DENGUE DETECTION AND PREDICTION SYSTEM USING DATA MINING WITH FREQUENCY ANALYSIS

Natarajan Meghanathan et al. (Eds) : ACITY, VLSI, AIAA, CNDC - 2016
pp. 53–67, 2016. © CS & IT-CSCP 2016 DOI : 10.5121/csit.2016.60906
DENGUE DETECTION AND PREDICTION
SYSTEM USING DATA MINING WITH
FREQUENCY ANALYSIS
Nandini. V1
and Sriranjitha. R2
and Yazhini. T. P3
Department of Computer Science and Engineering,
SSN College of Engineering, Kalavakkam
1
nandini.vishwa94@gmail.com
2
sriranjitha.raghuraman@gmail.com
3
tp.yazhini@gmail.com
ABSTRACT
Clinical documents are a repository of information about patients' conditions. However, this
wealth of data is not properly tapped by the existing analysis tools. Dengue is one of the most
widespread water borne diseases known today. Every year, dengue has been threatening lives
the world over. Systems already developed have concentrated on extracting disorder mentions
using dictionary look-up, or supervised learning methods. This project aims at performing
Named Entity Recognition to extract disorder mentions, time expressions and other relevant
features from clinical data. These can be used to build a model, which can in turn be used to
predict the presence or absence of the disease, dengue. Further, we perform a frequency
analysis which correlates the occurrence of dengue and the manifestation of its symptoms over
the months. The system produces appreciable accuracy and serves as a valuable tool for
medical experts.
KEYWORDS
Named Entity Recognition, Part of Speech tagging, Classification, Prediction, SMO
1. INTRODUCTION
Mining unstructured data is a very pressing issue in the field of text mining. This is especially a
major subject in the area of medicine. Clinical decisions are often made based on doctor's
intuition and experience rather than on the knowledge-rich data hidden in the database. Dengue is
attracting global concern from researchers and health care professionals over the world. Statistics
reveal that almost 25,000 people die from dengue every year. Timely detection of symptoms
associated with this deadly disease, and apt prevention measures will go a long way in bringing
down its effects on the world populace. Hence, we need a system that will first learn the
characteristics of people with dengue, and use this knowledge to predict dengue in new patients.
Over the years several NLP systems like cTakes, MetaMap, etc. [3] [2] were used to extract
medical concepts from clinical text. They focused on rule based, medical knowledge driven
dictionary lookup approaches. While some researchers have contributed to disease prediction,

54 Computer Science & Information Technology (CS & IT)
they have concentrated primarily around heart attacks [6] [7] [10] [11] [12]. Inspiration drawn
from such work, combined with the increasing rate of dengue cases around the world motivates
us to develop a system to model, predict and analyze dengue instances. The inability to extract
useful information from clinical documents may hamper the health care experts’ efforts from
understanding the relationship between the prevalence of diseases and the associated factors. The
frequency of diseases can also be allied with its time frame. This is especially true in the case of
water-borne and air-borne diseases. Addressing this task will be a major help to doctors, experts
and patients. This relation will enable health care connoisseurs to take preventive measures and
reduce the prevalence of these diseases.
2. PROBLEM STATEMENT
The knowledge available in medical repositories is effectively mined and analyzed using the
proposed system. The input is a set of annotated discharge summaries containing data pertaining
to the disease dengue. Disorder names are extracted from these summaries and looked up in a
summarized UMLS (Unified Medical Language System). The output produced in this step is
supplied to classifiers which then perform detection and prediction. Further, frequency
correlation is performed with the time frame.
3. OVERVIEW OF PROPOSED SYSTEM
Figure 1. Overview of the system
The annotated discharge summaries are supplied to feature extraction algorithms and the
extracted features are in turn used to generate a feature vector. This is supplied as input to a
classification algorithm and a prediction model is developed. The model generated can then be
used to detect and predict the presence of dengue. Finally, a correlation analysis is performed to
determine how the disease is spread over the months.
4. RELATED WORK
N. Aditya Sundar et al [5] use regular factors contributing to heart diseases, including age, sex,
blood sugar and blood pressure, to predict the likelihood of a patient getting a heart disease. Data
mining techniques of Naïve Bayesian classification and WAC (Weighted Associative Classifier)

