Uncertainty-Aware Temporal Learning for Early Prediction of Clinical Deterioration from Electronic Health Records

Abstract

Early prediction of clinical deterioration is essential for improving patient outcomes and supporting timely medical interventions. However, existing machine learning approaches often struggle with irregular clinical measurements, missing observations, and uncertainty in patient trajectories. This study proposes an uncertainty-aware temporal learning framework that integrates longitudinal electronic health records, laboratory measurements, vital signs, and patient demographics to predict clinical deterioration within a 24-hour prediction horizon. The framework combines temporal attention mechanisms with uncertainty estimation and adaptive feature weighting to capture evolving patient conditions while reducing sensitivity to incomplete observations. The proposed approach is designed for evaluation using the MIMIC-IV database, with a target experimental cohort of approximately 35,000 ICU admissions. Performance will be assessed against logistic regression, XGBoost, LSTM, and Transformer-based baselines using AUROC, AUPRC, sensitivity, and calibration error. The study will additionally examine prediction stability under varying missing-data rates and across clinically relevant patient subgroups. By jointly modeling temporal dependencies and predictive uncertainty, the proposed framework aims to provide more reliable early-warning predictions and support interpretable risk assessment in critical care environments.