Setting the Standard: Using the ABA Burn Registry to Benchmark Risk Adjusted Mortality

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Burn Surgery Outcomes

Background:

• Established predictors of burn center mortality include age, burn size, and inhalation injury.
• Previous analyses lacked the scope for effective cross-center benchmarking.
• There is a need for a reliable, risk-adjusted model for comparing burn center outcomes.

Purpose of the Study:

• To develop a robust statistical model for predicting burn patient mortality.
• To enable risk-adjusted benchmarking across multiple burn centers in the U.S.
• To leverage a large national dataset for improved mortality prediction.

Main Methods:

• Utilized the American Burn Association 2020 Full Burn Research Dataset from the Burn Center Quality Platform (BCQP).
• Included 130,729 subjects from 103 burn centers (July 2015 - June 2020).
• Employed gradient-boosted regression (CatBoost) and compared it with logistic regression, evaluating with AUC and PR curves.

Main Results:

• The CatBoost model achieved a superior test AUC of 0.980 and average precision of 0.800.
• Logistic regression yielded an AUC of 0.951 and average precision of 0.664.
• The CatBoost model demonstrated significantly higher sensitivity and precision compared to logistic regression.

Conclusions:

• Machine learning, specifically CatBoost, provides a highly accurate and sensitive method for predicting burn mortality.
• The developed model enables effective risk-adjusted benchmarking for burn centers participating in the BCQP.
• This approach enhances the ability to compare burn center performance using real-world data.