Computer Science & Information Technology (CS & IT) 55
are used to train a model on existing data. Subsequently, patients and nurses can use this model to
supply features and get a prediction on a possible heart attack. Oona Frunza et al [6] present a
machine learning approach that identifies semantic relations between treatments and diseases and
focuses on three semantic relations (prevent, cure and side effect). Later, features were extracted
from unstructured clinical text, and were used to classify the relationship between diseases and
associated treatments. Jyoti Soni et al [7] have developed a predictive data mining algorithm to
predict the presence of heart disease. Fifteen attributes were selected to perform the prediction
and Decision Tree was found to produce the best results. Classification based on clustering
algorithms was found to not perform well. [12] Proposes a Medical Diagnosis System for
predicting the risk of cardiovascular disease. It uses genetic algorithm to determine the weights
for a neural network. This feed forward neural network is subsequently used for classification and
prediction of heart diseases. A data set of 303 instances of heart disease with 14 attributes each is
used for training the system. Devendra Ratnaparkhi, Tushar Mahajan and Vishal Jadhav in their
paper [10] describe a system for prediction of heart disease using Naïve Bayes algorithms. They
further propose a web interface to help healthcare practitioners assess the possibility of a heart
problem in patients. A similar attempt proposes a heart disease prediction system [11] using
Decision Tree and Naïve Bayes and its implementation in .NET platform by I.S.Jenzi et al in
their paper. Some data mining techniques used for modeling and prediction of dengue include
SVM [13], decision tree [14] and neural network [15].
5. SYSTEM DESIGN
The system design is divided into 2 parts.
5.1. Feature Vector Generation
Figure 2. Feature Vector Generation

5.1.1. POS Tagging
The Stanford POS Tagger is used to tag the discharge summaries. An instance of the tagger class
is created. The input data is stored in a folder. The program iterates through the folder
tags all the input files using the tagger instance created. The tagged data is stored in a file.
5.1.2. Key Term Extraction
The key terms such as nouns and adjectives (specified by the tags NN NNP NNS NNPS JJ JJS
etc) are extracted from the tagged data and stored in a file.
Figure 3. Key term extraction pseudo code
5.1.3. Duplicates Removal
The file generated might contain redundant attributes. To avoid this, the duplicates are removed.
The pseudo code for the same is given below
Figure 4. Duplicates removal pseudo code
Computer Science & Information Technology (CS & IT)
is created. The input data is stored in a folder. The program iterates through the folder
Figure 2. POS tagging pseudo code
etc) are extracted from the tagged data and stored in a file.
Figure 3. Key term extraction pseudo code
same is given below
Figure 4. Duplicates removal pseudo code
is created. The input data is stored in a folder. The program iterates through the folder’s files and

Computer Science & Information Technology
5.1.4. Dictionary Look Up
UMLS (Unified Medical Language System) serves as a repository of mentions. The UMLS is
used to extract the relevant symptoms from the tagged file. FileSearcher
the FindWordInFile method is used to search for a word in a given file.
Figure 5. Dictionary lookup pseudo code
5.1.5. Temporal Data Extraction
The discharge summaries are fed as input to the temporal data extraction algorithm.
admission months are extracted using regular expressions.
Figure 6. Temporal data extraction pseudo code
5.1.6. Non-Symptomatic Feature Extraction
Non-Symptomatic features such as age, gender, marital status, family history and past medical
history are extracted using regular expressions from the annotated discharge summaries.
fied Medical Language System) serves as a repository of mentions. The UMLS is
used to extract the relevant symptoms from the tagged file. FileSearcher Class is imported and
the FindWordInFile method is used to search for a word in a given file.
Figure 5. Dictionary lookup pseudo code
5.1.5. Temporal Data Extraction
The discharge summaries are fed as input to the temporal data extraction algorithm.
admission months are extracted using regular expressions.
Figure 6. Temporal data extraction pseudo code
Symptomatic Feature Extraction
Symptomatic features such as age, gender, marital status, family history and past medical
ry are extracted using regular expressions from the annotated discharge summaries.
57
fied Medical Language System) serves as a repository of mentions. The UMLS is
Class is imported and
The discharge summaries are fed as input to the temporal data extraction algorithm. The
Symptomatic features such as age, gender, marital status, family history and past medical
ry are extracted using regular expressions from the annotated discharge summaries.

