Comparative Evaluation of Ensembles in Type 2 Diabetes Classification: Performance, Explainability, and Interpretability

Type 2 diabetes mellitus (T2DM) is a priority public health issue that requires accurate, efficient, and interpretable predictive tools to support the early identification of at-risk cases. Nevertheless, the comparison of machine learning models in this domain is frequently constrained by heterogeneous protocols, a reliance on overall accuracy, and an inadequate integration of metrics for agreement, calibration, robustness, and interpretability. This study employed a comparative approach to evaluate 10 ensemble and boosting models for the binary classification of T2DM under a unified experimental protocol. The evaluation employed nested cross-validation with 10 external folds and 3 internal folds, a 20% internal holdout, and complementary metrics of performance, discrimination, agreement, calibration, and computational cost. In the context of nested cross-validation, the Random Forest algorithm demonstrated a superior performance, attaining the highest average weighted F1-score (92.51 ± 5.80%), the highest Matthews' correlation coefficient (MCC) (0.840 ± 0.120%), and high discriminative power (weighted area under the receiver operating characteristic curve (ROC-AUC) = 97.46 ± 2.12%). Consequently, it was selected based on a predefined composite criterion. CatBoost achieved the highest weighted ROC-AUC (97.91 ± 2.39%) and area under the precision-recall curve (PR-AUC) (96.60 ± 3.61%), while Extra Trees demonstrated performance that was virtually equivalent to that of the selected model. The Friedman test revealed significant overall differences among models (χ² = 27.898607; p = 0.000992). However, the Nemenyi post-hoc test indicated that the leading models were statistically comparable to one another and that the significant differences were concentrated relative to AdaBoost. In the holdout set, the calibrated Random Forest attained an accuracy of 0.882, a balanced accuracy of 0.880, a weighted F1-score of 0.884, an ROC-AUC of 0.943, a PR-AUC of 0.879, a κ of 0.736, an MCC of 0.739, a Brier score of 0.091, and an expected calibration error of 0.130, with a recall of 0.88 for the positive class. Explainability analyses employing SHapley Additive Explanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) techniques have been demonstrated to offer traceability and predictive plausibility, whilst eschewing the assumption of causality. The findings of this study suggest that Random Forest, Extra Trees, and CatBoost represent robust alternatives for the tabular classification of T2DM. However, it is important to note that their clinical application requires multicenter external validation, local recalibration, and prospective evaluation.