An explainable and calibrated LightGBM model for hypertension risk stratification in older adults with diabetes

Hypertension is a common comorbidity in older adults with diabetes, increasing cardiovascular and renal risk. Accurate risk stratification is challenging due to heterogeneous clinical characteristics and complex interactions among demographic, metabolic, biochemical, and lifestyle factors. This study developed iLGBM-Hypertension, an interpretable gradient-boosting model for hypertension risk assessment in elderly diabetic patients. A retrospective cohort of 387 patients from Yaoshan Community Health Service Center China(January 2023-March 2025) was analyzed. Clinical and laboratory variables were modeled using LightGBM, with SHAP and LIME providing global and local interpretability. Model evaluation employed stratified 10-fold cross-validation, assessing discrimination, calibration, statistical agreement, and interpretability metrics. Feature-ablation and sensitivity analyses examined the impact of variable inclusion and probability calibration on predictive performance and explanation consistency. The interpretable gradient-boosting model achieved strong discrimination (AUC 0.984) and calibration (ECE 0.028). SHAP and LIME analyses identified body mass index, family history, and age as dominant risk drivers, while enabling individualized patient-specific explanations without implying causality. Individual-level explanations illustrated patient-specific risk profiles without implying causality. These results indicate that interpretable gradient-boosting models can provide accurate, well-calibrated hypertension risk stratification in elderly diabetic populations while maintaining transparency. External validation is required before broader clinical application, but this framework offers a practical approach for individualized risk assessment in routine care.