55
Journal of Advances in Computer Research Quarterly pISSN: 2345-606x eISSN: 2345-6078 Sari Branch, Islamic Azad University, Sari, I.R.Iran (Vol. 7, No. 4, November 2016), Pages: 55-66 www.jacr.iausari.ac.ir
Discrimination between Iron Deficiency Anaemia (IDA) and β-Thalassemia Trait (β-TT) Based on Pattern-
Based Input Selection Artificial Neural Network (PBIS-ANN)
Mehrzad Khaki Jamei, Khadijeh Mirzaei Talarposhti
Department of Computer Engineering, Sari Branch, Islamic Azad University, Sari, Iran
[email protected]; [email protected]
Received: 2016/04/16; Accepted: 2016/06/11
Abstract
Discrimination between iron deficiency (IDA) anemia and β-thalassemia trait
(β-TT) is a time consuming and costly problem. Because, they have approximate
similar effects on routine blood test indices, in some cases, the complementary tests,
which are expensive and time consuming, would be needed for differentiate the
anemia. Complete blood count (CBC) is a fast, inexpensive, and accessible medical
test that is used as a primary test for diagnosis anemia. However, when the CBC
indices cannot exactly state the subject, more advanced tests such as electrophoresis
of hemoglobin must be performed. In this study, the CBC indices have been
considered as the inputs of classifier and the chosen architecture is pattern-based
input selection artificial neural network (PBIS-ANN). For evaluation the proposed
method, traditional methods, which are still using for the problem such as Mentzer
Index (MI), and several automated anemia diagnostic systems such as artificial
neural networks (ANN), adaptive neuro-fuzzy inference system (ANFIS) and multi-
layer perceptron (MLP) have been compared with the proposed method. The results
indicate that the proposed method significantly outperforms the mentioned methods.
Keywords: Iron Deficiency Anemia, β-Thalassemia Trait, Artificial Neural Network, Complete
Blood Count, Hemoglobin
1. Introduction
Despite iron deficiency anemia (IDA) that is not a genetic disease, thalassemia is a
genetic trait that causes a reduction in the life span of a red blood cell [1]. Thalassemia
disease is a result of an abnormality in the genes that regulate the formation of
hemoglobin. Thalassemia screening is an action to detect the couples with the potential
of having a thalassemic infant. To achieve this goal the subjects with the thalassemia
trait of same type must be recognized. There are several types of thalassemia. When the
parents of a child have the thalassemia trait of the same type, he or she has the disease,
trait, or is normal with the probability of 25%, 50%, and 25%, respectively. One of the
most common types of thalassemia trait is β-thalassemia trait (β-TT). CBC, a fast an
inexpensive medical test, is the primary test in the thalassemia screening. The similarity
of CBC indices in IDA and β-TT makes discrimination between them so difficult.
Hence, in the subjects that couldn’t be stated with their CBC indices, the
complementary tests, which are time consuming and expensive, would be needed.
Additionally, these tests are not available in all laboratories and hospitals. As a result,
Discrimination between Iron Deficiency … M. Khaki Jamei, Kh. Mirzaei Talarposhti
56
many researchers have tried to find methods for discrimination between IDA and β-TT
with interpreting the CBC indices [2–14].
Historically, mathematical formulas were used for the problem [2–7,12,13,15]. In the
past few decades there has been increasing interested in the automated medical
diagnostic systems. Early attempts to formulate an automated diagnostic tool for anemia
classification employed image analysis [16], statistical [17], and clustering techniques
[18]. Later, the implementation protocol has shifted to the expert systems, in which both
rule-based [19–21] and hybrid neural network/rule-based [22] systems have been
successfully tested in clinical trials. The other studies on automated classification of
anemia was performed using image analysis [23] and artificial neural networks (ANN)
[24]. A comparison study on thalassemia screening in which support vector machine
(SVM), K-nearest neighbor (KNN), and an MLP were investigated [25], reported that
the MLP classifier gives slightly better results than the SVM. Another comparison study
on diagnosis IDA [26] was performed between ANN, ANFIS, and a logistic regression
model. The authors revealed that the ANN is superior to the others.
