Explainable Cross-Validated Greedy Stacking Ensemble with RFECV-Based Feature Selection for T2DM Prediction

Abstract

Type 2 Diabetes Mellitus (T2DM) is a significant health burden and causes a considerable burden of disease and mortality, especially in South Asians. This research introduces an Explainable Cross-Validated Greedy Stacking Ensemble (CV-GSE) technique for predicting T2DM on the Type-2 Diabetes Dataset Bangladesh (T2DDB), which contains a set of 1065 patient records containing demographic, anthropometric and biochemical features. The proposed framework first uses Recursive Feature Elimination with Cross-Validation (RFECV) to find an informative subset of predictors and then tests eight heterogeneous machine-learning classifiers via stratified 10-fold cross-validation with SMOTE only on the training folds. Out-of-fold (OOF) probability predictions are made for each candidate classifier and classifiers are added to the final model in a greedy forward fashion, based on OOF ROC-AUC. Then a tuned XGBoost meta-classifier is used to combine the selected base learners. A dual-layer explainability approach is also provided in which SHapley Additive exPlanations (SHAP) is used for feature and base-learner attribution and Local Interpretable Model-Agnostic Explanations (LIME) is used for instance-level explanations. In total, RFECV selected 6 informative predictors (Age, BMI, SBP, DPF, Insulin, and Glucose) with maximum cross-validated ROC-AUC of 0.9558. The stacking ensemble proposed in this study obtained 98.50% accuracy, 98.25% precision, 99.70% recall, 98.62% F1-score, and 96.85% ROC-AUC, which were better than the evaluated baseline classifiers. The outcomes show the capability of the proposed method to predict T2DM using feature selection, heterogeneous ensemble learning, OOF-based greedy learner selection, and Explainable Artificial Intelligence (XAI). But, before clinical applicability can be established, there needs to be external validation, with independent and diverse datasets.