Abstract
Introduction: Blunt thoracic aortic injury (BTAI) is a rare but lethal trauma mechanism. Mortality prediction has traditionally relied on logistic regression (LR), yet machine learning (ML) methods may better capture complex nonlinear relationships. Methods: Using the Aortic Trauma Foundation registry, we compared LR, extreme gradient boosting (XGB), and multilayer perceptron (MLP) for two tasks: (1) in-hospital mortality prediction and (2) classification of management strategy, open surgical repair (OSR), thoracic endovascular aortic repair (TEVAR), or medical management alone (MMA). Explanatory LR models identified independent predictors; predictive models targeted ~80% sensitivity in test sets. Performance metrics included area under the receiver operating characteristic curve (AUC), average precision, accuracy, F1 score, precision, recall, and Brier score. Results: Explanatory LR for mortality achieved excellent discrimination (AUC = 0.961), with higher Injury Severity Score, age, and ICU length of stay predicting increased mortality, and conversely, higher Glasgow Coma Scale, pH, and hospital stay protective. Predictive LR and XGB performed similarly (AUCs = 0.849 and 0.855; Brier = 0.083 and 0.101), both outperforming MLP (AUC = 0.757). For management strategy, explanatory LR identified limited predictors (e.g., ISS for OSR; age, heart rate, platelet count, and ISS for TEVAR). Predictive performance was modest: XGB (accuracy = 0.690; macro-AUC = 0.648) slightly outperformed MLP (accuracy = 0.667; macro-AUC = 0.565). Conclusions: Well-specified LR matched XGB for mortality prediction in BTAI, offering interpretability with high accuracy. Management strategy prediction remained limited across models, likely reflecting unmeasured institutional and procedural factors. Enriching datasets with imaging and practice-pattern variables may improve operative decision modeling.

This publication has 12 references indexed in Scilit: