Machine learning predicts postoperative mortality in a Latin American cardiovascular surgery multicenter cohort: Toward precision risk stratification

OBJECTIVE: A total of 28% of global cardiac surgeries are performed in Latin America; however, surgeons there are faced with many preventable deaths, partly as the result of the lack of accurate, region-specific risk assessment tools. We developed an explainable machine learning (ML) model tailored to a multicenter Latin American cohort that outperformed European System for Cardiac Operative Risk Evaluation II and Society of Thoracic Surgeons scores in predicting postoperative mortality. METHODS: We analyzed 8680 adult cardiovascular surgery cases (2010-2025) from 5 Latin American countries. Sixty perioperative variables trained 5 ML models-random forest, neural network, support vector machine, eXtreme Gradient Boosting, and logistic regression-in R. The models were trained with repeated 10-fold cross-validation (80%) and tested (20%). Performance (area under the receiver operating characteristic curve, sensitivity, specificity, calibration) was compared with traditional scores. Shapley Additive exPlanations values explained predictions at both the cohort and individual levels. RESULTS: The eXtreme Gradient Boosting model showed the greatest accuracy (area under the receiver operating characteristic curve, 0.80; sensitivity, 75.4%, specificity, 73.6%). Shapley Additive exPlanations analysis identified creatinine clearance (0.142) and mean arterial pressure (0.134) as top predictors. Risk factors included chronic kidney disease (+0.25) and malnutrition (+0.22); protective features were mean arterial pressure >80 mm Hg (-0.25) and high-volume centers (-0.20). Low-volume centers (<500 surgeries/year) had greater mortality (6.2% vs 4.2%, P < .001). Synergistic risk factors amplified effects, whereas protective factors mitigated them. High-risk patients were elderly (72.6 ± 10.4 years), with renal impairment (creatinine clearance, 48.3 mL/min) or emergency cases (23% mortality). Predicted high-risk profiles correlated strongly with observed outcomes (83%, P < .001) CONCLUSIONS: This study presents the first explainable ML tool for Latin American cardiovascular surgery, which outperforms existing models and reveals actionable risk factors to guide preoperative care and policy.