• The age is computed using the
• The gender can be either M or F (Male or Female)
• The marital status can be Y or N (Yes or No)
• The family history can be Y or N (Yes or No)
• The past medical history can be Y or N (Yes or No)
• The disease can be Y or N (Yes or No)
5.1.7. Feature Vector Generation
The feature vector is the input supplied to the classifier. The features extracted are combined in a
comma separated format and a feature vector is generated. The vector can be represented using
frequency value representation or using binary representation. The frequency value format
implies that, the frequency of occurrence of the feature in the documen
representation on the other hand only considers the presence or absence of the feature in concern.
Dengue has very few prominent symptoms and therefore it is not advisable to use the frequency
value representation to retrieve th
therefore preferred in this case. Non
Figure 7. Feature Vector Generation pseudo code
The feature vector generated is supplied to a
analysis is performed.
The age is computed using the date of birth of the patient.
The gender can be either M or F (Male or Female)
The marital status can be Y or N (Yes or No)
ory can be Y or N (Yes or No)
The past medical history can be Y or N (Yes or No)
The disease can be Y or N (Yes or No)
5.1.7. Feature Vector Generation
implies that, the frequency of occurrence of the feature in the document is considered. The binary
value representation to retrieve them from clinical text. Binary representation of symptoms is
therefore preferred in this case. Non-symptomatic features are represented as nominal attributes.
Figure 7. Feature Vector Generation pseudo code
The feature vector generated is supplied to a set of classifiers. To identify the best classifier an
t is considered. The binary
em from clinical text. Binary representation of symptoms is
symptomatic features are represented as nominal attributes.
set of classifiers. To identify the best classifier an

5.2. Classification and Analysis
Figure 8. Classification and analysis
5.2.1. Classification
The following is the gist of steps followed during classification process:
1. Prepare Training set
2. Supply training set to the classifiers
3. Build the classification models
4. Save the models that have been built
5. Prepare the test set
6. Evaluate the test set on the saved models

7. Analyze the performance of the classifiers
8. Choose the best classifier
9. Supply unlabeled dataset to the best classifier
10. Obtain the prediction
5.2.2. Frequency Analysis
Frequency analysis aims at correlating the frequency of occurrence of the disease over the months.
Eight most common and highly contributing symptoms for dengue have been chosen. The
occurrences of these symptoms over the months is represented using graphs to give a better
understanding of which symptom contributes the most to the presence of dengue.
6. IMPLEMENTATION
6.1. Data set used
We have used 100 samples of annotated discharge summaries as input to this system. The
personal details of the patients are already preprocessed to ensure patient confidentiality. They
contain details like age, date of birth, date of admission, patient's medical history, medication
administered to the patient during the period of stay in the hospital. And the final diagnosis of the
patient is also mentioned.
6.2. Tagged file
The above dataset is sent to a POS tagger to perform the part of speech tagging. An instance of
the tagger is created and its TagFile method is used to tag the data. This tagged file is sent to a
key term extraction algorithm and the relevant features are extracted. The duplicate terms are
removed from using the duplicates removal algorithm. These terms are stored in a file.
6.3. UMLS Look up
A subset of the UMLS containing terms relevant to the disease are used as basis to perform the
dictionary look up. The file containing the key terms is then compared with the thesaurus and
symptoms that contribute to dengue are stored in another file.
6.4. Feature Extraction and Vector Generation
6.4.1. Symptomatic features
To extract the symptomatic features, the following steps are performed:
1. A file reader object is created
2. The discharge summaries are read line by line
• Each line is split into words
• The words are compared with the file containing filtered output

