Abstract
INTRODUCTION: Blunt thoracic aortic injury (BTAI) is rare but highly lethal. Despite advances such as thoracic endovascular aortic repair (TEVAR), management remains challenging due to heterogeneous physiologic presentations. Injury grade alone may not capture early derangements influencing risk. In other acute syndromes, machine-learning-derived physiologic phenotypes have informed prognosis. We applied this approach to BTAI to identify phenotypes, examine outcomes, and use explainable AI (XAI) to characterize their physiologic drivers. METHODS: We analyzed 1,375 patients from the Aortic Trauma Foundation registry using eleven admission physiologic and injury severity variables. After plausibility checks and median imputation, K-means clustering (k = 3; silhouette = 0.23) defined physiologic phenotypes. Patients missing >50% physiologic data were excluded; the remaining 1,083 were assigned to the nearest phenotype centroid. Outcomes included in-hospital mortality, treatment modality, hospital and ICU length of stay (LOS), ventilator days, and time to repair. For XAI, an XGBoost classifier was trained to reproduce K-means labels, and SHapley Additive exPlanations (SHAP) quantified feature importance and stability. RESULTS: Three phenotypes emerged: "Stable" (n = 431; preserved hemodynamics, low lactate), "Shock" (n = 398; hypotension, tachycardia, high lactate), and "Neurologically Compromised" (n = 254; depressed Glasgow Coma Score with intermediate hemodynamics). Mortality was 7.2%, 18.9%, and 18.1% (p < 0.001), significantly higher in both high-risk groups vs. Stable. TEVAR predominated (58.2-63.0% across phenotypes); open repair was uncommon (≤6.3%), and medical management was more frequent in Stable (31.1% vs. Shock 19.1%; p = 0.0010). ICU LOS differed overall (p = 0.008), with Neurologically Compromised requiring longer stays. The XGBoost surrogate reproduced K-means phenotypes with high fidelity (accuracy 0.917; macro-F1 0.906; κ 0.873; adjusted Rand index 0.788). SHAP identified systolic blood pressure, lactate, and heart rate as dominant phenotype-defining features, with stable rankings across 200 bootstrap refits. CONCLUSION: Machine learning identified three physiologic BTAI phenotypes with distinct presentations and mortality despite similar management patterns. XAI showed that perfusion and metabolic markers, not anatomy alone, drive phenotype structure. These data reinforce the potential for physiologic phenotyping to enhance prognostication and guide BTAI decision-making.