507 minus143 minus24 722 minus29715 34228

Area

Left superior parietal 284 minus51 427 1702 minus25085 64276Left precentral 322 minus218 608 109 18969 5154Left paracentral 141 minus367 524 238 minus18959 7968

Right inferior parietal minus352 minus872 137 50 minus18099 3453Right precentral minus542 minus11 71 51 minus17244 2134

Right superior frontal minus74 4 665 34 16541 1786

Volume

Left superior parietal minus285 minus588 40 540 35868 2161Left superior frontal minus85 412 303 48 minus26334 3398

Left caudal anterior cingulate minus49 113 321 314 minus2549 15007Right entorhinal 215 minus89 minus295 207 minus31433 5569

Right lateral orbitofrontal 428 278 minus137 198 minus31034 12501Right superior parietal 321 minus442 412 115 minus25309 4029

HC vs MCI

ickness

Left insula minus364 minus94 minus117 811 minus35988 30994Left precuneus minus6 minus689 417 628 23682 32202Left lingual minus293 minus445 minus66 612 minus27175 27916Right insula 358 minus106 minus74 1154 minus31385 43662

Right pars triangularis 389 317 1 372 minus21264 20487Right inferior parietal 351 minus72 409 285 minus27121 1522

Area

Left superior frontal minus178 315 496 237 17293 15688Left postcentral minus521 minus227 507 264 17286 12275

Left superior parietal minus32 minus493 473 141 minus17572 6604Right supramarginal 534 minus474 35 1336 27163 67685

Right fusiform 347 minus12 minus341 551 20892 32284Right precuneus 182 minus773 277 482 17491 31336Right precentral 207 minus306 539 328 minus19165 11545

Volume

Left supramarginal 521 minus463 226 441 27521 21023Left cuneus 171 minus695 168 628 2611 48174

Left precentral 212 minus307 538 376 minus21546 12702Right parahippocampal minus238 minus361 minus156 278 minus18212 12768Right superior parietal minus91 minus743 465 598 25199 30367Right superior frontal minus181 333 396 210 26035 11472

8 Computational Intelligence and Neuroscience

orbitofrontal right rostral middle frontal and rightpars opercularis areas was lower in the AD groupFor AD vs MCI the left superior parietal left su-perior frontal left caudal anterior cingulate rightentorhinal and right superior parietal areas showedthe largest decreases in volume For HC vs MCI theleft supramarginal left cuneus left precentral rightparahippocampal and right superior parietal areasshowed the most volume atrophy Similarly forMCIs vs MCIc the left bankssts left precentral leftrostral middle frontal right precentral and rightfusiform areas had the largest decreases in volume

Moreover from this analysis we observe that areathickness and volume in the AD group were significantlydecreased in comparison with the HC group Similarly therewas significant atrophy in the MCIc group in comparisonwith the MCIs group and there was little difference in theatrophy pattern among AD and MCI patients

32 Feature Selection and Classification e individualfeature set was selected using an MRMR (filter) and a se-quential feature selection (wrapper) method to identify theoptimal feature set for different groups and improve theclassifier accuracy Similarly we combined all the feature setsfrom different measures and applied the MRMR and SFSmethod to select the optimal features Multiple measuresfeature sets were created by combining cortical thicknessarea and volume from sMRI three component measure-ments from CSF APOE ε4 status andMMSE score First weselected the top 60 features from theMRMR algorithm basedon the features score and then we applied a sequentialfeature selection algorithm on these top 60 ranked featureswhich gave the optimal feature set to achieve the maximumclassification accuracy on ELM classifiers Figure 3 shows the

number of optimal features sequence after sequential featureselection on the top 60 ranked features obtained from theMRMR algorithm using ELM classifier Figure 3 shows onlythe selected features for the combined feature set From thiswe can assume that the proposed feature selection withcross-validation provides the optimal feature vector forinput to the classifiers In this proposed method classifi-cation performance was quantified by the number of featuresselected versus accuracy and area under the ROC curve

To further analyze the effectiveness of the purposeclassification method combining different measures wecalculated the AUCs for the concatenation of all featuresFigure 4shows the receiver operating curves for individualfeatures and all feature (imaging and nonimaging bio-markers ie multiple features) combinations for eachclassification group using the ELM classifier

4 Discussion

41 Performance Analysis In this study we first performedstatistical analysis and pattern classification to differentiateand identify atrophy patterns for the four groups (AD HCMCIs and MCIc) Individual sMRI was preprocessed usingthe FreeSurfer tool After preprocessing the statisticalanalysis results of sMRI QDEC was applied and finally weperformed the classification task by the proposed featureselection and classification method respectively For brainatrophy analysis we used three types of sMRI corticalmetrics (cortical thickness gray matter volume and surfacearea) In comparing the AD group with the HC group theinsula pars opercularis parahippocampal and superiortemporal areas were severely affected in terms of corticalthickness gray matter volume and surface area e corticalthickness of the left hemisphere was thinner than that of theleft hemisphere Similarly for AD vs MCI the most atrophy

Table 2 Continued

Feature RegionCoordinates

Vertex Value Size (mm2)x y z

MCIs vs MCIc

ickness

Left inferior parietal minus463 minus60 111 2770 minus37613 142748Left superior temporal minus455 minus04 minus208 1067 minus26742 46624Left parahippocampal minus317 minus403 minus101 550 minus23925 23985Right temporal pole 292 9 minus38 1449 minus55983 82282

Right superior temporal 623 minus347 152 1292 minus31646 54997Right inferior temporal 516 minus566 minus37 882 minus30694 50791

Right precentral 488 minus65 405 858 minus28315 35193

Area

Left fusiform minus364 minus295 minus22 659 minus30958 31221Left lateral occipital minus194 minus99 minus151 326 26022 25061Left parahippocampal minus188 minus335 minus14 224 minus20347 9759Right superior temporal 481 minus329 22 1515 minus25815 5798Right superior parietal 106 minus527 651 1697 minus22367 64152

Right postcentral 338 minus29 519 799 minus20859 36723

Volume

Left bankssts minus579 minus469 minus1 2196 minus26263 116258Left precentral minus361 minus22 517 2736 minus25339 10731

Left rostral middle frontal minus224 279 328 582 minus23081 32686Right superior temporal 623 minus347 152 3392 minus44677 144996

Right precentral 495 minus62 411 4103 minus32472 181394Right fusiform 339 minus378 minus228 782 minus31822 42505

Computational Intelligence and Neuroscience 9

was seen in the left inferior parietal and right lateral occipitalareas For HC vs MCI the supramarginal and cuneus areasshowed the most atrophy For MCIs vs MCIc the superiortemporal and precentral areas showed the largest decreasesin thickness volume and area Based on these differencesthe majority of the decrease in cortical thickness gray mattervolume and surface area appears mostly in the frontal lobetemporal lobe occipital lobe cingulate gyrus and parietallobe is phenomenon strongly agrees with findings relatedto atrophy patterns seen in previous studies [14 66] eseregions are mainly involved in the expression of personalitymotor execution complex cognitive behavior and decisionmaking [50] In addition we present the less commonanalysis of MCIs vs MCIc For MCIc the most atrophy wasseen in the superior temporal bankssts precentral inferior

parietal and insula areas which show potential for the earlyrecognition of progression to Alzheimerrsquos disease

To test the effectiveness of our purposed featurescombination method (ie imaging and nonimaging fea-tures) we adapted the individual feature from sMRI and CSFseparately to conduct the study although regarding geneticand cognitive features we combined them to test the per-formance and then compared the accuracy with the accuracyof the combination of all features For the proposed featureselection we used a combination of filter and wrapper al-gorithms and compared the performance of the classifiers onthe selected feature set In this text SVM-linear SVM-RBFand ELM were used for classification e classificationresults for the compared methods are presented inTables 3ndash6 and also in graphical form in Figure 5 As shown

AD vs HC

Number of selected features = 27

05

101520253035404550556065707580859095

100

Accu

racy

()

5 10 15 20 25 30 35 40 45 50 55 600

Number of features sequence

(a)

AD vs MCI

Number of selected features = 17

05

101520253035404550556065707580859095

100

Accu

racy

()

5 10 15 20 25 30 35 40 45 50 55 600

Number of feature sequence

(b)HC vs MCI

Number of selected features = 25

5 10 15 20 25 30 35 40 45 50 55 600

Number of feature sequence

05

101520253035404550556065707580859095

100

Accu

racy

()

(c)

MCIs vs MCIc

Number of selected features = 13

05

101520253035404550556065707580859095

100Ac

cura

cy (

)

5 10 15 20 25 30 35 40 45 50 55 600

Number of features sequence

(d)

Figure 3 Number of feature sets selected from the sequential feature selection algorithm on the top 60 ranked features obtained using theMRMR algorithm Only the number of features versus accuracy for the combination of all feature sets using ELM classifiers is shown (a) ADvs HC feature subset (b) AD vs MCI feature subset (c) HC vs MCI feature subset (d) MCIs vs MCIc feature subset

10 Computational Intelligence and Neuroscience

AD vs HC

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(a)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

AD vs MCI

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(b)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

HC vs MCI

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(c)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

MCIs vs MCIc

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(d)

Figure 4 ROC curves for discriminating AD MCIs MCIc and HC (healthy controls) ROC curves are plotted for the five biomarkersseparately and for the concatenation of all features (ie all five biomarkers measures) (a) ROC curves for AD vs HC classification (b) forAD vs MCI classification (c) for HC vs MCI classification and (d) for MCIs vs MCIc classification using two-stage feature selection withELM classifiers

Computational Intelligence and Neuroscience 11

in the tables the performance accuracy of feature fusion isnoticeably improved when compared with that of the in-dividual feature set and there are distinct degrees of ele-vation in other indexes the specificity index is particularlymore noticeable In contrast performance accuracy basedon the nonimaging features is almost the same as the im-aging features for AD diagnosis but far superior for MCIdiagnosis For AD vs HC the accuracy obtained by CSFmeasures genetics and cognitive score showed a lowerincrease although for MCI vs HC AD vs MCI and MCIsvs MCIc the accuracy was higher e result showed thatthe ELM classifier achieves better classification scorescompared to the SVM classifier (linear and RBF-SVM) for

both single modality feature sets as well as all featurescombined (shown in Figure 6) e ELM is good machinelearning tool and the major strength is that the hiddenlayerrsquos learning parameters including input weight andbiases do not tune iteratively as in SLFN e ELM offersmany advantages over other learning algorithms [54] Fromthe results shown in Tables 3ndash6 for AD vs HC the clas-sification result increased by 5ndash7 as compared to SVMclassifiers with 9804 sensitivity and 9628 specificityFor AD vs MCI classification accuracy increased by 5ndash9with 92 sensitivity and 9733 accuracy For HC vs MCIclassification accuracy increased by 6ndash10 with 9123sensitivity and 9913 specificity Similarly for MCIs vs

Table 3 10-fold cross-validated classification performance for AD vs HC

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 8513 9207 8544 086 8068 8568 8381 7717 8317 7381Surface area 8230 8930 9154 085 7616 8616 7514 7867 8767 7262Volume 8790 9387 8475 088 7929 7429 7833 8000 7710 6324CSF 8973 9633 8738 089 7905 8905 7417 7533 8333 6857APOE+MMSE 8645 9513 8700 093 8203 8703 8467 8416 8305 7879Concatenation (all_features_set) 9731 9804 9628 097 9133 9333 8757 9350 955 9058

Table 4 10-fold cross-validated classification performance for AD vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7010 8130 6610 071 6401 8383 6233 6503 7880 7330Surface area 6160 6670 8430 063 6345 6810 8011 6067 7123 6285Volume 6780 8210 6710 069 5820 7792 6856 6005 7337 6373CSF 6470 6600 8140 073 5607 6275 7573 6123 6871 7937APOE+MMSE 8145 8440 8770 082 7681 6781 8548 7620 6694 8748Concatenation (all_features_set) 8791 9200 9733 089 8350 8864 7672 7831 7445 8262

Table 5 10-fold cross-validated classification performance for HC vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7540 8407 6790 083 7417 8081 7123 6734 7317 7793Surface area 7415 6520 8385 084 6354 6767 6647 6872 7443 6735Volume 7784 8841 7280 084 6547 7473 6875 7283 7827 6570CSF 8330 7393 8402 085 7339 7525 6683 7580 7338 7814APOE+MMSE 7937 9325 7297 089 6829 8173 7668 7023 8124 7805Concatenation (all_features_set) 9172 9123 9913 093 8170 8368 8749 8565 7854 8926

Table 6 10-fold cross-validated classification performance for MCIs vs MCIc

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 6371 7110 3837 064 5718 6553 6038 5005 7118 6813Surface area 6143 7228 7480 069 5425 6322 7617 5827 7223 6485Volume 6785 8440 8773 074 5917 7828 6755 6214 7445 5773CSF 7137 7884 6857 072 6473 6009 7278 6733 7882 6878APOE+MMSE 7609 8713 7209 079 6553 6881 6743 6808 7221 6633Concatenation (all_features_set) 8338 9301 7577 085 7137 8116 7675 7571 7838 8467

12 Computational Intelligence and Neuroscience

MCIc classification accuracy increased by 7ndash12 with9301 sensitivity and 7577 specificity but 8467 and7667 specificity was obtained for linear and RBF-SVMclassifiers respectively which is more than obtained by theELM classifier From our analyses we believe that the ELM

gives better performance as compared to SVM from alearning efficiency standpoint because the ELMrsquos originaldesign has high learning accuracy fast learning speedscalability and the least human intervention [54 55] Fromthe results shown in Table 7 we can see that our suggested

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(a)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(b)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(c)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(d)

Figure 5 Classification results for different Alzheimerrsquos groups using ELM classifiers (a) Classification results for AD vs HC groups (b) forAD vs MCI groups (c) for HC vs MCI groups and (d) for MCIs vs MCIc groups

Computational Intelligence and Neuroscience 13

method achieved better accuracy than other existingmethodsFor classifying AD and HC our method achieved a classi-fication accuracy of 9731 with a sensitivity of 9804 aspecificity of 9628 and an AUC value of 097 Forclassifying MCI and HC our method achieves a classificationaccuracy of 9172 with a sensitivity of 9123 a specificity of9913 and an AUC value of 093 For distinguishing ADfrom MCI our proposed method obtained a classificationaccuracy of 8791 with a sensitivity of 9200 a specificityof 9733 and an AUC value of 089 Similarly for MCIs vsMCIc we achieved outstanding performance as compared toprevious methods by combining multiple features with anaccuracy of 8338 a sensitivity of 9301 a specificity of7577 and an AUC value of 085

42ComparisonwithOtherMethods In the previous sectionwe discussed in detail the findings of the proposed featureselection and classification method In this section we

compare and discuss the findings of our method in com-parison with existing state-of-the-art methods Tables 7 and 8show a comparison of the classification performance of theproposed method with recently published studies which usedmultimodality data to distinguish individuals with AD andMCI from HC Westman et al [67] used MRI and CSFbiomarkers and obtained 918 accuracy for AD vs HC776 for HC vs MCI and 685 for pMCI vs MCIsclassification using the orthogonal partial least squares tolatent structures (OPLS) method Zhang and Shen [39] used amultimodal (MRI PET and CSF) classification method withmultitask feature selection and an SVM classifier for AD andMCI classification By combining MRI PET and CSF datathey achieved a higher accuracy of 933 for AD vs HC and832 accuracy for HC vs MCI classification Similarly theauthors achieved an accuracy of 739 on pMCI vs MCIssamples Hinrichs et al [69] obtained an accuracy of 924 forAD vs HC classification using MRI PET CSF APOE and

ELMSVM_RBFSVM-linear

0

20

40

60

80

100

ACC

()

AD vs HC HC vs MCI MCIs vs MCIcAD vs MCI

Figure 6 Performance comparison of different classifiers

Table 7 Comparison of classification of AD vs HC and HC vs MCI between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection AD vsHC ()

HC vsMCI

Westman et al[67] MRI +CSF OPLS mdash 918 776

Johnson et al [68] MRI + PET+CSF+ cognitive scores Stackedautoencoder

Sparse representationlearning 959 85

Hinrichs et al [69] MRI + PET+CSF+APOE+ cognitive scores MKL mdash 924 naZhang and Shenet al [39] MRI + PET+CSF SVM Multitask feature selection 933 832

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 9301 mdash

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 995 mdash

Maqsood et al[72] sMRI CNN mdash 9285 mdash

Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper(SFS) 9731 9172

14 Computational Intelligence and Neuroscience

cognitive score in multiple kernel learning Similarly Johnsonet al [68] proposed a sparse representation learning featureselection and stacked autoencoder for classifying AD usingMRI PET CSF and cognitive scores as features and obtained959 accuracy for AD vs HC classification and 85 accuracyfor HC vs MCI classification Maqsood et al [72] proposedthe transfer learning classification model of AD for bothbinary and multiclass problems e algorithm utilizes apretrained AlexNet network and fine-tuned the convolutionalneural network (CNN) for classificationis model was fine-tuned over both segmented and unsegmented 3D views of thehuman brain e MRI scans were segmented into thecharacteristic components of GM white matter and CSFeretrained CNNwas then validated using the test data to obtainaccuracies of 896 and 928 for binary and multiclassproblems respectively using the OASIS dataset Beheshtiet al [70] developed a computer-aided diagnosis (CAD)system composed of four systematic stages for analyzingglobal and local differences in the GM of AD patientscompared with HC using the voxel-based morphometry(VBM) methodey used feature ranking based on the t-testand a genetic algorithm with Fisherrsquos criteria as part of theobjective function in the genetic algorithm e authorsutilized the SVM classifier with 10-fold cross-validation toobtain accuracies of 9301 and 75 for AD vs HC andMCIsvs pMCI classification respectively Spasov et al [71] de-veloped a deep learning architecture based on dual learningand an ad hoc layer for 3D separable convolution to identifyMCI patients eir deep learning procedure combined MRIneuropsychological demographic and APOE ε4 data toachieve an accuracy of 86 ey achieved an accuracy of995 for AD vs HC classification In another study Moradiet al used VBM analysis of GM as a feature combined withage and cognitive measures ey achieved an accuracy of82 for pMCI vs MCIs classification From this comparisonwe can infer that the proposed feature combination (MRICSF APOE and MMSE data) is robust or comparable to theother multimodal biomarker methods reported in the liter-ature for both AD vs HC and MCIs vs MCIc classification