Computer Science & Information Technology
• If there is a match , 1 is written to the feature to the feature vector
3. If there is no match, 0 is written to the vector
6.4.2. Non- Symptomatic feature
The non-symptomatic features are extracted using regular expressions. The features are extracted
and written to the feature vector file.
Snapshot of the generated vector is as shown:
6.5. Classification
The training set is supplied as input to 6 classifiers. Classification analysis was performed on the
classifiers. The steps involved in this analysis are:
• Import the weka and java packages
• Call function useClassifier with the data to be classified as parameter
• Create the classifier object
• Build the classifier model
• Save the model
• Create an Evaluation object
• Cross validate using 10 fold cross validation
• Print the confusion matrix
The results of the analysis are discussed in the Results and Discussions section of the paper.
6.5.1. Prediction on Test Set
The test set contains the samples that aren’t known to the classification model yet. The saved
model is then evaluated on the test set and the accuracy is obtained.
If there is a match , 1 is written to the feature to the feature vector
, 0 is written to the vector
Symptomatic features
symptomatic features are extracted using regular expressions. The features are extracted
ten to the feature vector file. The feature vector is saved as an arff file.
of the generated vector is as shown:
Figure 9. Feature vector
classifiers. The steps involved in this analysis are:
and java packages
Call function useClassifier with the data to be classified as parameter
Create the classifier object
Build the classifier model
Create an Evaluation object
Cross validate using 10 fold cross validation
matrix
model is then evaluated on the test set and the accuracy is obtained.
61
symptomatic features are extracted using regular expressions. The features are extracted

6.5.2. Prediction on Unlabeled Dataset
Unlabeled dataset is fed to the saved model. The disease label is a "?" in this case. The model
then predicts the labels for these samples.
6.5.3. Graphical User Interface
A GUI was developed to simplify access to the dengue detection system. Separate panels, one for
researchers and another for common users were developed. Researchers can upload a folder
consisting of discharge summaries which will be used as the training set. Common users can
indicate which symptoms they are experiencing and get a prediction from the system.
Figure 10. Patient GUI
Figure 11. Researcher GUI

6.6. Frequency Analysis
To perform frequency analysis, we have used bar charts. The bar charts are generated using
JFreeCharts. The correlations of the spread of the symptoms and in turn the disease over the
months are reported briefly to give a clear picture to the researchers. This feature is only available
to the researchers.
Figure 12. Fever vs month
7. RESULTS AND DISCUSSIONS
The feature vector is supplied to various supervised learning algorithms and classifier models are
generated. LibSVM is integrated software for support vector classification, regression and
distribution estimation. It supports multi-class classification. Logistic regression classifier uses a
sigmoid function to perform the classification. Multilayer perceptron is a classifier based on
Artificial Neural Networks. Each layer is completely connected to the next layer in the network.
Naïve Bayes methods are a set of supervised learning methods based on applying Bayes theorem
with the naïve assumption of independence between every pair of features. The Sequential
Minimal Optimizer uses John Plat’s sequential minimal optimization algorithm for training a
support vector classifier. It also normalizes all attributes by default. The Simple Logistic Classifier
is used for building linear logistic regression models. These classifiers are subject to two types of
classifications – 10-fold cross-validation and percentage split (2/3rd
training and 1/3rd
test).
Accuracies obtained from the 2 methods are compared. In addition, accuracy of the various
classifiers are analyzed based on five performance metrics (Accuracy, Kappa statistics, Mean
absolute error, Root mean squared error, Relative absolute error) [16] and the best model is
chosen.
• Accuracy: The number of samples that are correctly classified from the given 100
input samples.
• Kappa Statistic: The Kappa Statistic can be defined as measuring degree of agreement
between two sets of categorized data. Kappa result varies between 0 to 1intervals.
Higher the value of Kappa means stronger the agreement/ bonding. If Kappa = 1, then
there is perfect agreement. If Kappa = 0, then there is no agreement. If values of Kappa
statics are varying in the range of 0.40 to 0.59 considered as moderate, 0.60 to 0.79
considered as substantial, and above 0.80 considered as outstanding.

