An Ensemble Machine Learning Framework for Early Prediction and Risk Classification of Diabetes Using Health Indicators

Abstract

Diabetes is a major chronic disease that, if identified early and appropriately risk-assessed, can be prevented from causing serious complications. In this study, an exhaustive machine learning framework to identify the risk of diabetes and stratification by physiological, metabolic and lifestyle parameters was developed. The data set was publicly available and was comprised of demographic, clinical and behavioral attributes which were used to label people as Low Risk, Prediabetes, and High Risk. It comes with data preprocessing, exploratory analysis, feature selection, dimensionality reduction, explainable artificial intelligence (XAI), and several machine learning and ensemble models. The body mass index (BMI) followed by Waist circumference, fasting glucose, HbA1c, triglycerides and physical activity were important features identified by feature importance analysis using the Mutual Information and SHAP methods. Classification accuracy, precision, recall, F1 score, log loss and ROC-AUC were used to assess several classifiers. All ensemble and boosting techniques were better than standard techniques, and they performed very well in comparison. CatBoost outperformed remaining models in the overall accuracy, precision, recall, F1 score, and log loss with the scores of 94.0%, 92%, 93%, 92% and 0.1501, respectively, and achieved the best weighted ROC-AUC of 99.25%. The results show the feasibility of using ensemble learning and XAI in early detection and personalized diabetes risk management.