A dynamic weighted ensemble learning framework for cardiovascular risk prediction in type 2 diabetes: a comparative study with SHAP-based interpretability

Diabetes mellitus is one of the major issues in global public health and its cardiovascular complications are the primary cause of death in patients. Traditional risk assessment models, such as the Framingham Risk Score and the UKPDS Risk Engine, are built primarily on linear hypotheses, so they fail to capture complex non-linear relationships and show poor generalization across different populations. In this study, a multi-index dynamic weighted ensemble model was built by innovatively integrating TCM tongue diagnosis indexes with modern medical biomarkers. Specifically, adopting a cross-sectional design, this study initially enrolled 3,111 Type 2 diabetes patients, of which 2,895 were included in the final analysis after excluding 216 participants with excessive missing data ([Formula: see text]), built base models using Random Forest (RF), Gradient Boosting Decision Tree (GBDT), K-nearest Neighbors (KNN), and eXtreme Gradient Boosting (XGBoost) algorithms, and developed a dynamic weighted ensemble framework for optimizing model integration. According to the study results, the ensemble model achieved an accuracy of 95.68%, a sensitivity of 94.92%, and a specificity of 96.21%, remarkably outperforming the existing models. The SHAP analysis revealed that indexes such as chest tightness, ESR, and tongue purple had a significant non-linear influence on diabetes risk. Moreover, the ablation test results further proved the superiority of the ensemble framework over single-algorithm frameworks. This study effectively integrates TCM diagnosis indexes with Western medical indexes, provides a reliable decision-making tool for early screening and personalized intervention in diabetic complications, and shows great values for clinical application and transformations.