Computational intelligence in early diabetes diagnosis: a review.
ABSTRACT The development of an effective diabetes diagnosis system by taking advantage of computational intelligence is regarded as a primary goal nowadays. Many approaches based on artificial network and machine learning algorithms have been developed and tested against diabetes datasets, which were mostly related to individuals of Pima Indian origin. Yet, despite high accuracies of up to 99% in predicting the correct diabetes diagnosis, none of these approaches have reached clinical application so far. One reason for this failure may be that diabetologists or clinical investigators are sparsely informed about, or trained in the use of, computational diagnosis tools. Therefore, this article aims at sketching out an outline of the wide range of options, recent developments, and potentials in machine learning algorithms as diabetes diagnosis tools. One focus is on supervised and unsupervised methods, which have made significant impacts in the detection and diagnosis of diabetes at primary and advanced stages. Particular attention is paid to algorithms that show promise in improving diabetes diagnosis. A key advance has been the development of a more indepth understanding and theoretical analysis of critical issues related to algorithmic construction and learning theory. These include tradeoffs for maximizing generalization performance, use of physically realistic constraints, and incorporation of prior knowledge and uncertainty. The review presents and explains the most accurate algorithms, and discusses advantages and pitfalls of methodologies. This should provide a good resource for researchers from all backgrounds interested in computational intelligencebased diabetes diagnosis methods, and allows them to extend their knowledge into this kind of research.
 [Show abstract] [Hide abstract]
ABSTRACT: Automated diagnosis of diseases has forever been of interest as an interdisciplinary study among computer and medical science researchers. Detection of hepatitis is really a big problem for general practitioners. An expert doctor commonly takes decisions by evaluating the current test results of a patient or by comparing the patient with other patients with the same condition with reference to the previous decisions. In this study, various models were generated by using mixture of experts as a classification method. Further, Model having very good accuracy of 97.37% with least minimum square error was selected for the prediction of disease. This approach can be used for easy diagnosis of hepatitis for a large number of populations by incorporating the profile of more samples in the training stage.IJCST. 06/2013; 4(2):280283.  SourceAvailable from: Shankaracharya[Show abstract] [Hide abstract]
ABSTRACT: BACKGROUND: The incidence of diabetes is increasing rapidly across the globe. India has the highest proportion of diabetic patients, earning it the doubtful distinction of the 'diabetes capital of the world'. Early detection of diabetes could help to prevent or postpone its onset by taking appropriate preventive measures, including the initiation of lifestyle changes. To date, early identification of prediabetes or type 2 diabetes has proven problematic, such that there is an urgent requirement for tools enabling easy, quick, and accurate diagnosis. AIM: To develop an easy, quick, and precise tool for diagnosing early diabetes based on machine learning algorithms. METHODS: The dataset used in this study was based on the health profiles of diabetic and nondiabetic patients from hospitals in India. A novel machine learning algorithm, termed "mixture of expert", was used for the determination of a patient's diabetic state. Out of a total of 1415 subjects, 1104 were used to train the mixture of expert system. The remaining 311 data sets were reserved for validation of the algorithm. Mixture of expert was implemented in matlab to train the data for the development of the model. The model with the minimum mean square error was selected and used for the validation of the results. RESULTS: Different combinations and numbers of hidden nodes and expectation maximization (EM) iterations were used to optimize the accuracy of the algorithm. The overall best accuracy of 99.36% was achieved with an iteration of 150 and 20 hidden nodes. Sensitivity, specificity, and total classification accuracy were calculated as 99.5%, 99.07%, and 99.36%, respectively. Furthermore, a graphical user interface was developed in java script such that the user can readily enter the variables and easily use the algorithm as a tool. CONCLUSIONS: This study describes a highly precise machine learning prediction tool for identifying prediabetic, diabetic, and nondiabetic individuals with high accuracy. The tool could be used for large scale screening in hopsitals or diabetes prevention programs.The Review of Diabetic Studies 05/2012; 1(9):5562.  SourceAvailable from: Shankaracharya[Show abstract] [Hide abstract]
ABSTRACT: The concept of classification of clinical data can be utilized in the development of an effective diagnosis system by taking the advantage of computational intelligence. Diabetes disease diagnosis via proper interpretation of the diabetes data is an important problem in neural networks. Unfortunately, although several classification studies have been carried out with significant performance, many of the current methods often fail to reach out to patients. Graphical user interfaceenabled tools need to be developed through which medical practitioners can simply enter the health profiles of their patients and receive an instant diabetes prediction with an acceptable degree of confidence. In this study, the neural network approach was used for a dataset of 768 persons from a Pima Indian population living near Phoenix, AZ. A neural network mixture of experts model was trained with these data using the expectationminimization algorithm. The mixture of experts method was used to train the algorithm with 97% accuracy. A graphical user interface was developed that would work in conjunction with the trained network to provide the output in a presentable format. This study provides a machineimplementable approach that can be used by physicians and patients to minimize the extent of error in diagnosis. The authors are hopeful that replication of results of this study in other populations may lead to improved diagnosis. Physicians can simply enter the health profile of patients and get the diagnosis for diabetes type 2.Diabetes Technology & Therapeutics 11/2011; 14(3):2516. · 2.21 Impact Factor
Page 1
R EVIEW
www.TheRDS.org 248 DOI 10.1900/ RDS.2010.7.248
DIABETIC
STUDIES
The Review of
Computational Intelligence in Early Diabetes Diagnosis:
A Review
Shankaracharya
1, Devang Odedra
and Ambarish S. Vidyarthi
1, Subir Samanta
1
2,
1 Department of Biotechnology, Birla Institute of Technology, Mesra, Ranchi 835215, India.
Birla Institute of Technology, Mesra, Ranchi 835215, India.
