Machine learning models for predicting myocardial injury after non-cardiac surgery showed variable discrimination (median AUROC 0.777 to 0.805) and lacked established clinical readiness.
Do machine learning models accurately predict myocardial injury after non-cardiac surgery in adult surgical patients?
Machine learning models for predicting myocardial injury after non-cardiac surgery show moderate discrimination but currently lack robust external validation and do not demonstrate clear superiority over standard regression models.
Effect estimate: Median AUROC 0.777 (internal validation) and 0.805 (external validation)
Myocardial injury after non-cardiac surgery (MINS) is common and often under detected. Machine learning (ML) has been proposed as a tool for perioperative risk stratification and to support targeted postoperative troponin monitoring. This scoping review summarizes current evidence on ML models developed to predict MINS. Five databases were searched in January 2025. Eligible studies applied at least one ML method to predict MINS in adult surgical patients and reported at least one performance metric. Findings were synthesized narratively. Of 2,463 records screened, nine studies met inclusion criteria. Six reported internal validation and three external validation. Median AUROC was 0.777 (IQR 0.770–0.788) for internally validated models and 0.805 (range 0.790–0.821) for externally validated models. Common predictors included age, hemoglobin, renal function markers, perioperative biomarkers, and intraoperative hemodynamic variables. Available supervised prediction models for MINS show variable discrimination, but the evidence base is small, heterogeneous, and largely at high risk of bias. Current studies do not establish clinical readiness or superiority of more complex ML approaches over regression-based models. Standardized outcome ascertainment, transparent reporting, clinically meaningful performance evaluation, and robust external validation are needed before implementation can be considered. • Nine ML studies for MINS prediction were identified in this scoping review • Models showed variable discrimination and limited external validation • Common predictors included age, biomarkers, renal function, and hemodynamics • Most studies were heterogeneous and at high risk of bias • Future MINS tools need standardized troponin surveillance and validation
Nadesan et al. (2026) conducted a review in Myocardial injury after non-cardiac surgery (MINS) (n=9). Machine learning prediction models vs. Regression-based models was evaluated on Model discrimination for predicting MINS (Median AUROC 0.777 (internal validation) and 0.805 (external validation)). Machine learning models for predicting myocardial injury after non-cardiac surgery showed variable discrimination (median AUROC 0.777 to 0.805) and lacked established clinical readiness.