An investigation on genetic programming (GP) in thalassemia classification [27]
discovered that a GP-based decision tree and an MLP with one hidden layer are
approximately equal in accuracy. But, the MLP with two hidden layers is superior.
Piroonratana et al. investigated three classifier in hemoglobin typing for thalassemia
screening [28] that are: C4.5, random forest and MLP. They revealed that C4.5 is more
suitable than the others when using high performance liquid chromatography (HPLC) as
the input. Although, the previous studies have been successfully tested, there are two
major reasons for further investigation into automated discrimination between IDA and
β-TT. First, an increase in precision is required, because a high accuracy in thalassemia
screening is vital for having healthy population in the next generation. Second,
reduction cost and time is achievable by using the CBC indices as the inputs of the
anemia classifier, while the most accurate previous studies used the time consuming and
expensive medical tests such as the electrophoresis or HPLC. This study has been
performed to satisfy these requirements. It compares the proposed method with the
existing methods of discrimination IDA and β-TT.
The remaining of the article is organized as follows. Section 2 explains some
scientific information on the structure of hemoglobin and its formation, and the relations
between the anemia and hemoglobin. In section 3, the collected data will be analyzed.
Section 4 introduces the proposed method and section 5 compares the obtained results
of the proposed method with the other methods in literature. Finally, section 6
concludes the study and gives some suggestions for further works
2. The Effects of Anemia
Hemoglobin, which is a core component of a red the blood cell, consists of four poly
peptide chains (globin) and the other component that is called ‘heme’ [1]. First
component is affected by thalassemia and the second one is affected by IDA. There are
four types of globin chains that are: alpha (α), beta (β), gamma (γ), and delta (δ).
Regarding to the types of globin chains in the structure of hemoglobin, it can belongs to
one of the three major types of hemoglobin, including HbA (α2β2), HbA2 (α2δ2), and
HbF (α2γ2). HbF or fetal hemoglobin is the main constructor of a fetus’ red blood cells,
which after born gradually exchanges with HbA. The proportions of the HbA, HbA2,
Journal of Advances in Computer Research (Vol. 7, No. 4, November 2016) 55-66
57
and HbF in a mature and healthy person’s blood are approximately, 97%, 2-3%, and
less than 1%, respectively.
Thalassemia disease stems from the defects on the genes that regulate the formation of
globin chains. The type of thalassemia and the severity of disease are depended on the
type and the number of defective genes, respectively. In fact, persons with thalassemia
trait do not have the disease but inherit genes that cause the disease. Generally, the
incidence of β-thalassemia is more than α-thalassemia. Therefore, detection of β-TT to
avoid of having the β-thalassemia infants is a major purpose of screening. In the other
hand, when the subject has IDA disease, because of deficiency of the iron, the
concentration of the component ‘heme’, which is an iron-based molecule, will be
reduced. Hence, the quantity and quality of hemoglobin will be affected in the presence
of IDA. In the next subsection more details of the effects of anemia on blood indices
will be explained.
2.1. CBC indices
Each red blood cell contains approximately 300 million molecules of hemoglobin.
Hence, a change in the structure of globin affects the structure and functionality of the
red blood cells. As previous discussed, IDA and β-TT both lead to decrease
hemoglobin’s quality and quantity. With the difference that the first one affects the
‘heme’ component and the later affects the globin formation. As a result, thalassemia
changes the CBC indices in the following manner: Hemoglobin concentration (Hb),
which denotes the quantity of hemoglobin in blood, reduces in presence of thalassemia
disease or trait. The red blood cell count (RBC) is the other index of CBC that in a
thalassemic subject might be increased for compensation of the hemoglobin reduction.
The other significant indices of CBC that are usually employed in diagnosis thalassemia
are: mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), mean
corpuscular hemoglobin concentration (MCHC), red blood cell distribution width
(RDW), and hematocrit (HCT). These indices would also be reduced in the presence of
thalassemia disease or β-TT.
In the other hand, IDA has similar effects on CBC indices except that in the β-TT
subjects, we have increased RBC while in the IDA subjects the RBC index would be
decreased. This is the most important difference between those anemia. In spite of the
difference on RBC volume on different cases, it is so difficult to discriminate IDA from
β-TT subjects.
