34 Using a National Burn Registry to Develop a Model for Risk-adjusted Length of Stay Benchmarking

Abstract

Introduction

Length of stay (LOS) is a frequently reported outcome after a burn injury. Previous literature estimates LOS at 1 day per % total burn surface area (TBSA) but this varies considerably across patients & centers. LOS benchmarking will benefit individual burn centers as a way to measure their performance & set expectations for patients. We sought to create a nationwide, risk-adjusted model to allow for LOS benchmarking based on data from a national burn registry.

Methods

Using data from a national burn registry, we queried admissions from 7/2015-6/2020 & identified 126,129 records with LOS data reported by 103 centers. We selected 23 predictor variables on the basis of completeness (min. 75% required) & clinical significance. Missing data were multiply imputed with a Bayesian Ridge Regression estimator. All statistics were calculated in Python using Numpy & Scikit-Learn libraries. Comparisons of unpenalized linear regression & Gradient boosted (CatBoost) regressor models were performed by measuring the R² & concordance correlation coefficient (CCC) on the application of the model to the test dataset. The CatBoost model applied to bootstrapped versions of the entire dataset was then used to calculate O/E ratios for individual burn centers. Confidence intervals (CI) for O/E ratios were calculated using a normal distribution parametric model. Analyses were run on 3 cohorts: all patients, 10-20% TBSA, >20% TBSA.

Results

The CatBoost model outperformed the linear regression model with a test R² of 0.68 & CCC of 0.81 compared to the regression model with R²=0.52, CCC=0.70. The CatBoost was also less biased for higher & lower LOS durations. Due to the CatBoost model’s superiority in predicting the outcome, this model alone was used for O/E ratio calculations. The O/E ratio data from the model for all 3 cohorts are shown in Figure 1.

Conclusions

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 & 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.