Machine Learning versus Penalized Logistic Regression for Predicting In-Hospital Mortality in Ruptured Abdominal Aortic Aneurysm: A Comparison Using a Uniform Validation Framework.

BACKGROUND: Ruptured abdominal aortic aneurysm (rAAA) remains associated with substantial in-hospital mortality. Although machine-learning methods can model complex nonlinear relationships between admission characteristics and outcome, their incremental value over conventional regression remains uncertain. We compared 3 prediction models using identical admission variables and a uniform validation framework.

METHODS: This retrospective single-center cohort included 196 unique patients with rAAA managed between 2006 and 2024. Five prespecified admission predictors were evaluated: age, hypovolaemic shock, maximal aneurysm diameter, hemoglobin, and systolic blood pressure. Missing data were imputed independently within each training partition. Penalized logistic regression, gradient boosting, and a multilayer perceptron (MLP) were compared using nested five-fold cross-validation repeated 5 times. Secondary temporal validation was performed in the 181 patients with a recoverable treatment year: models were developed in the 2006-2018 cohort (n = 110) and evaluated, without refitting or recalibration, in the 2019-2024 cohort (n = 71).

RESULTS: In-hospital mortality occurred in 99/196 patients (50.5%). Gradient boosting and penalized logistic regression achieved comparable moderate discrimination, with area under the curves (AUCs) of 0.739 (95% confidence interval [CI] 0.666 to 0.805) and 0.729 (95% CI, 0.654 to 0.796), respectively. Their paired AUC difference was 0.010 (95% CI, -0.024 to 0.045). The MLP showed lower discrimination (AUC, 0.643; 95% CI, 0.563-0.719), with a paired difference versus penalized logistic regression of -0.086 (95% CI, -0.160 to -0.013). In temporal validation, AUCs were 0.800 (95% CI, 0.686 to 0.904) for gradient boosting, 0.739 (95% CI, 0.607 to 0.858) for penalized logistic regression, and 0.632 (95% CI, 0.496 to 0.762) for the MLP. Temporal calibration demonstrated overprediction of absolute mortality risk, with observed mortality of 40.8% compared with mean predicted mortality of 49.9%, 50.0%, and 63.4%, respectively.

CONCLUSION: Using a uniform internal and temporal validation framework, gradient boosting achieved the highest numerical performance but did not demonstrate a conclusive advantage over penalized logistic regression. The MLP provided no incremental predictive benefit. These findings indicate that increasing algorithmic complexity does not necessarily improve mortality prediction in modest-sized emergency vascular datasets and support interpretable regression as an essential benchmark. External multicenter validation and more complete prospective data collection are required before clinical implementation.