Machine Learning-Based Prediction Of Mortality Risk In Type 2 Diabetes Patients Using Multi-Organ Biomarkers

Type 2 Diabetes Mellitus (T2DM) remains a major global health burden and a leading contributor to cardiovascular, renal, and hepatic mortality. Traditional risk assessment models rely on limited clinical parameters and fail to capture complex nonlinear interactions among multi-organ biomarkers. This study proposes a comprehensive machine learning (ML) and deep learning (DL)-based survival modeling framework to predict mortality risk in T2DM patients using multi-organ biomarkers, including fasting blood glucose, HbA1c, serum creatinine, triglycerides (TG), total cholesterol, LDL, HDL, liver enzymes (ALT, AST), and fatty liver indicators. Using the National Health and Nutrition Examination Survey (NHANES) linked mortality dataset, we compare Cox Proportional Hazards, Random Survival Forest (RSF), Gradient Boosting Survival (GBM), DeepSurv, and Long Short-Term Memory (LSTM) models. Performance was evaluated using Concordance Index (C-index), time-dependent Area Under Curve (AUC), Hazard Ratio (HR), and Brier score. Results show DeepSurv achieved the highest C-index (0.82), followed by RSF (0.79), outperforming traditional Cox regression (0.72). SHAP-based feature importance revealed HbA1c, creatinine, triglycerides, and ALT as dominant mortality predictors. Risk stratification analysis demonstrated clear separation between low-, medium-, and high-risk groups (log-rank p < 0.001). The findings highlight the superiority of nonlinear survival models for mortality prediction in T2DM and provide clinically interpretable insights for personalized risk management.