Predicting 10-Year Diabetes Risk Through Physiological Acceleration: A Longitudinal Deep Learning Ensemble Approach

Background/Objectives: Type 2 diabetes (T2D) develops gradually over many years through a prolonged preclinical phase, yet traditional static risk scores often fail to capture these dynamic metabolic trajectories. We propose a longitudinal deep learning framework to predict the 10-year risk of Type 2 diabetes onset defined by comprehensive ADA criteria by modeling the physiological acceleration of routine clinical biomarkers. Methods: Utilizing an 18-year longitudinal dataset from the community-based Korean Genome and Epidemiology Study (KoGES) cohort, we selected N=4354 participants with complete follow-up records, ensuring high data integrity without requiring synthetic data augmentation. We constructed a 3-dimensional tensor of 21 non-invasive clinical variables spanning a 6-year observation window. To resolve the inherent precision-recall trade-offs of individual models, we developed a stacking ensemble that integrates Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures via a logistic regression meta-learner. To evaluate the added value of longitudinal modeling, we compared this dynamic framework against a static XGBoost baseline that only saw the most recent data. Results: Evaluated on an independent test set (n=874), the ensemble significantly outperformed baseline models, achieving an overall accuracy of 0.90 (95% CI: 0.88-0.92) and an AUROC of 0.94 (95% CI: 0.93-0.95). By harmonizing LSTM's sensitivity and GRU's precision, the model yielded an exceptional Positive Predictive Value (PPV) of 0.97, a sensitivity of 0.80, and a specificity of 0.98. Conclusions: This framework provides a highly accurate, resource-efficient triage instrument for T2D screening, thereby reducing unnecessary clinical alerts and improving screening efficiency.