Address correspondence to: Shankaracharya, email: shankaracharya@bitmesra.ac.in
2 Department of Pharmaceutical Sciences,
Manuscript submitted December 8, 2010; resubmitted January 17, 2011; accepted January 19, 2011
■ Abstract
The development of an effective diabetes diagnosis system
by taking advantage of computational intelligence is re
garded as a primary goal nowadays. Many approaches based
on artificial network and machine learning algorithms have
been developed and tested against diabetes datasets, which
were mostly related to individuals of Pima Indian origin. Yet,
despite high accuracies of up to 99% in predicting the correct
diabetes diagnosis, none of these approaches have reached
clinical application so far. One reason for this failure may be
that diabetologists or clinical investigators are sparsely in
formed about, or trained in the use of, computational diag
nosis tools. Therefore, this article aims at sketching out an
outline of the wide range of options, recent developments,
and potentials in machine learning algorithms as diabetes
diagnosis tools. One focus is on supervised and unsuper
vised methods, which have made significant impacts in the
detection and diagnosis of diabetes at primary and advanced
stages. Particular attention is paid to algorithms that show
promise in improving diabetes diagnosis. A key advance has
been the development of a more indepth understanding and
theoretical analysis of critical issues related to algorithmic
construction and learning theory. These include tradeoffs
for maximizing generalization performance, use of physi
cally realistic constraints, and incorporation of prior knowl
edge and uncertainty. The review presents and explains the
most accurate algorithms, and discusses advantages and pit
falls of methodologies. This should provide a good resource
for researchers from all backgrounds interested in computa
tional intelligencebased diabetes diagnosis methods, and
allows them to extend their knowledge into this kind of re
search.
Keywords: diabetes diagnosis · computational · algorithm ·
artificial neural network · learning · logistic regression
Introduction
iabetes has been recognized as a continuing
health challenge for the twentyfirst cen
tury, both in developed and developing
countries. It is understood that diabetes preva
lence is increased because of modern lifestyles, ur
banization, and economic development [1]. It is a
global problem with devastating human, social,
and economic impact, affecting around 300 million
people worldwide [2].
Type 2 diabetes is a chronic disease that oc
curs either when the pancreas does not produce
enough insulin, or when the body cannot effec
tively use the insulin it produces. It is frequently
asymptomatic [3]. Although detection is improv
ing, the delay from disease onset to diagnosis may
exceed 10 years [4]. To diagnose diabetes, a physi
cian has to analyze many factors. Undoubtedly,
the evaluations of data obtained from patients and
expert decisions are critical for diagnosis. How
ever, factors such as lack of experience by the ex
Reprint from
The Review ofDIABETIC STUDIES
Vol 7 No 4 2010
Page 2
ComputerBased Diabetes Diagnosis
perts, or their fatigue, may lead to erroneous di
agnosis. Early intervention with lifestyle modifica
tions or pharmacotherapy has been shown to effec
tively delay or prevent type 2 diabetes and its
complications in adults [5].
The Review of DIABETIC STUDIES
Vol . 7 ⋅ No. 4 ⋅ 2010
249
www.TheRDS.org Rev Diabet Stud (2010) 7:248258
For prevention of type 2 diabetes, a compre
hensive guideline was issued specifying lifestyle
changes [6]. Various strategies have also been put
forward to reduce diabetes risk [7]. Naturally, pre
vention is preferable, but current treatment meth
ods are not yet fully adequate to reach this goal.
Hence, there is a growing need for early detection
of diabetes. To address this need, and to provide
more detailed and rapid analysis of medical data,
risk assessment tools and their various algorithms
have been widely investigated.
For early detection of diabetes, various risk
scores have been devised. A detailed survey of
these tools with their specificity and sensitivity
has been provided by Schwarz et al. in which the
authors found the Finnish Diabetes Risk Score as
the most convenient tool for early diagnosis of dia
betes [8]. However, as this method involves hu
man intervention in deciding criteria and score, it
may by be exposed to the human error. Therefore,
machine learning and statistical pattern recogni
tion has been the subject of tremendous interest in
the biomedical community as these approaches of
fer promise for improving the sensitivity and/or
specificity of detection and diagnosis of disease. At
the same time, these approaches reduce the poten
tial for human error in the decision making proc
ess [9]. In particular, further development of
methods that explicitly incorporate prior knowl
edge and uncertainty into the decisionmaking
process would be very important for diabetes de
tection. Extensive studies by many researchers
have demonstrated higher performance and accu
racy in predicting clinical outcomes of diabetes di
agnosis using neural network strategies (Table 1).
Advantages and pitfalls of using various algo
rithms in diabetes prediction are listed in Table 2.
Datasets for diabetes diagnosis
Significant work has been reported on Pima
Indian diabetes datasets (PID). These studies ap
plied different methods to the given problem, and
achieved high classification accuracies using the
dataset taken from the University of California,
Irvine (UCI) machine learning repository [10].
This database provides a well validated data re
source to explore the prediction of diabetes. The
eight variables in the dataset include:
 number of times pregnant,
 plasma glucose concentration at 2 hour in an
oral glucose tolerance test,
 diastolic blood pressure (mmHg),
 triceps skin fold thickness (mm),
 2h serum insulin (IU/ml),
 body mass index (weight in kg/height in m),
 diabetes pedigree function, and
 age (years).
While PID is one of the mostly used datasets
for prediction of type 2 diabetes, some researchers
prefer to investigate diagnosis using data from
hospitals, and to incorporate their own parameters
Abbreviations:
ADAP  adaptive learning routine
ANFIS  artificial neurofuzzy inference system
ANN  artificial neural network
ARTMAP  adaptive resonance theory mapping
ARTMAPIC  adaptive resonance theory mapping instance
counting
BPNN  backpropagation neural network
CART  classification and regression trees
CARTDB  classification and regression trees distribution
based
ESOM  evolving selforganizing maps
FIS  fuzzy inference system
GCS  growing cell structure
GDA  generalized discriminant analysis
GNG  growing neural gas
GRG2  generalized reduced gradient 2
GRNN  general regression neural network
kNN  knearest neighbor
LDA  linear discriminant analysis
LM  LevenbergMarquardt
LSSVM  least square support vector machine
LVQ  learning vector quantization
ME  mixture of experts
MEA  multimodal evolutionary algorithm
MFNNCA  modified feed forward neural network con
structive algorithm
MKS  multiple knot spline
MLP  multilayer perceptron
MLPNN  multilayer perceptron neural network
MLNN  multilayer neural networks
MME  modified mixture of experts
NFIS  neurofuzzy inference system
NG  neural gas
NHANES  National Health and Nutrition Examination
Survey
PC  principal components
PCA  principal component analysis
PID  Pima Indian diabetes dataset
PNN  probabilistic neural network
RBF  radial basis function
SOM  selforganizing map
SSVM  smooth support vector machines
SVM  support vector machine
UCI  University of California, Irvine
Page 3
250 The Review of DIABETIC STUDIES
Vol . 7 ⋅ No. 4 ⋅ 2010
Shankaracharya, Odedra, et al.
