Optimally Predicting Mortality in Patients with Abdominal Aortic Aneurysms

Abdominal aortic aneurysm (AAA) patients in the ICU represent a heterogeneous, high-risk population with mortality risk evolving across distinct clinical phases. Existing prognostic tools rely largely on Cox proportional hazards (Cox PH) nomograms with narrow predictor sets and single time horizons, leaving the value of modern machine learning, extended features, and external generalizability uncharacterized. We extracted an ICD-coded AAA ICU cohort from MIMIC-IV v2.2 (858 patients with complete six-predictor admission data: age, BUN, sepsis, antihypertensive use, anion gap, mean SpO2) using a 24-hour admission window. An extended feature set added hemodynamic, laboratory, and comorbidity variables, with feature selection via LASSO and SVM-RFE intersection. Six models (Cox PH, logistic regression, random forest, gradient boosting, XGBoost, MLP) were trained on a 70% split and evaluated at 7-, 14-, and 28-day horizons using ROC-AUC, C-index, Brier score, calibration, and SHAP. External validation used a harmonized eICU-CRD cohort. In-hospital mortality was [~]11.8%. On the six-predictor set, logistic regression led at 7 days (AUC 0.866) and 14 days (AUC 0.872), with XGBoost competitive. Extended features yielded modest gains; random forest achieved the best 28-day AUC (0.892). The MLP consistently underperformed. Discrimination declined monotonically with longer horizons. External validation showed expected attenuation (best 7-day AUC 0.771). SHAP consistently identified anion gap, BUN, and age as top contributors. We conclude that regularized linear models excel under data scarcity, while tree ensembles gain advantage as features and horizons expand. External results motivate local recalibration before deployment.