Predicting 72-hour and 9-day return to the emergency department using machine learning

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Predicting emergency department (ED) return visits is crucial for optimizing healthcare resource allocation and improving patient outcomes.
• Traditional methods often rely on limited administrative data, potentially missing key clinical insights for accurate prediction.

Purpose of the Study:

• To develop and evaluate gradient boosting models for predicting 72-hour and 9-day ED return visits.
• To compare the predictive performance of gradient boosting models against logistic regression using diverse clinical variables.

Main Methods:

• Retrospective analysis of over 330,000 adult ED discharges from March 2013 to July 2017.
• Extraction and utilization of approximately 1500 clinical variables from electronic health records.
• Training and validation of gradient boosting models using administrative data, triage data, and full discharge data; logistic regression served as a baseline.

Main Results:

• Gradient boosting models significantly outperformed logistic regression, achieving higher Area Under the Curve (AUC) values for both 72-hour and 9-day return predictions.
• Models utilizing data available at triage and the full discharge dataset demonstrated the highest predictive accuracy (AUCs up to 0.76).
• Reduced models using the top 20 variables also showed strong predictive capabilities, indicating feature importance for return prediction.

Conclusions:

• Gradient boosting models, leveraging comprehensive clinical data from electronic health records, offer superior performance in predicting ED return visits compared to logistic regression.
• The findings highlight the potential of machine learning for enhancing patient management and reducing avoidable ED readmissions.