Machine Learning Prediction of Postoperative Emergency Department Hospital Readmission

Area of Science:

• Medical informatics
• Machine learning in healthcare
• Surgical outcomes research

Background:

• Hospital readmission prediction is well-studied in medical patients but less so in surgical populations.
• Existing prediction models often rely on data only available at hospital discharge.
• Early prediction of readmissions in surgical patients is crucial for timely intervention.

Purpose of the Study:

• To investigate the efficacy of machine learning (ML) in predicting 30-day hospital readmissions for surgical patients originating from the emergency department.
• To determine if ML models can predict readmission risk earlier than the point of hospital discharge.
• To develop hospital-specific ML models for readmission prediction.

Main Methods:

• A cohort of 34,532 surgical patients from a tertiary academic medical center was analyzed.
• Data extracted from electronic health records included surgical, demographic, lab, medication, and procedural information.
• Various ML models were evaluated using area under the receiver-operator characteristic curve (AUC) for predicting 30-day readmissions, with predictions analyzed at different time points post-surgery.

Main Results:

• Surgical and demographic features provided moderate prediction accuracy (AUC: 0.74–0.76).
• Laboratory features significantly improved prediction, with gradient-boosted trees achieving the highest performance (AUC: 0.866).
• Predictions made 36 hours post-surgery demonstrated high accuracy (AUC: 0.88–0.89), comparable to predictions made at discharge.

Conclusions:

• Machine learning models can accurately predict 30-day postoperative readmissions from the emergency department.
• These ML models can be tailored to specific hospital data for improved accuracy.
• Readmission risk can be reliably assessed as early as 36 hours post-surgery, without needing discharge data.