Diabetes is the disease that causes severe harm to human
beings, which elevated sugar levels at a high rate [1]. It
causes severe continuing problems such as heart disease,
kidney disorders, ulcers, and spoil eyes. At present, the
kinds of diabetes, namely insipidus, and Mellitus. Insipidus
is due to turn out inadequate insulin. In Mellitus, the cells
dose not reacts to the creation of insulin. At present, the
diabetic patient uses a Fingerstick device with lab tests for
testing the elevated sugar level. However, this method is
more painful and it consumes more time to detect the
elevated sugar level of an individual. In order to defeat this
drawback of the existing model, Neural Network-based
classifiers [2] is introduced in the literature
A Multi-Layer Perception is often used for the
prediction. Multi-Layer uses supervised learning and back
propagation for training process. In the neural network, it
has layers and nonlinear activation, which distinguish the
linear perception. It can also distinguish whether it is
linearly or non-linearly independent. It also focused, mainly
on computer techniques to perform clinical diagnoses and
the prediction with suggestions for the treatment.
Several kinds of research shown attention for
diabetes prediction using machine learning and deep
learning methods. The following reviews were studied in the
literature. Thirugnana et al., [3] proposed improved diabetes
prediction using fuzzy neural networks. Afsaneh Morteza et
al., presented a neural network predicted albuminuria in
type II diabetes compared the condition logistic regression
[4].
Kevin et al.,[5] suggested a Machine Learning method
for diabetes treatment to Predict Blood Glucose Levels. The
proposed model has outperformed diabetes experts at
estimate blood glucose rates and it can forecast 23% of
hypoglycaemic cases 30 minutes. Sneha Joshi et al., [6]
introduced MATLAB built-in forecasting method that can
determine whether a in dividable is diabetes. The GUI is
designed to make application user friendly so that even in
the absence of a doctor, patients can get test result from
assistants. The BPNN results used for predicting diabetes is
76%, which indicates the progress in the previous research.
Zahed Soltani and Ahmad Jafarian [7] proposed a neural
network method for identifying the diabetes. The maximum
training accuracy is 89.56% and testing accuracy 81.49% is
obtained for the proposed framework.
Takoua Hamdi et.al.,[8] used an Neural Network for
predict insipidus diabetes in the blood sugar levels
Experimental tests showed that it was used for detect
hyperglycemia or hypoglycemia quarter-hour well in
advance. The key concept of ANN is to use the previous N
steps to forecast subsequent steps. The Predetermined
calculation is then used as reference with the previous (N-1)
measurements to estimate consequential meaning and so
forth. The calculation of the consequential values as am
benefit is cumulative, elastic and nonlinear.
Surajini et al., [9] proposed a prediction model of diabetes
with the support of the Probabilistic Neural Network. He
trained the prediction model using the Back propagation
algorithm. PNN achieved the prediction model with minimal
error and it shows the diabetic prediction.
Quanzou et.al. [10] used a decision tree and random
Performance Analysis of Neural Network Based
Classifiers for the Prediction of Diabetes
J. PRADEEP M. HARIKRISHNAN, K.VIJAYAKUMAR
Department of Electronics & Communication Engineering Sri Manakula Vinayagar Engineering College
Puducherry, INDIA
Abstract— Diabetes is the most harmful diseases to consider in recent years since it causes severe damage to
human beings in the form of elevated sugar levels. In a recent survey, it was projected that over 385 million
public were affected in the entire world. Several investigators were conducted various experiments for
prediction of diabetes using various classification techniques. This paper deals with a neural classifier based
prediction system to recognize diabetes. Two learning algorithms namely, Levenberg Marquardt back
propagation (LM), and gradient descent with variable learning rate are is investigated for different architecture
and the best architecture with good accuracy was identified. The data are together from the Government Hospital
of Pondicherry and it is formed as a database. Totally, datasets of 500 have been together, out of which 350
datasets as training sets for training process and 150 datasets as testing sets for the testing process. The
recognition accuracy is obtained. For comparison, k-Nearest Neigourhood and the K- nearest neighbor and
Radial Basis Function (RBF) network are also implemented and it is trained and tested with the same datasets.
