Predicting early complications after surgical aortic valve replacement: A machine learning-based approach

Introduction Postoperative complications are a frequent concern following surgical aortic valve replacement (SAVR) for severe aortic stenosis. These adverse events contribute significantly to early postoperative morbidity and mortality, despite advances in surgical techniques and perioperative care. Identifying early predictive factors remains essential to guide clinical decision-making and reduce risk. Objective To predict the occurrence of early postoperative complications or death after SAVR using a machine learning approach and to identify the most influential predictive parameters. Method We conducted a retrospective, observational study at our center, including patients operated on for severe aortic stenosis between January 2023 and September 2024. The primary endpoint was a composite of death or postoperative complication occurring within 30 days. Complications included stroke, atrial fibrillation, atrioventricular block, nosocomial infection, pericardial effusion, and surgical reintervention. Patients were divided into two groups: class 0 (no complication or death) and class 1 (at least one complication or death). Data included clinical, echocardiographic (pre- and postoperative), and intraoperative variables. A supervised k-nearest neighbors (k-NN) algorithm was used for classification. Additionally, a Mann-Whitney test was used to compare LV mass between groups, ( Figure 1 ). Results Among the variables tested, the k-NN model identified four main predictors: preoperative left ventricular (LV) mass, mean transprosthetic gradient, postoperative prosthetic valve area, and cardiopulmonary bypass time. In contrast, variables such as the presence of comorbidities, LV ejection fraction, and global longitudinal strain (GLS) did not improve prediction and were not retained by the model. The algorithm demonstrated excellent performance for the no-complication group prediction, with 89% accuracy, 100% precision, and 100% recall. However, due to class imbalance, it showed limited ability to predict complications. The Mann-Whitney test showed a significantly lower mean LV mass in patients with complications (136.97 g/m 2 ) compared to those without (174.56 g/m 2 ), with P = 0.042, ( Figure 1 ). Conclusion Preoperative LV mass, bypass duration, and postoperative prosthetic parameters are associated with early complications or death after SAVR. The k-NN algorithm identified patients at low risk, but failed to capture high-risk cases, likely due to dataset imbalance.