Abstract
This study introduces a novel Transformer-based time-series framework designed to revolutionize risk stratification in Intensive Care Units (ICUs) by predicting patient outcomes with high temporal precision. Leveraging sequential data from the eICU database, our two-stage architecture dynamically captures evolving health trajectories throughout a patient's ICU stay, enabling real-time identification of high-risk individuals and actionable insights for personalized interventions. The model demonstrated exceptional predictive power, achieving a progressive AUC increase from 0.87 (±0.021) on admission day to 0.92 (±0.009) by day 5, reflecting its capacity to assimilate longitudinal physiological patterns. Rigorous external validation across geographically diverse cohorts-including an 81.8% accuracy on Chinese sepsis data (AUC=0.73) and 76.56% accuracy on MIMIC-IV-3.1 (AUC=0.84)-confirmed robust generalizability. Crucially, SHAP-derived temporal heatmaps unveiled mortality-associated feature dynamics over time, bridging the gap between model predictions and clinically interpretable biomarkers. These findings establish a new paradigm for ICU prognostics, where data-driven temporal modeling synergizes with clinician expertise to optimize triage, reduce diagnostic latency, and ultimately improve survival outcomes in critical care.