Rev Diabet Stud (2010) 7:248258 Copyright © by Lab & Life Press/ SBDR
of interest. K azemnejad et al. used the Tehran
Lipid and Glucose Study dataset which consists of
variables like age, body mass index, waisttohip
ratio, gender, history of hyperlipidemia, and his
tory of hypertension [11]. In another study con
ducted by Dey et al. on data of 530 patients from
Sikkim Manipal Institute of Medical Sciences, risk
factors such as random blood sugar test results,
fasting blood sugar test results, post plasma blood
sugar tests, age, sex, and occupation were taken
into account [12].
The third National Health and Nutrition Ex
amination Survey (NHANES III, http://www.
cdc.gov/diabetes/) dataset resulted from a survey
conducted on a US population. The eighteen vari
ables identified as important for diabetes risk pre
diction include body mass index, height, weight,
waist circumference, waisttohip ratio, age, sex,
Table 1. Artificial intelligence approaches for early diabetes detection
Algorithm
Dataset
Accuracy
(% )
80.07
81.25
98.14
80.21
77.08
68.23
81.00
80.07
78.40
74.60
73.80
77.00
71.90
72.80
75.20
75.80
77.50
74.40
94.00
76.73
93.20
78.21
89.47
84.61
74.50
76.00
76.00
91.53
97.93
99.17
Specificity
(% )
84.38

98.58















94.00




85.18



91.19
98.01
99.43
Sensitivity
(% )
74.00

96.97















93.00




83.33



92.42
97.73
98.48
Reference
MFNNCA
GRG2
ANFIS
GRNN
MLP
RBF
ARTMAPIC
MEA
ESOM
GNG
GCS
kNN
kNN
CART
MLP
LVQ
LDA
CARTDB
SVM
SSVM
MKSSSVM
GDA and LSSVM
PCAANFIS
LDAANFIS
Naive Bayes
Seminaive Bayes
C4.5
MLPNN
ME
MME
PID
PID
PID
PID
PID
PID
PID
PID
PID
PID
PID
PID
PID
PID
PID
PID
PID
PID
Questionnaire
PID
PID
PID
PID
PID
PID
PID
PID
PID
PID
PID
Kamaruzzaman et al. [51]
Shanker et al. [15]
Ubeyli [52]
Kayaer, Yildirim [38]
Kayaer, Yildirim [38]
Kayaer, Yildirim [38]
Carpenter, Markuzon [47]
Stoean et al. [53]
Deng, Kasabov [23]
Deng, Kasabov [23]
Deng, Kasabov [23]
Kordos et al. [16]
Ster, Dobnikar [17]
Ster, Dobnikar [17]
Ster, Dobnikar [17]
Ster, Dobnikar [17]
Ster, Dobnikar [17]
Shang, Breiman [54]
Barakat et al. [25]
Purnami et al. [27]
Purnami et al. [27]
Polat et al. [44]
Polat, Gunes [55]
Dogantekin et al. [44]
Friedman [56]
Friedman [56]
Friedman [56]
Ubeyli [49]
Ubeyli [49]
Ubeyli [49]
Legend: PID: Pima Indian dataset. MFNNCA: modified feed forward neural network constructive algorithm. GRG2: genera
lized reduced gradient 2. ANFIS: adaptive neurofuzzy inference system. GRNN: general regression neural network. MLP:
multilayer perceptron. RBF: radial basis function. ARTMAPIC: adaptive resonance theory mapping instance counting.
MEA: multimodal evolutionary algorithm. ESOM: evolving selforganizing maps. GNG: growing neural gas. GCS: growing
cell structure. kNN: knearestneighbor. CART: classification and regression trees. LVQ: learning vector quantization. LDA:
linear discriminant analysis. CARTDB: classification and regression trees distributionbased. SVM: support vector machine.
SSVM: smooth support vector machine. MKSSSVM: multiple knot spline smooth support vector machine. GDA: general
ized discriminant analysis. LSSVM: least square support vector machine. PCAANFIS: principal component analysis and
adaptive neurofuzzy inference system. LDAANFIS: linear discriminant analysis and adaptive network based fuzzy inference
system. C4.5: sample class 4.5 algorithm. MLPNN: multilayer perceptron neural network. ME: mixture of experts. MME:
modified mixture of experts.
Page 4
ComputerBased Diabetes Diagnosis
race/ethnicity,
blood pressure medica
tion, taking cholesterol
medication, gestational
diabetes, high blood
pressure, high choles
terol, history of diabe
tes (any blood rela
tive), history of diabe
tes (parent or sibling),
history of
(parent), history
diabetes (sibling), and
exercise [13].
The Review of DIABETIC STUDIES
Vol . 7 ⋅ No. 4 ⋅ 2010
251
www.TheRDS.org Rev Diabet Stud (2010) 7:248258
taking
diabetes
of
Data
through logistic
regression
analysis
Logistic regression
can be applied when
the data consist of a
binary response and a
set of explanatory
variables [14]. At first,
the maximum likeli
hood estimates for the
parameters of the lo
gistic regression model
are estimated using an
iteratively reweighted
least squares
rithm. Then, it is pos
sible to calculate the
predicted probability of
an individual having
diabetes by using the
following logistic func
tion:
algo
()
nnxxx
e
ββββ
θ
++++
+
=
...
