See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/338581650

Diabetic Prediction System Using Data Mining

Conference Paper · September 2016

CITATIONS READS

2 922

6 authors, including:

Nandana Pathirage Kasun Kosala Jinasena

General Sir John Kotelawala Defence University University of Sri Jayewardenepura
7 PUBLICATIONS   3 CITATIONS    9 PUBLICATIONS   5 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Blockchain-based Secure Voting System View project

Blockchain-based Secure, Reliable, and Distributed Voting System for Decision Making in Government Policies and Projects View project

All content following this page was uploaded by Kasun Kosala Jinasena on 14 January 2020.

The user has requested enhancement of the downloaded file.

 Proceedings in Computing, 9th International Research Conference-KDU, Sri Lanka 2016

Diabetic Prediction System Using Data Mining

L.H.S De Silva1#, Nandana Pathirage2 and T.M.K.K Jinasena3

1,2Department of Computer Science, General Sir John Kotelawala Defence University, Sri Lanka
3Department of Computer Science, Faculty of Applied Sciences, University of Sri Jayewardenepura
# L.H.S De Silva; <[email protected]>
Type 1- This results when the body fails to produce insulin.
Abstract— Diabetes is one of deadliest diseases in the This form of diabetes was previously referred to as
world. As per the existing system in Sri Lanka, patients
have to visit a diagnostic center, consult their doctor and "insulin-dependent diabetes mellitus" (IDDM) diabetes".
wait for a day or more to get their result. Moreover, every
time they want to get their diagnosis report, they have to Type 2- This results because of insulin resistance, a
waste their money in vain. But with the rise of Machine condition in which cells fail to use insulin properly,
Learning approaches, we have been able to find a solution sometimes also with an absolute insulin deficiency. This
to this problem using data mining. Data mining is one of was earlier referred to as non-insulin-dependent diabetes
the key areas of Machine learning. It plays a significant mellitus (NIDDM) or "adult-onset diabetes".
role in diabetes research because It has the ability to
extract hidden knowledge from a huge amount of Type 3- This is the 3rd main form of diabetes and it called
diabetes related data. The aim of this research is to as Gestational diabetes. It occurs when pregnant women
develop a system which can predict whether the patient without an earlier diagnosis of diabetes develop a high
has diabetes or not. Furthermore, predicting the disease blood glucose level.
early leads to treatment of the patients before it becomes
critical. This research has focused on developing a system As we can realize from these facts, problems related to
based on three classification methods namely, Decision diabetes are many and quite costly. It is a very serious
Tree, Naïve Bayes and Support Vector Machine algorithms. disease because, if not treated properly and on time, it
Currently, the models give accuracies of 84.6667%, could lead to very serious complications, may be the
76.6667%, and 77.3333% for Decision Tree, Naïve Bayes, death of the patient. This makes diabetes one of the main
and SMO Support Vector Machine respectively. These priorities in medical science research. However, the
results have been verified using Receiver Operating country has not been utilized the strength of computer
Characteristic curves in a cost-sensitive manner. The technology to reduce the risk of diabetes yet. With the
developed ensemble method uses votes given by the other rise of the new knowledge, scientists have discovered
algorithms to produce the final result. This voting various kinds of new technologies that we could use to
mechanism eliminates the algorithm dependent solve this problem. One of the most popular technologies
misclassifications. Results show a significant improvement today Data Mining. It is capable of predicting the risk level
of accuracy of ensemble method compares to other of a patient with significantly higher accuracy by
methods. extracting hidden patterns from historical medical records.
This will help us to give timely treatment for patients by
Keywords— Data Mining, Diabetes, Machine Learning diagnosing disease early before it goes to a critical stage.

