Benchmarking Ensemble Models to Predict Prolonged Hospital Stay in Traumatic Brain Injury: A Single-Institution Study

Prolonged length of stay (PLOS) in hospitals is a critical metric representing quality and efficiency of care, especially for patients with traumatic brain injury (TBI). Machine learning offers the potential to predict PLOS, although class imbalance, limited sample size, or lack of generalizability impact their real-world application. This study benchmarks machine learning models from prior studies and explores ensemble models to predict PLOS in TBI patients and address domain adaptation concerns in surgical settings.

An anonymized dataset of 263 adult TBI patients admitted to a single level 1 trauma intensive care unit was used. Nine features were used across basic demographic, clinical, and procedure status variables. Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and Artificial Neural Network (ANN) machine learning algorithms were used, with hyperparameters optimized using GridSearchCV. The SA2DELA framework was used to build ensemble models. Performance metrics were analyzed holistically using a critical difference diagram and post hoc analyses.

The ensemble models combining XGBoost, ANN, and SVM, as well as XGBoost with a snapshot ANN, outperformed the base models (are under the curve: 0.87, critical difference rank: 2.3, Conover post hoc P < 0.05). Bias-variance-diversity decomposition highlighted the complementary strengths of XGBoost and ANN, whereas SVM added incremental improvements. Feature importance identified age, body mass index, and injury severity score as predictors of PLOS.

This study is the first to benchmark established machine learning models and implement ensemble techniques for predicting PLOS in TBI patients. Combining complementary algorithms in a standardized framework can improve robustness in data-constrained and diverse settings. Future studies should incorporate multiclass or regression models and use stronger domain adaptation analyses.