5 Conclusion

In conclusion we demonstrated that a combination of threesMRImeasures cortical thickness cortical area cortical volumeand three nonimaging measures CSF components APOE ε4status and MMSE score improves AD diagnosis Furthermore

this combination shows great potential for the early identifi-cation of mild cognitive impairment (the prodromal stage ofAlzheimerrsquos disease) In this method we proposed filter andwrapper feature selection with an ELM classifier for multiplebiomarker-based AD diagnosis which significantly improvedthe classifierrsquos performance Moreover the results were betterthan or comparable with those previously reported particularlyfor the most challenging classification such as HC vs MCI andMCIs vs MCIc e added value of combining different ana-tomical MRI measures should be considered in AD scanningprotocols Only using specific aspects or a single measure ofwhole brain atrophy for AD diagnosis is still common practiceOur results show that clinical AD diagnosis could benefit fromthe combination of multiple measures from an anatomical MRIscan and other nonimaging biomarkers incorporated into anautomated machine learning system Our suggested methodeffectively enhances the diagnostic accuracy ofADandMCI butstill has some drawbacks Future work will include variousimprovements First we will optimize the parameter obtainingprocess Second in order to enhance the effectiveness of thesuggested method the dataset will be increased in terms of thefollowing extension of the longitudinal dataset for better un-derstanding of the progression from MCI and inclusion ofmultimodal data such as PET and fMRI data which providesdifferent insights into the characteristics of AD

Data Availability

e Alzheimerrsquos Disease Neuroimaging Initiative (ADNI)dataset (httpadniloniuscedu) was used in this studyComplete information regarding ADNI investigators can befound at httpadniloniusceduwp_contentuploadshow_to_applyADNI_Acknowledgement_istpd

Conflicts of Interest

e authors declare no conflicts of interest relating to thiswork

Acknowledgments

is work was supported by the National Research Foun-dation of Korea Grant funded by the Korean Government(NRF-2019R1F1A1060166 and NRF-2019R1A4A1029769)Data collection and sharing for this project was funded bythe ADNI (National Institutes of Health grant number U01

Table 8 Comparison of classification of MCIs vs MCIc between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection MCIs vs MCIc()

Westman et al [67] MRI +CSF OPLS mdash 685Zhang and Shen et al[39] MRI + PET+CSF SVM Multitask feature selection 7390

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 7500

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 86

Moradi et al [73] MRI + age and cognitive mdash 82Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper (SFS) 8338

Computational Intelligence and Neuroscience 15

Table 2 Cluster differences in cortical thickness area and volume in AD MCI and MCIc patients

Feature RegionCoordinates

Vertex Value Size (mm2)x y z

AD vs HC

ickness

Left insula minus295 179 119 3165 minus23512 361947Left parahippocampal minus313 minus417 minus87 557 minus22303 1073627

Left cuneus minus479 minus209 428 1039 minus23188 282428Right rostral middle frontal 385 43 7 1224 minus21737 148Right superior temporal 533 minus52 minus52 468 minus28676 25628Right pars opercularis 503 102 88 1242 minus36071 5815525

Area

Left paracentral minus158 minus355 494 1197 minus22767 701872Left lateral orbitofrontal minus533 minus2 75 489 minus21962 321823

Right paracentral 125 minus371 554 1233 minus22555 4255Right inferior parietal 34 minus519 374 1133 minus21026 721305

Right posterior cingulate 14 minus304 366 1201 minus20979 141667Right inferior parietal 395 minus447 346 7 minus20263 5076

Volume

Left caudal anterior cingulate minus5 184 264 2073 minus32264 83461Left orbitalis minus287 minus602 417 1103 23437 120255

Left lateral orbitofrontal minus594 minus69 94 412 minus20146 63567Right lateral orbitofrontal 304 229 minus203 631 minus27411 20256Right rostral middle frontal 348 51 47 198 minus2717 109882

Right pars opercularis 468 152 87 136 minus2717 80771AD vs MCI

ickness

Left inferior parietal minus436 minus615 33 939 27022 43937Left fusiform minus40 minus537 minus203 254 2459 14984

Left superior temporal minus477 minus25 minus91 210 minus25577 32803Right lateral occipital 261 minus939 37 363 34944 363Right inferior parietal 377 minus717 428 1129 33654 64624Right superior temporal 507 minus143 minus24 722 minus29715 34228

Area

Left superior parietal 284 minus51 427 1702 minus25085 64276Left precentral 322 minus218 608 109 18969 5154Left paracentral 141 minus367 524 238 minus18959 7968

Right inferior parietal minus352 minus872 137 50 minus18099 3453Right precentral minus542 minus11 71 51 minus17244 2134

Right superior frontal minus74 4 665 34 16541 1786

Volume

Left superior parietal minus285 minus588 40 540 35868 2161Left superior frontal minus85 412 303 48 minus26334 3398

Left caudal anterior cingulate minus49 113 321 314 minus2549 15007Right entorhinal 215 minus89 minus295 207 minus31433 5569

Right lateral orbitofrontal 428 278 minus137 198 minus31034 12501Right superior parietal 321 minus442 412 115 minus25309 4029

HC vs MCI

ickness

Left insula minus364 minus94 minus117 811 minus35988 30994Left precuneus minus6 minus689 417 628 23682 32202Left lingual minus293 minus445 minus66 612 minus27175 27916Right insula 358 minus106 minus74 1154 minus31385 43662

Right pars triangularis 389 317 1 372 minus21264 20487Right inferior parietal 351 minus72 409 285 minus27121 1522

Area

Left superior frontal minus178 315 496 237 17293 15688Left postcentral minus521 minus227 507 264 17286 12275

Left superior parietal minus32 minus493 473 141 minus17572 6604Right supramarginal 534 minus474 35 1336 27163 67685

Right fusiform 347 minus12 minus341 551 20892 32284Right precuneus 182 minus773 277 482 17491 31336Right precentral 207 minus306 539 328 minus19165 11545

Volume

Left supramarginal 521 minus463 226 441 27521 21023Left cuneus 171 minus695 168 628 2611 48174

Left precentral 212 minus307 538 376 minus21546 12702Right parahippocampal minus238 minus361 minus156 278 minus18212 12768Right superior parietal minus91 minus743 465 598 25199 30367Right superior frontal minus181 333 396 210 26035 11472

8 Computational Intelligence and Neuroscience

orbitofrontal right rostral middle frontal and rightpars opercularis areas was lower in the AD groupFor AD vs MCI the left superior parietal left su-perior frontal left caudal anterior cingulate rightentorhinal and right superior parietal areas showedthe largest decreases in volume For HC vs MCI theleft supramarginal left cuneus left precentral rightparahippocampal and right superior parietal areasshowed the most volume atrophy Similarly forMCIs vs MCIc the left bankssts left precentral leftrostral middle frontal right precentral and rightfusiform areas had the largest decreases in volume

Moreover from this analysis we observe that areathickness and volume in the AD group were significantlydecreased in comparison with the HC group Similarly therewas significant atrophy in the MCIc group in comparisonwith the MCIs group and there was little difference in theatrophy pattern among AD and MCI patients

32 Feature Selection and Classification e individualfeature set was selected using an MRMR (filter) and a se-quential feature selection (wrapper) method to identify theoptimal feature set for different groups and improve theclassifier accuracy Similarly we combined all the feature setsfrom different measures and applied the MRMR and SFSmethod to select the optimal features Multiple measuresfeature sets were created by combining cortical thicknessarea and volume from sMRI three component measure-ments from CSF APOE ε4 status andMMSE score First weselected the top 60 features from theMRMR algorithm basedon the features score and then we applied a sequentialfeature selection algorithm on these top 60 ranked featureswhich gave the optimal feature set to achieve the maximumclassification accuracy on ELM classifiers Figure 3 shows the

number of optimal features sequence after sequential featureselection on the top 60 ranked features obtained from theMRMR algorithm using ELM classifier Figure 3 shows onlythe selected features for the combined feature set From thiswe can assume that the proposed feature selection withcross-validation provides the optimal feature vector forinput to the classifiers In this proposed method classifi-cation performance was quantified by the number of featuresselected versus accuracy and area under the ROC curve

To further analyze the effectiveness of the purposeclassification method combining different measures wecalculated the AUCs for the concatenation of all featuresFigure 4shows the receiver operating curves for individualfeatures and all feature (imaging and nonimaging bio-markers ie multiple features) combinations for eachclassification group using the ELM classifier

4 Discussion

41 Performance Analysis In this study we first performedstatistical analysis and pattern classification to differentiateand identify atrophy patterns for the four groups (AD HCMCIs and MCIc) Individual sMRI was preprocessed usingthe FreeSurfer tool After preprocessing the statisticalanalysis results of sMRI QDEC was applied and finally weperformed the classification task by the proposed featureselection and classification method respectively For brainatrophy analysis we used three types of sMRI corticalmetrics (cortical thickness gray matter volume and surfacearea) In comparing the AD group with the HC group theinsula pars opercularis parahippocampal and superiortemporal areas were severely affected in terms of corticalthickness gray matter volume and surface area e corticalthickness of the left hemisphere was thinner than that of theleft hemisphere Similarly for AD vs MCI the most atrophy

Table 2 Continued

Feature RegionCoordinates

Vertex Value Size (mm2)x y z

MCIs vs MCIc

ickness

Left inferior parietal minus463 minus60 111 2770 minus37613 142748Left superior temporal minus455 minus04 minus208 1067 minus26742 46624Left parahippocampal minus317 minus403 minus101 550 minus23925 23985Right temporal pole 292 9 minus38 1449 minus55983 82282

Right superior temporal 623 minus347 152 1292 minus31646 54997Right inferior temporal 516 minus566 minus37 882 minus30694 50791

Right precentral 488 minus65 405 858 minus28315 35193

Area

Left fusiform minus364 minus295 minus22 659 minus30958 31221Left lateral occipital minus194 minus99 minus151 326 26022 25061Left parahippocampal minus188 minus335 minus14 224 minus20347 9759Right superior temporal 481 minus329 22 1515 minus25815 5798Right superior parietal 106 minus527 651 1697 minus22367 64152

Right postcentral 338 minus29 519 799 minus20859 36723

Volume

Left bankssts minus579 minus469 minus1 2196 minus26263 116258Left precentral minus361 minus22 517 2736 minus25339 10731

Left rostral middle frontal minus224 279 328 582 minus23081 32686Right superior temporal 623 minus347 152 3392 minus44677 144996

Right precentral 495 minus62 411 4103 minus32472 181394Right fusiform 339 minus378 minus228 782 minus31822 42505

Computational Intelligence and Neuroscience 9

was seen in the left inferior parietal and right lateral occipitalareas For HC vs MCI the supramarginal and cuneus areasshowed the most atrophy For MCIs vs MCIc the superiortemporal and precentral areas showed the largest decreasesin thickness volume and area Based on these differencesthe majority of the decrease in cortical thickness gray mattervolume and surface area appears mostly in the frontal lobetemporal lobe occipital lobe cingulate gyrus and parietallobe is phenomenon strongly agrees with findings relatedto atrophy patterns seen in previous studies [14 66] eseregions are mainly involved in the expression of personalitymotor execution complex cognitive behavior and decisionmaking [50] In addition we present the less commonanalysis of MCIs vs MCIc For MCIc the most atrophy wasseen in the superior temporal bankssts precentral inferior

parietal and insula areas which show potential for the earlyrecognition of progression to Alzheimerrsquos disease

To test the effectiveness of our purposed featurescombination method (ie imaging and nonimaging fea-tures) we adapted the individual feature from sMRI and CSFseparately to conduct the study although regarding geneticand cognitive features we combined them to test the per-formance and then compared the accuracy with the accuracyof the combination of all features For the proposed featureselection we used a combination of filter and wrapper al-gorithms and compared the performance of the classifiers onthe selected feature set In this text SVM-linear SVM-RBFand ELM were used for classification e classificationresults for the compared methods are presented inTables 3ndash6 and also in graphical form in Figure 5 As shown

AD vs HC

Number of selected features = 27

05

101520253035404550556065707580859095

100

Accu

racy

()

5 10 15 20 25 30 35 40 45 50 55 600

Number of features sequence

(a)

AD vs MCI

Number of selected features = 17

05

101520253035404550556065707580859095

100

Accu

racy

()

5 10 15 20 25 30 35 40 45 50 55 600

Number of feature sequence

(b)HC vs MCI

Number of selected features = 25

5 10 15 20 25 30 35 40 45 50 55 600

Number of feature sequence

05

101520253035404550556065707580859095

100

Accu

racy

()

(c)

MCIs vs MCIc

Number of selected features = 13

05

101520253035404550556065707580859095

100Ac

cura

cy (

)

5 10 15 20 25 30 35 40 45 50 55 600

Number of features sequence

(d)

Figure 3 Number of feature sets selected from the sequential feature selection algorithm on the top 60 ranked features obtained using theMRMR algorithm Only the number of features versus accuracy for the combination of all feature sets using ELM classifiers is shown (a) ADvs HC feature subset (b) AD vs MCI feature subset (c) HC vs MCI feature subset (d) MCIs vs MCIc feature subset

10 Computational Intelligence and Neuroscience

AD vs HC

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(a)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

AD vs MCI

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(b)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

HC vs MCI

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(c)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

MCIs vs MCIc

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(d)

Figure 4 ROC curves for discriminating AD MCIs MCIc and HC (healthy controls) ROC curves are plotted for the five biomarkersseparately and for the concatenation of all features (ie all five biomarkers measures) (a) ROC curves for AD vs HC classification (b) forAD vs MCI classification (c) for HC vs MCI classification and (d) for MCIs vs MCIc classification using two-stage feature selection withELM classifiers

Computational Intelligence and Neuroscience 11

in the tables the performance accuracy of feature fusion isnoticeably improved when compared with that of the in-dividual feature set and there are distinct degrees of ele-vation in other indexes the specificity index is particularlymore noticeable In contrast performance accuracy basedon the nonimaging features is almost the same as the im-aging features for AD diagnosis but far superior for MCIdiagnosis For AD vs HC the accuracy obtained by CSFmeasures genetics and cognitive score showed a lowerincrease although for MCI vs HC AD vs MCI and MCIsvs MCIc the accuracy was higher e result showed thatthe ELM classifier achieves better classification scorescompared to the SVM classifier (linear and RBF-SVM) for

both single modality feature sets as well as all featurescombined (shown in Figure 6) e ELM is good machinelearning tool and the major strength is that the hiddenlayerrsquos learning parameters including input weight andbiases do not tune iteratively as in SLFN e ELM offersmany advantages over other learning algorithms [54] Fromthe results shown in Tables 3ndash6 for AD vs HC the clas-sification result increased by 5ndash7 as compared to SVMclassifiers with 9804 sensitivity and 9628 specificityFor AD vs MCI classification accuracy increased by 5ndash9with 92 sensitivity and 9733 accuracy For HC vs MCIclassification accuracy increased by 6ndash10 with 9123sensitivity and 9913 specificity Similarly for MCIs vs

Table 3 10-fold cross-validated classification performance for AD vs HC

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 8513 9207 8544 086 8068 8568 8381 7717 8317 7381Surface area 8230 8930 9154 085 7616 8616 7514 7867 8767 7262Volume 8790 9387 8475 088 7929 7429 7833 8000 7710 6324CSF 8973 9633 8738 089 7905 8905 7417 7533 8333 6857APOE+MMSE 8645 9513 8700 093 8203 8703 8467 8416 8305 7879Concatenation (all_features_set) 9731 9804 9628 097 9133 9333 8757 9350 955 9058

Table 4 10-fold cross-validated classification performance for AD vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7010 8130 6610 071 6401 8383 6233 6503 7880 7330Surface area 6160 6670 8430 063 6345 6810 8011 6067 7123 6285Volume 6780 8210 6710 069 5820 7792 6856 6005 7337 6373CSF 6470 6600 8140 073 5607 6275 7573 6123 6871 7937APOE+MMSE 8145 8440 8770 082 7681 6781 8548 7620 6694 8748Concatenation (all_features_set) 8791 9200 9733 089 8350 8864 7672 7831 7445 8262

Table 5 10-fold cross-validated classification performance for HC vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7540 8407 6790 083 7417 8081 7123 6734 7317 7793Surface area 7415 6520 8385 084 6354 6767 6647 6872 7443 6735Volume 7784 8841 7280 084 6547 7473 6875 7283 7827 6570CSF 8330 7393 8402 085 7339 7525 6683 7580 7338 7814APOE+MMSE 7937 9325 7297 089 6829 8173 7668 7023 8124 7805Concatenation (all_features_set) 9172 9123 9913 093 8170 8368 8749 8565 7854 8926

Table 6 10-fold cross-validated classification performance for MCIs vs MCIc

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 6371 7110 3837 064 5718 6553 6038 5005 7118 6813Surface area 6143 7228 7480 069 5425 6322 7617 5827 7223 6485Volume 6785 8440 8773 074 5917 7828 6755 6214 7445 5773CSF 7137 7884 6857 072 6473 6009 7278 6733 7882 6878APOE+MMSE 7609 8713 7209 079 6553 6881 6743 6808 7221 6633Concatenation (all_features_set) 8338 9301 7577 085 7137 8116 7675 7571 7838 8467

12 Computational Intelligence and Neuroscience

MCIc classification accuracy increased by 7ndash12 with9301 sensitivity and 7577 specificity but 8467 and7667 specificity was obtained for linear and RBF-SVMclassifiers respectively which is more than obtained by theELM classifier From our analyses we believe that the ELM

gives better performance as compared to SVM from alearning efficiency standpoint because the ELMrsquos originaldesign has high learning accuracy fast learning speedscalability and the least human intervention [54 55] Fromthe results shown in Table 7 we can see that our suggested

