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

Area of Science:

• Critical Care Medicine
• Health Informatics
• Rehabilitation Science

Background:

• Intensive care unit (ICU) survivors frequently experience new or worsening functional impairment post-discharge.
• Early identification of at-risk patients can facilitate timely preventive interventions.
• High-dimensional nursing data within electronic health records (EHRs) can enhance risk prediction models.

Purpose of the Study:

• To develop and validate predictive models for functional impairment at hospital discharge using early ICU admission EHR data.
• To identify key predictors of functional impairment using machine learning and explainability techniques.
• To assess the performance of streamlined models based on the most critical predictors.

Main Methods:

• Retrospective analysis of 799 sepsis survivors' EHR data from the first 48 hours of ICU admission.
• Utilized CTAB-GAN for data synthesis to address limited real-world EHR data.
• Developed and validated eXtreme Gradient Boosting (XGBoost) models, employing SHapley Additive exPlanations (SHAP) for feature importance.

Main Results:

• Models incorporating patient characteristics and nursing assessments accurately predicted functional impairment (Activity Measure for Post Acute Care [AMPAC] score <18).
• Key predictors included initial AMPAC score, age, mobility, Braden Scale, walking device, BMI, and SOFA score.
• Streamlined models using top predictors achieved high predictive accuracy (AUCs of 0.83 at 24h and 48h).

Conclusions:

• Early ICU admission data, including nursing assessments, can effectively identify patients at high risk for post-discharge functional impairment.
• The developed models demonstrate potential for integration into EHR systems for clinical decision support.
• Streamlined predictive models offer a practical and efficient approach to improving patient outcomes by enabling early interventions.