Comparison Between Artificial Intelligence-Based Models and Traditional Risk Scores for Predicting Risks in Adult Cardiothoracic Surgery: A Systematic Review.

INTRODUCTION: Conventional risk scores like EuroSCORE II and Society of Thoracic Surgeons models, derived from logistic regression, may not fully represent complex interactions within cardiac surgery cohorts. Using nonlinear modeling, machine learning (ML) may improve risk prediction by capturing complex relationships. OBJECTIVE: To conduct a systematic review of studies (2020-2026) that compare ML models with traditional risk scores for predicting perioperative mortality or significant adverse events in adult cardiac surgery. METHODS: Following the PRISMA 2020 guidelines, PubMed, Google Scholar, and Cochrane were searched from January 2020 to January 2026. The protocol was registered in PROSPERO (CRD420261295268; registered January 28, 2026). Inclusion necessitated direct comparisons between ML and traditional scoring methodologies with reported performance metrics, such as the area under the curve (AUC). Prediction model Risk Of Bias ASsessment Tool+ artificial intelligence was utilized for bias risk assessment. RESULTS: Thirteen studies were included (N = 308-647,726). Studies originated from China (n = 5), United States (n = 3), United Kingdom (n = 2), Colombia, Saudi Arabia, and Turkey (n = 1 each). Algorithms included extreme gradient boosting, random forest, and ensembles. ML models showed improved or comparable performance, with AUC differences of 0.006-0.42, improved calibration, and reclassification (net reclassification improvement 0.550). Extreme gradient boosting AUC was 0.96 for postoperative infection, and random forest AUC was 0.975 for major adverse events in type A dissection. Prediction model Risk Of Bias ASsessment Tool+ artificial intelligence indicated low risk of bias; external validation was limited, and some analyses raised concerns regarding overfitting. Key methodological limitations included data leakage, overfitting risk, and limited temporal and external validation. CONCLUSIONS: ML models may improve risk prediction in cardiac surgery relative to traditional scores, especially through ensembles and advanced validation techniques. Prospective multicenter validation and evaluation of clinical integration and algorithmic fairness are needed before widespread implementation, with attention to data leakage prevention, overfitting mitigation, temporal drift, and calibration on independent validation sets.