Parsimonious machine learning models to predict resource use in cardiac surgery across a statewide collaborative

Area of Science:

• Cardiothoracic Surgery
• Medical Informatics
• Machine Learning

Background:

• Predicting resource utilization and clinical outcomes after cardiac surgery is crucial for efficient healthcare management.
• Preoperative factors are valuable predictors of postoperative events and resource needs.

Purpose of the Study:

• To develop parsimonious machine learning models for predicting resource utilization and clinical outcomes after cardiac operations using only preoperative data.
• To identify key preoperative predictors of increased resource use and adverse clinical outcomes.

Main Methods:

• Utilized data from the 2015-2021 University of California Cardiac Surgery Consortium repository for patients undergoing coronary artery bypass grafting and/or valve operations.
• Developed and compared linear regression, gradient boosted machines, random forest, and extreme gradient boosting models.
• Evaluated model performance using coefficient of determination for length of stay (LOS) and intensive care unit LOS (ICU LOS), and area under the receiver operating characteristic (AUC) for mortality and other endpoints.

Main Results:

• Gradient boosted machines demonstrated superior performance in predicting LOS (R-squared: 0.42), ICU LOS (R-squared: 0.23), and 30-day mortality (AUC: 0.69).
• The gradient boosted machine model also showed the best prediction for acute kidney injury (AUC: 0.76), while random forest excelled in predicting postoperative transfusion (AUC: 0.73).
• Advancing age, reduced hematocrit, and multiple-valve procedures were significant predictors of increased LOS and ICU LOS.

Conclusions:

• Machine learning models effectively estimate resource use and clinical outcomes in cardiac surgery.
• Identified preoperative risk factors can inform hospital resource allocation and case scheduling, especially during periods of high demand.