An Improved Diabetes Risk Prediction Framework : An Indonesian Case Study

• Main Author: Sutanto, Daniel Hartono
• Format: Thesis
• Language: English; English
• Published: 2018
• Subjects: Q Science (General); QA76 Computer software
• Online Access: http://eprints.utem.edu.my/id/eprint/23340/1/An%20Improved%20Diabetes%20Risk%20Prediction%20Framework%20An%20Indonesian%20Case%20Study.pdf; http://eprints.utem.edu.my/id/eprint/23340/2/An%20Improved%20Diabetes%20Risk%20Prediction%20Framework%20An%20Indonesian%20Case%20Study.pdf; http://eprints.utem.edu.my/id/eprint/23340/; http://plh.utem.edu.my/cgi-bin/koha/opac-detail.pl?biblionumber=113297

Lack of diagnosis for diabetes often transpire in some ASEAN countries with relatively diminutive doctor to patient ratio.Essentially,it is believed that a systematic framework to predict diabetes risk factors is crucial for refining diagnostics and improving accuracy. However,there is the issue of noisy dataset detected as incomplete data and the outlier class problem that affects sampling bias.Existing frameworks were deemed difficult in identifying the critical risk factors of diabetes;some of which were considerably inaccurate and consume substantial computation time.The purpose of this study is to develop a suitable framework for predicting diabetes risks.From a complete blood test,the framework can predict and classify the output of either having diabetes risk or no diabetes risk.A Diabetes Risk Prediction Framework (DRPF) was developed from the literature review and case studies were afterwards conducted in three private hospitals in Semarang.Analyses were conducted to find a suitable component of the framework—due to lack of comparison and analysis on the combination of feature selection and classification algorithm.DRPF comprises four main sections: pre-processing,outlier detection,risk weighting,and learning. Pre-processing resolves the issue of missing data and hence normalizes the data.Outlier treatment employs k-mean clustering to validate the class.Suitable components were selected through comparison of classifier algorithms and feature selection.Attribute weighting based feature selection was selected for assigning weightage.Weighted risk factor was used on training dataset in order to improve accuracy and computation time of the prediction. In the learning section,Support Vector Machine and Artificial Neural Network were selected as suitable classification algorithms,while Gradient Boosted Tree was employed to interpret the rule based on the black box classifiers.Testing the framework involved Pima Indian Dataset as public dataset and Semarang Hospital Dataset as private dataset (800 patients’ data).In validating the DRPF,four case studies investigated Subject Matter Expert (SME) groups based on the agreement level.The questionnaire consists of a DRPF component,implementation of DRPF,and viability of DRPF.DRPF components were validated by the SMEs,whereby the group ascertained five highest risk factors:HbA1c,systole/diastole,blood glucose,and creatinine and blood urea nitrogen that were assigned by attribute weighting.Results from the questionnaire revealed an average agreement level of 80%. In conclusion,DRPF is implementable as prototype and has been highly accepted by Indonesian practitioners as aid for the diagnostics of diabetes.