22110
1
1
Here X is a vector of variables and β is the re
gression coefficient estimated by using maximum
likelihood methods. Shanker applied logistic re
gression on eight variables in PID and obtained a
significant accuracy of 79.17% [15]. Statistically
least significant (at 0.05 level) variables were de
leted sequentially in the training sample. Logistic
regression with the remaining four statistically
significant parameters, e.g. number of times preg
nant, glucose tolerance test, body mass index, and
diabetes pedigree function, resulted in an overall
classification accuracy of 80.21%. Heikes et al.
have developed a diabetes risk calculator tool
based on logistic regression function to identify
people at high risk of diabetes [13]. It was built
upon NHANES III dataset with a sensitivity of
75%.
Clustering techniques
Most quality prediction models are based on
clustering techniques that make use of kmeans,
mixtureofGaussians, selforganizing map (SOM)
and neural gas (NG) for diagnosis. According to
Table 2. Advantages and disadvantages of algorithms commonly used in diabetes prediction
Algorithm
Advatages
Disadvantages
Back propagation
LM
Better error minimization
Fast convergence rate
Slow convergence rate
Memorization effect on over
training
No specific rule to choose a kernel
that will give better classification
Low interpretability of learned in
formation, computationally expen
sive
Sensitive to dimensionality of data
SVM
Guaranteed global minimum
ANFIS
Fast convergence rate
RBF
Uses small numbers of locally
tuned units and is adaptive in na
ture
Fast convergence rate
ARTMAPIC
Tends to be conservative which
reduces sensitivity
Topology mismatch leads to poor
classification
Poor adaptability to input data
Poor response to changing inputs
SOM
Little computational and memory
requirements
Shorter learning process than SOM
Can adaptively determine the num
ber of connections
Good choice when there is no prior
knowledge of data distribution
Little computational and memory
requirements
Works best when class has Gaus
sian density
Requires only small number of
connections in neural network
Requires only small number of
connections in neural network.
Faster than ME
ESOM
GNG
kNN
Requires rigorous tuning to opti
mally fit the real world data
Less accurate with high dimen
sional data
Less accurate with small sample
size
Learns only static inputoutput
mappings (i.e. no feedback)
Learns only static inputoutput
mappings
LVQ
LDA
ME
MME
Legend: SVM: support vector machine. ANFIS: adaptive neurofuzzy inference system. RBF: radial
basis function. ARTMAPIC: adaptive resonance theory mapping instance counting. SOM: self
organizing maps. ESOM: evolving selforganizing maps. GNG: growing neural gas. kNN: k
nearestneighbor. LVQ: learning vector quantization. LDA: linear discriminant analysis. ME: mix
ture of experts. MME: modified mixture of experts.
Page 5
252 The Review of DIABETIC STUDIES
Vol . 7 ⋅ No. 4 ⋅ 2010
Shankaracharya, Odedra, et al.
Rev Diabet Stud (2010) 7:248258 Copyright © by Lab & Life Press/ SBDR
the knearest neighbor (kNN) algorithm, a new
input pattern x is assigned to the class voted by
the majority of its knearest training patterns [16].
The weight change in kNN is given by:
(
==∆
0
( )
x
)
,
()
=
otherwise ,
ifXijWX
fW
j
j
γ
where γ is the learning rate and i(x) is the win
ning node. While the accuracy of kNN on diabetes
detection problem ranges between 7178% [16, 17],
a more sensitive performance with accuracy of
92.38% was achieved with a hybrid model of kNN
and C4.5 algorithms [18, 19].
SOM is a sheetlike artificial neural network
(ANN). Cells of this ANN become specifically
tuned to input patterns [20]. In order to overcome
the topology mismatches that occur with the origi
nal SOM algorithm, and to achieve an optimal use
of the neurons, the geometry of the lattice has to
match with the data manifold. For this purpose,
several socalled growing (incremental) SOM algo
rithms have been developed. The growing neural
gas (GNG) algorithms start with two randomly
placed, connected neurons [21]. After a fixed num
ber of time steps, the neuron i with the largest ac
cumulated error is determined, and a new neuron
inserted between i and one of its neighbors. It does
not require predetermination of the neuron quan
tity or topology of structure to be used. It starts
with a minimal neuron structure that is incre
mented during training until it reaches a maxi
mum number limit for clusters defined by the
user.
The growing cell structure (GCS) algorithm
assumes a fixed dimensionality for the lattice [22].
It is well suited for generating a dimensionality
reducing mapping from the input space to the lat
tice space. Deng and Kasabov applied GNG and
GCS algorithms to the diabetes diagnosis problem,
and reported accuracies of 74.6% and 73.8%, re
spectively [23]. Both GNG and GCS need to calcu
late local resources for prototypes, which intro
duces extra computational effort and reduces their
efficiency. Deng and K asabov proposed the evolv
ing selforganizing maps (ESOM) network struc
ture, which is similar to that of GNG [21]. When
applied to diabetes diagnosis, they obtained 78.4%
classification accuracy using ESOM.
Support vector machine (SVM)
Support vector machine (SVM) operates by
finding a linear hyperplane that separates the
positive and negative examples with a maximum
interclass distance [24]. We can define zi as an in
dicator variable which specifies whether a data
vector xi is in class diabetics or nondiabetics (e.g.,
zi = 1 if xi is in the diabetic class and zi = 1 if xi is
in the nondiabetic class). The distance of a hyper
plane w to a (transformed) data vector y is defined
as  f (y)/ w. Together with the fact that the
separating hyperplane ensures zi f(yi) ≥ 1 for all n
data vectors i, we can express the condition on the
margin m as:
( )
m where ,
≥
w
The goal of SVM training is to find the weight
vector w that maximizes the margin m. Barakat et
al. employed SVM to process the inputs, and ex
tracted the rules using an electic approach [25].
This approach was then used to predict the diag
nosis of diabetes using a questionnaire based on
demographic, historic, and anthropometric meas
ures. The authors achieved a prediction accuracy
of 94%.
A cascade learning system based on general
ized discriminant analysis (GDA) and least square
support vector machine (LSSVM) has been pro
posed for early diagnosis of Pima Indian diabetes
disease [26]. The accuracy reported in this study
was 78.21% with 10fold crossvalidation. Purnami
et al. applied smooth support vector machines
(SSVM) to the diabetes detection problem [27].
