Adaptive Feature-Weighted Stacking Ensemble for Short-Term Risk Prediction of Prolonged Length of Stay in Elderly Trauma Patients

The Adaptive Feature-Weighted Stacking Ensemble (AFWSE) model is presented here as a new machine learning method that provides staged prediction of prolonged length of stay in trauma patients. We used a retrospective dataset of elderly trauma patients using simulated data from publicly available feature characteristics. AFWSE is designed to combine different base learners like logistic regression, random forest, and gradient boosting, into a stacked ensemble with an adaptive feature weighting mechanism that allows researchers to identify complex patterns while emphasizing clinically relevant features. The AFWSE model was compared to standard machine learning methods, specifically, logistic regression, random forest, gradient boosting, and neural networks, demonstrating consistently better predictive validity and accuracy. The AFWSE identified important features which included older age, injury site, injury mechanism, or type of trauma, and Glasgow Coma Scale, which contributes to the existing body of clinical evidence. The AFWSE model mechanism and its potential clinical interpretation and implications are discussed, along with addressing the recognized limitations of using simulated datasets.