A Multi-Modal Machine Learning Framework for Type 2 Diabetes Risk Stratification and Progression Time Prediction

Abstract

The growing global burden of Type 2 Diabetes (T2D) urgently requires advanced predictive tools that go beyond traditional binary risk classification. This study presents an innovative machine learning framework that integrates multimodal data, including clinical parameters, lifestyle factors, Nuclear Magnetic Resonance (NMR) metabolomics data, and 12-lead Electrocardiogram (ECG) signals, to simultaneously predict T2D risk and disease progression time. This study employs a two-stage prediction framework: an AdaBoost classifier optimized by Optuna is used for risk stratification (AUROC=0.9468, F1=0.8857), combined with an Elastic Net regression model for progression time estimation (R²=0.7649, MSE=15.1535). This framework achieves superior predictive performance compared to single-source models by synergistically integrating complementary data modalities, guides clinical intervention timing through quantitative estimation of individualized progression time, and enhances model interpretability through rigorous feature importance analysis, demonstrating key advantages. This method, which combines high accuracy with clinical utility, fills a critical gap in current T2D prediction methodologies and provides a practical tool for personalized prevention strategies and optimized healthcare resource allocation.