Interpretable Machine Learning-Based Multiclass Prediction of Vascular Complication Profiles in Type 2 Diabetes: A Comparative Study

Abstract

Vascular complications are common and heterogeneous in type 2 diabetes, yet multiclass classification is challenging because of substantial class imbalance. This study evaluated interpretable machine-learning approaches for characterizing microvascular and macrovascular complication profiles. After removing duplicated clinical profiles, 1,068 profiles were classified into four categories: no complications, microvascular only, macrovascular only, and both. Multiple machine-learning models and imbalance-handling strategies were compared using stratified five-fold cross-validation and an independent test set. Performance was assessed using macro-F1, balanced accuracy, class-specific metrics, bootstrap confidence intervals, and SHAP-based feature interpretation. SMOTENC achieved the strongest overall balance between minority-class performance and discrimination, with a cross-validation macro-F1 of 0.273 ± 0.011 and balanced accuracy of 0.294 ± 0.026. On the independent test set, macro-F1 was 0.277 (95% CI, 0.213–0.343) and balanced accuracy was 0.299 (95% CI, 0.220–0.385). The macrovascular-only class remained particularly difficult to distinguish. Medication adherence, smoking, physical activity, sex, and medication use were among the most influential features. Machine-learning classification of vascular complication profiles remains challenging in this imbalanced dataset. Nevertheless, explicit imbalance handling and class-balanced evaluation provided a more informative assessment of model performance than accuracy alone. These findings support the use of interpretable, imbalance-aware frameworks for future studies and highlight the need for larger, longitudinal, and externally validated datasets before clinical application.