3. Data Analysis
The samples were collected from the archives of the several thalassemia screening
centers where situated in five northern cities of Iran. Total number of collected CBC
tests is 750, which were obtained from the blood specimen of 390 males in ages 20-35
and 360 females in ages 17-32 years old. The information has been organized as
follows.
I. The samples that their CBC indices obviously are in the normal ranges (218
samples).
II. The samples that their CBC indices are abnormal and definitely show
β-TT or IDA on them (98 samples, IDA = 38, β-TT = 60).
Discrimination between Iron Deficiency … M. Khaki Jamei, Kh. Mirzaei Talarposhti
58
III. The samples that their CBC indices cannot exactly determine them.
They were stated by complementary tests such as electrophoresis
(434 samples, IDA = 231, β-TT = 203).
According to the above issues total number of β-TT and number of IDA subjects are
269 and 263, respectively. Table 1 shows some examples of different categories of
collected samples. Abnormal values are shown in bold face (Table 1). It is undeniable
that the samples in the cat. III are so similar and there is a complex relation that a simple
mathematical formula cannot distinguish between IDA and β-TT subjects in this
category as well. Because, the study’s purpose is differentiate between IDA and β-TT
subjects, we excluded the normal subjects from the collected samples and the remaining
samples (532 samples including IDA and β-TT) were randomly divided into the training
and the testing patterns with 132 and 400 samples, respectively.
Table 1. Some examples of different classes of collected samples.
4. Proposed Method
ANN, which is inspired by human organism, first considered by Warren McCulloch
and Walter Pitts [29] in decade 40. After that, Donald Hebb introduced a training
algorithm for ANN. Classification, pattern recognition, prediction, function
approximation and data processing are some applications for ANNs [30–33]. In this
study a new method based on ANN has been proposed to differentiate between IDA and
β-TT subjects. In the next subsections we explain the architecture of the proposed
method.
4.1. The architecture of the proposed method
To understand the idea behind the proposed method, consider the abnormal values in
Table 1, which are marked with the symbols “↑” or “↓”. It is clear that some indices
Sample
number
RBC(M/mm3) Hb(g/dL) HCT (%) MCV(fL) MCH(pg) MCHC (%) Category
number
Type (4.5-6.3)* (13.5-18) (39-50) (80-96) (27-32) (32-38)
1 5.43 10.2 ↓ 34 ↓ 62.6 ↓ 18.8 ↓ 30 ↓ II β-TT
2 6.13 12.5 ↓ 40.1 65.4 ↓ 20.4 ↓ 31.2 ↓ II β-TT
3 6.8 ↑ 12.6 ↓ 43.4 63.8 ↓ 18.5 ↓ 29 ↓ II β-TT
4 4.73 10.1 ↓ 38.7 ↓ 69.9 ↓ 19.7 ↓ 30.3 ↓ II IDA
5 4.13 ↓ 8.2 ↓ 38.4 ↓ 71.8 ↓ 19.9 ↓ 28.9 ↓ II IDA
6 4.61 12.8 ↓ 38.9 ↓ 84.4 27.8 32.9 I normal
7 4.36 ↓ 13.1 ↓ 39.7 91.1 30 33 I normal
8 4.77 13.3 ↓ 39.7 83.2 27.9 33.5 I normal
9 3.99 ↓ 11.4 ↓ 35.1 ↓ 88 28.6 32.5 I normal
10 4.4 ↓ 13.7 40.5 92 31.1 33.8 I normal
11 6.62 ↑ 13.1 ↓ 43.9 65.9 ↓ 19.7 ↓ 29.8 ↓ III β-TT
12 4.56 10.3 ↓ 34.7 ↓ 76.1 ↓ 22.6 ↓ 29.7 ↓ III β-TT
13 5.95 12.5 ↓ 43.4 72.8 ↓ 21 ↓ 28.9 ↓ III β-TT
14 4.65 10.5 ↓ 35.6 ↓ 76.6 ↓ 22.6 ↓ 29.5 ↓ III β-TT
15 5.12 10.6 ↓ 36.7 ↓ 71.7 ↓ 20.7 ↓ 28.9 ↓ III β-TT
16 4.81 11.6 ↓ 37.6 ↓ 78.2 ↓ 24.1 ↓ 30.9 ↓ III IDA
17 5.05 11.5 ↓ 38.8 ↓ 76.8 ↓ 22.8 ↓ 29.6 ↓ III IDA
18 4.88 12.1 ↓ 38.8 ↓ 79.5 ↓ 24.8 ↓ 31.2 ↓ III IDA
19 5.03 12.7 ↓ 39.9 79.3 ↓ 25.2 ↓ 31.8 ↓ III IDA
20 5.1 13.2 ↓ 39.7 77.8 ↓ 25.9 ↓ 30.7 ↓ III IDA *Normal ranges are shown in the below of each column’s title.