The result shows that Neural Network outperforms well with other classifiers.
Keywords: Neural Network, Gradient descent with variable rate Sigmoid Activation Function, Prediction,
Diabetes, k- Nearest Neigourhood.
Received: May 15, 2021. Revised: February 18, 2022. Accepted: March 20, 2022. Published: April 26, 2022.
1. Introduction
MOLECULAR SCIENCES AND APPLICATIONS
DOI: 10.37394/232023.2022.2.4
J. Pradeep M. Harikrishnan, K. Vijayakumar
E-ISSN: 2732-9992
24
Volume 2, 2022
forest to predict diabetes mellitus. They randomly selected
68994 healthy people and diabetic patient's data as the
training set. In this study, the proposed utilized principal
component analysis and minimum redundancy maximum
relevance to reducing the dimensionality.
Suresh Kumar et.al.,[11] implemented data mining
strategies to determine the type of diabetes and its intensity
degree for each individual from the data gathered including
clustering and grouping. A base k-means algorithm is used
to segment the whole dataset, classifiers the risk level of
each patient as mild, moderate and server.
Vrushali Balpande et al., [12] discussed the detailed
review of existing data mining methods used for the
prediction of diabetes. The K- Nearest Neighbor Algorithm,
Bayesian Classifier, Naïve Bayesian Classifier methods are
used for the prediction of diabetes, which gives patient's
condition of Normal, Pre-diabetes, and diabetes.
Suyash Srivastava et al.,[13] proposed and presented a
diabetes prediction with the help of Neural Network method
and it archives 92% accuracy for predicting diabetes.
The above-discussed kinds of literature are the inspiration to
initiate this paperwork. From the studies of various related
and existing models, the idea for creating a prediction model
of diabetes with the help of an artificial neural network is
achieved with valuable knowledge. This paper proposes a
diabetes prediction with a neural network classifier. The
datasets are collected and created as a database. The datasets
are used for the trained and tested process. The result of this
prediction method is obtained and it is evaluated with the
existing systems.
This paper is organized as follows. Section II describes
the pre-processing of the data and the benefits of the
proposed system, overcoming the disadvantages of the
previous system. Section III confers the results of the
prediction method. Section IV describes the comparisons of
the proposed prediction model. Section V ends with a
conclusion and Section VI provides the acknowledgment
respectively.
TABLE.1 SAMPLE OF COLLECTED DATASET
Sl .No
NAME
GENDER
RBS
FBS
PPBS
UREA
CREATIVE
HBAIC
OUTCOME
1
Rajendiren
1
120
89
108
28
0.8
5.2
0
2
Soundarajan
1
174
92
126
34
2.4
5.5
0
3
Devanathan
1
145
150
281
77
2.6
7.2
1
4
Krishna
1
234
150
276
70
1
8.3
1
5
Velu
1
138
82
114
24
0.4
5.4
0
6
Rajaramam
1
210
84
135
38
1.6
7.3
1
7
Nedunchezian
1
113
84
154
20
0.9
6.2
0
8
Sanjeevi
0
84
82
172
20
0.9
5.8
0
In the proposed model, real-time diabetes dataset is
collected from the Government Hospital of Pondicherry.
The data consists of medical details of 500 instances, out of
which 350 datasets are used as training sets for the training
process and 150 datasets are used as testing sets for the
testing process. The collected datasets consist of 10
attributes namely Random Blood Sugar, Fasting Blood
sugar, Pre/Post Pradinal Blood Sugar, Urea, Creatinine,
Glycated haemoglobin, Age, Gender, and Outcome. The
value of Outcome '0' is considered as non-diabetic and the
value of Outcome '1' is considered as diabetic. The collected
dataset samples are shown in below Table.1
For enhanced perceptive about the dataset and to obtain
a high-quality result with a low error rate as possible from
the prediction model, the data pre-processing and data
visualizations are done. The data pre-processing are used
on the dataset is listed below.