I. INTRODUCTION Data mining or the Knowledge Discovery in Data(KDD) is

According to the World Health Organization (WHO), about the process of exploration huge amount of data in order
347 million people worldwide have diabetes. By the year to discover new patterns or the trends. It is far beyond
of 2030, it has predicted to become the 7th leading simple analytical techniques. Data mining uses many
reason for deaths in the world(“WHO | Diabetes,” n.d.). In sophisticated machine-learning algorithms to discover
2012, diabetes was the direct reason for more than 1.5 hidden patterns in a large data set automatically. Later,
million deaths in the world. The Diabetes Association of such identified patterns can be used to predict future
Sri Lanka (DASL) statistics reveals that by 2016, there are events. To do a proper diagnose on diabetic, doctors need
nearly four million diabetics in Sri Lanka(Thilakarathna, to gather a huge amount of data about the patient.
n.d.). According to DASL, almost one-fifth of the world’s Obviously, it is harder for a human to analysis such a data
people with diabetes lives in the South-East Asia Region. volume manually. Using data-mining techniques for this is
More and more young people were being suffered by the one of the best ways to enhance the accuracy and the
disease and still the number of people affected by efficiency of such process. Through both predictive
diabetes is increasing every day. Mainly there are three (classification) and descriptive (clustering and
types of diabetes in the world(Ross, 2010). association),data-mining techniques can be applied for
this. The present study is focused on developing a diabetic

66
 Proceedings in Computing, 9th International Research Conference-KDU, Sri Lanka 2016

prediction system based on classification(predictive) data algorithms for diabetes dataset. They have used 70:30
mining methods, namely Decision Tree algorithm, Naïve percentage split and 10-fold cross validation techniques to
Bayes algorithm and SMO Support Vector Machine build their model. When using Decision Tree, they have
algorithms. got 76.9565% accuracy and for the Naïve Bayes, they have
got 79.5652% accuracy. These results assure that
II. LITERATURE REVIEW classification data mining methods are better for
A systematic review of research findings and applications prediction of diabetes. But there is no evidence that they
of data mining techniques in the field of diabetes has have developed a system that can predict risk level of a
been done in order to identify the present status of the patient in real time. They just have analysed those two
research question, the research gap, and the alternatives. algorithms using classifier models. But we have developed
Here our main objectives were to identify research goals, a system using classification data mining techniques which
diabetes types, data-mining methods, data-mining can diagnose the diabetic risk level of a patient.
software’s and their technologies, data sets and outcomes.
Based on that, we have developed a novel approach to Thirumal and Nagarajan have used Fuzzy, Neural Network,
predict diabetes using data mining technologies. Case-Based (FNC) approach to predict rate of Diabetes.
(Thirugnanam et al., 2012). They present a novel approach
for the computational intelligence and knowledge
Huge amounts of data produced by healthcare
engineering techniques as neural network (N), fuzzy logic
transactions are too complex and voluminous to be
(F), and case-based reasoning(C) as an individual approach
processed and analysed by conventional methods
(FCN). At the final prediction stage, they have applied the
(Cunningham and Holmes, 1999). However, the data
rule-based algorithm to the values obtained from the
mining is capable of extracting hidden knowledge from
initial stage. They position as the benefits of applying that
complex data repositories such as- research reports, flow
is the accuracy of predication rate is higher than other
charts, evidence tables and medical reports, and
diabetes prediction algorithms. But using Neural network
transform into useful information for decision making.
is somewhat slow because it does require more time to
train the network. But this diabetic disease will be critical
Breault and colleagues applied a “Classification and in some stages. So we have to find a quicker solution for
regression tree (CART) using the CART data-mining” this. That is one of the reasons we moved to classification
software on data of 15,902 diabetes patients and based approach. Because there if we know the problem,
detected that most important variable related to bad we can get results quickly.
glycemic control (HbA1c >9.5) is age (Marinov et al., 2011).
Patients below the threshold of 65.6 years old have worse A Case Study on “UTILIZATION OF DATA MINING
glycemic control than older people, which was very TECHNIQUES FOR DIAGNOSIS OF DIABETES MELLITUS”
surprising to clinicians. Using this knowledge, they have have done by Coimbatore Institute of Engineering and
targeted the specific age groups that are more likely to Technology (Thirumal and Nagarajan, 2006). This research
have poor glycemic control. However, they have found has based on old diabetes patients. They have found that
age is the most valuable variable for glycemic control risk of diabetes will be low when patients are often given
using the CART algorithm. There may be other important assessment and treatment plans that suit their wants and
variables too. Thus, more methods have to be used to lifestyle. Straight forward awareness measures like low
discover those. sugar diet, correct diet will avoid fatness. The Goal of this
study was to urge best algorithms that describe given
Myiaki and colleagues (Mehrpoor et al., 2014) conducted knowledge in multiple aspects. In this paper, several data
a study to find the best predictors of diabetes vascular mining algorithms have been used for test the dataset.
complications using CART on data from 165 type 2 Naïve Bayes, Decision trees, k nearest neighbor and SVM
diabetes mellitus (T2DM) patients. The authors found that are discussed and tested with Pima Indian polygenic
age (cut-off: 65.4 years) was the best predictor, and disease dataset. Accuracy of these models are needed to
depending on the age, the second best predictor was be evaluated before it is being used. If the available data
body weight (cut-off: 53.9kg) for the group above 65.4 or are limited, it makes estimating accuracy a difficult task.
systolic blood pressure for the group below 65.4. Here Table 1 shows accuracies of the algorithms given by the
they have gone more steps further. confusion matrix.
Aiswarya and colleagues have done a research on Table 1 accuracies of algorithms
“DIAGNOSIS OF DIABETES USING CLASSIFICATION MINING
Algori- Accuracy TP FP Preci- Recall
TECHNIQUES”(Iyer et al., 2015). According them, diabetes sion
thm (%)
has affected over 246 million individuals worldwide and
most of them were women. According to the WHO report, Naïve bayes 77.8646 0.83 0.317 0.83 0.83
by 2025, this variety is anticipated to rise to over 380
million. This paper focused on analyzing the patterns C4.5 78.2552 0.864 0.369 0.814 0.864
based Decision Tree and Naïve Bayes data mining