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(a)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(b)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(c)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(d)

Figure 5 Classification results for different Alzheimerrsquos groups using ELM classifiers (a) Classification results for AD vs HC groups (b) forAD vs MCI groups (c) for HC vs MCI groups and (d) for MCIs vs MCIc groups

Computational Intelligence and Neuroscience 13

method achieved better accuracy than other existingmethodsFor classifying AD and HC our method achieved a classi-fication accuracy of 9731 with a sensitivity of 9804 aspecificity of 9628 and an AUC value of 097 Forclassifying MCI and HC our method achieves a classificationaccuracy of 9172 with a sensitivity of 9123 a specificity of9913 and an AUC value of 093 For distinguishing ADfrom MCI our proposed method obtained a classificationaccuracy of 8791 with a sensitivity of 9200 a specificityof 9733 and an AUC value of 089 Similarly for MCIs vsMCIc we achieved outstanding performance as compared toprevious methods by combining multiple features with anaccuracy of 8338 a sensitivity of 9301 a specificity of7577 and an AUC value of 085

42ComparisonwithOtherMethods In the previous sectionwe discussed in detail the findings of the proposed featureselection and classification method In this section we

compare and discuss the findings of our method in com-parison with existing state-of-the-art methods Tables 7 and 8show a comparison of the classification performance of theproposed method with recently published studies which usedmultimodality data to distinguish individuals with AD andMCI from HC Westman et al [67] used MRI and CSFbiomarkers and obtained 918 accuracy for AD vs HC776 for HC vs MCI and 685 for pMCI vs MCIsclassification using the orthogonal partial least squares tolatent structures (OPLS) method Zhang and Shen [39] used amultimodal (MRI PET and CSF) classification method withmultitask feature selection and an SVM classifier for AD andMCI classification By combining MRI PET and CSF datathey achieved a higher accuracy of 933 for AD vs HC and832 accuracy for HC vs MCI classification Similarly theauthors achieved an accuracy of 739 on pMCI vs MCIssamples Hinrichs et al [69] obtained an accuracy of 924 forAD vs HC classification using MRI PET CSF APOE and

ELMSVM_RBFSVM-linear

0

20

40

60

80

100

ACC

()

AD vs HC HC vs MCI MCIs vs MCIcAD vs MCI

Figure 6 Performance comparison of different classifiers

Table 7 Comparison of classification of AD vs HC and HC vs MCI between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection AD vsHC ()

HC vsMCI

Westman et al[67] MRI +CSF OPLS mdash 918 776

Johnson et al [68] MRI + PET+CSF+ cognitive scores Stackedautoencoder

Sparse representationlearning 959 85

Hinrichs et al [69] MRI + PET+CSF+APOE+ cognitive scores MKL mdash 924 naZhang and Shenet al [39] MRI + PET+CSF SVM Multitask feature selection 933 832

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 9301 mdash

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 995 mdash

Maqsood et al[72] sMRI CNN mdash 9285 mdash

Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper(SFS) 9731 9172

14 Computational Intelligence and Neuroscience

cognitive score in multiple kernel learning Similarly Johnsonet al [68] proposed a sparse representation learning featureselection and stacked autoencoder for classifying AD usingMRI PET CSF and cognitive scores as features and obtained959 accuracy for AD vs HC classification and 85 accuracyfor HC vs MCI classification Maqsood et al [72] proposedthe transfer learning classification model of AD for bothbinary and multiclass problems e algorithm utilizes apretrained AlexNet network and fine-tuned the convolutionalneural network (CNN) for classificationis model was fine-tuned over both segmented and unsegmented 3D views of thehuman brain e MRI scans were segmented into thecharacteristic components of GM white matter and CSFeretrained CNNwas then validated using the test data to obtainaccuracies of 896 and 928 for binary and multiclassproblems respectively using the OASIS dataset Beheshtiet al [70] developed a computer-aided diagnosis (CAD)system composed of four systematic stages for analyzingglobal and local differences in the GM of AD patientscompared with HC using the voxel-based morphometry(VBM) methodey used feature ranking based on the t-testand a genetic algorithm with Fisherrsquos criteria as part of theobjective function in the genetic algorithm e authorsutilized the SVM classifier with 10-fold cross-validation toobtain accuracies of 9301 and 75 for AD vs HC andMCIsvs pMCI classification respectively Spasov et al [71] de-veloped a deep learning architecture based on dual learningand an ad hoc layer for 3D separable convolution to identifyMCI patients eir deep learning procedure combined MRIneuropsychological demographic and APOE ε4 data toachieve an accuracy of 86 ey achieved an accuracy of995 for AD vs HC classification In another study Moradiet al used VBM analysis of GM as a feature combined withage and cognitive measures ey achieved an accuracy of82 for pMCI vs MCIs classification From this comparisonwe can infer that the proposed feature combination (MRICSF APOE and MMSE data) is robust or comparable to theother multimodal biomarker methods reported in the liter-ature for both AD vs HC and MCIs vs MCIc classification

5 Conclusion

In conclusion we demonstrated that a combination of threesMRImeasures cortical thickness cortical area cortical volumeand three nonimaging measures CSF components APOE ε4status and MMSE score improves AD diagnosis Furthermore

this combination shows great potential for the early identifi-cation of mild cognitive impairment (the prodromal stage ofAlzheimerrsquos disease) In this method we proposed filter andwrapper feature selection with an ELM classifier for multiplebiomarker-based AD diagnosis which significantly improvedthe classifierrsquos performance Moreover the results were betterthan or comparable with those previously reported particularlyfor the most challenging classification such as HC vs MCI andMCIs vs MCIc e added value of combining different ana-tomical MRI measures should be considered in AD scanningprotocols Only using specific aspects or a single measure ofwhole brain atrophy for AD diagnosis is still common practiceOur results show that clinical AD diagnosis could benefit fromthe combination of multiple measures from an anatomical MRIscan and other nonimaging biomarkers incorporated into anautomated machine learning system Our suggested methodeffectively enhances the diagnostic accuracy ofADandMCI butstill has some drawbacks Future work will include variousimprovements First we will optimize the parameter obtainingprocess Second in order to enhance the effectiveness of thesuggested method the dataset will be increased in terms of thefollowing extension of the longitudinal dataset for better un-derstanding of the progression from MCI and inclusion ofmultimodal data such as PET and fMRI data which providesdifferent insights into the characteristics of AD

Data Availability

e Alzheimerrsquos Disease Neuroimaging Initiative (ADNI)dataset (httpadniloniuscedu) was used in this studyComplete information regarding ADNI investigators can befound at httpadniloniusceduwp_contentuploadshow_to_applyADNI_Acknowledgement_istpd

Conflicts of Interest

e authors declare no conflicts of interest relating to thiswork

Acknowledgments

is work was supported by the National Research Foun-dation of Korea Grant funded by the Korean Government(NRF-2019R1F1A1060166 and NRF-2019R1A4A1029769)Data collection and sharing for this project was funded bythe ADNI (National Institutes of Health grant number U01

Table 8 Comparison of classification of MCIs vs MCIc between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection MCIs vs MCIc()

Westman et al [67] MRI +CSF OPLS mdash 685Zhang and Shen et al[39] MRI + PET+CSF SVM Multitask feature selection 7390

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 7500

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 86

Moradi et al [73] MRI + age and cognitive mdash 82Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper (SFS) 8338

Computational Intelligence and Neuroscience 15

orbitofrontal right rostral middle frontal and rightpars opercularis areas was lower in the AD groupFor AD vs MCI the left superior parietal left su-perior frontal left caudal anterior cingulate rightentorhinal and right superior parietal areas showedthe largest decreases in volume For HC vs MCI theleft supramarginal left cuneus left precentral rightparahippocampal and right superior parietal areasshowed the most volume atrophy Similarly forMCIs vs MCIc the left bankssts left precentral leftrostral middle frontal right precentral and rightfusiform areas had the largest decreases in volume

Moreover from this analysis we observe that areathickness and volume in the AD group were significantlydecreased in comparison with the HC group Similarly therewas significant atrophy in the MCIc group in comparisonwith the MCIs group and there was little difference in theatrophy pattern among AD and MCI patients

32 Feature Selection and Classification e individualfeature set was selected using an MRMR (filter) and a se-quential feature selection (wrapper) method to identify theoptimal feature set for different groups and improve theclassifier accuracy Similarly we combined all the feature setsfrom different measures and applied the MRMR and SFSmethod to select the optimal features Multiple measuresfeature sets were created by combining cortical thicknessarea and volume from sMRI three component measure-ments from CSF APOE ε4 status andMMSE score First weselected the top 60 features from theMRMR algorithm basedon the features score and then we applied a sequentialfeature selection algorithm on these top 60 ranked featureswhich gave the optimal feature set to achieve the maximumclassification accuracy on ELM classifiers Figure 3 shows the

number of optimal features sequence after sequential featureselection on the top 60 ranked features obtained from theMRMR algorithm using ELM classifier Figure 3 shows onlythe selected features for the combined feature set From thiswe can assume that the proposed feature selection withcross-validation provides the optimal feature vector forinput to the classifiers In this proposed method classifi-cation performance was quantified by the number of featuresselected versus accuracy and area under the ROC curve

To further analyze the effectiveness of the purposeclassification method combining different measures wecalculated the AUCs for the concatenation of all featuresFigure 4shows the receiver operating curves for individualfeatures and all feature (imaging and nonimaging bio-markers ie multiple features) combinations for eachclassification group using the ELM classifier

4 Discussion

41 Performance Analysis In this study we first performedstatistical analysis and pattern classification to differentiateand identify atrophy patterns for the four groups (AD HCMCIs and MCIc) Individual sMRI was preprocessed usingthe FreeSurfer tool After preprocessing the statisticalanalysis results of sMRI QDEC was applied and finally weperformed the classification task by the proposed featureselection and classification method respectively For brainatrophy analysis we used three types of sMRI corticalmetrics (cortical thickness gray matter volume and surfacearea) In comparing the AD group with the HC group theinsula pars opercularis parahippocampal and superiortemporal areas were severely affected in terms of corticalthickness gray matter volume and surface area e corticalthickness of the left hemisphere was thinner than that of theleft hemisphere Similarly for AD vs MCI the most atrophy

Table 2 Continued

Feature RegionCoordinates

Vertex Value Size (mm2)x y z

MCIs vs MCIc

ickness

Left inferior parietal minus463 minus60 111 2770 minus37613 142748Left superior temporal minus455 minus04 minus208 1067 minus26742 46624Left parahippocampal minus317 minus403 minus101 550 minus23925 23985Right temporal pole 292 9 minus38 1449 minus55983 82282

Right superior temporal 623 minus347 152 1292 minus31646 54997Right inferior temporal 516 minus566 minus37 882 minus30694 50791

Right precentral 488 minus65 405 858 minus28315 35193

Area

Left fusiform minus364 minus295 minus22 659 minus30958 31221Left lateral occipital minus194 minus99 minus151 326 26022 25061Left parahippocampal minus188 minus335 minus14 224 minus20347 9759Right superior temporal 481 minus329 22 1515 minus25815 5798Right superior parietal 106 minus527 651 1697 minus22367 64152

Right postcentral 338 minus29 519 799 minus20859 36723

Volume

Left bankssts minus579 minus469 minus1 2196 minus26263 116258Left precentral minus361 minus22 517 2736 minus25339 10731

Left rostral middle frontal minus224 279 328 582 minus23081 32686Right superior temporal 623 minus347 152 3392 minus44677 144996

Right precentral 495 minus62 411 4103 minus32472 181394Right fusiform 339 minus378 minus228 782 minus31822 42505

Computational Intelligence and Neuroscience 9

was seen in the left inferior parietal and right lateral occipitalareas For HC vs MCI the supramarginal and cuneus areasshowed the most atrophy For MCIs vs MCIc the superiortemporal and precentral areas showed the largest decreasesin thickness volume and area Based on these differencesthe majority of the decrease in cortical thickness gray mattervolume and surface area appears mostly in the frontal lobetemporal lobe occipital lobe cingulate gyrus and parietallobe is phenomenon strongly agrees with findings relatedto atrophy patterns seen in previous studies [14 66] eseregions are mainly involved in the expression of personalitymotor execution complex cognitive behavior and decisionmaking [50] In addition we present the less commonanalysis of MCIs vs MCIc For MCIc the most atrophy wasseen in the superior temporal bankssts precentral inferior

parietal and insula areas which show potential for the earlyrecognition of progression to Alzheimerrsquos disease

To test the effectiveness of our purposed featurescombination method (ie imaging and nonimaging fea-tures) we adapted the individual feature from sMRI and CSFseparately to conduct the study although regarding geneticand cognitive features we combined them to test the per-formance and then compared the accuracy with the accuracyof the combination of all features For the proposed featureselection we used a combination of filter and wrapper al-gorithms and compared the performance of the classifiers onthe selected feature set In this text SVM-linear SVM-RBFand ELM were used for classification e classificationresults for the compared methods are presented inTables 3ndash6 and also in graphical form in Figure 5 As shown

AD vs HC

Number of selected features = 27

05

101520253035404550556065707580859095

100

Accu

racy

()

5 10 15 20 25 30 35 40 45 50 55 600

Number of features sequence

(a)

AD vs MCI

Number of selected features = 17

05

101520253035404550556065707580859095

100

Accu

racy

()

5 10 15 20 25 30 35 40 45 50 55 600

Number of feature sequence

(b)HC vs MCI

Number of selected features = 25

5 10 15 20 25 30 35 40 45 50 55 600

Number of feature sequence

05

101520253035404550556065707580859095

100

Accu

racy

()

(c)

MCIs vs MCIc

Number of selected features = 13

05

101520253035404550556065707580859095

100Ac

cura

cy (

)

5 10 15 20 25 30 35 40 45 50 55 600

Number of features sequence

(d)

Figure 3 Number of feature sets selected from the sequential feature selection algorithm on the top 60 ranked features obtained using theMRMR algorithm Only the number of features versus accuracy for the combination of all feature sets using ELM classifiers is shown (a) ADvs HC feature subset (b) AD vs MCI feature subset (c) HC vs MCI feature subset (d) MCIs vs MCIc feature subset

10 Computational Intelligence and Neuroscience

AD vs HC

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(a)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

AD vs MCI

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(b)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

HC vs MCI

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(c)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

MCIs vs MCIc

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(d)

Figure 4 ROC curves for discriminating AD MCIs MCIc and HC (healthy controls) ROC curves are plotted for the five biomarkersseparately and for the concatenation of all features (ie all five biomarkers measures) (a) ROC curves for AD vs HC classification (b) forAD vs MCI classification (c) for HC vs MCI classification and (d) for MCIs vs MCIc classification using two-stage feature selection withELM classifiers

Computational Intelligence and Neuroscience 11

in the tables the performance accuracy of feature fusion isnoticeably improved when compared with that of the in-dividual feature set and there are distinct degrees of ele-vation in other indexes the specificity index is particularlymore noticeable In contrast performance accuracy basedon the nonimaging features is almost the same as the im-aging features for AD diagnosis but far superior for MCIdiagnosis For AD vs HC the accuracy obtained by CSFmeasures genetics and cognitive score showed a lowerincrease although for MCI vs HC AD vs MCI and MCIsvs MCIc the accuracy was higher e result showed thatthe ELM classifier achieves better classification scorescompared to the SVM classifier (linear and RBF-SVM) for

both single modality feature sets as well as all featurescombined (shown in Figure 6) e ELM is good machinelearning tool and the major strength is that the hiddenlayerrsquos learning parameters including input weight andbiases do not tune iteratively as in SLFN e ELM offersmany advantages over other learning algorithms [54] Fromthe results shown in Tables 3ndash6 for AD vs HC the clas-sification result increased by 5ndash7 as compared to SVMclassifiers with 9804 sensitivity and 9628 specificityFor AD vs MCI classification accuracy increased by 5ndash9with 92 sensitivity and 9733 accuracy For HC vs MCIclassification accuracy increased by 6ndash10 with 9123sensitivity and 9913 specificity Similarly for MCIs vs

Table 3 10-fold cross-validated classification performance for AD vs HC

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 8513 9207 8544 086 8068 8568 8381 7717 8317 7381Surface area 8230 8930 9154 085 7616 8616 7514 7867 8767 7262Volume 8790 9387 8475 088 7929 7429 7833 8000 7710 6324CSF 8973 9633 8738 089 7905 8905 7417 7533 8333 6857APOE+MMSE 8645 9513 8700 093 8203 8703 8467 8416 8305 7879Concatenation (all_features_set) 9731 9804 9628 097 9133 9333 8757 9350 955 9058

Table 4 10-fold cross-validated classification performance for AD vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7010 8130 6610 071 6401 8383 6233 6503 7880 7330Surface area 6160 6670 8430 063 6345 6810 8011 6067 7123 6285Volume 6780 8210 6710 069 5820 7792 6856 6005 7337 6373CSF 6470 6600 8140 073 5607 6275 7573 6123 6871 7937APOE+MMSE 8145 8440 8770 082 7681 6781 8548 7620 6694 8748Concatenation (all_features_set) 8791 9200 9733 089 8350 8864 7672 7831 7445 8262

Table 5 10-fold cross-validated classification performance for HC vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7540 8407 6790 083 7417 8081 7123 6734 7317 7793Surface area 7415 6520 8385 084 6354 6767 6647 6872 7443 6735Volume 7784 8841 7280 084 6547 7473 6875 7283 7827 6570CSF 8330 7393 8402 085 7339 7525 6683 7580 7338 7814APOE+MMSE 7937 9325 7297 089 6829 8173 7668 7023 8124 7805Concatenation (all_features_set) 9172 9123 9913 093 8170 8368 8749 8565 7854 8926