Journal of Advances in Computer Research (Vol. 7, No. 4, November 2016) 55-66
59
have approximately the same patterns of abnormal values and supposed to be redundant.
Thus, to choose the most suitable indices as the inputs of the ANN, we proposed an
architecture naming pattern-based input selection artificial neural network (PBIS-ANN).
A briefly illustration of the proposed method has been shown in Fig. 1.
The dataset that were used for implementing the proposed method has been explained
in the previous section. This dataset contains 532 CBC samples from IDA and β-TT
subjects.
Figure 1. Briefly illustration of the proposed method
Table 2. The algorithm for finding the pair of CBC indices with the most similarity.
Inputs : Indices as matrix Am×n 1 & Array B in size of m1 (Bi is the lower band of normal ranges for 𝑖’𝑡ℎ
CBC index)
Output : R1 , R2 : the indices for the pair of rows with the most similarity;
1 Max = 0
2 FOR i =1 TO m-1 DO
3 FOR j=i+1 TO m DO
4 COS = Calculate similarity2 between rows i and j ;
5 IF COS > max THEN
6 max = COS ;
7 R1= i , R2 = j;
8 END IF
9 END FOR
10 END FOR 1 m is the number of indices & n is the number of samples; 2 calculate similarity regarding Eqs. (1,2);
As seen in Fig. 1, the most suitable indices of CBC are achieved with an iterative
algorithm. To achieve the goal, first, the pair of indices with the most similarity (with
Discrimination between Iron Deficiency … M. Khaki Jamei, Kh. Mirzaei Talarposhti
60
respect to the abnormal pattern) must be found with the algorithm that is shown in Table
2. Then, one CBC index between that pair must be eliminated. To find the best choice
for elimination a coefficient that is called coefficient of interference (COI) will be
calculated with the algorithm that is shown in Table 3. The index with higher COI from
the most similar pair must be eliminated.
Table 3. The algorithm to calculate coefficient of interference (COI) for an index of the CBC tests.
Input: the values of an index of the collected CBC tests as a one dimensional array ‘I’ with the
size of N
Output: the coefficient of interference (COI) for the input ( I)
1 Insert the values of ‘I’ that are related to IDA and β-TT subjects into the
arrays A & B, respectively;
2 N1 = size of (A) & N2= size of (B) ;
4 Max1 = max of (A) & Max2 = max of (B) ;
5 Min1 = min of (A) & Min2 = min of (B) ;
6 S = 0 ;
7 IF Min1 < Min2 THEN
8 FOR i = 1 TO N1 DO
9 IF A(i) > Min2 THEN
10 S=S+1;
11 END IF
12 END FOR
13 FOR i =1 TO N2 DO
14 IF B(i) < Max1 THEN
15 S=S+1;
16 END IF
17 END FOR
18 ELSE
19 FOR i =1 TO N2 DO
20 IF B(i)> Min1 THEN
21 S=S+1;
22 END IF
23 END FOR
24 FOR i =1 TO N1 DO
25 IF A(i)< Max2 THEN
26 S=S+1;
27 END IF
28 END FOR
29 END IF
30 COI= S / N;
31 END.
As seen in Table 2, to find the most similar pair of indices, a coefficient that is called
coefficient of similarity (COS) must be calculated. The COS could be obtained by Eq.
(1) as follows.
1
( ) ( )
1
nk k
i i j j
k
LB Row LB Row
COSn
(1)
Where n is the number of samples, k
iRow is 'k th value for 'i th CBC index, and
( )iLB r is a function that denotes: whether r is an abnormal value or not, with respect
to the lower band of normal ranges for 'i th CBC index, and it can be obtained by Eq.
(2) as follows.