67
 Proceedings in Computing, 9th International Research Conference-KDU, Sri Lanka 2016

SVM 77.474 0.775 0.309 0.77 0.775 model. It’s a java based algorithm, it works as follows. In
order to classify a new item, it first creates a decision tree
kNN 77.7344 0.892 0.437 0.792 0.892 based on the attribute values of the available training
data set. Every node of the decision tree is generated by
calculating the highest information gain for all attributes.
From the experiments, it's complete that kNN provides
If any attribute gives an unambiguous end result (explicit
lower accuracy when putting next to alternative
classification of class attribute), the branch of that
algorithms because it stores training examples and delays
attribute will be terminated and then target value is
the processing until a new instance is classified. The speed
assigned to it. We have used 12-fold cross validation
of the algorithm is also important when we decide the
technique to build the model using this algorithm. It’s
efficiency of an algorithm. Which tells that classification
simply as follows.
algorithms are better that these KNN algorithms when to
compare with this problem domain. Here also they only
have done an analysis using weka data mining tool  Break data into 12 sets of size n/12.
whether what algorithm gives better results. But this is  Train on 11 datasets and test on 1.
also a case study that focused on finding best data mining  Repeat 12 times and take a mean accuracy.
algorithms for diabetic related data. So considering these
facts and previous diabetic related research, we have In 12-fold cross-validation, the original sample is randomly
developed a system which gives a real-time prediction partitioned into 12 equal sized subsamples. Of
about whether the patient has diabetes or not. the 12 subsamples, a single subsample is retained as the
validation data for test the model, and the remaining (12−
III. METHODOLOGY 1) subsamples are used as training data.
In previous studies, they have used only single approach
to identify the disease. But we have combined three B. Naïve Bayes Algorithm
classification algorithms through a voting mechanism to Naïve Bayes classifier algorithm has been created based
increase the accuracy level of the model. So if one on the Bayes rule of conditional probability. It uses all the
algorithm does not predict it correctly, it doesn’t affect to attributes contained in the data, and then analyses them
the final prediction because the system considers the individually as though they are equally important and
predictions of other two algorithms too. It gives the independent of each other. There are various data mining
majorities decision. Thus ensures more accuracy than a existing solutions exists to find relations between the
single algorithm. diseases and their symptoms also the medications for
them. But these algorithms have their own limitations like
A. Decision Tree J48 Algorithm binning of the continuous arguments, numerous
A Decision tree is basically a tree structure(Han and iterations, high computational time, etc. But Naïve Bayes
Kamber, 2006), which has the form of a flowchart. It can classifier affords fast, highly scalable model building and
be used as a method for classification and prediction with scoring. The build process for Naive Bayes is parallelized.
a representation using nodes and internodes. Root and It overcomes various limitations like the omission of
internal nodes are the test cases. Leaf nodes considered complex iterative estimations of the parameter because it
as class variables. Figure 1 shows a sample decision tree can be applied to a large dataset in real time. The formula
structure. used for that algorithm is simply showed here.

