Abstract
<title>Abstract</title> <p>Background ICH has high ICU mortality, yet conventional scores rely on static data and miss early dynamics. We developed and validated machine learning models using 24-hour time-series features to predict in-hospital mortality. Methods This TRIPOD + AI retrospective study used MIMIC-IV for development (n = 1,962) and internal validation (n = 842), and eICU-CRD for external validation (n = 3,470). The first 24 hours were divided into four 6-hour windows; statistical features of eight physiological parameters were extracted per window. MICE and SMOTE addressed missing data and class imbalance. LASSO selected 29 features for four models (LR, RF, XGBoost, LightGBM), benchmarked against a GCS-only baseline. Performance was assessed by AUC, calibration, DCA, and SHAP. Results Mortality was 21.3%, 21.5%, and 17.3% across cohorts. Twenty-two of 29 features (76%) were dynamic. LightGBM achieved the best performance: internal AUC 0.823 and external AUC 0.813, significantly exceeding the GCS-only baseline (0.778). Calibration was satisfactory; DCA demonstrated net benefit. SHAP identified GCS and BUN as top predictors, with dynamic vital sign features contributing substantially. Conclusions Dynamic physiological trajectories encode prognostic information beyond static snapshots, underscoring the value of continuous ICU monitoring. Our study offers an interpretable, imaging-free tool for early risk stratification, with potential for integration into clinical systems to guide triage and surveillance in high-risk patients.</p>