SSVM, developed by Lee et al., is an extension to
SVM in which smoothing function is applied to
solve the problem [28]. With SSVM, the investiga
tors achieved a 76.73% accuracy. To improve effi
ciency, they proposed a new multiple knot spline
(MKS) smoothing function for SSVM. Replacing
the defaultplus function of SSVM by MKS, they
enhanced the automated diagnosis performance of
SSVM with an accuracy of 93.2%.
ni
fzi
,...,1
=
y
Neural networks
Multilayer neural networks
Multilayer neural networks (MLNN) are com
posed of one or more hidden layers between input
and output (Figure 1) [29]. In the training phase,
the training data is fed through the input layer.
The data is propagated from the hidden layer to
the output layer (Figure 2), which is called for
ward pass. During this phase, each node in the
hidden layer gets input from all the input layer
Page 6
ComputerBased Diabetes Diagnosis
The Review of DIABETIC STUDIES
Vol . 7 ⋅ No. 4 ⋅ 2010
253
www.TheRDS.org Rev Diabet Stud (2010) 7:248258
nodes, which are then multiplied by the randomly
assigned weights before summing up. Similarly,
the output layer node receives inputs from all
nodes of the hidden layer, which are then multi
plied by the randomly assigned weights and
summed up. This forms the output of the output
layer.
The input to each hidden layer is calculated
by:
input *
∑
=
i
wy
where wi is the weight for neuron i. The output
of the hidden layer is calculated by using an acti
vation function. The activation function acts as a
squashing function, such that the output of a neu
ron in a neural network is between certain values
(usually 0 and 1 for sigmoid, or 1 and 1 for hyper
bolic tangent). Common activation functions used
in diabetes diagnosis are the sigmoid (a) and hy
perbolic tangent (b) function:
1
a)
+
with sigmoid range = [0,1], and hyperbolic
range = [1,1]. Error rates are calculated as fol
lows:
( )( )
[
xfxf1* Error =
value
( )
x
( )
x
xx
xx
x
ee
ee
f
e
f




b)
1
+
==
]
( )
x
[]
f  valuetarget*
Backpropagation neural networks
The backpropagation neural network (BPNN)
algorithm is widely recognized as a powerful tool
for training of the MLNN. In this algorithm, er
rors are backpropagated to the hidden layers,
weights are reassigned, and the process continues
until the error rate is a minimum. The new
weights are calculated based on the following
equation:
weight (new) Weight
=
where η is the learning rate. However, since it
applies the steepest descent method to update the
weights, it suffers from a slow convergence rate,
and often yields suboptimal solutions [30, 31].
J aafar et al. used the back propagation neural
network algorithm for diagnosing diabetes [32].
The inputs to the system were glucose tolerance
test, diastolic blood pressure, triceps skin fold
thickness, serum insulin, body mass index, diabe
tes pedigree function, number of times pregnant,
and age. BPNN was used to predict the glucose
level [33], and also to train and test its perform
ance using diabetes patients [12].
Although the BPNN algorithm is widely used,
one major drawback is that it requires a complete
set of input data. However, most diabetes datasets
are often incomplete in the one respect or another.
Back propagation algorithm cannot interpret the
missing values (if any) which may prevent the
identification of factors leading to rare outputs. To
overcome this problem, J ayalakshmi and Santha
kumaran proposed a new approach to deal with
the missing values [34]. They achieved an accu
racy of 99.9% by replacing the missing values with
its mean, and then normalizing the data with a
principal component analysis (PCA) technique
[35]. PCA is an extraction method aimed at de
scribing the data variance by constructing a set of
new orthogonal features called principal compo
nents (PCs). The PCs are a lin
ear combination of the data
variables that are mutually or
thogonal. Every new PC de
scribes a part of the data vari
ance not explained by compo
nents used previously. Due to
this fact, a few first PCs are
usually enough to represent
the data variance well.
It was reported that the
LevenbergMarquardt (LM) al
)( * error * xf
η+
Activation
function
∑wi
Input
Output
Figure 1. Architecture of a single neuron.
Input
(parameter
values)
Output
Figure 2. Multilayer neural network with 3 neuron layers.
Page 7
254 The Review of DIABETIC STUDIES
Vol . 7 ⋅ No. 4 ⋅ 2010
Shankaracharya, Odedra, et al.
Rev Diabet Stud (2010) 7:248258 Copyright © by Lab & Life Press/ SBDR
gorithm [36] provides generally faster convergence
and better estimation results than other training
algorithms [37]. However, this method can cause a
memorization effect when overtraining occurs. If a
neural network starts to memorize the training
set, its generalization starts to decrease, and its
performance may not be improved for untrained
test sets. Kayaer and Yildirim used the LM algo
rithm on a Pima Indian dataset, and achieved an
accuracy of 77.08% [38], which was lower than
other algorithms. Temurtas et al. trained the neu
ral network optimally with a probabilistic neural
network (PNN) along with a LM algorithm [39,
40]. They achieved an 82.37% accuracy with this
approach.
Radial basis function (RBF)
In neural networks, radial basis functions
(RBFs) are used as a replacement for the sigmoi
dal hidden layer transfer function in multilayer
perceptrons (MLP) [41]. The only parameters ad
justed in the learning process are the linear map
ping from the hidden layer to the output layer.
Hence, RBF networks have the advantage of not
suffering from local minima.
RBF shows good performance in regression
applications where the input space dimension is
relatively small. However, in prediction problems
like diabetes diagnosis, only 68.23% efficiency has
been reported, which is far less than other algo
rithms. RBF networks have the disadvantage of
requiring good coverage of the input space by ra
dial basis functions. Determination of RBF centers
is heavily dependent on the distribution of the in
put data without reference to the prediction task.
General regression neural network (GRNN)
The general regression neural network
(GRNN) is related to the radial basis function
network and is based on a standard statistical
technique called Kernel regression [42]. It ap
proximates any arbitrary function between input
and output vectors, and draws the function esti
mate directly from the training data. It does not
require an iterative training procedure, as in
MLP. For an input estimator ‘x’, corresponding to
diabetes risk factor variables, GRNN produces an
output estimator ‘y’ which minimizes the estima
tion error. GRNN works on following formula:
(
)
x yfxyE,(
∫
∞
dyyxf dyy),(/)

∫
∞
∞∞
=
where E[yx] is the expected value of output y,
given the input vector x, and f(x, y) the joint prob
ability density function of x and y.