Table 6 10-fold cross-validated classification performance for MCIs vs MCIc

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 6371 7110 3837 064 5718 6553 6038 5005 7118 6813Surface area 6143 7228 7480 069 5425 6322 7617 5827 7223 6485Volume 6785 8440 8773 074 5917 7828 6755 6214 7445 5773CSF 7137 7884 6857 072 6473 6009 7278 6733 7882 6878APOE+MMSE 7609 8713 7209 079 6553 6881 6743 6808 7221 6633Concatenation (all_features_set) 8338 9301 7577 085 7137 8116 7675 7571 7838 8467

12 Computational Intelligence and Neuroscience

MCIc classification accuracy increased by 7ndash12 with9301 sensitivity and 7577 specificity but 8467 and7667 specificity was obtained for linear and RBF-SVMclassifiers respectively which is more than obtained by theELM classifier From our analyses we believe that the ELM

gives better performance as compared to SVM from alearning efficiency standpoint because the ELMrsquos originaldesign has high learning accuracy fast learning speedscalability and the least human intervention [54 55] Fromthe results shown in Table 7 we can see that our suggested

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(a)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(b)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(c)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(d)

Figure 5 Classification results for different Alzheimerrsquos groups using ELM classifiers (a) Classification results for AD vs HC groups (b) forAD vs MCI groups (c) for HC vs MCI groups and (d) for MCIs vs MCIc groups

Computational Intelligence and Neuroscience 13

method achieved better accuracy than other existingmethodsFor classifying AD and HC our method achieved a classi-fication accuracy of 9731 with a sensitivity of 9804 aspecificity of 9628 and an AUC value of 097 Forclassifying MCI and HC our method achieves a classificationaccuracy of 9172 with a sensitivity of 9123 a specificity of9913 and an AUC value of 093 For distinguishing ADfrom MCI our proposed method obtained a classificationaccuracy of 8791 with a sensitivity of 9200 a specificityof 9733 and an AUC value of 089 Similarly for MCIs vsMCIc we achieved outstanding performance as compared toprevious methods by combining multiple features with anaccuracy of 8338 a sensitivity of 9301 a specificity of7577 and an AUC value of 085

42ComparisonwithOtherMethods In the previous sectionwe discussed in detail the findings of the proposed featureselection and classification method In this section we

compare and discuss the findings of our method in com-parison with existing state-of-the-art methods Tables 7 and 8show a comparison of the classification performance of theproposed method with recently published studies which usedmultimodality data to distinguish individuals with AD andMCI from HC Westman et al [67] used MRI and CSFbiomarkers and obtained 918 accuracy for AD vs HC776 for HC vs MCI and 685 for pMCI vs MCIsclassification using the orthogonal partial least squares tolatent structures (OPLS) method Zhang and Shen [39] used amultimodal (MRI PET and CSF) classification method withmultitask feature selection and an SVM classifier for AD andMCI classification By combining MRI PET and CSF datathey achieved a higher accuracy of 933 for AD vs HC and832 accuracy for HC vs MCI classification Similarly theauthors achieved an accuracy of 739 on pMCI vs MCIssamples Hinrichs et al [69] obtained an accuracy of 924 forAD vs HC classification using MRI PET CSF APOE and

ELMSVM_RBFSVM-linear

0

20

40

60

80

100

ACC

()

AD vs HC HC vs MCI MCIs vs MCIcAD vs MCI

Figure 6 Performance comparison of different classifiers

Table 7 Comparison of classification of AD vs HC and HC vs MCI between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection AD vsHC ()

HC vsMCI

Westman et al[67] MRI +CSF OPLS mdash 918 776

Johnson et al [68] MRI + PET+CSF+ cognitive scores Stackedautoencoder

Sparse representationlearning 959 85

Hinrichs et al [69] MRI + PET+CSF+APOE+ cognitive scores MKL mdash 924 naZhang and Shenet al [39] MRI + PET+CSF SVM Multitask feature selection 933 832

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 9301 mdash

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 995 mdash

Maqsood et al[72] sMRI CNN mdash 9285 mdash

Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper(SFS) 9731 9172

14 Computational Intelligence and Neuroscience

cognitive score in multiple kernel learning Similarly Johnsonet al [68] proposed a sparse representation learning featureselection and stacked autoencoder for classifying AD usingMRI PET CSF and cognitive scores as features and obtained959 accuracy for AD vs HC classification and 85 accuracyfor HC vs MCI classification Maqsood et al [72] proposedthe transfer learning classification model of AD for bothbinary and multiclass problems e algorithm utilizes apretrained AlexNet network and fine-tuned the convolutionalneural network (CNN) for classificationis model was fine-tuned over both segmented and unsegmented 3D views of thehuman brain e MRI scans were segmented into thecharacteristic components of GM white matter and CSFeretrained CNNwas then validated using the test data to obtainaccuracies of 896 and 928 for binary and multiclassproblems respectively using the OASIS dataset Beheshtiet al [70] developed a computer-aided diagnosis (CAD)system composed of four systematic stages for analyzingglobal and local differences in the GM of AD patientscompared with HC using the voxel-based morphometry(VBM) methodey used feature ranking based on the t-testand a genetic algorithm with Fisherrsquos criteria as part of theobjective function in the genetic algorithm e authorsutilized the SVM classifier with 10-fold cross-validation toobtain accuracies of 9301 and 75 for AD vs HC andMCIsvs pMCI classification respectively Spasov et al [71] de-veloped a deep learning architecture based on dual learningand an ad hoc layer for 3D separable convolution to identifyMCI patients eir deep learning procedure combined MRIneuropsychological demographic and APOE ε4 data toachieve an accuracy of 86 ey achieved an accuracy of995 for AD vs HC classification In another study Moradiet al used VBM analysis of GM as a feature combined withage and cognitive measures ey achieved an accuracy of82 for pMCI vs MCIs classification From this comparisonwe can infer that the proposed feature combination (MRICSF APOE and MMSE data) is robust or comparable to theother multimodal biomarker methods reported in the liter-ature for both AD vs HC and MCIs vs MCIc classification

5 Conclusion

In conclusion we demonstrated that a combination of threesMRImeasures cortical thickness cortical area cortical volumeand three nonimaging measures CSF components APOE ε4status and MMSE score improves AD diagnosis Furthermore

this combination shows great potential for the early identifi-cation of mild cognitive impairment (the prodromal stage ofAlzheimerrsquos disease) In this method we proposed filter andwrapper feature selection with an ELM classifier for multiplebiomarker-based AD diagnosis which significantly improvedthe classifierrsquos performance Moreover the results were betterthan or comparable with those previously reported particularlyfor the most challenging classification such as HC vs MCI andMCIs vs MCIc e added value of combining different ana-tomical MRI measures should be considered in AD scanningprotocols Only using specific aspects or a single measure ofwhole brain atrophy for AD diagnosis is still common practiceOur results show that clinical AD diagnosis could benefit fromthe combination of multiple measures from an anatomical MRIscan and other nonimaging biomarkers incorporated into anautomated machine learning system Our suggested methodeffectively enhances the diagnostic accuracy ofADandMCI butstill has some drawbacks Future work will include variousimprovements First we will optimize the parameter obtainingprocess Second in order to enhance the effectiveness of thesuggested method the dataset will be increased in terms of thefollowing extension of the longitudinal dataset for better un-derstanding of the progression from MCI and inclusion ofmultimodal data such as PET and fMRI data which providesdifferent insights into the characteristics of AD

Data Availability

e Alzheimerrsquos Disease Neuroimaging Initiative (ADNI)dataset (httpadniloniuscedu) was used in this studyComplete information regarding ADNI investigators can befound at httpadniloniusceduwp_contentuploadshow_to_applyADNI_Acknowledgement_istpd

Conflicts of Interest

e authors declare no conflicts of interest relating to thiswork

Acknowledgments

is work was supported by the National Research Foun-dation of Korea Grant funded by the Korean Government(NRF-2019R1F1A1060166 and NRF-2019R1A4A1029769)Data collection and sharing for this project was funded bythe ADNI (National Institutes of Health grant number U01

Table 8 Comparison of classification of MCIs vs MCIc between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection MCIs vs MCIc()

Westman et al [67] MRI +CSF OPLS mdash 685Zhang and Shen et al[39] MRI + PET+CSF SVM Multitask feature selection 7390

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 7500

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 86

Moradi et al [73] MRI + age and cognitive mdash 82Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper (SFS) 8338

Computational Intelligence and Neuroscience 15

was seen in the left inferior parietal and right lateral occipitalareas For HC vs MCI the supramarginal and cuneus areasshowed the most atrophy For MCIs vs MCIc the superiortemporal and precentral areas showed the largest decreasesin thickness volume and area Based on these differencesthe majority of the decrease in cortical thickness gray mattervolume and surface area appears mostly in the frontal lobetemporal lobe occipital lobe cingulate gyrus and parietallobe is phenomenon strongly agrees with findings relatedto atrophy patterns seen in previous studies [14 66] eseregions are mainly involved in the expression of personalitymotor execution complex cognitive behavior and decisionmaking [50] In addition we present the less commonanalysis of MCIs vs MCIc For MCIc the most atrophy wasseen in the superior temporal bankssts precentral inferior

parietal and insula areas which show potential for the earlyrecognition of progression to Alzheimerrsquos disease

To test the effectiveness of our purposed featurescombination method (ie imaging and nonimaging fea-tures) we adapted the individual feature from sMRI and CSFseparately to conduct the study although regarding geneticand cognitive features we combined them to test the per-formance and then compared the accuracy with the accuracyof the combination of all features For the proposed featureselection we used a combination of filter and wrapper al-gorithms and compared the performance of the classifiers onthe selected feature set In this text SVM-linear SVM-RBFand ELM were used for classification e classificationresults for the compared methods are presented inTables 3ndash6 and also in graphical form in Figure 5 As shown

AD vs HC

Number of selected features = 27

05

101520253035404550556065707580859095

100

Accu

racy

()

5 10 15 20 25 30 35 40 45 50 55 600

Number of features sequence

(a)

AD vs MCI

Number of selected features = 17

05

101520253035404550556065707580859095

100

Accu

racy

()

5 10 15 20 25 30 35 40 45 50 55 600

Number of feature sequence

(b)HC vs MCI

Number of selected features = 25

5 10 15 20 25 30 35 40 45 50 55 600

Number of feature sequence

05

101520253035404550556065707580859095

100

Accu

racy

()

(c)

MCIs vs MCIc

Number of selected features = 13

05

101520253035404550556065707580859095

100Ac

cura

cy (

)

5 10 15 20 25 30 35 40 45 50 55 600

Number of features sequence

(d)

Figure 3 Number of feature sets selected from the sequential feature selection algorithm on the top 60 ranked features obtained using theMRMR algorithm Only the number of features versus accuracy for the combination of all feature sets using ELM classifiers is shown (a) ADvs HC feature subset (b) AD vs MCI feature subset (c) HC vs MCI feature subset (d) MCIs vs MCIc feature subset

10 Computational Intelligence and Neuroscience

AD vs HC

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(a)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

AD vs MCI

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(b)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

HC vs MCI

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(c)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

MCIs vs MCIc

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(d)

Figure 4 ROC curves for discriminating AD MCIs MCIc and HC (healthy controls) ROC curves are plotted for the five biomarkersseparately and for the concatenation of all features (ie all five biomarkers measures) (a) ROC curves for AD vs HC classification (b) forAD vs MCI classification (c) for HC vs MCI classification and (d) for MCIs vs MCIc classification using two-stage feature selection withELM classifiers

Computational Intelligence and Neuroscience 11

in the tables the performance accuracy of feature fusion isnoticeably improved when compared with that of the in-dividual feature set and there are distinct degrees of ele-vation in other indexes the specificity index is particularlymore noticeable In contrast performance accuracy basedon the nonimaging features is almost the same as the im-aging features for AD diagnosis but far superior for MCIdiagnosis For AD vs HC the accuracy obtained by CSFmeasures genetics and cognitive score showed a lowerincrease although for MCI vs HC AD vs MCI and MCIsvs MCIc the accuracy was higher e result showed thatthe ELM classifier achieves better classification scorescompared to the SVM classifier (linear and RBF-SVM) for

both single modality feature sets as well as all featurescombined (shown in Figure 6) e ELM is good machinelearning tool and the major strength is that the hiddenlayerrsquos learning parameters including input weight andbiases do not tune iteratively as in SLFN e ELM offersmany advantages over other learning algorithms [54] Fromthe results shown in Tables 3ndash6 for AD vs HC the clas-sification result increased by 5ndash7 as compared to SVMclassifiers with 9804 sensitivity and 9628 specificityFor AD vs MCI classification accuracy increased by 5ndash9with 92 sensitivity and 9733 accuracy For HC vs MCIclassification accuracy increased by 6ndash10 with 9123sensitivity and 9913 specificity Similarly for MCIs vs

Table 3 10-fold cross-validated classification performance for AD vs HC

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 8513 9207 8544 086 8068 8568 8381 7717 8317 7381Surface area 8230 8930 9154 085 7616 8616 7514 7867 8767 7262Volume 8790 9387 8475 088 7929 7429 7833 8000 7710 6324CSF 8973 9633 8738 089 7905 8905 7417 7533 8333 6857APOE+MMSE 8645 9513 8700 093 8203 8703 8467 8416 8305 7879Concatenation (all_features_set) 9731 9804 9628 097 9133 9333 8757 9350 955 9058

Table 4 10-fold cross-validated classification performance for AD vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7010 8130 6610 071 6401 8383 6233 6503 7880 7330Surface area 6160 6670 8430 063 6345 6810 8011 6067 7123 6285Volume 6780 8210 6710 069 5820 7792 6856 6005 7337 6373CSF 6470 6600 8140 073 5607 6275 7573 6123 6871 7937APOE+MMSE 8145 8440 8770 082 7681 6781 8548 7620 6694 8748Concatenation (all_features_set) 8791 9200 9733 089 8350 8864 7672 7831 7445 8262

Table 5 10-fold cross-validated classification performance for HC vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7540 8407 6790 083 7417 8081 7123 6734 7317 7793Surface area 7415 6520 8385 084 6354 6767 6647 6872 7443 6735Volume 7784 8841 7280 084 6547 7473 6875 7283 7827 6570CSF 8330 7393 8402 085 7339 7525 6683 7580 7338 7814APOE+MMSE 7937 9325 7297 089 6829 8173 7668 7023 8124 7805Concatenation (all_features_set) 9172 9123 9913 093 8170 8368 8749 8565 7854 8926

Table 6 10-fold cross-validated classification performance for MCIs vs MCIc

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 6371 7110 3837 064 5718 6553 6038 5005 7118 6813Surface area 6143 7228 7480 069 5425 6322 7617 5827 7223 6485Volume 6785 8440 8773 074 5917 7828 6755 6214 7445 5773CSF 7137 7884 6857 072 6473 6009 7278 6733 7882 6878APOE+MMSE 7609 8713 7209 079 6553 6881 6743 6808 7221 6633Concatenation (all_features_set) 8338 9301 7577 085 7137 8116 7675 7571 7838 8467

12 Computational Intelligence and Neuroscience

MCIc classification accuracy increased by 7ndash12 with9301 sensitivity and 7577 specificity but 8467 and7667 specificity was obtained for linear and RBF-SVMclassifiers respectively which is more than obtained by theELM classifier From our analyses we believe that the ELM

gives better performance as compared to SVM from alearning efficiency standpoint because the ELMrsquos originaldesign has high learning accuracy fast learning speedscalability and the least human intervention [54 55] Fromthe results shown in Table 7 we can see that our suggested

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(a)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(b)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(c)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(d)

Figure 5 Classification results for different Alzheimerrsquos groups using ELM classifiers (a) Classification results for AD vs HC groups (b) forAD vs MCI groups (c) for HC vs MCI groups and (d) for MCIs vs MCIc groups

Computational Intelligence and Neuroscience 13

method achieved better accuracy than other existingmethodsFor classifying AD and HC our method achieved a classi-fication accuracy of 9731 with a sensitivity of 9804 aspecificity of 9628 and an AUC value of 097 Forclassifying MCI and HC our method achieves a classificationaccuracy of 9172 with a sensitivity of 9123 a specificity of9913 and an AUC value of 093 For distinguishing ADfrom MCI our proposed method obtained a classificationaccuracy of 8791 with a sensitivity of 9200 a specificityof 9733 and an AUC value of 089 Similarly for MCIs vsMCIc we achieved outstanding performance as compared toprevious methods by combining multiple features with anaccuracy of 8338 a sensitivity of 9301 a specificity of7577 and an AUC value of 085

42ComparisonwithOtherMethods In the previous sectionwe discussed in detail the findings of the proposed featureselection and classification method In this section we

compare and discuss the findings of our method in com-parison with existing state-of-the-art methods Tables 7 and 8show a comparison of the classification performance of theproposed method with recently published studies which usedmultimodality data to distinguish individuals with AD andMCI from HC Westman et al [67] used MRI and CSFbiomarkers and obtained 918 accuracy for AD vs HC776 for HC vs MCI and 685 for pMCI vs MCIsclassification using the orthogonal partial least squares tolatent structures (OPLS) method Zhang and Shen [39] used amultimodal (MRI PET and CSF) classification method withmultitask feature selection and an SVM classifier for AD andMCI classification By combining MRI PET and CSF datathey achieved a higher accuracy of 933 for AD vs HC and832 accuracy for HC vs MCI classification Similarly theauthors achieved an accuracy of 739 on pMCI vs MCIssamples Hinrichs et al [69] obtained an accuracy of 924 forAD vs HC classification using MRI PET CSF APOE and

ELMSVM_RBFSVM-linear

0

20

40

60

80

100

ACC

()

AD vs HC HC vs MCI MCIs vs MCIcAD vs MCI

Figure 6 Performance comparison of different classifiers

Table 7 Comparison of classification of AD vs HC and HC vs MCI between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection AD vsHC ()