Here we have used 70:30 percentage split technique to

build the model using Naïve Bayes algorithm. This means
70 percent of the data set have been used to train the
data and other 30 percent of the data set have been used
to test the model.
Figure 1: sample decision tree structure
C. SMO (Sequential Minimal Optimization)
Among classification data mining methods, decision tree
This algorithm is commonly used for solving the quadratic
algorithm provides powerful techniques for prediction.
programming problems that arise during the training of
Among ID3, C4.5, C5, J48 and CHAIAD decision tree SVM (Support Vector Machines). SMO uses heuristics to
algorithms, we have selected J48 algorithm to develop our partition the training problem into smaller problems that
68
 Proceedings in Computing, 9th International Research Conference-KDU, Sri Lanka 2016

can be solved analytically. SMO algorithm it replaces all  Total Number of Instances
missing values and transforms nominal attributes into
binary ones. It also normalizes all attributes by default E. Procedure:
which helps to speed up the training process. We have
 Load previous data sets to the system (768 test
used 70:30 percentage split technique to train and test
cases).
the data set using this model. Here we are not only
 Data pre-processing has done using integrating
considering the accuracy but it should have the ability to
WEKA tool(Witten et al., 2011). Following
handle missing values well. This algorithm does that very
operations are performed on the dataset after
accurately because it uses heuristics to partition the
that.
training problem into smaller problems. That’s the main
a. Replace Missing Values
reason we have selected this algorithm.
b. Normalization of values.

III. EXPERIMENTAL DESIGN  Then User inputs data to the system in order to
This section explains the overall design of the system and diagnose whether he has the disease or not.
what is the process it has followed in order to get the  Build a model using J48 Decision Tree Algorithm
prediction. and train the data set.
 Build a model using Naïve Bayes Algorithm and
train the data set.
D. Dataset Used:
 Build a model using SMO Support Vector
The data set we have used is a benchmarked dataset Machine Algorithm and train the data set.
which can be used for comparing the accuracy and the  Test the data set using these three models.
efficiency of our model. Data has been obtained from  Get the evaluation results.
Pima Indians Diabetes Database, National Institute of  Finally, get the predicted voting from all
Diabetes and Digestive and Kidney Diseases. classifiers and gives the diagnostic result.

Number of Instances: 600 Following diagram shows the overall procedure of

Number of Attributes: 8 + (1 class attribute). this system.
For Each Attribute: (all numeric-valued).

1) Inputs: .
 Number of times pregnant
 Plasma glucose concentration 2 hours in an oral
glucose tolerance test
 Diastolic blood pressure (mm Hg)
 Triceps skin fold thickness (mm)
 2-Hour serum insulin (mu U/ml)
 Body mass index (weight in kg/ (height in m) ^2)
 Diabetes pedigree function
 Age (years)
 Class variable (0 or 1)

Missing Attribute Values: None

Relabelled values in attribute 'class'

 From: 0 To: tested negative
 From: 1 To: tested positive

2) Outputs:
 Predicted Results (Diagnosed State)
 Evaluation Results
 Correctly Classified Instances
 Incorrectly Classified Instances
 Kappa statistic
 Mean absolute error
 Root mean squared error
 Relative absolute error
 Root relative squared error
Figure 2: Overview of procedure
69
 Proceedings in Computing, 9th International Research Conference-KDU, Sri Lanka 2016