GRNNs produce a realvalued prediction be
tween 0 and 1. A cutoff value decides the criteria
to identify positive prediction. The best result
achieved by GRNN on PID is 80.21% using 0.5 as
cutoff value for the decision [38].
Neurofuzzy inference systems (NFIS)
A neurofuzzy network is a fuzzy inference
system in an artificial neural network [43]. De
pending on the fuzzy inference system (FIS) type,
there are several layers that simulate the proc
esses involved in a fuzzy inference like fuzzifica
tion, inference, aggregation, and defuzzification.
Embedding a FIS in the general structure of an
artificial neural network (ANN) has the benefit of
using ANN training methods to find the parame
ters of a fuzzy system. Linear discriminant analy
sis (LDA) is used to separate the two types of fea
ture variables in a given dataset [44]. Dogantekin
et al. used LDA along with artificial neuro FIS
(ANFIS) for the detection of diabetes [45]. In this
method, LDA is used to separate feature variables
between healthy and diabetes data. In the second
phase, both the healthy and diabetes features ob
tained in the first phase are given to inputs of the
ANFIS classifier. They achieved an 84.61% accu
racy with this approach.
Smith et al. used the PID data set to evaluate
the perceptronlike adaptive learning routine
(ADAP), and achieved a prediction accuracy of
76% [46]. The performance of fuzzy adaptive reso
nance theory mapping (ARTMAP) on the same da
tabase was 66% [47]. ARTMAP is a supervised
learning algorithm for input binary vectors. How
ever, the ARTMAP algorithm required fewer rules
and was comparatively faster. Carpenter and
Markuzon have presented an instance counting
algorithm (ARTMAPIC) and obtained an 81% ac
curacy against the test set [47].
E xpert systems
In real world problems like diabetes detection,
a simple classifier is too weak for accurate predic
tion. The use of expert systems and different arti
ficial intelligence techniques for classification sys
tems in medical diagnosis is increasing gradually.
Mixture of experts and modified mixture of ex
perts have been successfully implemented to the
problem of diabetes diagnosis prediction.
Page 8
ComputerBased Diabetes Diagnosis
Mixture of experts
The Review of DIABETIC STUDIES
Vol . 7 ⋅ No. 4 ⋅ 2010
255
www.TheRDS.org Rev Diabet Stud (2010) 7:248258
The new supervised learning algorithm called
mixture of experts (ME) was proposed by J acobs et
al. [48]. This algorithm divides a learning task
into appropriate subtasks, each of which can be
solved by simple expert network. The global out
put of the ME system is derived as a convex com
bination of the outputs from a set of N experts, in
which the overall predictive performance of the
system is generally superior to any of the individ
ual experts.
ME architecture is composed of several expert
networks and a gating network (Figure 3). The
gating network produces a scalar output from a
vector input X. The gating network operates on a
generalized linear
function where the
output for i
input
variable is given by:
e
vxg
,
th
()
∑
=
k
=
n
i
k
i
e
1
ξ
ξ
where ξi = Vi
and Vi is the weight
vector. Each
network produces an
output vector for an
input vector based on
the following general
ized linear equation:
T x,
expert
)()(xWfx
ii
=σ
where Wi is a weight matrix. The final output
of ME is the sum of multiplications of the outputs
from gating and expert networks:
∑
=
k
1
Ubeyli presented an approach to test the per
formance of ME on PID with a classification accu
racy of 97.93% [49], which was better than conven
tional MLNN. Moreover, the computational time
required for classification using ME was compara
tively small.
=
n
kk
xvxgx)(),()(
σσ
Modified mixture of experts (MME)
Ubeyli [49] employed a new, fast, and effective
modified mixture of experts (MME) approach pro
posed by Chen [50] to further improve the classifi
cation accuracy of ME.
The MME architecture is composed of an as
sembly of N expert networks and a gatebank
(Figure 4). For k different features, expert net
works are divided into k groups, each comprising
of N expert networks. Similarly, the gatebank is
composed of k gating networks. The resultant out
put of the gatebank is a convex weighted sum of
outputs produced by all the gating networks. Fi
nally, the overall output of MME is obtained by
linear combination of outputs of all N expert net
works weighted by the output of the gatebank.
)(x
σ
Expert
network 1
Expert
network 2
Expert
network n
Gating
network
x
xx
Figure 3. General architecture of mixture of experts.
σ
Expert
network (1,1)
Expert
network (1,n1)
Expert
network (k,1)
Gating
network 1
Gating
network 2
Gate bank
x1
xk
x1
Expert
network (k,nk)
xk
Figure 4. General architecture of modified mixture of experts.
Page 9
256 The Review of DIABETIC STUDIES
Vol . 7 ⋅ No. 4 ⋅ 2010
Shankaracharya, Odedra, et al.
Rev Diabet Stud (2010) 7:248258 Copyright © by Lab & Life Press/ SBDR
Ubeyli applied the MME algorithm to the dia
betes diagnosis problem and achieved an accuracy
of 99.17% [49]. Apart from outperforming all other
algorithms, the computational time required for
classification was very small.
Conclusions
Despite of the rapid development of theories
for computational intelligence, application to dia
betes diagnosis remains a challenge. This is due to
specific problems of data use. These problems
arise when statistical models of data are unknown
or timedependent, or when the parameters of the
learning system need to be updated incrementally,
while only a partial glimpse of incoming data is
available. Based on the promising outcomes of
studies applying computational algorithms to the
problem of diabetes diagnosis, it is clear that a
more sophisticated risk score could be developed.
This would significantly decrease healthcare costs
via early prediction and diagnosis of type 2 diabe
tes.
Some algorithms work better on the diabetes
diagnosis problem than others. It will be impor
tant to compare outcomes further to find the most
reliable algorithm for clinical application. Neural
network methodology has outperformed classical
statistical methods in cases where input variables
are interrelated. Because clinical measurements
are usually derived from multiple interrelated sys
tems, it is evident that neural networks might be
more accurate than classical methods in multi
variate analysis of clinical data.
