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

Area of Science:

• Critical Care Medicine
• Health Informatics
• Gerontology

Background:

• Sleep fragmentation and frailty are significant, yet understudied, factors impacting critically ill patients.
• Existing models lack integration of objective sleep data and frailty indices for predicting adverse outcomes upon ICU admission.

Purpose of the Study:

• To develop and compare machine learning models predicting adverse outcomes at ICU admission.
• To identify key predictors of 30-day mortality, prolonged mechanical ventilation, or discharge to a skilled nursing facility.

Main Methods:

• Utilized 31,139 ICU admissions from the MIMIC-IV database.
• Developed five supervised machine learning models using admission-time data.
• Included nocturnal chartevent frequency, RASS scores, Hospital Frailty Risk Score (HFRS), comorbidities, and demographics.

Main Results:

• Gradient boosting models demonstrated strong predictive performance (ROC AUC ≈ 0.82-0.83).
• Key predictors identified by SHAP analysis included age, HFRS, nocturnal chartevent count, and nighttime RASS scores.
• Frailty and sleep fragmentation were confirmed as independent risk contributors.

Conclusions:

• EHR-derived frailty signals and nocturnal monitoring intensity offer actionable risk stratification at ICU admission.
• These data provide a foundation for developing embedded decision-support tools for critical care.
• No additional assessment burden is required for this risk stratification.