Journal of Advances in Computer Research (Vol. 7, No. 4, November 2016) 55-66
61
1 '( )
0 'i
if m lower bound of normal ranges for i th CBC indexLB m
if m lower bound of normal ranges for i th CBC index
(2)
After running the proposed method, four CBC indices have survived in the final
ANN that are: Hb, RBC, HCT, and MCV. For implementation the ANN, the collected
samples were randomly divided into two groups with 132 and 400 samples for training
and test, respectively. The chosen training algorithm was Levenberg–Marquardt and the
error function was sum square errors (SSE), which could be obtained by Eq. (3) as
follows.
( ) ( ) 2
1 1
( )n m
i i
j j
i j
SSE a d
(3)
Where n and m are the number of training patterns and the number of network’s
outputs, respectively. ( )i
ja denotes the 'j th output of the network for 'i th training
pattern, and ( )i
jd represents its corresponding desired output.
4.2. Training the employed ANN
In this subsection the structure of ANN and the designing issues are explained. As
previous discussed the methodology for choosing the inputs of ANN is performed by
the PBIS algorithm. The algorithm includes several iterations. In each iteration of the
algorithm, whole procedure of designing and the test of ANN will be performed.
As previous studies have proven, the most suitable neural network for the anemia
classification problem is multi-layer perceptron (MLP) [22,25–28,34,35]. To choose the
number of layers and neurons in each layer, we used trial and error method. Finally, the
MLP with one hidden layer as shown in Fig. 2. Has been chosen as discriminator
between IDA and β-TT.
Figure 2. The employed ANN in the proposed method
As seen in Fig. 2, signed sigmoid is used as the activation function in the input layer
and the hidden layer of MLP. The output layer takes a linier function as the activation
function. Each MLP has trained several times. Fig. 3 shows the best obtained error
diagram, which is SSE function, in the training of the designed MLP.
Discrimination between Iron Deficiency … M. Khaki Jamei, Kh. Mirzaei Talarposhti
62
Figure 3. SSE diagram for training of the proposed MLP
5. Results and Comparisons
For evaluation the proposed method, we compared the obtained results with the
existing methods of discrimination IDA and β-TT. This section has been organized in
two subsections. In the first one, some traditional methods, which are still used for the
problem, have been introduced and the CBC samples were applied to them. In the
second subsection several recent works have been compared with the proposed method.
5.1. Comparison with traditional methods
Several popular traditional methods, which is still using in discrimination IDA and β-
TT, have been shown in Table 4. For better comparison, the samples that used in
evaluation traditional methods were same as the test samples of our proposed method.
Table 4. Mathematical formulas for discrimination between IDA and β-TT.
Authors Symbol Year Formula IDA β-TT
Mentzer et al. [3] MI 1973 MCV RBC 13 13
England & Fraser [2] E&FI 1973 5 3.4MCV RBC Hb 0 0
Srivastava & Bevington [4] S&BI 1973 MCH RBC 3.8 3.8
Shine & Lal [5] S&LI 1977 2 0.01MCV MCH 1530 1530
Green & King [6] G&KI 1989 2 0.01MCV RDW Hb 72 72
Sirdah et al. [12] SI 2008 3MCV RBC Hb 27 27
Ehsani et al. [13] EI 2009 10MCV RBC 15 15
In all of the experiments six popular medical indices namely: sensitivity (SENS),
specificity (SPEC), positive predictive value (PPV), negative predictive value (NPV),
accuracy (ACC) and Youden’s index (YI), were used to comparison. The mentioned
indices are obtained by Eqs. (4 – 9) as follows:
100TP
SENSTP FN
(4)
Journal of Advances in Computer Research (Vol. 7, No. 4, November 2016) 55-66
63
100TN
SPECTN FP
(5)
100TP
PPVTP FP
(6)
100TN
NPVTN FN
(7)
100TP TN
ACCTP TN FP FN
(8)
100YI SENS SPEC (9)
Where TP, TN, FP and FN are true positive, true negative, false positive and false
negative, respectively. Currently, several well-known traditional methods in the field of
diagnosis anemia (Table 4) have been compared with the proposed method. Table 5
indicates that the proposed method gives significantly better results than traditional
methods.