■ References
1. Lieberman LS. Dietary, evolutionary, and modernizing
influences on the prevalence of type 2 diabetes. Annu Rev
Nutr 2003. 23:345377.
2. Z hang Y , Dall T , Mann SE, Chen Y , Martin J,
Moore V, Baldwin A, Reidel VA, Quick WW. The
economic costs of undiagnosed diabetes. Popul Health Manag
2009. 12(2):95101.
3. Jackson DM, Wills R, Davies J, Meadows K, Singh
BM, Wise PH. Public awareness of the symptoms of diabe
tes mellitus. Diabet Med 1991. 8:971972.
4. Harris MI, Klein R, Wellborn T A, Knuiman MW.
Onset of NIDDM occurs at least 47 yr before clinical diag
nosis. Diabetes Care 1992. 15:815819.
5. Knowler WC, BarrettConnor E, Fowler SE, Ham
man RF, Lachin JM, Walker EA, Nathan DM. R educ
tion in the incidence of type 2 diabetes with lifestyle inter
vention or metformin. N Engl J Med 2002. 346(6):393403.
6. Paulweber B, Valensi P, Lindström J Lalic NM,
Greaves CJ, McKee M, KissimovaSkarbek K, Liatis
S, Cosson E, Szendroedi J, et al. A European evidence
Trained models of diabetes risk factors should
be incorporated into easytouse software solutions
such that medical practitioners, who are not ex
perts in artificial intelligence and computational
techniques, may apply them easily. For this pur
pose, graphical user interfaceenabled tools need
to be developed by which medical practitioners can
simply enter health profiles of their patients and
receive an instant diabetes prediction with an ac
ceptable degree of confidence. If the ANNbased
prediction approach shows improved medical di
agnosis, then it may become more widely accepted
as a means to assist patient care in more hospitals
and clinics.
Though the PID dataset provides a well vali
dated data for predicting diabetes diagnosis, it is
possible that models trained on such a dataset
may not perform equally well on profiles of pa
tients from other ethnic group. Therefore, it is rec
ommended that models of choice must be trained
on a dataset that closely represents patient pro
files of medical practitioners within specific geo
graphic regions.
Acknowledgements: The authors are grateful to the
SubDistributed Information Center (BTISnet SubDIC),
Department of Biotechnology (No. BT/BI/04/065/04),
New Delhi, India, and to the Department of Biotechnol
ogy, Birla Institute of Technology, Mesra, Ranchi, for
providing access to software and infrastructure facility
for the present study.
Disclosures (conflict of interests statement): The
authors report no conflict of interests.
based guideline for the prevention of type 2 diabetes. Horm
Metab Res 2010. 42(Suppl 1):S3S36.
7. Lindstrom J, Neumann A, Sheppard KE, Gilis
Januszewska A, Greaves CJ, Handke U, Pajunen P,
Puhl S, Pölönen A, Rissanen A, et al. Take action to
prevent diabetes  the IMAGE toolkit for the prevention of
type 2 diabetes in Europe. Horm Metab Res 2010. 42 (Suppl
1):S37S55.
8. Schwarz PE, Li J, Lindstorm J, T uomilehto J. Tools
for predicting the risk of type 2 diabetes in daily practice.
Horm Metab Res 2009. 41(2):8697.
9. Sajda P. Machine learning for detection and diagnosis of
disease. Annu Rev Biom ed Eng 2006. 8:537565.
10. Frank A, Asuncion A. UCI machine learning repository.
Irvine, CA, University of California, School of Information
and Computer Science, 2010.
11. Kazemnejad A, Batvandi Z , Faradmal J. Comparison
of artificial neural network and binary logistic regression for
determination of impaired glucose tolerance/diabetes. East
Mediterr Health J 2010. 16(6):615620.
12. Dey R, Bajpai V, Gandhi G, et al. Application of artifi
cial neural network technique for diagnosing diabetes melli
Page 10
ComputerBased Diabetes Diagnosis
tus. IEEE Third International Conference on Industrial and
Information Systems, Kharagpur, India, 2008, 14.
13. Heikes KE, Eddy DM, Arondekar B, Schlessinger L.
Diabetes risk calculator: a simple tool for detecting undiag
nosed diabetes and prediabetes. Diabetes Care 2008.
31:10401045.
14. Abbot RD. Logistic regression in survival analysis. Am J
Epidem iol 1985. 121(3):465471.
15. Shanker MS. Using neural networks to predict the onset of
diabetes mellitus. J Chem Inf Com put Sci 1996. 36:3541.
16. Kordos M, Blachnik M, Strzempa D. Do we need
whatever more than kNN? In: Proceedings of the 10th In
ternational Conference on Artificial Intelligence and Soft
Computing, Part I, SpringerVerlag Berlin, 2010, 414421.
17. Ster B, Dobnikar A. Neural networks in medical diagno
sis: comparison with other methods. In: Proceedings of the
International Conference on Engineering Applications with
Neural Networks, London, 1996, 427430.
18. Patil BM, Joshi RC, T oshniwal D. Hybrid prediction
model for type2 diabetic patients. Exp Syst Appl 2010.
37:81028108.
19. Jantan H, Hamdan AR, Othman Z A. Human talent
prediction in HR M using C4.5 classification algorithm. Int J
Com p Sci Engin 2010. 2:25262534.
20. Kohonen T . Selforganizing formation of topologically
correct feature maps. Biol Cybern 1982. 43:5969.
21. Fritzke B. A growing neural gas network learns topologies.
Adv Neural Inf Process Syst 1995. 7:625632.
22. Fritzke B. Growing cell structures  a selforganizing net
work for unsupervised and supervised learning. Neural Netw
1994. 7:14411460.
23. Deng D, Kasabov N. Online pattern analysis by evolving
selforganizing maps. Proceedings of the 5th Biannual Con
ference on Aritificial Neural Networks and Expert Systems
(ANNES), Dunedin, 2001, 4651.
24. Ali S, Abraham A. An empirical comparison of kernel
selection for support vector machines. 2nd International
Conference on Hybrid Intelligent Systems, Soft Computing
systems: Design, Management and Applications, IOS Press,
The Netherlands, 2002, 321330.
