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

Reports of single center experience and studies of larger databases have identified several predictors of burn center mortality, including age, burn size, and inhalation injury. None of these analyses has been broad enough to allow benchmarking across burn centers. The purpose of this study was to derive a reliable, risk-adjusted, statistical model of mortality based on real-life experience at many burn centers in the U.S. We used the American Burn Association 2020 Full Burn Research Dataset, from the Burn Center Quality Platform (BCQP) to identify 130,729 subjects from July 2015 through June 2020 across 103 unique burn centers. We selected 22 predictor variables, from over 50 recorded in the dataset, based on completeness (at least 75% complete required) and clinical significance. We used gradient-boosted regression, a form of machine learning, to predict mortality and compared this to traditional logistic regression. Model performance was evaluated with AUC and PR curves. The CatBoost model achieved a test AUC of 0.980 with an average precision of 0.800. The logistic regression produced an AUC of 0.951 with an average precision of 0.664. While AUC, the measure most reported in the literature, is high for both models, the CatBoost model is markedly more sensitive, leading to a substantial improvement in precision. Using BCQP data, we can predict burn mortality allowing comparison across burn centers participating in BCQP.

Using a National Burn Registry to Develop a Model for Risk-Adjusted Length of Stay Benchmarking.

Length of stay (LOS) is a frequently reported outcome after a burn injury. LOS benchmarking will benefit individual burn centers as a way to measure their performance and set expectations for patients. We sought to create a nationwide, risk-adjusted model to allow for LOS benchmarking based on the data from a national burn registry. Using data from the American Burn Association's Burn Care Quality Platform, we queried admissions from 7/2015 to 6/2020 and identified 130,729 records reported by 103 centers. Using 22 predictor variables, comparisons of unpenalized linear regression and Gradient boosted (CatBoost) regressor models were performed by measuring the R2 and concordance correlation coefficient on the application of the model to the test dataset. The CatBoost model applied to the bootstrapped versions of the entire dataset was used to calculate O/E ratios for individual burn centers. Analyses were run on 3 cohorts: all patients, 10-20% TBSA, >20% TBSA. The CatBoost model outperformed the linear regression model with a test R2 of 0.67 and CCC of 0.81 compared with the linear model with R2=0.50, CCC=0.68. The CatBoost was also less biased for higher and lower LOS durations. Gradient-boosted regression models provided greater model performance than traditional regression analysis. Using national burn data, we can predict LOS across contributing burn centers while accounting for patient and center characteristics, producing more meaningful O/E ratios. These models provide a risk-adjusted LOS benchmarking using a robust data source, the first of its kind, for burn centers.