HC vsMCI

Westman et al[67] MRI +CSF OPLS mdash 918 776

Johnson et al [68] MRI + PET+CSF+ cognitive scores Stackedautoencoder

Sparse representationlearning 959 85

Hinrichs et al [69] MRI + PET+CSF+APOE+ cognitive scores MKL mdash 924 naZhang and Shenet al [39] MRI + PET+CSF SVM Multitask feature selection 933 832

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 9301 mdash

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 995 mdash

Maqsood et al[72] sMRI CNN mdash 9285 mdash

Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper(SFS) 9731 9172

14 Computational Intelligence and Neuroscience

cognitive score in multiple kernel learning Similarly Johnsonet al [68] proposed a sparse representation learning featureselection and stacked autoencoder for classifying AD usingMRI PET CSF and cognitive scores as features and obtained959 accuracy for AD vs HC classification and 85 accuracyfor HC vs MCI classification Maqsood et al [72] proposedthe transfer learning classification model of AD for bothbinary and multiclass problems e algorithm utilizes apretrained AlexNet network and fine-tuned the convolutionalneural network (CNN) for classificationis model was fine-tuned over both segmented and unsegmented 3D views of thehuman brain e MRI scans were segmented into thecharacteristic components of GM white matter and CSFeretrained CNNwas then validated using the test data to obtainaccuracies of 896 and 928 for binary and multiclassproblems respectively using the OASIS dataset Beheshtiet al [70] developed a computer-aided diagnosis (CAD)system composed of four systematic stages for analyzingglobal and local differences in the GM of AD patientscompared with HC using the voxel-based morphometry(VBM) methodey used feature ranking based on the t-testand a genetic algorithm with Fisherrsquos criteria as part of theobjective function in the genetic algorithm e authorsutilized the SVM classifier with 10-fold cross-validation toobtain accuracies of 9301 and 75 for AD vs HC andMCIsvs pMCI classification respectively Spasov et al [71] de-veloped a deep learning architecture based on dual learningand an ad hoc layer for 3D separable convolution to identifyMCI patients eir deep learning procedure combined MRIneuropsychological demographic and APOE ε4 data toachieve an accuracy of 86 ey achieved an accuracy of995 for AD vs HC classification In another study Moradiet al used VBM analysis of GM as a feature combined withage and cognitive measures ey achieved an accuracy of82 for pMCI vs MCIs classification From this comparisonwe can infer that the proposed feature combination (MRICSF APOE and MMSE data) is robust or comparable to theother multimodal biomarker methods reported in the liter-ature for both AD vs HC and MCIs vs MCIc classification

5 Conclusion

In conclusion we demonstrated that a combination of threesMRImeasures cortical thickness cortical area cortical volumeand three nonimaging measures CSF components APOE ε4status and MMSE score improves AD diagnosis Furthermore

this combination shows great potential for the early identifi-cation of mild cognitive impairment (the prodromal stage ofAlzheimerrsquos disease) In this method we proposed filter andwrapper feature selection with an ELM classifier for multiplebiomarker-based AD diagnosis which significantly improvedthe classifierrsquos performance Moreover the results were betterthan or comparable with those previously reported particularlyfor the most challenging classification such as HC vs MCI andMCIs vs MCIc e added value of combining different ana-tomical MRI measures should be considered in AD scanningprotocols Only using specific aspects or a single measure ofwhole brain atrophy for AD diagnosis is still common practiceOur results show that clinical AD diagnosis could benefit fromthe combination of multiple measures from an anatomical MRIscan and other nonimaging biomarkers incorporated into anautomated machine learning system Our suggested methodeffectively enhances the diagnostic accuracy ofADandMCI butstill has some drawbacks Future work will include variousimprovements First we will optimize the parameter obtainingprocess Second in order to enhance the effectiveness of thesuggested method the dataset will be increased in terms of thefollowing extension of the longitudinal dataset for better un-derstanding of the progression from MCI and inclusion ofmultimodal data such as PET and fMRI data which providesdifferent insights into the characteristics of AD

Data Availability

e Alzheimerrsquos Disease Neuroimaging Initiative (ADNI)dataset (httpadniloniuscedu) was used in this studyComplete information regarding ADNI investigators can befound at httpadniloniusceduwp_contentuploadshow_to_applyADNI_Acknowledgement_istpd

Conflicts of Interest

e authors declare no conflicts of interest relating to thiswork

Acknowledgments

is work was supported by the National Research Foun-dation of Korea Grant funded by the Korean Government(NRF-2019R1F1A1060166 and NRF-2019R1A4A1029769)Data collection and sharing for this project was funded bythe ADNI (National Institutes of Health grant number U01

Table 8 Comparison of classification of MCIs vs MCIc between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection MCIs vs MCIc()

Westman et al [67] MRI +CSF OPLS mdash 685Zhang and Shen et al[39] MRI + PET+CSF SVM Multitask feature selection 7390

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 7500

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 86

Moradi et al [73] MRI + age and cognitive mdash 82Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper (SFS) 8338

Computational Intelligence and Neuroscience 15

AD vs HC

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(a)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

AD vs MCI

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(b)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

HC vs MCI

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(c)

ThicknessAreaVolume

CSFAPOE + MMSEConcatenation(all_features_set)

MCIs vs MCIc

0

20

40

60

80

100

True

-pos

itive

rate

20 40 60 80 1000False-positive rate

(d)

Figure 4 ROC curves for discriminating AD MCIs MCIc and HC (healthy controls) ROC curves are plotted for the five biomarkersseparately and for the concatenation of all features (ie all five biomarkers measures) (a) ROC curves for AD vs HC classification (b) forAD vs MCI classification (c) for HC vs MCI classification and (d) for MCIs vs MCIc classification using two-stage feature selection withELM classifiers

Computational Intelligence and Neuroscience 11

in the tables the performance accuracy of feature fusion isnoticeably improved when compared with that of the in-dividual feature set and there are distinct degrees of ele-vation in other indexes the specificity index is particularlymore noticeable In contrast performance accuracy basedon the nonimaging features is almost the same as the im-aging features for AD diagnosis but far superior for MCIdiagnosis For AD vs HC the accuracy obtained by CSFmeasures genetics and cognitive score showed a lowerincrease although for MCI vs HC AD vs MCI and MCIsvs MCIc the accuracy was higher e result showed thatthe ELM classifier achieves better classification scorescompared to the SVM classifier (linear and RBF-SVM) for

both single modality feature sets as well as all featurescombined (shown in Figure 6) e ELM is good machinelearning tool and the major strength is that the hiddenlayerrsquos learning parameters including input weight andbiases do not tune iteratively as in SLFN e ELM offersmany advantages over other learning algorithms [54] Fromthe results shown in Tables 3ndash6 for AD vs HC the clas-sification result increased by 5ndash7 as compared to SVMclassifiers with 9804 sensitivity and 9628 specificityFor AD vs MCI classification accuracy increased by 5ndash9with 92 sensitivity and 9733 accuracy For HC vs MCIclassification accuracy increased by 6ndash10 with 9123sensitivity and 9913 specificity Similarly for MCIs vs

Table 3 10-fold cross-validated classification performance for AD vs HC

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 8513 9207 8544 086 8068 8568 8381 7717 8317 7381Surface area 8230 8930 9154 085 7616 8616 7514 7867 8767 7262Volume 8790 9387 8475 088 7929 7429 7833 8000 7710 6324CSF 8973 9633 8738 089 7905 8905 7417 7533 8333 6857APOE+MMSE 8645 9513 8700 093 8203 8703 8467 8416 8305 7879Concatenation (all_features_set) 9731 9804 9628 097 9133 9333 8757 9350 955 9058

Table 4 10-fold cross-validated classification performance for AD vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7010 8130 6610 071 6401 8383 6233 6503 7880 7330Surface area 6160 6670 8430 063 6345 6810 8011 6067 7123 6285Volume 6780 8210 6710 069 5820 7792 6856 6005 7337 6373CSF 6470 6600 8140 073 5607 6275 7573 6123 6871 7937APOE+MMSE 8145 8440 8770 082 7681 6781 8548 7620 6694 8748Concatenation (all_features_set) 8791 9200 9733 089 8350 8864 7672 7831 7445 8262

Table 5 10-fold cross-validated classification performance for HC vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7540 8407 6790 083 7417 8081 7123 6734 7317 7793Surface area 7415 6520 8385 084 6354 6767 6647 6872 7443 6735Volume 7784 8841 7280 084 6547 7473 6875 7283 7827 6570CSF 8330 7393 8402 085 7339 7525 6683 7580 7338 7814APOE+MMSE 7937 9325 7297 089 6829 8173 7668 7023 8124 7805Concatenation (all_features_set) 9172 9123 9913 093 8170 8368 8749 8565 7854 8926

Table 6 10-fold cross-validated classification performance for MCIs vs MCIc

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 6371 7110 3837 064 5718 6553 6038 5005 7118 6813Surface area 6143 7228 7480 069 5425 6322 7617 5827 7223 6485Volume 6785 8440 8773 074 5917 7828 6755 6214 7445 5773CSF 7137 7884 6857 072 6473 6009 7278 6733 7882 6878APOE+MMSE 7609 8713 7209 079 6553 6881 6743 6808 7221 6633Concatenation (all_features_set) 8338 9301 7577 085 7137 8116 7675 7571 7838 8467

12 Computational Intelligence and Neuroscience

MCIc classification accuracy increased by 7ndash12 with9301 sensitivity and 7577 specificity but 8467 and7667 specificity was obtained for linear and RBF-SVMclassifiers respectively which is more than obtained by theELM classifier From our analyses we believe that the ELM

gives better performance as compared to SVM from alearning efficiency standpoint because the ELMrsquos originaldesign has high learning accuracy fast learning speedscalability and the least human intervention [54 55] Fromthe results shown in Table 7 we can see that our suggested

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(a)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(b)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(c)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(d)

Figure 5 Classification results for different Alzheimerrsquos groups using ELM classifiers (a) Classification results for AD vs HC groups (b) forAD vs MCI groups (c) for HC vs MCI groups and (d) for MCIs vs MCIc groups

Computational Intelligence and Neuroscience 13

method achieved better accuracy than other existingmethodsFor classifying AD and HC our method achieved a classi-fication accuracy of 9731 with a sensitivity of 9804 aspecificity of 9628 and an AUC value of 097 Forclassifying MCI and HC our method achieves a classificationaccuracy of 9172 with a sensitivity of 9123 a specificity of9913 and an AUC value of 093 For distinguishing ADfrom MCI our proposed method obtained a classificationaccuracy of 8791 with a sensitivity of 9200 a specificityof 9733 and an AUC value of 089 Similarly for MCIs vsMCIc we achieved outstanding performance as compared toprevious methods by combining multiple features with anaccuracy of 8338 a sensitivity of 9301 a specificity of7577 and an AUC value of 085

42ComparisonwithOtherMethods In the previous sectionwe discussed in detail the findings of the proposed featureselection and classification method In this section we

compare and discuss the findings of our method in com-parison with existing state-of-the-art methods Tables 7 and 8show a comparison of the classification performance of theproposed method with recently published studies which usedmultimodality data to distinguish individuals with AD andMCI from HC Westman et al [67] used MRI and CSFbiomarkers and obtained 918 accuracy for AD vs HC776 for HC vs MCI and 685 for pMCI vs MCIsclassification using the orthogonal partial least squares tolatent structures (OPLS) method Zhang and Shen [39] used amultimodal (MRI PET and CSF) classification method withmultitask feature selection and an SVM classifier for AD andMCI classification By combining MRI PET and CSF datathey achieved a higher accuracy of 933 for AD vs HC and832 accuracy for HC vs MCI classification Similarly theauthors achieved an accuracy of 739 on pMCI vs MCIssamples Hinrichs et al [69] obtained an accuracy of 924 forAD vs HC classification using MRI PET CSF APOE and

ELMSVM_RBFSVM-linear

0

20

40

60

80

100

ACC

()

AD vs HC HC vs MCI MCIs vs MCIcAD vs MCI

Figure 6 Performance comparison of different classifiers

Table 7 Comparison of classification of AD vs HC and HC vs MCI between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection AD vsHC ()

HC vsMCI

Westman et al[67] MRI +CSF OPLS mdash 918 776

Johnson et al [68] MRI + PET+CSF+ cognitive scores Stackedautoencoder

Sparse representationlearning 959 85

Hinrichs et al [69] MRI + PET+CSF+APOE+ cognitive scores MKL mdash 924 naZhang and Shenet al [39] MRI + PET+CSF SVM Multitask feature selection 933 832

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 9301 mdash

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 995 mdash

Maqsood et al[72] sMRI CNN mdash 9285 mdash

Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper(SFS) 9731 9172

14 Computational Intelligence and Neuroscience

cognitive score in multiple kernel learning Similarly Johnsonet al [68] proposed a sparse representation learning featureselection and stacked autoencoder for classifying AD usingMRI PET CSF and cognitive scores as features and obtained959 accuracy for AD vs HC classification and 85 accuracyfor HC vs MCI classification Maqsood et al [72] proposedthe transfer learning classification model of AD for bothbinary and multiclass problems e algorithm utilizes apretrained AlexNet network and fine-tuned the convolutionalneural network (CNN) for classificationis model was fine-tuned over both segmented and unsegmented 3D views of thehuman brain e MRI scans were segmented into thecharacteristic components of GM white matter and CSFeretrained CNNwas then validated using the test data to obtainaccuracies of 896 and 928 for binary and multiclassproblems respectively using the OASIS dataset Beheshtiet al [70] developed a computer-aided diagnosis (CAD)system composed of four systematic stages for analyzingglobal and local differences in the GM of AD patientscompared with HC using the voxel-based morphometry(VBM) methodey used feature ranking based on the t-testand a genetic algorithm with Fisherrsquos criteria as part of theobjective function in the genetic algorithm e authorsutilized the SVM classifier with 10-fold cross-validation toobtain accuracies of 9301 and 75 for AD vs HC andMCIsvs pMCI classification respectively Spasov et al [71] de-veloped a deep learning architecture based on dual learningand an ad hoc layer for 3D separable convolution to identifyMCI patients eir deep learning procedure combined MRIneuropsychological demographic and APOE ε4 data toachieve an accuracy of 86 ey achieved an accuracy of995 for AD vs HC classification In another study Moradiet al used VBM analysis of GM as a feature combined withage and cognitive measures ey achieved an accuracy of82 for pMCI vs MCIs classification From this comparisonwe can infer that the proposed feature combination (MRICSF APOE and MMSE data) is robust or comparable to theother multimodal biomarker methods reported in the liter-ature for both AD vs HC and MCIs vs MCIc classification

5 Conclusion

In conclusion we demonstrated that a combination of threesMRImeasures cortical thickness cortical area cortical volumeand three nonimaging measures CSF components APOE ε4status and MMSE score improves AD diagnosis Furthermore

this combination shows great potential for the early identifi-cation of mild cognitive impairment (the prodromal stage ofAlzheimerrsquos disease) In this method we proposed filter andwrapper feature selection with an ELM classifier for multiplebiomarker-based AD diagnosis which significantly improvedthe classifierrsquos performance Moreover the results were betterthan or comparable with those previously reported particularlyfor the most challenging classification such as HC vs MCI andMCIs vs MCIc e added value of combining different ana-tomical MRI measures should be considered in AD scanningprotocols Only using specific aspects or a single measure ofwhole brain atrophy for AD diagnosis is still common practiceOur results show that clinical AD diagnosis could benefit fromthe combination of multiple measures from an anatomical MRIscan and other nonimaging biomarkers incorporated into anautomated machine learning system Our suggested methodeffectively enhances the diagnostic accuracy ofADandMCI butstill has some drawbacks Future work will include variousimprovements First we will optimize the parameter obtainingprocess Second in order to enhance the effectiveness of thesuggested method the dataset will be increased in terms of thefollowing extension of the longitudinal dataset for better un-derstanding of the progression from MCI and inclusion ofmultimodal data such as PET and fMRI data which providesdifferent insights into the characteristics of AD

Data Availability

e Alzheimerrsquos Disease Neuroimaging Initiative (ADNI)dataset (httpadniloniuscedu) was used in this studyComplete information regarding ADNI investigators can befound at httpadniloniusceduwp_contentuploadshow_to_applyADNI_Acknowledgement_istpd

Conflicts of Interest

e authors declare no conflicts of interest relating to thiswork

Acknowledgments

is work was supported by the National Research Foun-dation of Korea Grant funded by the Korean Government(NRF-2019R1F1A1060166 and NRF-2019R1A4A1029769)Data collection and sharing for this project was funded bythe ADNI (National Institutes of Health grant number U01

Table 8 Comparison of classification of MCIs vs MCIc between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection MCIs vs MCIc()

Westman et al [67] MRI +CSF OPLS mdash 685Zhang and Shen et al[39] MRI + PET+CSF SVM Multitask feature selection 7390

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 7500

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 86

Moradi et al [73] MRI + age and cognitive mdash 82Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper (SFS) 8338

Computational Intelligence and Neuroscience 15

in the tables the performance accuracy of feature fusion isnoticeably improved when compared with that of the in-dividual feature set and there are distinct degrees of ele-vation in other indexes the specificity index is particularlymore noticeable In contrast performance accuracy basedon the nonimaging features is almost the same as the im-aging features for AD diagnosis but far superior for MCIdiagnosis For AD vs HC the accuracy obtained by CSFmeasures genetics and cognitive score showed a lowerincrease although for MCI vs HC AD vs MCI and MCIsvs MCIc the accuracy was higher e result showed thatthe ELM classifier achieves better classification scorescompared to the SVM classifier (linear and RBF-SVM) for

both single modality feature sets as well as all featurescombined (shown in Figure 6) e ELM is good machinelearning tool and the major strength is that the hiddenlayerrsquos learning parameters including input weight andbiases do not tune iteratively as in SLFN e ELM offersmany advantages over other learning algorithms [54] Fromthe results shown in Tables 3ndash6 for AD vs HC the clas-sification result increased by 5ndash7 as compared to SVMclassifiers with 9804 sensitivity and 9628 specificityFor AD vs MCI classification accuracy increased by 5ndash9with 92 sensitivity and 9733 accuracy For HC vs MCIclassification accuracy increased by 6ndash10 with 9123sensitivity and 9913 specificity Similarly for MCIs vs