25. Barakat NH, Bradley AP, Barakat MB. Intelligible sup
port vector machines for diagnosis of diabetes mellitus. Trans
Inf Technol Biom ed 2010. 14:11141120.
26. Gunes PK, Aslan A. A cascade learning system for classifi
cation of diabetes disease: generalized discriminant analysis
and least square support vector machine. Exp Syst Appl
2008. 34:214221.
27. Purnami SW, Embong A, Z ain JM. A New smooth
support vector machine and its applications in diabetes dis
ease diagnosis. J Com p Sci 2009. 5:10061011.
28. Lee Y J, Mangasarian OL. A smooth support vector ma
chine. J Com p Optim Appl 2001. 20:522.
29. Basheer IA, Hajmeer M. Artificial neural networks: fun
damentals, computing, design, and application. J Microbiol
Meth 2000. 43:331.
30. Brent RP. Fast training algorithms for multilayer neural
nets. IEEE Trans Neural Netw 1991. 2(3):346354.
31. Gori M, T esi A. On the problem of local minima in back
propagation. IEEE Trans Pattern Anal Mach Intell 1992.
14:7685.
32. Jaafar, SF, Ali DM. Diabetes mellitus forecast using artifi
cial neural networks. Asian conference of paramedical re
search proceedings, Kuala Lumpur, Malaysia, 2005, 57.
The Review of DIABETIC STUDIES
Vol . 7 ⋅ No. 4 ⋅ 2010
257
www.TheRDS.org Rev Diabet Stud (2010) 7:248258
33. Eskaf EK, Badawi O, Ritchings T . Predicting blood
glucose levels in diabetes using feature extraction and artifi
cial neural networks. Third ICTTA conference, Damascus,
2008, 16.
34. Jayalakshmi T , Santhakumaran A. A novel classification
method for classification of diabetes mellitus using artificial
neural networks. International Conference on Data Storage
and Data Engineering, Bangalore, 2010, 159163.
35. Chen LH, Chang S. An adaptive learning algorithm for
principal component analysis. IEEE Trans Neural Netw 1995.
6:12551263.
36. Hagan MT , Menhaj M. Training feed forward networks
with the Marquardt algorithm. IEEE Trans Neural Netw
1994. 5:989993.
37. Gulbag A, T emurtas F. A study on quantitative classifica
tion of binary gas mixture using neural networks and adap
tive neuro fuzzy inference systems. Sens Actuators B Chem
2006. 115:252262.
38. Kayaer K, Y ildirim T . Medical diagnosis on Pima Indian
diabetes using general regression neural networks. Proceed
ings of the international conference on artificial neural net
works and neural information processing, Istanbul, 2003,
181184.
39. T emurtas H, Y umusak N, T emurtas F. A comparative
study on diabetes disease diagnosis using neural networks.
Expert Syst Appl 2009. 36:86108615.
40. Specht DF. Probabilistic neural networks. Neural Netw
1990. 3:109118.
41. Buhmann, Martin D. R adial basis functions: theory and
implementations. Cambridge University Press, 2003. pp 54
78.
42. Hagan MT , Demuth HB, Beale M. Neural network de
sign. PWS Publishing Company, 1996. pp 102108.
43. Bart K. Neural networks and fuzzy systems: a dynamical
systems approach to machine intelligence. Prentice Hall,
1992. pp 3649.
44. Polat K, Gunes S, Arslan A. A cascade learning system
for classification of diabetes disease: Generalized discriminant
analysis and least square support vector machine. Exp Syst
Appl 2008. 34:482487.
45. Dogantekin E, Dogantekin A, Avci D, et al. An intel
ligent diagnosis system for diabetes on Linear Discriminant
Analysis and Adaptive Network Based Fuzzy Inference Sys
tem: LDAANFIS. Digit Signal Process 2009. 20:12481255.
46. Smith JW, Everhart JE, Dickson WC, et al. Using the
ADAP learning algorithm to forecast the onset of diabetes
mellitus. Proceedings of the Symposium on Computer Ap
plications and Medical Care, IEEE Computer Society Press,
1988, 261265.
47. Carpenter GA, Markuzon N. AR TMAPIC and medical
diagnosis: instance counting and inconsistent cases. Neural
Netw 1998. 11:323336.
48. Jacobs RA, Jordan MI, Nowlan SJ, et al. Adaptive mix
tures of local experts. Neural Com put 1991. 3:7987.
49. Ubeyli ED. Modified mixture of experts for diabetes diag
nosis. J Med Syst 2009. 33:299305.
50. Chen K. A connectionist method for pattern classification
with diverse features. Pattern Recognit Lett 1998. 19:7545
558.
51. Kamruzzaman SM, Hasan AR, Siddiquee AB, et al.
Medical diagnosis using neural network. Proceedings of the
3rd International Conference on Electrical and Computer
Engineering, Dhaka, Bangladesh, 2004. 537540.
Page 11
258 The Review of DIABETIC STUDIES
Vol . 7 ⋅ No. 4 ⋅ 2010
Shankaracharya, Odedra, et al.
Rev Diabet Stud (2010) 7:248258 Copyright © by Lab & Life Press/ SBDR
52. Ubeyli ED. Automatic diagnosis of diabetes using adaptive
neurofuzzy inference systems. Expert Syst 2010. 27:259
266.
53. Stoean C, Stoean R, Preuss M, et al. Diabetes diagnosis
through the means of a multimodal evolutionary algorithm.
Proceedings of the 1st East European Conference on Health
Care Modelling and Computation, Craiova, R omania,
Craiova, Medical University Press, 2005, 277289.
54. Shang N, Breiman L. Distribution based trees are more
accurate. Proceedings of ICONIP 96, Springer, Hong Kong,
1996, 133138.
55. Polat K, Gunes S. An expert system approach based on
principal component analysis and adaptive neurofuzzy in
ference system to diagnosis of diabetes disease. Digit Sign
Proc 2007. 17:702710.
56. Friedman N, Geiger D, Goldszmit M. Bayesian net
works classifiers. Mach Learn 1997. 29:131163.