Predicting diabetes status using ensemble algorithms with hyperparameter tuning

Sonjit Mondol; Dr. Ajit Kumar Majumder; Dr. Mohammad Alamgir Kabir

doi:10.37591/rrjocb.v13i2.3303

Abstract

Diabetes is a condition in which the body is unable to produce enough insulin to keep blood sugar levels under control. If diabetes is not properly identified and treated, it can lead to kidney failure, nerve damage, blindness, and coronary heart disease. A healthy lifestyle, therefore, depends on the early identification of diabetes diseases. However, it can be difficult to assess a person's diabetic status if they live in remote areas or in other places where there is little chance of detecting or testing diabetes. In addition, scheduling an appointment at a diagnostic center and consulting with a doctor adds time and cost to the process of monitoring diabetes in urban regions. Machine learning may be utilized in this situation to overcome these concerns as there are numerous strategies available for resolving classification challenges. The objective of this work is to design an ensemble algorithm and optimize its hyperparameters to accurately identify diabetes from a patient's early symptoms without doing a diagnostic test. Consequently, three ensemble algorithms—Boosting, Bagging, and Random Forest—as well as a grid search hyperparameter tuning strategy—are used on the Bangladesh Demographic and Health Survey (BDHS) 2017–18 dataset. The effectiveness of these algorithms is assessed using the metrics accuracy, sensitivity, specificity, kappa, and ROC curve. The Boosting algorithm has the highest accuracy, at 77.65%, compared to the other two algorithms, with a 7.06% improvement brought on by hyperparameter adjustment.