A Neural Network (NN) technique shows the potential
solution for the classifying for the prediction of diabetes. The
features are the input to the different classifiers. The ability of
the classification is determined from architecture of the
network and the rule of learning. The architectures used in this
paper are feed-forward, radial basis function and nearest
neighborhood architecture. The prediction of diabetes is
evaluated using the NN based classifier technique. Totally
500 datasets were collected, out of which 350 datasets are
used as training sets for the training process, and 250 datasets
are used as testing sets for the testing process. The prediction
model of diabetes has been implemented using Matlab
software. The feed- forward back propagation classifier is
introduced and investigated. For comparison of accuracy, the
K- nearest neighbor and Radial Basis Function (RBF) network
are also designed with the help of a real-time diabetes dataset.
The prediction models have been designed using all the
various system as mentioned below.
2. Proposed Model
2.1 Database Description
2.2 Data Preprocessing
2.3 Neural Network Based Classifiers
MOLECULAR SCIENCES AND APPLICATIONS
DOI: 10.37394/232023.2022.2.4
J. Pradeep M. Harikrishnan, K. Vijayakumar
E-ISSN: 2732-9992
25
Volume 2, 2022
In order to obtain the maximum recognition accuracy,
different neural network architecture with two learning
algorithms namely, Levenberg Marquardt back propagation
(LM), and gradient descent with variable learning rate are
investigated. It is observed from Table.2, that the hidden
layer with 65 neurons gives the result with maximum
accuracy. Thus, the two hidden layers with 65 neurons in
each are used. For the testing process, the testing dataset is
given to the trained neural network Architecture and it has
perform is obtained. From the result, the Recognition
accuracy is also determined.
TABLE.2 RECOGNITION PERFORMANCE OF NEURAL NETWORK
ARCHIECTURE FOR DIFFERRENT LEARNING ALGORITHM
Sl.No
NN
Architecture
Training
Algorithm
Recognition
accuracy (%)
1
8:30:2
LM
76
GD with VLR
79
2
8:30:30:2
LM
82
GD with VLR
84
3
8:40:2
LM
79
GD with VLR
82
4
8:40:40:2
LM
65
GD with VLR
75
5
8:65:2
LM
75
GD with VLR
86
6
8:65:65:2
LM
82
GD with VLR
96.47
7
8:75:2
LM
79
GD with VLR
82
8
8:75:75:2
LM
69
GD with VLR
65
The k-nearest neighbor algorithm is a technique used for
classifying the neighborhood in the feature space [15]. The
training stage consists of storing only data from the
function vectors with class labels. The same features are
computed from the test data at the classification level. To
get closest neighbours, the Euclidean distance between the
test data and the entire cumulative vector is measured
input together, and the distances obtained are listed in
ascending order the smallest distance is taken.
The k-nearest neighbor algorithm is applied in this
paper. The Simulation results are obtained for the 3rd nearest
neighbor which yields better accuracy and the results are
tabulated in the subsequent section.
Radial Basis Function (RBF) network has better quality
and it is used in wide variety of functions [16]. RBF
network has Gaussian function as nonlinearity for the
transmission elements of hidden layers. The Gaussian
function only refers to a specific area where the Gaussian is
located in the input space. The key to successful
implementation of these networks is to find appropriate centres
for the Gussian functions. The basic architecture for an RBF is
a 3-layer network that is investigated and the best architecture
is obtained. The hidden layer neurons are 100 in the RBF
network. For the classifying the prediction of the diabetes, two
neurons are used in the output layer. The feed- forward neural
network, RBF and the k-nearest neighbor network classifier are
used for investigation and the performance study is carried out
in the next section.
A real-time database is generated and the data’s are collected
from diabetic patients in the Government Hospital of
Pondicherry. Totally 500 datasets were collected, out of which
350 datasets are used as training sets for the training process,
and 250 datasets are used as testing sets for the testing process.
From the proposed 8x65x65x2 neural network Architecture,
the output is obtained and the accuracy is 96.47%.
