Predicting Intensive Care Unit Readmission with Machine Learning Using Electronic Health Record Data.

Area of Science:

• Critical Care Medicine
• Health Informatics
• Machine Learning in Healthcare

Background:

• Patients transferred from intensive care units (ICUs) to general wards who are readmitted to the ICU face worse outcomes.
• Accurate risk stratification for these patients is crucial for improving care and reducing mortality.
• Current prediction methods have limitations in identifying patients at high risk for ICU readmission.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for predicting ICU readmission.
• To utilize real-time electronic health record (EHR) data for model development.
• To compare the ML model's performance against established scoring systems.

Main Methods:

• An observational cohort study involving 24,885 ICU transfers to wards at an academic hospital.
• A gradient-boosted machine model was trained using comprehensive EHR data.
• Model performance was validated internally and externally using the MIMIC-III database, comparing against SWIFT and MEWS scores.

Main Results:

• The ML model achieved a significantly higher area under the receiver operating curve (0.76) compared to SWIFT (0.65) and MEWS (0.58).
• At 95% specificity, the ML model's sensitivity (28%) was substantially better than SWIFT (15%) and MEWS (7%).
• Similar performance improvements were observed in the external MIMIC-III validation cohort.

Conclusions:

• A machine learning approach significantly enhances the accuracy of predicting ICU readmissions compared to existing algorithms.
• This ML model offers a promising tool for identifying patients who may benefit from extended ICU care or enhanced ward monitoring.
• Implementation of this predictive model can potentially improve patient outcomes and resource allocation in critical care settings.