Early postoperative prediction of gastrointestinal bleeding after surgery for acute type A aortic dissection: a five-variable logistic regression model benchmarked against machine learning

Abstract

Gastrointestinal bleeding is a clinically important complication after emergency surgery for acute type A aortic dissection. Dedicated postoperative risk-stratification tools for this high-risk population remain limited. We aimed to develop and internally validate an early postoperative prediction model for gastrointestinal bleeding after acute type A aortic dissection repair and to compare logistic regression with commonly used machine-learning classifiers. We conducted a single-centre retrospective cohort study of 1,345 consecutive patients who underwent emergency acute type A aortic dissection repair between 5 December 2017 and 4 August 2022. Comprehensive clinical data, laboratory findings, imaging results, operative variables, transfusion data, and postoperative outcomes were collected. The cohort was divided into training and validation sets at a 70:30 ratio using stratified sampling. Candidate predictors were screened using least absolute shrinkage and selection operator logistic regression, and a clinically interpretable five-variable logistic regression model was developed. Model performance was assessed by discrimination, calibration, decision-curve analysis, repeated-split analysis, temporal validation, and comparison with tuned machine-learning classifiers. Overall, 108 patients (8.0%) developed postoperative gastrointestinal bleeding, which was associated with higher in-hospital mortality (26.9% vs. 8.2%; p < 0.001). The final model included red blood cell transfusion volume, lactate, cardiac troponin I, coronary artery disease, and continuous renal replacement therapy. The logistic regression model achieved an area under the receiver operating characteristic curve of 0.834 in the training set and 0.832 in the validation set, with a validation Brier score of 0.061. Calibration estimates were imprecise, supporting local recalibration before implementation. Temporal validation showed a test area under the curve of 0.869. Tuned machine-learning classifiers showed similar discrimination but no clinically meaningful advantage in calibration, decision-curve performance, or interpretability. A five-variable logistic regression model provided stable and clinically interpretable early postoperative risk stratification for gastrointestinal bleeding after acute type A aortic dissection repair. In this structured, low-event clinical dataset, machine-learning classifiers did not materially outperform a transparent regression model. External multicentre validation and local recalibration remain warranted before routine clinical implementation.