Early ICU Mortality Prediction Using Statistical and Gradient Boosting Models on Physiological Time-Series Data

Early prediction of mortality risk in Intensive Care Unit (ICU) patients is critical for improving clinical decision-making and resource allocation. This study develops and evaluates statistical and machine learning models for early ICU mortality prediction using physiological time-series data from the PhysioNet Challenge 2012 dataset. Logistic Regression was implemented as a statistical baseline model, while XGBoost was applied to capture nonlinear relationships and complex feature interactions. The XGBoost model achieved an AUROC of 0.8395 and an AUPRC of 0.4556, demonstrating improved discrimination compared to Logistic Regression. Calibration analysis indicated reliable probability estimation with a Brier Score of 0.1023. The findings highlight the effectiveness of gradient boosting models for early clinical risk stratification using structured physiological features.