Predicting Post-ICU Functional Impairment During Early ICU Admission Using Real-world Electronic Health Record Data

Intensive care unit (ICU) survivors increasingly report new or worsening functional impairment at hospital discharge. Early risk identification models that include high-dimensional nursing data may improve the delivery of preventive interventions. This study aims to develop and validate models predicting functional impairment at hospital discharge (Activity Measure for Post Acute Care [AMPAC] score <18) using electronic health record (EHR) data from the first 48 h of ICU admission. We identified 799 sepsis survivors hospitalized in the ICU (April 2016-May 2020) from a Midwestern health system's data warehouse. We extracted demographics, illness severity, nursing assessments, and ICU interventions. Given the limited availability of real-world EHR data, we employed CTAB-GAN, a generative adversarial network, to synthesize training data, enabling more robust model development. After feature engineering, 53 of 99 features were selected. We trained an eXtreme Gradient Boosting (XGBoost) classification model and used SHapley Additive exPlanations (SHAP) analysis to identify key predictors. Model performance was evaluated using the area under the receiver operating characteristic curves (AUC). For the 24-h model, the most critical features were first documented AMPAC score, age, mobility level, Braden Scale score, and walking device, while the 48-h model added body mass index and sequential organ failure assessment (SOFA) score as key predictors. Leveraging these findings, lightweight models were constructed using only the most important (top 5/10) predictors, which achieved results comparable to the full predictor model, with AUCs of 0.83 (24 h) and 0.83 (48 h), respectively. Our model, which includes patient characteristics and nurse assessments, can identify patients during early ICU admission who are at high risk for functional impairment at hospital discharge. Our streamlined modeling approach highlights the potential for integration into EHR systems, providing a practical and efficient tool for clinical decision support while maintaining predictive accuracy.