Researchers have developed and internally validated a machine learning model capable of predicting 30-day mortality and major postoperative complications in adult patients undergoing surgery for left-sided infective endocarditis. The study, published in BMC Surgery, addresses a clinical scenario where surgical intervention is frequently essential but preoperative risk assessment remains challenging. Using retrospective data from adult patients treated between May 2009 and May 2025, the research team constructed and evaluated multiple predictive pipelines to support surgical risk stratification.

The retrospective cohort included 992 adult patients, among whom 100 individuals, or 10.1 percent, experienced the primary study outcome. This primary endpoint was defined as a composite measure encompassing all-cause death or major postoperative complications occurring within 30 days of surgery. The observed complications included postoperative stroke, malignant arrhythmia, prolonged mechanical ventilation, acute kidney injury requiring dialysis, shock, and multiple organ dysfunction syndrome.

To construct the predictive framework, the investigators implemented a machine learning-based feature selection pipeline that identified ten key predictors from patient data. The development process evaluated eight distinct machine learning algorithms with six strategies designed to address class imbalance. Candidate performance was initially assessed in the training cohort using out-of-fold predictions to evaluate discrimination, calibration, and overall prediction error.

Out of the 56 evaluated candidate models, 13 met predefined screening criteria during out-of-fold evaluation. Among these qualifying approaches, a categorical boosting algorithm trained with random over-sampling, designated CatBoost-RO, demonstrated the lowest out-of-fold Brier score. Based on this superior overall performance in mitigating prediction error, the research team selected the CatBoost-RO architecture as their final predictive model for subsequent validation.

The model also exhibited favorable calibration metrics, recording a calibration intercept of -0.036 with a 95 percent confidence interval between -0.474 and 0.401, alongside a calibration slope of 0.995 with a 95 percent confidence interval from 0.658 to 1.331.

In addition to discrimination and calibration metrics, the CatBoost-RO model registered a low overall prediction error in the test dataset, reflected by a Brier score of 0.065 and a corresponding 95 percent confidence interval of 0.053 to 0.078.

To provide model interpretability for clinical practice, the researchers conducted SHapley Additive exPlanations analysis on the final model architecture. This feature attribution analysis identified circulating total bilirubin levels and N-terminal pro-B-type natriuretic peptide concentrations as the most influential predictors driving postoperative risk estimates. To facilitate practical deployment, the authors also developed an online computational tool based on the final algorithm.

The authors concluded that the CatBoost model trained with random over-sampling provides favorable discrimination, reliable calibration, and low prediction error for estimating 30-day death or major surgical complications in patients with left-sided infective endocarditis. While the algorithm may offer valuable support for preoperative clinical decision-making and risk stratification, the investigators emphasized that external validation in independent cohorts remains warranted.

Subject of Research: Machine learning prediction of 30-day postoperative outcomes in left-sided infective endocarditis

Article Title: Machine learning-based prediction of 30-day outcomes after surgery for left-sided infective endocarditis

Article References: Tang, H., Wang, B., Li, T., An, Q., Tang, M., & Ling, Y. (2026). Machine learning-based prediction of 30-day outcomes after surgery for left-sided infective endocarditis. BMC Surgery. https://doi.org/10.1186/s12893-026-04287-z

Image Credits: AI Generated

DOI: 10.1186/s12893-026-04287-z

Keywords: infective endocarditis, cardiac surgery, machine learning, CatBoost, random over-sampling, postoperative complications, 30-day mortality, total bilirubin, N-terminal pro-B-type natriuretic peptide, stroke, malignant arrhythmia, acute kidney injury, prolonged ventilation, multiple organ dysfunction syndrome, SHAP analysis, Brier score, preoperative risk assessment