Fig.3 Performance Illustration of Gradient Descent
Optimization
Fig. 3 shows that accuracy is steadily increased, when the
epoch rate increases. It states that the epoch and the accuracy
are directly proportional to each other. The accuracy
achieved for the proposed prediction model is shown in
Table.3 and it is compared with the other two
classifiers.Table.3 shows the reduction of error rate concerning
the epochs for the gradient descent optimization. The error rate
was obtained for every 1000 iterations in the training process
and it is shown as Table.3.
For the performance comparison, the k-nearest classifier
and the logistic regression classifier is used. The training and
testing process is carried with the help of the created same real-
time database. After the training, the classifiers tested with the
testing samples, and the results are obtained. The result obtained
for the k-nearest and logistic regression classifiers are
illustrated in the below table. This reveals from the table that
the k-Nearest Neighbor classification offers 85.65 %accuracy
and 89 % accuracy for the RBF network in classification.
Table.3 shows that the average accuracy obtained for the
neural network classifier of architecture 8:65:65:2 with
2.3.1. Feed Forward Neural Network
2.3.2. k- Nearest Neighbor Network
2.3.3. Radial Basis Function Network
3. Result and Discussion
3.1 Performance Comparison of the Classifier
MOLECULAR SCIENCES AND APPLICATIONS
DOI: 10.37394/232023.2022.2.4
J. Pradeep M. Harikrishnan, K. Vijayakumar
E-ISSN: 2732-9992
26
Volume 2, 2022
Gradient descent and variable rate accuracy is 96.47 % and
the proposed neural classifier leads 10% of accuracy when it
compared with other classifiers. Therefore, the neural
network classifier outperforms well compared with the other
two classifiers in terms of accuracy. Moreover, the proposed
system is more suitable and efficient for real-time
applications.
TABLE.3 DIABETES ACCURACY WITHDIFFERENT NN
CLASSIFERS
S. No
Dataset
Classifiers
Accuracy
1
Real time dataset
(AGE, GENDER, RBS,
FBS, PPBS, UREA,
CREATINE, HBA1C)
Neural
Network
96.47%
2
Real time dataset
(AGE, GENDER, RBS,
FBS, PPBS, UREA,
CREATINE, HBA1C)
Radial basis
Function
network
89%
3
Real time dataset
(AGE, GENDER, RBS,
FBS, PPBS, UREA,
CREATINE, HBA1C)
k-Nearest
Neighbour
85.65%
The prediction of diabetes using a Neural Network
classifier is proposed. For the proposed system, the data
collected from the Government Hospital of Pondicherry and
the database is created using the data collected. Totally 500
datasets were collected, out of which 350 datasets are used
as training sets for the training process and1 50 datasets are
used as the testing set for the testing process. After the
training process, the classifier is tested. The NN classifier
with 65 neurons in the two hidden layers gives the result
with maximum accuracy finally; the NN is obtained with
maximum accuracy. The obtained result of the neural
network classifier is evaluated with the otherclassifiers.
The best performance among these classifiers is found and
the proposed neural network classifier with 96.47% of
accuracy is obtained. It shows that the performance of the
proposed NN classifier outperforms well with the remaining
classifier with respect to accuracy.
Our thanks to Dr. M. Sivakamy (General Duty Medical
Officer) belonging to Government Hospital of Pondicherry
for helping us by providing the relevant and most required
real time diabetes dataset is used for train and test in the
proposed system.
[1]
Ishwarya, R., Gayathri, P., Jaisankar, N., 2013. A
Method for Classification Using Machine Learning
Technique for Diabetes. International Journal of
Engineering and Technology (IJET) 5, 2903–2908
[2]
Temurtas, H., Yumusak, N., Temurtas, F., "A
comparative study on diabetes disease diagnosis using
neural networks", Expert Syst, Vol. 36, pp. 8610–15
2009.
