# 14 Ensemble Machine Learning Models for Burn Mortality Prediction Using the WHO Global Burn Registry

## TL;DR

This study uses machine learning to predict burn mortality using global data, showing that XGBoost performs best and could help improve patient care and resource allocation.

## Contribution

The study introduces and compares three ensemble machine learning models for predicting burn mortality using the WHO Global Burn Registry.

## Key findings

- XGBoost outperformed RF and GBM in predicting burn mortality with the highest accuracy and specificity.
- TBSA (%), length of stay, Baux score, age, and management in low-resource settings were the top predictors of mortality.
- The models showed strong performance across global settings, suggesting potential for optimizing resource allocation in burn care.

## Abstract

Burn injuries represent a significant global health challenge, with mortality prediction being a critical component that can dictate patient care and resource allocation. This study aims to apply ensemble machine learning models to predict burn mortality using a large database that captures burns from diverse global settings. We identified the most important predictors of burn mortality.

The dataset comprised patient records from the WHO Global Burn Registry from its inception. The primary outcome of interest was mortality after burn injury. We used random forest (RF), gradient boosting machine (GBM), and extreme gradient boosting (XGBoost) to develop predictive models. Their performance was assessed based on accuracy, area under the curve (AUC), F1 score, sensitivity, specificity, and balanced accuracy. Variable importance was also analyzed to determine the most influential factors in predicting burn mortality.

The models were evaluated using data from 9,130 burn patients and 35 predictors. The RF model achieved an accuracy of 91.84% (95%CI: 90.48–93.05%), an AUC of 0.843, and an F1 score of 0.9503, with a sensitivity of 96.54% and specificity of 72.00%. The GBM model yielded an accuracy of 91.73% (95%CI: 90.37–92.95%), an AUC of 0.9503, and an F1 score of 0.9497, with a sensitivity of 96.61% and specificity of 71.14%. The XGBoost model demonstrated the best overall performance, with an accuracy of 91.95% (95%CI: 90.60–93.15%), an AUC of 0.9518, and an F1 score of 0.9505, along with a sensitivity of 95.73% and the highest specificity at 76.00%. The balanced accuracy for XGBoost was 85.86%, outperforming both the RF and GBM. The top 5 predictors of mortality for the XGBoost model were: 1) TBSA (%), 2) length of stay (days), 3) Baux score, 4) age, and 5) management in a low-resource setting.

All three ensemble learning models—RF, GBM, and XGBoost—demonstrated strong predictive ability for burn mortality on a global scale, with XGBoost performing the best across all metrics. These findings suggest that machine learning models, especially XGBoost, can serve as valuable tools for burn providers in predicting outcomes across resource levels. The development of a risk stratification tool based on these models could potentially optimize resource allocation for burns.

This study captures the potential of ensemble machine learning models to enhance predictive accuracy in burn mortality outcomes using real-world data from the World Health Organization. These models can hopefully assist in triaging burn patients, optimizing resource allocation, and ultimately improving decision-making in critical care settings in high- and low-resource regions.

N/A