A Dataset-Specific Machine Learning Study for Predicting Diabetes (Type-2) in a Developing Country Context

Objectives: Diabetes become more prevalent across the globe, understanding their sources and causes are more important than ever. This study uses machine learning techniques to efficiently detect Diabetic patients from many features. Methods: The purpose of this paper is to conduct a dataset-specific machine learning study for predicting diabetes in Bangladesh. Classification is used with 18 features including demographic characteristics, family history, dieting habit, clinical features, physical activities, and life quality. Five different classifiers are used. Findings: Based on using five different classifiers, results suggest that the Logistic Regression performed the best in predicting diabetes for this dataset. The accuracy of the logistic regression classifier exceeds 83.8%. Novelty: Unlike other studies, the authors combine eating habits with demographic and health features to enhance the performance of the classifiers. The result suggests that while addition of factors or features related to eating habits and lifestyle can increase the accuracy of prediction, the inclusion of more clinical features is more important to increase the accuracy. The authors believe that this finding is significant in the context of developing countries like Bangladesh considering the limited health-resource available as well as the fact of fast-changing of eating habits and lifestyle. Keywords: Machine Learning; chronic disease; classification; logistic regression; and diabetes