V. RESULTS AND DISCUSSION

Number of times pregnant 8
Although it could produce different results for local data
Plasma glucose concentration 2 hours in an oral 183 set due to varies differences such as DNA, foods they are
glucose tolerance test eating or may be because of the life style of local people.
We have chosen a benchmark diabetic dataset in order to
Diastolic blood pressure (mm Hg) 64 compare our results with the other results.
Triceps skin fold thickness (mm) 0 Following are input readings from diabetic affected
2-Hour serum insulin (mu U/ml) person.
0
You can see that this automated system also has
Body mass index (weight in kg/ (height in m) ^2) 23. diagnosed it correctly. You can see the diagnosed results
3 from Figure 4 user interface also. It gives final result as
Diabetes pedigree function 0.6 this:
72
Age (years) 32 Voted final results: Patient has diabetes.
Voted final result is the final results generated after voting
Class variable (0 or 1) ? of all algorithms. If two or more models gives same
diagnostic results it will identify that as the final result.
That way the system ensures it always produces a correct
diagnose. Because even though one algorithm blindly
gives a wrong prediction, other two algorithms also check
Followings are some of the user interfaces of developed their results and prevents providing wrong diagnostic
system results.

We have produced evaluation results of this system after

testing it using 600 diabetic records. Obtained results are
as follows
Table 2: Evaluation results of three classifier models
Evaluation J48 Decision Naïve Bayes SMO
Results Tree Support
Vector
Machine
Predicted tested_posit tested_positive tested_posit
results ive ive
Correctly 508(84.6667 460(76.6667%) 464(77.3333
Classified %) %)
Instances
Incorrectly 92(15.3333 140(23.3333%) 136(22.6667
Classified %) %)
Instances
Kappa 0.6343 0.4718 0.4593
statistic
Figure 3 : User interface for the input attributes Mean 0.2225 0.2824 0.2267
absolute
error
Root mean 0.3335 0.4156 0.4761
squared
error
Relative 49.0928% 62.3167% 50.022%
absolute
error
Root 70.0783% 86.6965% 100.0392
relative
squared
error
Total 600 600 600
Number of
Instances
These results show that most higher accurate results are
Figure 4: prediction and evaluation result interface
given by the J48 Decision Tree and SVM Support vector
70
 Proceedings in Computing, 9th International Research Conference-KDU, Sri Lanka 2016

machine algorithms. J48 has more than 84% accuracy and X axis: True positive rate
other two also have more than 76% accuracy. So it has Y axis: False positive rate
more accuracy when comparing with most of other
systems that have developed. Furthermore, because the You can see in all the ROC curves, that they are skewed to
voting process that we have used in this system, it the True positive side. Which proves that our accuracies
ensures that it gives higher accurate results than when of the all three classifier models are high. Which conclude
considering accuracies of the classifiers separately. that our all three classifiers are appropriate ones and also
Because it first considers all the diagnosed results of three have good accuracies.
classifiers and gives the final prediction results after that.

F. Confusion Matrix
Confusion Matrix has the information about predicted
and actual classification results of the classifiers. The
Performance of the classifiers has been evaluated using
Predicted

Positive Negative

Actual True TP FN

Actual False FP TN

this matrix.
Figure 5: ROC Curve of J48 Decision Tree model
Table 3: Confusion Matrix

TP- Positive tuples that were correctly labelled by the

classifier
TN-True Negative tuples that were correctly labelled by
the classifier.
FP- False Positive tuples that were incorrectly labelled as
positive.
FN- False Negative tuples that were mislabelled as
negative.

