EARLY-STAGE DIABETES RISK PREDICTION USING MACHINE LEARNING TECHNIQUES BASED ON ENSEMBLE APPROACH

Abstract

Diabetes Mellitus which is considered as one of the deadliest is a common, chronic disease. It also causes the emergence of many diseases, especially neuropathy, nephropathy and retinopathy. In this context, early diagnosis of the disease by accurately evaluating the symptoms and initiating a rapid treatment process is very important. The aim of this study is to present an effective model that can determine the diabetes risk in eary-stage with the best accuracy. To do so, the classification algorithms that are frequently used in diabetes risk estimation are supported with ensemble approaches. Firstly, the performance of Naive Bayes (NB), Trees-J48, k Nearest Neighbor (kNN) and Sequential Minimal Optimization (SMO) classifiers is analyzed separately by using a dataset of 520 samples collected with direct questionnaires from Sylhet Diabetes Hospital patients in Sylhet, Bangladesh. Then, the effects of Adabost, Bagging and Random Sub-Space (RSS) algorithms on classifier success are investigated and it is shown that the j48 classifier based on Adabost approach has the best accuracy in this dataset. Finally, the Wrapper Subset Eval (WSE) feature extraction algorithm is applied to reduce the estimation cost of diabetes and increase classification success. Thus, the best accuracy at 99% is achieved using reduced data set with proposed classifier method.