Table 5. Comparison between the proposed method and traditional methods for diagnosis IDA.
YI (%)
ACC (%)
NPV (%)
PPV (%)
SPEC (%)
SENS (%)
FN FP TN TP Method
84.6 92.2 90.6 93.8 93.0 91.6 18 13 173 196 MI [3]
81.2 90.5 88.1 92.7 91.9 89.3 23 15 171 191 E&FI [2]
85.6 92.7 91.1 94.3 93.5 92.1 17 12 174 197 S&BI [4]
91.1 95.5 94.2 96.7 96.2 94.9 11 7 179 203 S&LI [5]
81.6 90.7 88.6 92.8 91.9 89.7 22 15 171 192 SI [12]
84.6 92.2 90.2 94.2 93.5 91.1 19 12 174 195 EI [13]
88.5 94.2 93.6 94.8 94.1 94.4 12 11 175 202 G&KI [6]
96.1 98.0 97.3 98.6 98.4 97.7 5 3 183 209 Proposed method
Total number of samples 400 (214 IDA & 186 β-TT).
The results that are shown in Table 5 give the performance indices of diagnosis IDA.
Because, the number of β-TT subjects is depended on the number of IDA in Table 5, we
can obtain the performance indices for β-TT diagnosis with Eqs. (10 – 15) as follows:
SENS SPECTT IDA (10)
SPEC SENSTT IDA (11)
PPV NPVTT IDA (12)
NPV PPVTT IDA (13)
ACC ACCTT IDA (14)
Y I Y ITT IDA (15)
5.2. More comparisons
The second part of the tests is aimed to compare the performance of the proposed
method with some recent works. As can be observed in Table 6, the proposed method is
more accurate than the other works in this area.
Discrimination between Iron Deficiency … M. Khaki Jamei, Kh. Mirzaei Talarposhti
64
Table 6. Comparison between the proposed method and the other works.
Accuracy Specificity Sensitivity Method Year Authors
90.7 95.6 87.1 ANFIS1 2012 Azarkhish et al. [26]
96.3 95.6 96.8 ANN2 2012 Azarkhish et al. [26]
93.5 92.0 95.0 MLP3 2002 Amendolia et al. [24]
89.0 83.0 95.0 SVM4 2003 Amendolia et al. [25]
85.0 93.0 77.0 KNN5 2003 Amendolia et al. [25]
----- 91.0 93.0 RBF6 2013 Masala et. al. [35]
----- 73.0 89.0 PNN7 2013 Masala et. al. [35]
----- 91.0 80.0 KNN5 2013 Masala et. al. [35]
94.3 87.9 84.7 Math8 2015 Bordbar et al. [36]
99.5 98.4 97.7 PBIS-ANN9 Proposed method 1 adaptive neuro-fuzzy inference system; 2 artificial neural network; 3 multi-layer perceptron; 4 support vector
machine; 5 K-nearest neighbor; 6 Radial Basis Function; 7 Probabilistic Neural Network; 8 mathematical formula; 9 pattern-based input selection artificial neural network;
6. Conclusion
In this study, based entirely on fast and inexpensive blood test, we developed an
accurate medical diagnostic system for discrimination between iron deficiency anemia
(IDA) and β-thalassemia trait (β-TT). The model that is called pattern-based input
selection artificial neural network (PBIS-ANN) has been compared with traditional
methods of discrimination diagnosis IDA and β-TT and the recently published works
such as ANFIS, ANN, MLP, SVM, and KNN. The results indicate that the proposed
method significantly outperforms the mentioned methods. The proposed method stems
from a combination of human expert decision making and ANN. Before running the
proposed method the patterns of abnormal values in each input must be stated by an
expert human, on a set of inputs. The method could be employed in many medical
problems in which the medical indices are used for diagnosis a disease. Since
nationality and the ages of population, which is considered in the tests, is severely
limited. Hence, it is suggested that a wider population would be considered for the
future works.
References
[1] D.J. Weatherall, J.B. Clegg, "The Thalassemia Syndromes," Malden: MA: Blackwell Science,
2001.
[2] J.M. England, P.M. Fraser, "Differentiation of iron deficiency from thalassemia trait by routine
blood–count," Lancent, Vol. 1, No. 7801, pp.447-449, 1973.