• Mean Absolute Error:
divided by number of predictions. It is measure set of predicted value to actual value
i.e. how close a predicted model to actual model. The lower the value of MAE the
better the classification.
• Root Mean Squared Error :
squares error divided by number of predictions. It is measure the differences
values predicted by a model and the values actually observed. Small value of RMSE
means better accuracy of model. Lower the value of RMSE, better the prediction and
accuracy.
• Relative Absolute Error:
measurement to the accepted measurement. A lower percentage indicated better
prediction and accuracy.
Figure 13. Classifier analysis using 10
Based on the above analysis, SMO is identified to be the most optimal classifier
7.1 Analysis and correlation
The predicted results are visualized in graphical form subsequent to prediction. Counts of
occurrences of various symptoms over the months are depicted using bar charts, and these values
are compared with the graphs
maximum manifestation of all symptoms was found to be September. This was also the month
with maximum cases of dengue, according to the prediction. This inference was also corroborated
by the graph generated from the initial training dataset, and we gather from these graphs that
August, September and October are the months most vulnerable to dengue.
Mean Absolute Error: Mean absolute error can be defined as sum of absolute errors
a predicted model to actual model. The lower the value of MAE the
cation.
Root Mean Squared Error : Root mean square error is defined as square root of sum of
squares error divided by number of predictions. It is measure the differences
Relative Absolute Error: Relative error is the ratio of the absolute erro
prediction and accuracy.
Figure 13. Classifier analysis using 10-fold cross validation
Based on the above analysis, SMO is identified to be the most optimal classifier.
Analysis and correlation
are compared with the graphs generated for the original training dataset. The month with
h generated from the initial training dataset, and we gather from these graphs that
August, September and October are the months most vulnerable to dengue.
ned as sum of absolute errors
a predicted model to actual model. The lower the value of MAE the
ned as square root of sum of
squares error divided by number of predictions. It is measure the differences between
Relative error is the ratio of the absolute error of the
generated for the original training dataset. The month with
h generated from the initial training dataset, and we gather from these graphs that

Figure 14. Overview of all symptoms spread over the months
Figure 15. Occurrences of all symptoms over the months
8. CONCLUSION
To conclude, we have discussed, in this report, the detailed design and related algorithms for a
system to identify disorder mentions from clinical text and correlate its frequency with the time
frame. The annotated discharge summaries are tagged and feature extraction algorithms are used
to obtain the features relevant to the disease, Dengue. This is followed by the generation of a
feature vector (Binary representation). This vector is then used to train and build various
classification models and SMO is found to produce the best results. The model generated further
aids in the prediction of the disease. Bar graphs are then used to succinctly represent this
correlation. Additionally the correlation of training samples with time frame was compared with
the correlation obtained from predicted results and the disease occurrence was found to
concentrate in the months of August, September and October in both the cases.
9. LIMITATIONS
Our system uses only 15 features. Extracting more features might increase the accuracy of the
model. The feature vector is depicted using the binary representation. Using the frequency value
representation might improve overall classification.
10. FUTURE WORK
As a part of our future work, we intend to write an implementation to produce bag of words and
extract more features to produce an extensive analysis. Further, we also intend to implement