Accuracy = (TP+TN)/ (TP+FP+TN+FN)

But if your data set is heavily skewed, there is a problem

in the confusion matrix. E.g. 90% of its instances are
positive instances and 10% of instances are negative
instances, in that case, any classifier you are producing Figure 6: ROC Curve of Naïve Bayes mode
that will be very bad. Because it doesn’t know the
difference between positive and negative instances. It just
blindly tells everything as positive. So even though it has
90% of accuracy, it’s just wrong because that is not Figure 7: ROC Curve of Naïve Bayes mode
reflecting the appropriate evaluation of your classifier.
Because of that, we have used ROC (Receiver Operating
Characteristic) curves in order to justify it has the correct
accuracy. ROC curves are visualization tools which you can
tell easily that in a cost sensitive manner whether your
classifier is really appropriate or not. Because of that, we
have generated ROC curves for each of our classifier
models to test the accuracies in cost sensitive manner.

Here in Figure 6, 7 and 8 you can see the ROC curves

generated for our J48 Decision Tree algorithm, Naïve
Bayes and SVM Support Vector algorithm. Here X and Y
axis are representing followings. Figure 8: ROC Curve of SMO model

71
 Proceedings in Computing, 9th International Research Conference-KDU, Sri Lanka 2016

Furthermore, we have planned to gather more data from

different locales over the country and develop more
precise and general prescient model because increasing
the data set also cause to increase the accuracy of the
results. This data set that we have used is a benchmark
dataset, which is gathered from other countries. It will be
very helpful to build a precise model when we can use a
data set from our country because the DNA patterns are
different for every region.

V. CONCLUSION
Although all the methods have given more than 75%
accuracy, the Decision Tree and the SMO Support Vector
Machine give more accurate results than the Naïve Bayes
algorithm. However, the ensemble method gives the
highest accuracy from all due to the voting process of all
the algorithms.

References
Cunningham, S.J., Holmes, G., 1999. Developing innovative
applications in agriculture using data mining, in: The
Proceedings of the Southeast Asia Regional Computer
Confederation Conference.
Han, J., Kamber, M., 2006. Data mining: concepts and techniques,
2nd ed. ed, The Morgan Kaufmann series in data
management systems. Elsevier ; Morgan Kaufmann,
Amsterdam ; Boston : San Francisco, CA.
Iyer, A., S, J., Sumbaly, R., 2015. Diagnosis of Diabetes Using
Classification Mining Techniques. Int. J. Data Min.
Knowl. Manag. Process 5, 01–14.
doi:10.5121/ijdkp.2015.5101
Marinov, M., Mosa, A.S.M., Yoo, I., Boren, S.A., 2011. Data-
mining technologies for diabetes: a systematic review.
J. Diabetes Sci. Technol. 5, 1549–1556.
Mehrpoor, G., Azimzadeh, M.M., Monfared, A., 2014. Data
Mining: A Novel Outlook to Explore Knowledge in
Health and Medical Sciences. Int. J. Travel Med. Glob.
Health 2, 87–90.
Ross, T.J., 2010. Fuzzy logic with engineering applications, 3. ed.
ed. Wiley, Chichester.
SureshikaThilakarathna, n.d. Nearly four Million Diabetics in Sri
Lanka [WWW Document]. URL
http://www.news.lk/news/business/item/5701-nearly-
four-million-diabetics-in-sri-lanka (accessed 1.26.16).
Thirugnanam, M., Kumar, P., Srivatsan, S.V., Nerlesh, C.R., 2012.
Improving the Prediction Rate of Diabetes Diagnosis
Using Fuzzy, Neural Network, Case Based (FNC)
Approach. Procedia Eng. 38, 1709–1718.
doi:10.1016/j.proeng.2012.06.208
Thirumal, P.C., Nagarajan, N., 2006. UTILIZATION OF DATA
MINING TECHNIQUES FOR DIAGNOSIS OF DIABETES
MELLITUS-A CASE STUDY.
WHO | Diabetes [WWW Document], n.d. URL
http://www.who.int/mediacentre/factsheets/fs312/en
/ (accessed 5.22.16).
Witten, I.H., Frank, E., Hall, M.A., 2011. Data mining: practical
machine learning tools and techniques, 3rd ed. ed,
Morgan Kaufmann series in data management systems.
Morgan Kaufmann, Burlington, MA.

72

View publication stats