Predicting Adverse ICU Outcomes from Admission-Time Frailty and Nocturnal Sleep Fragmentation Using Explainable AI

Background. Sleep fragmentation and frailty are interrelated yet underexplored determinants of adverse outcomes in critically ill patients, and no published model has combined objective nocturnal fragmentation metrics with administrative frailty indices to predict adverse outcomes at ICU admission. Methods. Using 31,139 first ICU admissions from the MIMIC-IV (v3.1) database, we developed and compared five supervised machine learning models to predict a composite outcome of 30-day mortality, prolonged mechanical ventilation (>7 days), or discharge to a skilled nursing or rehabilitation facility (outcome prevalence 56.1%). Predictors were drawn exclusively from routinely available admission-time data, including nocturnal chartevent frequency, nighttime RASS scores, the Hospital Frailty Risk Score (HFRS), comorbidity and severity indices, and standard demographic variables. Results. The three gradient boosting models achieved statistically equivalent discrimination (ROC AUC ≈ 0.82-0.83) with strong calibration. SHAP analysis of the best-performing model (CatBoost) identified age, HFRS, nocturnal chartevent count, and nighttime RASS as the most influential predictors, confirming that frailty and sleep fragmentation contribute independently to risk. Conclusions. These findings show that EHR-derived frailty signals and nocturnal monitoring intensity, available without additional assessment burden, provide clinically actionable risk stratification at ICU admission and a viable foundation for embedded decision-support tools.