Table 3 10-fold cross-validated classification performance for AD vs HC

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 8513 9207 8544 086 8068 8568 8381 7717 8317 7381Surface area 8230 8930 9154 085 7616 8616 7514 7867 8767 7262Volume 8790 9387 8475 088 7929 7429 7833 8000 7710 6324CSF 8973 9633 8738 089 7905 8905 7417 7533 8333 6857APOE+MMSE 8645 9513 8700 093 8203 8703 8467 8416 8305 7879Concatenation (all_features_set) 9731 9804 9628 097 9133 9333 8757 9350 955 9058

Table 4 10-fold cross-validated classification performance for AD vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7010 8130 6610 071 6401 8383 6233 6503 7880 7330Surface area 6160 6670 8430 063 6345 6810 8011 6067 7123 6285Volume 6780 8210 6710 069 5820 7792 6856 6005 7337 6373CSF 6470 6600 8140 073 5607 6275 7573 6123 6871 7937APOE+MMSE 8145 8440 8770 082 7681 6781 8548 7620 6694 8748Concatenation (all_features_set) 8791 9200 9733 089 8350 8864 7672 7831 7445 8262

Table 5 10-fold cross-validated classification performance for HC vs MCI

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 7540 8407 6790 083 7417 8081 7123 6734 7317 7793Surface area 7415 6520 8385 084 6354 6767 6647 6872 7443 6735Volume 7784 8841 7280 084 6547 7473 6875 7283 7827 6570CSF 8330 7393 8402 085 7339 7525 6683 7580 7338 7814APOE+MMSE 7937 9325 7297 089 6829 8173 7668 7023 8124 7805Concatenation (all_features_set) 9172 9123 9913 093 8170 8368 8749 8565 7854 8926

Table 6 10-fold cross-validated classification performance for MCIs vs MCIc

Features measureELM SVM-RBF SVM-linear

ACC SEN SPE AUC ACC SEN SPE ACC SEN SPECortical thickness 6371 7110 3837 064 5718 6553 6038 5005 7118 6813Surface area 6143 7228 7480 069 5425 6322 7617 5827 7223 6485Volume 6785 8440 8773 074 5917 7828 6755 6214 7445 5773CSF 7137 7884 6857 072 6473 6009 7278 6733 7882 6878APOE+MMSE 7609 8713 7209 079 6553 6881 6743 6808 7221 6633Concatenation (all_features_set) 8338 9301 7577 085 7137 8116 7675 7571 7838 8467

12 Computational Intelligence and Neuroscience

MCIc classification accuracy increased by 7ndash12 with9301 sensitivity and 7577 specificity but 8467 and7667 specificity was obtained for linear and RBF-SVMclassifiers respectively which is more than obtained by theELM classifier From our analyses we believe that the ELM

gives better performance as compared to SVM from alearning efficiency standpoint because the ELMrsquos originaldesign has high learning accuracy fast learning speedscalability and the least human intervention [54 55] Fromthe results shown in Table 7 we can see that our suggested

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(a)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(b)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(c)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(d)

Figure 5 Classification results for different Alzheimerrsquos groups using ELM classifiers (a) Classification results for AD vs HC groups (b) forAD vs MCI groups (c) for HC vs MCI groups and (d) for MCIs vs MCIc groups

Computational Intelligence and Neuroscience 13

method achieved better accuracy than other existingmethodsFor classifying AD and HC our method achieved a classi-fication accuracy of 9731 with a sensitivity of 9804 aspecificity of 9628 and an AUC value of 097 Forclassifying MCI and HC our method achieves a classificationaccuracy of 9172 with a sensitivity of 9123 a specificity of9913 and an AUC value of 093 For distinguishing ADfrom MCI our proposed method obtained a classificationaccuracy of 8791 with a sensitivity of 9200 a specificityof 9733 and an AUC value of 089 Similarly for MCIs vsMCIc we achieved outstanding performance as compared toprevious methods by combining multiple features with anaccuracy of 8338 a sensitivity of 9301 a specificity of7577 and an AUC value of 085

42ComparisonwithOtherMethods In the previous sectionwe discussed in detail the findings of the proposed featureselection and classification method In this section we

compare and discuss the findings of our method in com-parison with existing state-of-the-art methods Tables 7 and 8show a comparison of the classification performance of theproposed method with recently published studies which usedmultimodality data to distinguish individuals with AD andMCI from HC Westman et al [67] used MRI and CSFbiomarkers and obtained 918 accuracy for AD vs HC776 for HC vs MCI and 685 for pMCI vs MCIsclassification using the orthogonal partial least squares tolatent structures (OPLS) method Zhang and Shen [39] used amultimodal (MRI PET and CSF) classification method withmultitask feature selection and an SVM classifier for AD andMCI classification By combining MRI PET and CSF datathey achieved a higher accuracy of 933 for AD vs HC and832 accuracy for HC vs MCI classification Similarly theauthors achieved an accuracy of 739 on pMCI vs MCIssamples Hinrichs et al [69] obtained an accuracy of 924 forAD vs HC classification using MRI PET CSF APOE and

ELMSVM_RBFSVM-linear

0

20

40

60

80

100

ACC

()

AD vs HC HC vs MCI MCIs vs MCIcAD vs MCI

Figure 6 Performance comparison of different classifiers

Table 7 Comparison of classification of AD vs HC and HC vs MCI between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection AD vsHC ()

HC vsMCI

Westman et al[67] MRI +CSF OPLS mdash 918 776

Johnson et al [68] MRI + PET+CSF+ cognitive scores Stackedautoencoder

Sparse representationlearning 959 85

Hinrichs et al [69] MRI + PET+CSF+APOE+ cognitive scores MKL mdash 924 naZhang and Shenet al [39] MRI + PET+CSF SVM Multitask feature selection 933 832

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 9301 mdash

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 995 mdash

Maqsood et al[72] sMRI CNN mdash 9285 mdash

Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper(SFS) 9731 9172

14 Computational Intelligence and Neuroscience

cognitive score in multiple kernel learning Similarly Johnsonet al [68] proposed a sparse representation learning featureselection and stacked autoencoder for classifying AD usingMRI PET CSF and cognitive scores as features and obtained959 accuracy for AD vs HC classification and 85 accuracyfor HC vs MCI classification Maqsood et al [72] proposedthe transfer learning classification model of AD for bothbinary and multiclass problems e algorithm utilizes apretrained AlexNet network and fine-tuned the convolutionalneural network (CNN) for classificationis model was fine-tuned over both segmented and unsegmented 3D views of thehuman brain e MRI scans were segmented into thecharacteristic components of GM white matter and CSFeretrained CNNwas then validated using the test data to obtainaccuracies of 896 and 928 for binary and multiclassproblems respectively using the OASIS dataset Beheshtiet al [70] developed a computer-aided diagnosis (CAD)system composed of four systematic stages for analyzingglobal and local differences in the GM of AD patientscompared with HC using the voxel-based morphometry(VBM) methodey used feature ranking based on the t-testand a genetic algorithm with Fisherrsquos criteria as part of theobjective function in the genetic algorithm e authorsutilized the SVM classifier with 10-fold cross-validation toobtain accuracies of 9301 and 75 for AD vs HC andMCIsvs pMCI classification respectively Spasov et al [71] de-veloped a deep learning architecture based on dual learningand an ad hoc layer for 3D separable convolution to identifyMCI patients eir deep learning procedure combined MRIneuropsychological demographic and APOE ε4 data toachieve an accuracy of 86 ey achieved an accuracy of995 for AD vs HC classification In another study Moradiet al used VBM analysis of GM as a feature combined withage and cognitive measures ey achieved an accuracy of82 for pMCI vs MCIs classification From this comparisonwe can infer that the proposed feature combination (MRICSF APOE and MMSE data) is robust or comparable to theother multimodal biomarker methods reported in the liter-ature for both AD vs HC and MCIs vs MCIc classification

5 Conclusion

In conclusion we demonstrated that a combination of threesMRImeasures cortical thickness cortical area cortical volumeand three nonimaging measures CSF components APOE ε4status and MMSE score improves AD diagnosis Furthermore

this combination shows great potential for the early identifi-cation of mild cognitive impairment (the prodromal stage ofAlzheimerrsquos disease) In this method we proposed filter andwrapper feature selection with an ELM classifier for multiplebiomarker-based AD diagnosis which significantly improvedthe classifierrsquos performance Moreover the results were betterthan or comparable with those previously reported particularlyfor the most challenging classification such as HC vs MCI andMCIs vs MCIc e added value of combining different ana-tomical MRI measures should be considered in AD scanningprotocols Only using specific aspects or a single measure ofwhole brain atrophy for AD diagnosis is still common practiceOur results show that clinical AD diagnosis could benefit fromthe combination of multiple measures from an anatomical MRIscan and other nonimaging biomarkers incorporated into anautomated machine learning system Our suggested methodeffectively enhances the diagnostic accuracy ofADandMCI butstill has some drawbacks Future work will include variousimprovements First we will optimize the parameter obtainingprocess Second in order to enhance the effectiveness of thesuggested method the dataset will be increased in terms of thefollowing extension of the longitudinal dataset for better un-derstanding of the progression from MCI and inclusion ofmultimodal data such as PET and fMRI data which providesdifferent insights into the characteristics of AD

Data Availability

e Alzheimerrsquos Disease Neuroimaging Initiative (ADNI)dataset (httpadniloniuscedu) was used in this studyComplete information regarding ADNI investigators can befound at httpadniloniusceduwp_contentuploadshow_to_applyADNI_Acknowledgement_istpd

Conflicts of Interest

e authors declare no conflicts of interest relating to thiswork

Acknowledgments

is work was supported by the National Research Foun-dation of Korea Grant funded by the Korean Government(NRF-2019R1F1A1060166 and NRF-2019R1A4A1029769)Data collection and sharing for this project was funded bythe ADNI (National Institutes of Health grant number U01

Table 8 Comparison of classification of MCIs vs MCIc between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection MCIs vs MCIc()

Westman et al [67] MRI +CSF OPLS mdash 685Zhang and Shen et al[39] MRI + PET+CSF SVM Multitask feature selection 7390

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 7500

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 86

Moradi et al [73] MRI + age and cognitive mdash 82Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper (SFS) 8338

Computational Intelligence and Neuroscience 15

MCIc classification accuracy increased by 7ndash12 with9301 sensitivity and 7577 specificity but 8467 and7667 specificity was obtained for linear and RBF-SVMclassifiers respectively which is more than obtained by theELM classifier From our analyses we believe that the ELM

gives better performance as compared to SVM from alearning efficiency standpoint because the ELMrsquos originaldesign has high learning accuracy fast learning speedscalability and the least human intervention [54 55] Fromthe results shown in Table 7 we can see that our suggested

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(a)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(b)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(c)

ACCSENSPE

Cor

tical

thic

knes

s

Surfa

ce ar

ea

Volu

me

CSF

APO

E +

MM

SE

All_

feat

ures

_set

0

20

40

60

80

100

(d)

Figure 5 Classification results for different Alzheimerrsquos groups using ELM classifiers (a) Classification results for AD vs HC groups (b) forAD vs MCI groups (c) for HC vs MCI groups and (d) for MCIs vs MCIc groups

Computational Intelligence and Neuroscience 13

method achieved better accuracy than other existingmethodsFor classifying AD and HC our method achieved a classi-fication accuracy of 9731 with a sensitivity of 9804 aspecificity of 9628 and an AUC value of 097 Forclassifying MCI and HC our method achieves a classificationaccuracy of 9172 with a sensitivity of 9123 a specificity of9913 and an AUC value of 093 For distinguishing ADfrom MCI our proposed method obtained a classificationaccuracy of 8791 with a sensitivity of 9200 a specificityof 9733 and an AUC value of 089 Similarly for MCIs vsMCIc we achieved outstanding performance as compared toprevious methods by combining multiple features with anaccuracy of 8338 a sensitivity of 9301 a specificity of7577 and an AUC value of 085

42ComparisonwithOtherMethods In the previous sectionwe discussed in detail the findings of the proposed featureselection and classification method In this section we

compare and discuss the findings of our method in com-parison with existing state-of-the-art methods Tables 7 and 8show a comparison of the classification performance of theproposed method with recently published studies which usedmultimodality data to distinguish individuals with AD andMCI from HC Westman et al [67] used MRI and CSFbiomarkers and obtained 918 accuracy for AD vs HC776 for HC vs MCI and 685 for pMCI vs MCIsclassification using the orthogonal partial least squares tolatent structures (OPLS) method Zhang and Shen [39] used amultimodal (MRI PET and CSF) classification method withmultitask feature selection and an SVM classifier for AD andMCI classification By combining MRI PET and CSF datathey achieved a higher accuracy of 933 for AD vs HC and832 accuracy for HC vs MCI classification Similarly theauthors achieved an accuracy of 739 on pMCI vs MCIssamples Hinrichs et al [69] obtained an accuracy of 924 forAD vs HC classification using MRI PET CSF APOE and

ELMSVM_RBFSVM-linear

0

20

40

60

80

100

ACC

()

AD vs HC HC vs MCI MCIs vs MCIcAD vs MCI

Figure 6 Performance comparison of different classifiers

Table 7 Comparison of classification of AD vs HC and HC vs MCI between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection AD vsHC ()

HC vsMCI

Westman et al[67] MRI +CSF OPLS mdash 918 776

Johnson et al [68] MRI + PET+CSF+ cognitive scores Stackedautoencoder

Sparse representationlearning 959 85

Hinrichs et al [69] MRI + PET+CSF+APOE+ cognitive scores MKL mdash 924 naZhang and Shenet al [39] MRI + PET+CSF SVM Multitask feature selection 933 832

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 9301 mdash

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 995 mdash

Maqsood et al[72] sMRI CNN mdash 9285 mdash

Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper(SFS) 9731 9172

14 Computational Intelligence and Neuroscience

cognitive score in multiple kernel learning Similarly Johnsonet al [68] proposed a sparse representation learning featureselection and stacked autoencoder for classifying AD usingMRI PET CSF and cognitive scores as features and obtained959 accuracy for AD vs HC classification and 85 accuracyfor HC vs MCI classification Maqsood et al [72] proposedthe transfer learning classification model of AD for bothbinary and multiclass problems e algorithm utilizes apretrained AlexNet network and fine-tuned the convolutionalneural network (CNN) for classificationis model was fine-tuned over both segmented and unsegmented 3D views of thehuman brain e MRI scans were segmented into thecharacteristic components of GM white matter and CSFeretrained CNNwas then validated using the test data to obtainaccuracies of 896 and 928 for binary and multiclassproblems respectively using the OASIS dataset Beheshtiet al [70] developed a computer-aided diagnosis (CAD)system composed of four systematic stages for analyzingglobal and local differences in the GM of AD patientscompared with HC using the voxel-based morphometry(VBM) methodey used feature ranking based on the t-testand a genetic algorithm with Fisherrsquos criteria as part of theobjective function in the genetic algorithm e authorsutilized the SVM classifier with 10-fold cross-validation toobtain accuracies of 9301 and 75 for AD vs HC andMCIsvs pMCI classification respectively Spasov et al [71] de-veloped a deep learning architecture based on dual learningand an ad hoc layer for 3D separable convolution to identifyMCI patients eir deep learning procedure combined MRIneuropsychological demographic and APOE ε4 data toachieve an accuracy of 86 ey achieved an accuracy of995 for AD vs HC classification In another study Moradiet al used VBM analysis of GM as a feature combined withage and cognitive measures ey achieved an accuracy of82 for pMCI vs MCIs classification From this comparisonwe can infer that the proposed feature combination (MRICSF APOE and MMSE data) is robust or comparable to theother multimodal biomarker methods reported in the liter-ature for both AD vs HC and MCIs vs MCIc classification

5 Conclusion

In conclusion we demonstrated that a combination of threesMRImeasures cortical thickness cortical area cortical volumeand three nonimaging measures CSF components APOE ε4status and MMSE score improves AD diagnosis Furthermore

this combination shows great potential for the early identifi-cation of mild cognitive impairment (the prodromal stage ofAlzheimerrsquos disease) In this method we proposed filter andwrapper feature selection with an ELM classifier for multiplebiomarker-based AD diagnosis which significantly improvedthe classifierrsquos performance Moreover the results were betterthan or comparable with those previously reported particularlyfor the most challenging classification such as HC vs MCI andMCIs vs MCIc e added value of combining different ana-tomical MRI measures should be considered in AD scanningprotocols Only using specific aspects or a single measure ofwhole brain atrophy for AD diagnosis is still common practiceOur results show that clinical AD diagnosis could benefit fromthe combination of multiple measures from an anatomical MRIscan and other nonimaging biomarkers incorporated into anautomated machine learning system Our suggested methodeffectively enhances the diagnostic accuracy ofADandMCI butstill has some drawbacks Future work will include variousimprovements First we will optimize the parameter obtainingprocess Second in order to enhance the effectiveness of thesuggested method the dataset will be increased in terms of thefollowing extension of the longitudinal dataset for better un-derstanding of the progression from MCI and inclusion ofmultimodal data such as PET and fMRI data which providesdifferent insights into the characteristics of AD

Data Availability

e Alzheimerrsquos Disease Neuroimaging Initiative (ADNI)dataset (httpadniloniuscedu) was used in this studyComplete information regarding ADNI investigators can befound at httpadniloniusceduwp_contentuploadshow_to_applyADNI_Acknowledgement_istpd

Conflicts of Interest

e authors declare no conflicts of interest relating to thiswork

Acknowledgments

is work was supported by the National Research Foun-dation of Korea Grant funded by the Korean Government(NRF-2019R1F1A1060166 and NRF-2019R1A4A1029769)Data collection and sharing for this project was funded bythe ADNI (National Institutes of Health grant number U01