tagging of the discharge summaries using BIOS tagging [5]. Whenever hospitals receive new
samples showing a tendency for dengue, those samples must be integrated with the existing
training set. This was, the training and predictive capacity of the model will grow, possible giving
better results in the future. To provide and up-to date analysis, we could extend the project to be
used as a desktop app or browser plugin which will automatically synchronize with new data
received from the hospitals' end.
REFERENCES
[1] Sameer Pradhan, Noemie Elhadad, Wendy Chapman, Suresh Manandhar & Guergana Savova, (July
2014) “Analysis of Clinical Text”, SemEval-2014 Task 7.
[2] Melinda Katona & Richard Farkas, (June 2014) “SZTE-NLP: Clinical Text Analysis with Named
Entity Recognition”, SemEval-2014.
[3] Koldo Gojenola, Maite Oronoz, Alicia Perez & Arantza Casillas, (December 2014) “IxaMed:
Applying Freeling and a Perceptron Sequential Tagger at the Shared Task on Analyzing Clinical
Texts”, SemEval-2014.
[4] Parth Pathak, Pinal Patel, Vishal Panchal, Narayan Choudhary, Amrish Patel & Gautam Joshi, (July
2014) “ezDI: A Hybrid CRF and SVM based Model for Detecting and Encoding Disorder Mentions
in Clinical Notes”, SemEval-2014.
[5] Oana Frunza, Diana Inkpen & Thomas Tran, (June 2011) “A Machine LearningApproach for
Identifying Disease-Treatment Relations in Short Texts”, IEEE transactions on knowledge and data
engineering, vol. 23, Issue no. 6.
[6] Deepali Chandna, (2014) “Diagnosis of Heart Disease Using Data Mining Algorithm”, International
Journal of Computer Science and Information Technologies, Vol. 5 (2), pp1678-1680.
[7] Jyoti Soni, Ujma Ansari, Dipesh Sharma & Sunita Soni , (March 2011) “Predictive Data Mining for
Medical Diagnosis: An Overview of Heart Disease Prediction”, International Journal of Computer
Applications, vol 17.
[8] Smitha T & Dr.V Sundaram, (2012) “Knowledge Discovery from Real Time Database using Data
Mining Technique”, International Journal of Scientific and Research Publications, Volume 2, Issue 4.
[9] M.A.Nishara Banu & B Gomathy, (Nov-Dec 2013) “Disease Predicting System Using Data Mining
Techniques”, International Journal of Technical Research and Applications, Volume 1, Issue 5, PP.
41-45
[10] Devendra Ratnaparkhi, Tushar Mahajan & Vishal Jadhav, (November 2015) “Heart Disease
Prediction System Using Data Mining Technique”, International Research Journal of Engineering and
Technology, Volume: 02 Issue: 08.
[11] I.S.Jenzi, P.Priyanka & Dr.P.Alli, (March 2013) “A Reliable Classifier Model Using Data Mining
Approach for Heart Disease Prediction”, International Journal of Advanced Research in Computer
Science and Software Engineering, Volume 3, Issue 3.
[12] N. G. Bhuvaneswari Amma, (February 2012) “Cardiovascular Disease Prediction System using
Genetic algorithm and Neural Network”, International Conference on Computing, Communication
and Applications, IEEE, pp1-5.

[13] A.Shameem Fathima & D.Manimeglai, (March 2012) “Predictive Analysis for the Arbovirus-Dengue
using SVM Classification”, International Journal of Engineering and Technology, Volume 2 No. 3
[14] Daranee Thitipryoonwongse, Prapat Suriyaphol & Nuanwan Soonthornphisaj, (2012) “A Data
Mining Framework for Building Dengue Infection Disease Model”, 26thAnnual Conference of the
Japanese Society for Artificial Intelligence
[15] N.Subitha & Dr.A.Padmapriya, (August 2013) “Diagnosis for Dengue Fever Using Spatial Data
Mining”, International Journal of Computer Trends and Technology, Volume 4 Issue 8
[16] Yugal kumar & G. Sahoo, (July 2012) “Analysis of Parametric & Non Parametric Classifiers for
Classification Technique using WEKA”, I.J. Information Technology and Computer Science, Volume
7, pp43-49.
AUTHORS
Nandini V is currently pursuing her final year, Computer Science and Engineering in
SSN College of Engineering. She has published a paper on Machine Vision in the
ARPN Journal of Engineering and Applied Sciences. Her research interests include
Artificial Intelligence, Robotics, Machine Learning, Machine Vision and Data Mining.
Sriranjitha R is currently pursuing her final year, Computer Science and Engineering
in SSN College of Engineering. She is a member of CSI (Computer Society of India).
Her research interests include Machine Learning, Artificial Intelligence, Data Mining
and Data Structures.
Yazhini T P is currently pursuing her final year, Computer Science and Engineering in
SSN College of Engineering. Her research interests include Computer Networks, Data
Mining, Artificial Intelligence and Web Technology.

DENGUE DETECTION AND PREDICTION SYSTEM USING DATA MINING WITH FREQUENCY ANALYSIS

Recommended

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to DENGUE DETECTION AND PREDICTION SYSTEM USING DATA MINING WITH FREQUENCY ANALYSIS (20)

Recently uploaded (20)

DENGUE DETECTION AND PREDICTION SYSTEM USING DATA MINING WITH FREQUENCY ANALYSIS