[3]
Thirugnanam, Mythili, et al., "Improving the Prediction
Rate of Diabetes Diagnosis Using Fuzzy, Neural
Network, Case Based (FNC) Approach."Procedia
Engineering, Vol.38, pp. 1709-118, 201
[4]
Morteza, Afsaneh, et al., "Inconsistency in albuminuria
predictors in type 2 diabetes: a comparison between neural
network and conditional logistic regression", Translational
Research, Vol.161, No.5, pp. 397-405, 2013.
[5]
Kevin phlis, Razvan Bunescu, and Cindy Marling ―A
Machine Learning Approach in Predicting Blood Glucose
Levels for Diabetes Management‖, Modern Artificial
Intelligence for Health Analytics, pp: 35-37, 2014.
[6]
Sneha Joshi and MeghaBorse, ―Detection and Prediction
of Diabetes Mellitus Using Back- Propagation Neural
Network‖, Proceedings International Conference on
Microelectronics and Telecommunication Engineering,
pp: 110-113, 2016.
[7]
ZahedSoltani and Ahmad Jafarian, ―A New Artificial
Neural Networks Approach for Diagnosing Diabetes
Disease Type II‖, International Journal of Advanced
Computer Science and Applications, Vol. 7, No. 6, pp: 85-
94, 2016.
[8]
Takoua Hamdi and Jaouher Ben Ali, ―Artificial Neural
Network for Blood Glucose Level Prediction‖,
Proceedings International Conference on Smart,
Monitored and Controlled Cities, pp: 17-19, 2017.
[9]
P. Sujarani and Dr. K. Kalaiselvi, ―Prediction of Diabetes
Using Artificial Neural Networks: A Review,
―International Journal of Computer Applications &
Information Technology, Vol. 10, No. 4, pp: 67-72, 2018.
[10]
Quanzou, Kaiyang Quo and Yamei Luo, ―Predicting
Diabetes Mellitus with Machine Learning Techniques‖,
Frontiers in Genetics, Vol.9, No.10 pp:1-9,2018.
[11]
P.Suresh Kumar and V. Umatejaswi, ―Diagnosing
Diabetes using Data Mining Techniques‖, International
Journal of Scientific and Research Publications, Vol.7,
No.6, pp: 705-707, 2017.
[12]
Vrushali Balpande and Rakhi Wajgi, ―Review on
Prediction of Diabetes using Data Mining Technique‖,
International Journal of Research and Scientific
Innovation, Vol.4, No.1, pp: 43-46, January 2017.
[13]
Parastoo Rahimloo and Ahmad Jafarian, ―Prediction of
Diabetes by Using Artificial Neural Network‖ Logistic
Regression Statistical Model and Combination of Them,
Vol. 85, pp: 1148-1164, 2016.
4. Conclusion
5. Acknowledgement
References
MOLECULAR SCIENCES AND APPLICATIONS
DOI: 10.37394/232023.2022.2.4
J. Pradeep M. Harikrishnan, K. Vijayakumar
E-ISSN: 2732-9992
27
Volume 2, 2022
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the Creative
Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en_US
[14]
Islam MR, Kamal ARM, Sultana N, Islam R, Moni MA
et al, ―Detecting depression using k-nearest neighbors
(KNN) classification technique.‖, International
conference on computer, communication, chemical,
material and electronic engineering (IC4ME2). IEEE, pp
1–4. 2018.
[15]
Amal H. Khaleel et al, ―A Weighted Voting of K- Nearest
Neighbor Algorithm for Diabetes Mellitus,‖ International
Journal of Computer Science and Mobile Computing,
Vol.6 Issue.1, January- 2017, pg. 43-51
[16]
Cheruku R., Edla D., and Kuppili V., ―Diabetes
Classification using Radial Basis Function Network by
Combining Cluster Validity Index and BAT Optimization
with Novel Fitness Function,‖ International Journal of
Computational Intelligence Systems, vol. 10, no. 1, pp.
247-265, 2017.
MOLECULAR SCIENCES AND APPLICATIONS
DOI: 10.37394/232023.2022.2.4
J. Pradeep M. Harikrishnan, K. Vijayakumar
E-ISSN: 2732-9992
28
Volume 2, 2022