Table 8 Comparison of classification of MCIs vs MCIc between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection MCIs vs MCIc()

Westman et al [67] MRI +CSF OPLS mdash 685Zhang and Shen et al[39] MRI + PET+CSF SVM Multitask feature selection 7390

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 7500

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 86

Moradi et al [73] MRI + age and cognitive mdash 82Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper (SFS) 8338

Computational Intelligence and Neuroscience 15

method achieved better accuracy than other existingmethodsFor classifying AD and HC our method achieved a classi-fication accuracy of 9731 with a sensitivity of 9804 aspecificity of 9628 and an AUC value of 097 Forclassifying MCI and HC our method achieves a classificationaccuracy of 9172 with a sensitivity of 9123 a specificity of9913 and an AUC value of 093 For distinguishing ADfrom MCI our proposed method obtained a classificationaccuracy of 8791 with a sensitivity of 9200 a specificityof 9733 and an AUC value of 089 Similarly for MCIs vsMCIc we achieved outstanding performance as compared toprevious methods by combining multiple features with anaccuracy of 8338 a sensitivity of 9301 a specificity of7577 and an AUC value of 085

42ComparisonwithOtherMethods In the previous sectionwe discussed in detail the findings of the proposed featureselection and classification method In this section we

compare and discuss the findings of our method in com-parison with existing state-of-the-art methods Tables 7 and 8show a comparison of the classification performance of theproposed method with recently published studies which usedmultimodality data to distinguish individuals with AD andMCI from HC Westman et al [67] used MRI and CSFbiomarkers and obtained 918 accuracy for AD vs HC776 for HC vs MCI and 685 for pMCI vs MCIsclassification using the orthogonal partial least squares tolatent structures (OPLS) method Zhang and Shen [39] used amultimodal (MRI PET and CSF) classification method withmultitask feature selection and an SVM classifier for AD andMCI classification By combining MRI PET and CSF datathey achieved a higher accuracy of 933 for AD vs HC and832 accuracy for HC vs MCI classification Similarly theauthors achieved an accuracy of 739 on pMCI vs MCIssamples Hinrichs et al [69] obtained an accuracy of 924 forAD vs HC classification using MRI PET CSF APOE and

ELMSVM_RBFSVM-linear

0

20

40

60

80

100

ACC

()

AD vs HC HC vs MCI MCIs vs MCIcAD vs MCI

Figure 6 Performance comparison of different classifiers

Table 7 Comparison of classification of AD vs HC and HC vs MCI between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection AD vsHC ()

HC vsMCI

Westman et al[67] MRI +CSF OPLS mdash 918 776

Johnson et al [68] MRI + PET+CSF+ cognitive scores Stackedautoencoder

Sparse representationlearning 959 85

Hinrichs et al [69] MRI + PET+CSF+APOE+ cognitive scores MKL mdash 924 naZhang and Shenet al [39] MRI + PET+CSF SVM Multitask feature selection 933 832

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 9301 mdash

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 995 mdash

Maqsood et al[72] sMRI CNN mdash 9285 mdash

Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper(SFS) 9731 9172

14 Computational Intelligence and Neuroscience

cognitive score in multiple kernel learning Similarly Johnsonet al [68] proposed a sparse representation learning featureselection and stacked autoencoder for classifying AD usingMRI PET CSF and cognitive scores as features and obtained959 accuracy for AD vs HC classification and 85 accuracyfor HC vs MCI classification Maqsood et al [72] proposedthe transfer learning classification model of AD for bothbinary and multiclass problems e algorithm utilizes apretrained AlexNet network and fine-tuned the convolutionalneural network (CNN) for classificationis model was fine-tuned over both segmented and unsegmented 3D views of thehuman brain e MRI scans were segmented into thecharacteristic components of GM white matter and CSFeretrained CNNwas then validated using the test data to obtainaccuracies of 896 and 928 for binary and multiclassproblems respectively using the OASIS dataset Beheshtiet al [70] developed a computer-aided diagnosis (CAD)system composed of four systematic stages for analyzingglobal and local differences in the GM of AD patientscompared with HC using the voxel-based morphometry(VBM) methodey used feature ranking based on the t-testand a genetic algorithm with Fisherrsquos criteria as part of theobjective function in the genetic algorithm e authorsutilized the SVM classifier with 10-fold cross-validation toobtain accuracies of 9301 and 75 for AD vs HC andMCIsvs pMCI classification respectively Spasov et al [71] de-veloped a deep learning architecture based on dual learningand an ad hoc layer for 3D separable convolution to identifyMCI patients eir deep learning procedure combined MRIneuropsychological demographic and APOE ε4 data toachieve an accuracy of 86 ey achieved an accuracy of995 for AD vs HC classification In another study Moradiet al used VBM analysis of GM as a feature combined withage and cognitive measures ey achieved an accuracy of82 for pMCI vs MCIs classification From this comparisonwe can infer that the proposed feature combination (MRICSF APOE and MMSE data) is robust or comparable to theother multimodal biomarker methods reported in the liter-ature for both AD vs HC and MCIs vs MCIc classification

5 Conclusion

In conclusion we demonstrated that a combination of threesMRImeasures cortical thickness cortical area cortical volumeand three nonimaging measures CSF components APOE ε4status and MMSE score improves AD diagnosis Furthermore

this combination shows great potential for the early identifi-cation of mild cognitive impairment (the prodromal stage ofAlzheimerrsquos disease) In this method we proposed filter andwrapper feature selection with an ELM classifier for multiplebiomarker-based AD diagnosis which significantly improvedthe classifierrsquos performance Moreover the results were betterthan or comparable with those previously reported particularlyfor the most challenging classification such as HC vs MCI andMCIs vs MCIc e added value of combining different ana-tomical MRI measures should be considered in AD scanningprotocols Only using specific aspects or a single measure ofwhole brain atrophy for AD diagnosis is still common practiceOur results show that clinical AD diagnosis could benefit fromthe combination of multiple measures from an anatomical MRIscan and other nonimaging biomarkers incorporated into anautomated machine learning system Our suggested methodeffectively enhances the diagnostic accuracy ofADandMCI butstill has some drawbacks Future work will include variousimprovements First we will optimize the parameter obtainingprocess Second in order to enhance the effectiveness of thesuggested method the dataset will be increased in terms of thefollowing extension of the longitudinal dataset for better un-derstanding of the progression from MCI and inclusion ofmultimodal data such as PET and fMRI data which providesdifferent insights into the characteristics of AD

Data Availability

e Alzheimerrsquos Disease Neuroimaging Initiative (ADNI)dataset (httpadniloniuscedu) was used in this studyComplete information regarding ADNI investigators can befound at httpadniloniusceduwp_contentuploadshow_to_applyADNI_Acknowledgement_istpd

Conflicts of Interest

e authors declare no conflicts of interest relating to thiswork

Acknowledgments

is work was supported by the National Research Foun-dation of Korea Grant funded by the Korean Government(NRF-2019R1F1A1060166 and NRF-2019R1A4A1029769)Data collection and sharing for this project was funded bythe ADNI (National Institutes of Health grant number U01

Table 8 Comparison of classification of MCIs vs MCIc between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection MCIs vs MCIc()

Westman et al [67] MRI +CSF OPLS mdash 685Zhang and Shen et al[39] MRI + PET+CSF SVM Multitask feature selection 7390

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 7500

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 86

Moradi et al [73] MRI + age and cognitive mdash 82Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper (SFS) 8338

Computational Intelligence and Neuroscience 15

cognitive score in multiple kernel learning Similarly Johnsonet al [68] proposed a sparse representation learning featureselection and stacked autoencoder for classifying AD usingMRI PET CSF and cognitive scores as features and obtained959 accuracy for AD vs HC classification and 85 accuracyfor HC vs MCI classification Maqsood et al [72] proposedthe transfer learning classification model of AD for bothbinary and multiclass problems e algorithm utilizes apretrained AlexNet network and fine-tuned the convolutionalneural network (CNN) for classificationis model was fine-tuned over both segmented and unsegmented 3D views of thehuman brain e MRI scans were segmented into thecharacteristic components of GM white matter and CSFeretrained CNNwas then validated using the test data to obtainaccuracies of 896 and 928 for binary and multiclassproblems respectively using the OASIS dataset Beheshtiet al [70] developed a computer-aided diagnosis (CAD)system composed of four systematic stages for analyzingglobal and local differences in the GM of AD patientscompared with HC using the voxel-based morphometry(VBM) methodey used feature ranking based on the t-testand a genetic algorithm with Fisherrsquos criteria as part of theobjective function in the genetic algorithm e authorsutilized the SVM classifier with 10-fold cross-validation toobtain accuracies of 9301 and 75 for AD vs HC andMCIsvs pMCI classification respectively Spasov et al [71] de-veloped a deep learning architecture based on dual learningand an ad hoc layer for 3D separable convolution to identifyMCI patients eir deep learning procedure combined MRIneuropsychological demographic and APOE ε4 data toachieve an accuracy of 86 ey achieved an accuracy of995 for AD vs HC classification In another study Moradiet al used VBM analysis of GM as a feature combined withage and cognitive measures ey achieved an accuracy of82 for pMCI vs MCIs classification From this comparisonwe can infer that the proposed feature combination (MRICSF APOE and MMSE data) is robust or comparable to theother multimodal biomarker methods reported in the liter-ature for both AD vs HC and MCIs vs MCIc classification

5 Conclusion

In conclusion we demonstrated that a combination of threesMRImeasures cortical thickness cortical area cortical volumeand three nonimaging measures CSF components APOE ε4status and MMSE score improves AD diagnosis Furthermore

this combination shows great potential for the early identifi-cation of mild cognitive impairment (the prodromal stage ofAlzheimerrsquos disease) In this method we proposed filter andwrapper feature selection with an ELM classifier for multiplebiomarker-based AD diagnosis which significantly improvedthe classifierrsquos performance Moreover the results were betterthan or comparable with those previously reported particularlyfor the most challenging classification such as HC vs MCI andMCIs vs MCIc e added value of combining different ana-tomical MRI measures should be considered in AD scanningprotocols Only using specific aspects or a single measure ofwhole brain atrophy for AD diagnosis is still common practiceOur results show that clinical AD diagnosis could benefit fromthe combination of multiple measures from an anatomical MRIscan and other nonimaging biomarkers incorporated into anautomated machine learning system Our suggested methodeffectively enhances the diagnostic accuracy ofADandMCI butstill has some drawbacks Future work will include variousimprovements First we will optimize the parameter obtainingprocess Second in order to enhance the effectiveness of thesuggested method the dataset will be increased in terms of thefollowing extension of the longitudinal dataset for better un-derstanding of the progression from MCI and inclusion ofmultimodal data such as PET and fMRI data which providesdifferent insights into the characteristics of AD

Data Availability

e Alzheimerrsquos Disease Neuroimaging Initiative (ADNI)dataset (httpadniloniuscedu) was used in this studyComplete information regarding ADNI investigators can befound at httpadniloniusceduwp_contentuploadshow_to_applyADNI_Acknowledgement_istpd

Conflicts of Interest

e authors declare no conflicts of interest relating to thiswork

Acknowledgments

is work was supported by the National Research Foun-dation of Korea Grant funded by the Korean Government(NRF-2019R1F1A1060166 and NRF-2019R1A4A1029769)Data collection and sharing for this project was funded bythe ADNI (National Institutes of Health grant number U01

Table 8 Comparison of classification of MCIs vs MCIc between the proposed method and existing state-of-the-art methods

Author Data Classifier Feature selection MCIs vs MCIc()

Westman et al [67] MRI +CSF OPLS mdash 685Zhang and Shen et al[39] MRI + PET+CSF SVM Multitask feature selection 7390

Beheshti et al [70] sMRI SVM Feature ranking + geneticalgorithm 7500

Spasov et al [71] sMRI + cognitivemeasures +APOE+demographic CNN mdash 86

Moradi et al [73] MRI + age and cognitive mdash 82Proposed method MRI +CSF+APOE+MMSE ELM Filter (MRMR) +wrapper (SFS) 8338

Computational Intelligence and Neuroscience 15

AG024904) and the DOD ADNI (Department of Defensegrant number W81XWH-12-2-0012) e ADNI is fundedby the National Institute on Aging the National Institute ofBiomedical Imaging and Bioengineering and the generouscontributions from the following AbbVie AlzheimerrsquosAssociation Alzheimerrsquos Drug Discovery FoundationAraclon Biotech BioClinica Inc Biogen Bristol-MyersSquibb Company CereSpir Inc CogstateEisai Inc ElanPharmaceuticals Inc Eli Lilly and Company EuroImmunF Hofmann-La Roche Ltd and its affiliated companyGenentech Inc Fujirebio GE Healthcare IXICO LtdJanssen Alzheimer Immunotherapy Research amp Develop-ment LLC Johnson amp Johnson Pharmaceutical Research ampDevelopment LLC Lumosity Lundbeck Merck amp Co IncMeso Scale Diagnostics LLC NeuroRx Research Neuro-track Technologies Novartis Pharmaceuticals CorporationPfzer Inc Piramal Imaging Servier Takeda PharmaceuticalCompany and Transition erapeutics e Canadian In-stitutes of Health Research provides funding to supportADNI clinical sites in Canada Private sector contributionsare facilitated by the Foundation for the National Institutesof Health (httpwwwfnihorg) e Northern CaliforniaInstitute for Research and Education is the grantee orga-nization and the study was coordinated by the Alzheimerrsquoserapeutic Research Institute at the University of SouthernCalifornia ADNI data are disseminated by the Laboratoryfor Neuroimaging at the University of Southern California

References

[1] A Serrano-Pozo M P Frosch E Masliah and B T HymanldquoNeuropathological alterations in alzheimer diseaserdquo ColdSpring Harbor Perspectives inMedicine vol 1 no 1 Article IDa006189 2011

[2] Alzheimerrsquos Association ldquo2015 Alzheimerrsquos disease facts andfiguresrdquo Alzheimerrsquos amp Dementia vol 11 no 3 pp 332ndash3842015

[3] S C Johnson ldquoe influence of Alzheimer disease familyhistory and apolipoprotein e epsilon4 onmesial temporal lobeactivationrdquo Journal of Neuroscience vol 26 no 22pp 6069ndash6076 2006

[4] P M ompson and L G Apostolova ldquoComputationalanatomical methods as applied to ageing and dementiardquo CeBritish Journal of Radiology vol 80 no 2 pp S78ndashS91 2007

[5] J L Whitwell S A Przybelski S D Weigand et al ldquo3Dmapsfrom multiple MRI illustrate changing atrophy patterns assubjects progress from mild cognitive impairment to Alz-heimerrsquos diseaserdquo Brain vol 130 no 7 pp 1777ndash1786 2007

[6] E M Reiman R J Caselli S Lang et al ldquoPreclinical evidenceof Alzheimerrsquos disease in persons homozygous for the epsilon4 allele for apolipoprotein erdquo New England Journal of Med-icine vol 334 no 12 pp 752ndash758 1996

[7] E J Golob R Irimajiri and A Starr ldquoAuditory corticalactivity in amnestic mild cognitive impairment relationshipto subtype and conversion to dementiardquo Brain vol 130 no 3pp 740ndash752 2007

[8] M S Albert S T DeKosky D Dickson et al ldquoe diagnosisof mild cognitive impairment due to Alzheimerrsquos diseaserecommendations from the national institute on aging-Alz-heimerrsquos association workgroups on diagnostic guidelines forAlzheimerrsquos diseaserdquo Alzheimerrsquos amp Dementia vol 7 no 3pp 270ndash279 2011

[9] R C Petersen G E Smith S C Waring R J IvnikE G Tangalos and E Kokmen ldquoMild cognitive impairmentrdquoArchives of Neurology vol 56 no 3 pp 303ndash308 1999

[10] R A Sperling P S Aisen L A Beckett et al ldquoToward de-fining the preclinical stages of Alzheimerrsquos disease recom-mendations from the national institute on aging-alzheimerrsquosassociation workgroups on diagnostic guidelines for Alz-heimerrsquos diseaserdquo Alzheimerrsquos amp Dementia vol 7 no 3pp 280ndash292 2011

[11] M Tabuas-Pereira I Baldeiras D Duro et al ldquoPrognosis ofearly-onset vs late-onset mild cognitive impairment com-parison of conversion rates and its predictorsrdquo Geriatricsvol 1 no 2 p 11 2016

[12] L Mosconi R Mistur R Switalski et al ldquoFDG-PET changesin brain glucose metabolism from normal cognition topathologically verified Alzheimerrsquos diseaserdquo European Journalof Nuclear Medicine and Molecular Imaging vol 36 no 5pp 811ndash822 2009

[13] G M McKhann D S Knopman H Chertkow et al ldquoediagnosis of dementia due to Alzheimerrsquos disease recom-mendations from the national institute on aging-Alzheimerrsquosassociation workgroups on diagnostic guidelines for Alz-heimerrsquos diseaserdquo Alzheimerrsquos amp Dementia vol 7 no 3pp 263ndash269 2011

[14] A-T Du N Schuff J H Kramer et al ldquoDifferent regionalpatterns of cortical thinning in Alzheimerrsquos disease and fronto-temporal dementiardquo Brain vol 130 no 4 pp 1159ndash1166 2007

[15] A Tessitore G Santangelo R De Micco et al ldquoCorticalthickness changes in patients with Parkinsonrsquos disease andimpulse control disordersrdquo Parkinsonism amp Related Disor-ders vol 24 pp 119ndash125 2016

[16] R K Lama J Gwak J-S Park and S-W Lee ldquoDiagnosis ofAlzheimerrsquos disease based on structural MRI images using aregularized extreme learning machine and PCA featuresrdquoJournal of Healthcare Engineering vol 2017 Article ID5485080 11 pages 2017

[17] O Querbes F Aubry J Pariente et al ldquoEarly diagnosis ofAlzheimerrsquos disease using cortical thickness impact of cog-nitive reserverdquo Brain vol 132 no 8 pp 2036ndash2047 2009

