Development and validation of machine learning nomograms for predicting mortality after cardiac valve surgery

To develop and validate machine learning (ML) models for predicting mortality after heart valve surgery and compare their performance to the conventional EuroSCORE II, with the final goal of creating clinically accessible and easy-to-use nomograms.

This multicenter, retrospective cohort study included 935 adult patients who underwent heart valve surgery. All-cause mortality at in-hospital, 30-day, and 365-day post-operative intervals were the primary outcomes. The Boruta algorithm was employed for feature selection. Five models, Logistic Regression, XGBoost, Random Forest, Extra Trees, and EuroSCORE II (as a benchmark), were developed on a 70% training set and validated on a 30% hold-out test set. Model performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (ROC AUC), sensitivity, and specificity. The best-performing model for each endpoint was converted into a nomogram.

Machine learning models demonstrated strong discriminative performance across all endpoints. For in-hospital mortality, the Extra Trees model achieved the highest discrimination (ROC-AUC 0.858). For 30-day mortality, Logistic Regression showed the best performance (ROC-AUC 0.800), substantially exceeding EuroSCORE II (ROC-AUC 0.610). For 365-day mortality, predictive performance was comparable across models, with EuroSCORE II demonstrating similar discrimination (ROC-AUC 0.787). Key predictive features consistently included age and biomarkers reflecting cardiac stress, renal function, and hepatic function. The derived nomograms exhibited good discrimination and calibration in both internal and external validation cohorts.

Machine learning models, particularly ensemble approaches, improved short-term mortality prediction following valve surgery compared with EuroSCORE II. The developed nomograms provide a practical and interpretable tool for individualized perioperative risk stratification.