Predicting individual patient and hospital-level discharge using machine learning.

Area of Science:

• Health Informatics
• Machine Learning Applications
• Healthcare Operations

Background:

• Accurate prediction of hospital discharge events is crucial for optimizing patient flow and healthcare delivery efficiency.
• The application of machine learning (ML) and diverse electronic health record (EHR) data for discharge prediction is an area with significant unexplored potential.

Purpose of the Study:

• To develop and evaluate ML models for predicting hospital discharge within 24 hours.
• To assess the performance of ML models using EHR data for both elective and emergency admissions.
• To identify key predictors influencing discharge events and evaluate model robustness.

Main Methods:

• Utilized EHR data from February 2017 to January 2020 in Oxfordshire, UK.
• Developed extreme gradient boosting models for elective and emergency admissions, trained on two years of data and tested on the final year.
• Examined individual and hospital-level performance, impact of data size, recency, and prediction time.

Main Results:

• Models achieved high performance with AUROCs of 0.87 (elective) and 0.86 (emergency), outperforming logistic regression models.
• Daily discharge estimates showed high accuracy with mean absolute errors of 8.9% (elective) and 4.9% (emergency).
• Antibiotic prescriptions, medications, and hospital capacity were key predictors; performance was robust across subgroups but lower for longer admissions.

Conclusions:

• ML models demonstrate significant potential for optimizing hospital patient flow.
• These predictive capabilities can facilitate improved patient care and recovery processes.
• The study underscores the value of EHR data and ML in enhancing healthcare operational efficiency.