[18] P P D M Oliveira R Nitrini G Busatto C BuchpiguelJ R Sato and E Amaro ldquoUse of SVM methods with surface-based cortical and volumetric subcortical measurements todetect Alzheimerrsquos diseaserdquo Journal of Alzheimerrsquos Diseasevol 19 no 4 pp 1263ndash1272 2010

[19] S F Eskildsen P Coupe D Garcıa-Lorenzo V FonovJ C Pruessner and D L Collins ldquoPrediction of Alzheimerrsquosdisease in subjects with mild cognitive impairment from theADNI cohort using patterns of cortical thinningrdquo Neuro-Image vol 65 pp 511ndash521 2013

[20] S Kloppel C M Stonnington C Chu et al ldquoAutomaticclassification of MR scans in Alzheimerrsquos diseaserdquo Brainvol 131 no 3 pp 681ndash689 2008

[21] M Liu D Zhang and D Shen ldquoHierarchical fusion offeatures and classifier decisions for Alzheimerrsquos disease di-agnosisrdquo Human Brain Mapping vol 35 no 4 pp 1305ndash1319 2014

[22] T Tong R Wolz Q Gao R Guerrero J V Hajnal andD Rueckert ldquoMultiple instance learning for classification ofdementia in brain MRIrdquo Medical Image Analysis vol 18no 5 pp 808ndash818 2014

[23] P Coupe S F Eskildsen J V Manjon V S Fonov andD L Collins ldquoSimultaneous segmentation and grading ofanatomical structures for patientrsquos classification application

16 Computational Intelligence and Neuroscience

to Alzheimerrsquos diseaserdquo NeuroImage vol 59 no 4pp 3736ndash3747 2012

[24] R Wolz V Julkunen J Koikkalainen et al ldquoMulti-methodanalysis of MRI images in early diagnostics of Alzheimerrsquosdiseaserdquo PLoS One vol 6 no 10 Article ID e25446 2011

[25] D Zhang Y Wang L Zhou H Yuan and D Shen ldquoMul-timodal classification of Alzheimerrsquos disease and mild cog-nitive impairmentrdquo NeuroImage vol 55 no 3 pp 856ndash8672011

[26] R Stephen T Ngandu Y Liu et al ldquoBrain volumes andcortical thickness on MRI in the finnish geriatric interventionstudy to prevent cognitive impairment and disability (finger)rdquoAlzheimerrsquos Research amp Cerapy vol 11 no 1 2019

[27] S F Eskildsen P Coupe V S Fonov J C Pruessner andD L Collins ldquoStructural imaging biomarkers of Alzheimerrsquosdisease predicting disease progressionrdquo Neurobiology ofAging vol 36 no 1 pp S23ndashS31 2015

[28] J Hwang C M Kim S Jeon et al ldquoPrediction of Alzheimerrsquosdisease pathophysiology based on cortical thickness patternsrdquoAlzheimerrsquos amp Dementia Diagnosis Assessment amp DiseaseMonitoring vol 2 no 1 pp 58ndash67 2016

[29] E Challis P Hurley L Serra M Bozzali S Oliver andM Cercignani ldquoGaussian process classification of Alzheimerrsquosdisease and mild cognitive impairment from resting-statefMRIrdquo NeuroImage vol 112 pp 232ndash243 2015

[30] Y Zhang Z Dong P Phillips et al ldquoDetection of subjects andbrain regions related to Alzheimerrsquos disease using 3D MRIscans based on eigenbrain and machine learningrdquo Frontiers inComputational Neuroscience vol 9 p 66 2015

[31] J B S Langbaum K Chen W Lee et al ldquoCategorical andcorrelational analyses of baseline fluorodeoxyglucose positronemission tomography images from the Alzheimerrsquos diseaseneuroimaging initiative (ADNI)rdquo NeuroImage vol 45 no 4pp 1107ndash1116 2009

[32] K R Gray P Aljabar R A Heckemann A Hammers andD Rueckert ldquoRandom forest-based similarity measures formulti-modal classification of Alzheimerrsquos diseaserdquo Neuro-Image vol 65 pp 167ndash175 2013

[33] J B Langbaum A S Fleisher K Chen et al ldquoUshering in thestudy and treatment of preclinical Alzheimerrsquos diseaserdquoNature Reviews Neurology vol 9 no 7 pp 371ndash381 2013

[34] H Hampel K Burger S J Teipel A L W BokdeH Zetterberg and K Blennow ldquoCore candidate neuro-chemical and imaging biomarkers of Alzheimerrsquos diseaserdquoAlzheimerrsquos amp Dementia vol 4 no 1 pp 38ndash48 2008

[35] M Vounou E Janousova R Wolz et al ldquoSparse reduced-rank regression detects genetic associations with voxel-wiselongitudinal phenotypes in Alzheimerrsquos diseaserdquoNeuroImagevol 60 no 1 pp 700ndash716 2012

[36] B Jie D Zhang B Cheng and D Shen ldquoManifold regu-larized multitask feature learning for multimodality diseaseclassificationrdquo Human Brain Mapping vol 36 no 2pp 489ndash507 2015

[37] F Liu C-Y Wee H Chen and D Shen ldquoInter-modalityrelationship constrained multi-modality multi-task featureselection for Alzheimerrsquos disease and mild cognitive im-pairment identificationrdquo NeuroImage vol 84 pp 466ndash4752014

[38] L Yuan Y Wang P Mompson V A Narayan and J YeldquoMulti-source feature learning for joint analysis of incompletemultiple heterogeneous neuroimaging datardquo NeuroImagevol 61 no 3 pp 622ndash632 2012

[39] D Zhang and D Shen ldquoMulti-modal multi-task learning forjoint prediction of multiple regression and classification

variables in Alzheimerrsquos diseaserdquo NeuroImage vol 59 no 2pp 895ndash907 2012

[40] O Kohannim X Hua D P Hibar et al ldquoBoosting power forclinical trials using classifiers based on multiple biomarkersrdquoNeurobiology of Aging vol 31 no 8 pp 1429ndash1442 2010

[41] C Davatzikos Y Fan X Wu D Shen and S M ResnickldquoDetection of prodromal Alzheimerrsquos disease via patternclassification of magnetic resonance imagingrdquo Neurobiologyof Aging vol 29 no 4 pp 514ndash523 2008

[42] Y Fan S M Resnick X Wu and C Davatzikos ldquoStructuraland functional biomarkers of prodromal Alzheimerrsquos diseasea high-dimensional pattern classification studyrdquo NeuroImagevol 41 no 2 pp 277ndash285 2008

[43] C Hinrichs V Singh L Mukherjee G Xu M K Chung andS C Johnson ldquoSpatially augmented lpboosting for ADclassification with evaluations on the ADNI datasetrdquo Neu-roImage vol 48 no 1 pp 138ndash149 2009

[44] B Cheng M Liu D Zhang B C Munsell and D ShenldquoDomain transfer learning for MCI conversion predictionrdquoIEEE Transactions on Biomedical Engineering vol 62 no 7pp 1805ndash1817 2015

[45] C-Y Wee P-T Yap and D Shen ldquoPrediction of Alzheimerrsquosdisease and mild cognitive impairment using cortical mor-phological patternsrdquo Human Brain Mapping vol 34 no 12pp 3411ndash3425 2013

[46] W Wang B C Ooi X Yang D Zhang and Y ZhuangldquoEffective multi-modal retrieval based on stacked auto-en-codersrdquo Proceedings of the VLDB Endowment vol 7 no 8pp 649ndash660 2014

[47] N Srivastava and R Salakhutdinov ldquoMultimodal learningwith deep boltzmann machinesrdquo Journal of Machine LearningResearch vol 15 no 1 pp 2949ndash2980 2014

[48] W Ouyang X Chu and X Wang ldquoMulti-source deeplearning for human pose estimationrdquo in Proceedings of theIEEE Conference on Computer Vision and Pattern Recognitionpp 2337ndash2344 Columbus OH USA September 2014

[49] H-I Suk and D Shen ldquoDeep learning-based feature repre-sentation for ADMCI classificationrdquo Advanced InformationSystems Engineering Springer Berlin Germany pp 583ndash5902013

[50] G Huang H Zhou X Ding and R Zhang ldquoExtreme learningmachine for regression and multiclass classificationrdquo IEEETransactions on Systems Man and Cybernetics Part B (Cy-bernetics) vol 42 no 2 pp 513ndash529 2012

[51] G-B Huang Q-Y Zhu and C-K Siew ldquoExtreme learningmachine a new learning scheme of feedforward neural net-worksrdquo in Proceedings of the IEEE International Joint Con-ference on Neural Networks Budapest Hungary July 2004

[52] D Jha S Alam J-Y Pyun K H Lee and G-R KwonldquoAlzheimerrsquos disease detection using extreme learning ma-chine complex dual tree wavelet principal coefficients andlinear discriminant analysisrdquo Journal of Medical Imaging andHealth Informatics vol 8 no 5 pp 881ndash890 2018

[53] D T Nguyen S Ryu M N I Qureshi M Choi K H Leeand B Lee ldquoHybrid multivariate pattern analysis combinedwith extreme learning machine for Alzheimerrsquos dementiadiagnosis using multi-measure rs-fMRI spatial patternsrdquoPLoS One vol 14 no 2 Article ID e0212582 2019

[54] G Huang G-B Huang S Song and K You ldquoTrends inextreme learning machines a reviewrdquo Neural Networksvol 61 pp 32ndash48 2015

[55] X Liu C Gao and P Li ldquoA comparative analysis of supportvector machines and extreme learning machinesrdquo NeuralNetworks vol 33 pp 58ndash66 2012

Computational Intelligence and Neuroscience 17

[56] B Fischl ldquoFreesurferrdquo NeuroImage vol 62 no 2pp 774ndash781 2012

[57] R S Desikan F Segonne B Fischl et al ldquoAn automatedlabeling system for subdividing the human cerebral cortex onMRI scans into gyral based regions of interestrdquo NeuroImagevol 31 no 3 pp 968ndash980 2006

[58] L Yaddanapudi ldquoe American statistical association state-ment on p values explainedrdquo Journal of AnaesthesiologyClinical Pharmacology vol 32 no 4 pp 421ndash423 2016

[59] M Radovic M Ghalwash N Filipovic and Z ObradovicldquoMinimum redundancy maximum relevance feature selectionapproach for temporal gene expression datardquo BMC Bio-informatics vol 18 no 1 2017

[60] S H Hojjati A Ebrahimzadeh A Khazaee and A Babajani-Feremi ldquoPredicting conversion from MCI to AD usingresting-state fMRI graph theoretical approach and SVMrdquoJournal of Neuroscience Methods vol 282 pp 69ndash80 2017

[61] C Cortes and V Vapnik ldquoSupport-vector networksrdquo Ma-chine Learning vol 20 no 3 pp 273ndash297 1995

[62] M N I Qureshi J Oh B Min H J Jo and B Lee ldquoMulti-modal multi-measure and multi-class discrimination ofADHD with hierarchical feature extraction and extremelearning machine using structural and functional brain MRIrdquoFrontiers in Human Neuroscience vol 11 p 157 2017

[63] A Akusok Y Miche J Karhunen K-M Bjork R Nian andA Lendasse ldquoArbitrary category classification of websitesbased on image contentrdquo IEEE Computational IntelligenceMagazine vol 10 no 2 pp 30ndash41 2015

[64] J Cao and Z Lin ldquoExtreme learning machines on high di-mensional and large data applications a surveyrdquo Mathe-matical Problems in Engineering vol 2015 Article ID 10379613 pages 2015

[65] Y Chen E Yao and A Basu ldquoA 128-channel extremelearning machine-based neural decoder for brain machineinterfacesrdquo IEEE Transactions on Biomedical Circuits andSystems vol 10 no 3 pp 679ndash692 2016

[66] B A Richards H Chertkow V Singh et al ldquoPatterns ofcortical thinning in Alzheimerrsquos disease and frontotemporaldementiardquo Neurobiology of Aging vol 30 no 10 pp 1626ndash1636 2009

[67] E Westman J-S Muehlboeck and A Simmons ldquoCombiningMRI and CSF measures for classification of Alzheimerrsquosdisease and prediction of mild cognitive impairment con-versionrdquo NeuroImage vol 62 no 1 pp 229ndash238 2012

[68] H-I Johnson S-W Lee S-W Lee and D Shen ldquoLatentfeature representation with stacked auto-encoder for ADMCI diagnosisrdquo Brain Structure and Function vol 220 no 2pp 841ndash859 2015

[69] C Hinrichs V Singh G Xu and S C Johnson ldquoPredictivemarkers for AD in a multi-modality framework an analysis ofMCI progression in the ADNI populationrdquo NeuroImagevol 55 no 2 pp 574ndash589 2011

[70] I Beheshti H Demirel and H Matsuda ldquoClassification ofAlzheimerrsquos disease and prediction of mild cognitive im-pairment-to-Alzheimerrsquos conversion from structural mag-netic resource imaging using feature ranking and a geneticalgorithmrdquo Computers in Biology and Medicine vol 83pp 109ndash119 2017

[71] S Spasov L Passamonti A Duggento P Lio and N ToschildquoA parameter-efficient deep learning approach to predictconversion from mild cognitive impairment to Alzheimerrsquosdiseaserdquo NeuroImage vol 189 pp 276ndash287 2019

[72] M Maqsood F Nazir U Khan et al ldquoTransfer learningassisted classification and detection of Alzheimerrsquos disease

stages using 3D MRI scansrdquo Sensors vol 19 no 11 p 26452019

[73] E Moradi A Pepe C Gaser H Huttunen and J TohkaldquoMachine learning framework for early MRI-based Alz-heimerrsquos conversion prediction in MCI subjectsrdquo Neuro-Image vol 104 pp 398ndash412 2015

18 Computational Intelligence and Neuroscience

[56] B Fischl ldquoFreesurferrdquo NeuroImage vol 62 no 2pp 774ndash781 2012

[57] R S Desikan F Segonne B Fischl et al ldquoAn automatedlabeling system for subdividing the human cerebral cortex onMRI scans into gyral based regions of interestrdquo NeuroImagevol 31 no 3 pp 968ndash980 2006

[58] L Yaddanapudi ldquoe American statistical association state-ment on p values explainedrdquo Journal of AnaesthesiologyClinical Pharmacology vol 32 no 4 pp 421ndash423 2016

[59] M Radovic M Ghalwash N Filipovic and Z ObradovicldquoMinimum redundancy maximum relevance feature selectionapproach for temporal gene expression datardquo BMC Bio-informatics vol 18 no 1 2017

[60] S H Hojjati A Ebrahimzadeh A Khazaee and A Babajani-Feremi ldquoPredicting conversion from MCI to AD usingresting-state fMRI graph theoretical approach and SVMrdquoJournal of Neuroscience Methods vol 282 pp 69ndash80 2017

[61] C Cortes and V Vapnik ldquoSupport-vector networksrdquo Ma-chine Learning vol 20 no 3 pp 273ndash297 1995

[62] M N I Qureshi J Oh B Min H J Jo and B Lee ldquoMulti-modal multi-measure and multi-class discrimination ofADHD with hierarchical feature extraction and extremelearning machine using structural and functional brain MRIrdquoFrontiers in Human Neuroscience vol 11 p 157 2017

[63] A Akusok Y Miche J Karhunen K-M Bjork R Nian andA Lendasse ldquoArbitrary category classification of websitesbased on image contentrdquo IEEE Computational IntelligenceMagazine vol 10 no 2 pp 30ndash41 2015

[64] J Cao and Z Lin ldquoExtreme learning machines on high di-mensional and large data applications a surveyrdquo Mathe-matical Problems in Engineering vol 2015 Article ID 10379613 pages 2015

[65] Y Chen E Yao and A Basu ldquoA 128-channel extremelearning machine-based neural decoder for brain machineinterfacesrdquo IEEE Transactions on Biomedical Circuits andSystems vol 10 no 3 pp 679ndash692 2016

[66] B A Richards H Chertkow V Singh et al ldquoPatterns ofcortical thinning in Alzheimerrsquos disease and frontotemporaldementiardquo Neurobiology of Aging vol 30 no 10 pp 1626ndash1636 2009

[67] E Westman J-S Muehlboeck and A Simmons ldquoCombiningMRI and CSF measures for classification of Alzheimerrsquosdisease and prediction of mild cognitive impairment con-versionrdquo NeuroImage vol 62 no 1 pp 229ndash238 2012

[68] H-I Johnson S-W Lee S-W Lee and D Shen ldquoLatentfeature representation with stacked auto-encoder for ADMCI diagnosisrdquo Brain Structure and Function vol 220 no 2pp 841ndash859 2015

[69] C Hinrichs V Singh G Xu and S C Johnson ldquoPredictivemarkers for AD in a multi-modality framework an analysis ofMCI progression in the ADNI populationrdquo NeuroImagevol 55 no 2 pp 574ndash589 2011

[70] I Beheshti H Demirel and H Matsuda ldquoClassification ofAlzheimerrsquos disease and prediction of mild cognitive im-pairment-to-Alzheimerrsquos conversion from structural mag-netic resource imaging using feature ranking and a geneticalgorithmrdquo Computers in Biology and Medicine vol 83pp 109ndash119 2017

[71] S Spasov L Passamonti A Duggento P Lio and N ToschildquoA parameter-efficient deep learning approach to predictconversion from mild cognitive impairment to Alzheimerrsquosdiseaserdquo NeuroImage vol 189 pp 276ndash287 2019

[72] M Maqsood F Nazir U Khan et al ldquoTransfer learningassisted classification and detection of Alzheimerrsquos disease

stages using 3D MRI scansrdquo Sensors vol 19 no 11 p 26452019

[73] E Moradi A Pepe C Gaser H Huttunen and J TohkaldquoMachine learning framework for early MRI-based Alz-heimerrsquos conversion prediction in MCI subjectsrdquo Neuro-Image vol 104 pp 398ndash412 2015

18 Computational Intelligence and Neuroscience