Predicting Trauma Severity from Imbalanced Data Using Ensemble Regression and Generative Models

Abstract

The Injury Severity Score (ISS) is a critical metric in trauma care, widely used for assessing injury severity, guiding clinical decisions, and evaluating patient outcomes. Despite the practical challenges of computing the score, due to its clinical significance, we propose a machine learning (ML) framework to predict ISS using structured clinical, demographic, and vehicle data, routinely documented in trauma registries and hospitals. We evaluate four ensemble-based regression models such as Random Forest Regressor (RFR), Gradient Boosting Regressor (GBR), eXtreme Gradient Boosting (XGBoost) regressor, and Gradient Boosting Machine (LightGBM). We identify GBR as the best performer when applied on a dataset generated at our clinical site, with coefficient of determination R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> =0.78. However, the original dataset exhibits substantial imbalance, with most cases concentrated in low-severity scores. To address the challenges of skewed ISS distribution, we implement various data augmentation techniques including transformations of target, resampling, interpolation, and noise-based strategies. Moreover, we develop two generative models Conditional Variational Autoencoder (cVAE) and Conditional Generative Adversarial Network (cGAN) to synthesize data from underrepresented severity ranges. The cVAE-augmented model achieves the highest performance of R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> =0.94, demonstrating the value of generative augmentation in enhancing regression accuracy under data imbalance.

Ähnliche Arbeiten