[3] W.C. Mentzer, "Differentiation of iron deficiency from thalassemia trait," Lancent, Vol. 1, No.
7808, pp.882, 1973.
[4] P.C. Srivastava, J.M. Bevington, "No Iron deficiency and–or thalassemia trait," Lancent, Vol. 1,
No. 7807, pp.832, 1973.
[5] I. Shine, S. Lal, "A strategy to detect beta thalassemia minor," Lancent, Vol. 1, No. 8013,
pp.692-694, 1977.
[6] R. Green, R. King, "A new red blood cell discriminant incorporating volume dispersion for
differentiating iron deficiency anemia from thalassemia minor," Blood Cells, Vol. 15, pp.481-
495, 1989.
[7] S. Ayabose, J. Giavanelli, I.O. Levendoglu-Tuga, C. Sandoval, F. Özkaynak, P. Visintainer,
"Differentiating iron deficiency anemia from Thalassemia minor by using an RDW-based
Journal of Advances in Computer Research (Vol. 7, No. 4, November 2016) 55-66
65
index," J Pediatr Hematol, Vol. 21, pp.314, 1999.
[8] A. Demir, N. Yarali, T. Fisgin, F. Duru, A. Kara, "Most reliable indices in differentiation
between thalassemia trait and iron deficiency anemia," Pediatr Int, Vol. 44, No. 6, pp.612-616,
2002.
[9] T. Higgins, Y. Brierley, A. Khajuria, "Use of mathematical indices based on CBC data to
identify patients with β thalassemia minor," Rev Român Ăde Med Ăde Lab, Vol. 8, No. 3, pp.7-
12, 2007.
[10] G. Ntaios, A. Chatzinikolaou, Z. Saouli, F. Girtovitis, M. Tsapanidou, G. Kaiafa, et al.,
"Discrimination indices as screening tests for β-thalassemic trait," Ann Hematol, Vol. 86, No. 7,
pp.487-491, 2007.
[11] A.D. Rathod, A. Kaur, V. Patel, K. Patel, R. Kabrawala, V. Patel, et al., "Usefulness of cell
counter-based parameters and formulas in detection of β-thalassemia trait in areas of high
prevalence," Am J Clin Pathol, Vol. 128, No. 4, pp.585-589, 2007.
[12] M. Sirdah, I. Tarazi, E. Al Najjar, R. Al Haddad, "Evaluation of the diagnostic reliability of
different RBC indices and formulas in the differentiation of the β-thalassemia minor from iron
deficiency in Palestinian population," Int J Lab Hematol, Vol. 30, No. 4, pp.324-330, 2008.
[13] M.A. Ehsani, E. Shahgholi, M.S. Rahiminejad, F. Seighali, A. Rashidi, "A new index for
discrimination between iron deficiency anemia and beta-thalassemia minor: results in 284
patients," Pak J Biol Sci, Vol. 12, No. 5, pp.473-475, 2009.
[14] E. Miri-Moghaddam, N. Sargolzaie, "Cut off determination of discrimination indices in
differential diagnosis between iron deficiency anemia and beta -thalassemia minor," Int J
Hematol Stem Cell Res, Vol. 8, No. 2, pp.27-32, 2014.
[15] C.A. Sahli, A. Bibi, F. Ouali, S.H. Fredj, B. Dakhlaoui, R. Othmani, et al., "Red cell indices:
Differentiation between β-thalassemia trait and iron deficiency anemia and application to sickle-
cell disease and sickle-cell thalassemia," Clin Chem Lab Med, Vol. 51, No. 11, pp.2115-24,
2013.
[16] P.R. Lund, R.D. Barnes, "Automated classification of anaemia using image analysis," Lancent,
Vol. 300, No. 7775, pp.463-464, 1972.
[17] R.L. Engle, B.J. Flehinger, S. Allen, R. Friedman, M. Lipkin, B.J. Davis, et al., "HEME: a
computer aid to diagnosis of hematologic disease," Bull N Y Acad Med, Vol. 52, pp.584-600,
1976.
[18] G.M. Barosi, Cazzola, C. Berzuini, S. Quaglini, M. Stefanelli, "Classification of anemia on the
basis of ferrokinetic parameters," Br J Haematol, Vol. 61, pp.357-370, 1985.
