Unlocking the potential of real-time ICU mortality prediction: redefining risk assessment with continuous data recovery.

Real-time prediction of short-term mortality risk in the intensive care unit (ICU) is often hampered by missing medical data. To address this, we developed RealMIP, an end-to-end framework leveraging generative model for the dynamic imputation of missing values and continuous mortality risk assessment. The model was trained on data from 188 centers in the eICU Collaborative Research Database (eICU-CRD), and internally validated on 20 held-out centers. External validation was performed using the Medical Information Mart for Intensive Care IV (MIMIC-IV) and Salzburg Intensive Care Database (SICdb). RealMIP's predictive performance was compared with nine established approaches. RealMIP achieved robust predictive performance, with AUCs of 0.957 (95% CI, 0.956-0.957) internally, 0.968 (95% CI, 0.968-0.968) in MIMIC-IV, and 0.932 (95% CI, 0.932-0.933) in SICdb, outperforming comparator models (p < 0.05). RealMIP unlocks the potential of real-time ICU mortality prediction by effectively handling missing data and delivering continuous, interpretable risk assessments.