[19] S. Quaglini, M. Stefanelli, G. Barosi, A. Berzuini, "ANEMIA: an expert consultation system,"
Comput Biomed Res, Vol. 19, No. 1, pp.13-27, 1986.
[20] S. Quaglini, M. Stefanelli, G. Barosi, A. Berzuini, "A performance evaluation of the expert
system ANEMIA," Comput Biomed Res, Vol. 21, pp.307-323, 1988.
[21] G. Lanzola, M. Stefanelli, G. Barosi, L. Magnani, "NEOANEMIA: a knowledge-based system
emulating diagnostic reasonin," Comput Biomed Res, Vol. 23, pp.560-582, 1990.
[22] N.I. Birndorf, J.O. Pentecost, J.R. Coakley, K.A. Spackman, "An expert system to diagnose
anemia and report results directly on hematology forms," Com Biomed Res, Vol. 26, No. 1,
pp.16-26, 1996.
[23] N. Nikolaev, L.M. de Menezes, H. Iba, "Automated classification of anaemia using image
analysis," In: 2002 Congress on Evolutionary Computation, pp.1209-14, 2002.
[24] S.R. Amendolia, A. Brunetti, P. Carta, G. Cossu, M.L. Ganadu, B. Golosio, et al., "A real-time
classification system of thalassemic pathologies based on artificial neural networks," Med Decis
Making, Vol. 18, No. 1, pp.18-26, 2002.
[25] S.R. Amendolia, G. Cossu, M.L. Ganadu, B. Golosio, G.L. Masala, G.M. Mura, "A comparative
Discrimination between Iron Deficiency … M. Khaki Jamei, Kh. Mirzaei Talarposhti
66
study of k-nearest neighbour, support vector machine and multi-layer perceptron for thalassemia
screening," Chemom Intell Lab Syst, Vol. 1, pp.13-20, 2003.
[26] I. Azarkhish, M.R. Raoufy, S. Gharibzadeh, "Artificial intelligence models for predicting iron
deficiency anemia and iron serum level based on accessible laboratory data," Journal of Medical
Systems, Vol. 36, No. 3, pp.2057-61, 2012.
[27] W. Wongseree, N. Chaiyaratana, K. Vichittumaros, P. Winichagoon, S. Fucharoen,
"Thalassaemia classification by neural networks and genetic programming," Information
Sciences, Vol. 177, No. 3, pp.771-786, 2007.
[28] T. Piroonratana, W. Wongseree, A. Assawamakin, N. Paulkhaolarn, C. Kanjanakorn, M.
Sirikong, et al., "Classification of haemoglobin typing chromatograms by neural networks and
decision trees for thalassaemia screening," Chemom Intell Lab Syst, Vol. 99, No. 2, pp.101-110,
2009.
[29] W. McCulloch, W. Pitts, "A logical calculus of the ideas immanent in nervous activity," Bull
Math Biophys, Vol. 7, pp.115-133, 1943.
[30] A.V. Holden, V.I. Kryokuv, "Neural networks - Theory and architecture," manchester university
press, NewYourk, 1990.
[31] K.N. Gurney, "Neural Networks," Taylor & Francis Group, 1997.
[32] O. Omidvar, J. Dayhoff, "neural network and pattern recognition," Academic Press, London,
1998.
[33] M.B. Menhaj, "Fundumentals of Neural Networks," Amir Kabir University, Tehran, 2002.
[34] D.E. Bonds, "Three decades of innovation in the management of sickle cell disease: the road to
understanding the sickle cell disease clinical phenotype," Blood Rev, Vol. 19, No. 2, pp.99-110,
2004.
[35] G.L. Masala, B. Golosio, R. Cutzu, R. Pola, "A two-layered classifier based on the radial basis
function for the screening of thalassaemia," Comput Biol Med, Vol. 43, pp.1724-31, 2013.
[36] E. Bordbar, M. Taghipour, B.E. Zucconi, "Reliability of Different RBC Indices and Formulas in
Discriminating between β-Thalassemia Minor and other Microcytic Hypochromic Cases,"
Mediterr J Hematol Infect Dis, Vol. 7, No.1, pp. 1-5, 2015.
Top Related