Key result
Machine learning integrating CPET parameters predicts early postoperative pulmonary complications with ~0.75 AUC.
Why the study?
Postoperative pulmonary complications are common after cardiac valve surgery, motivating the development and internal validation of an interpretable machine learning model that integrates cardiopulmonary functional assessments to predict early postoperative pulmonary complications.
Can a machine learning model integrating cardiopulmonary functional assessments and perioperative biomarkers accurately predict early postoperative pulmonary complications in adult patients undergoing cardiac valve surgery?
Cohort (n=818)
No
Can a machine learning model integrating cardiopulmonary functional assessments and perioperative biomarkers accurately predict early postoperative pulmonary complications in adult patients undergoing cardiac valve surgery?
Effect estimate: AUC 0.746 (95% CI 0.670-0.823)
An interpretable machine learning model integrating CPET-derived functional parameters and perioperative biomarkers provides moderate accuracy for predicting early postoperative pulmonary complications following cardiac valve surgery.
May support preoperative risk stratification in valve surgery; leaves open external validation before clinical adoption.
Background Postoperative pulmonary complications (PPCs) are common after cardiac valve surgery. This study aimed to develop and internally validate an interpretable machine learning (ML) model integrating cardiopulmonary functional assessments to predict early PPCs. Methods A total of 818 adult patients undergoing isolated valve surgery were randomly split into training (n = 655) and validation (n = 163) cohorts. Predictors were selected using LASSO regression, and six ML models were compared. Model performance was assessed by discrimination (AUC), Brier score, and decision curve analysis. Shapley Additive Explanations (SHAP) was used for model interpretation. Results PPCs occurred in 43.8% of patients. The training cohort included 283 PPC events, while 75 occurred in the validation cohort. LASSO regression identified 12 predictors: cardiopulmonary bypass (CPB) duration, cystatin C, age, blood glucose, Physical Component Summary (PCS) score, oxygen uptake per work rate (VO2/WR), ventilatory equivalent for carbon dioxide (VE/VCO2 slope), serum potassium, left ventricular ejection fraction (LVEF), C-reactive protein (CRP), fibrinogen, and history of valve replacement. The LightGBM model showed the best performance in the validation cohort (AUC = 0.746, 95% CI: 0.670–0.823; sensitivity = 0.653; specificity = 0.705; F1 score = 0.653). SHAP analysis highlighted CPB duration, cystatin C, PCS score, and CPET metrics as key contributors to PPC risk. Conclusions This study established a procedure-specific, interpretable ML-based prediction model for early PPCs following cardiac valve surgery. By integrating CPET-derived cardiopulmonary reserve assessment with perioperative biomarkers, the model may facilitate improved perioperative risk stratification and support targeted clinical management.
No takes yet. Share an insight, caveat, or question.
Jiang et al. (2026) conducted a cohort in Cardiac valve surgery (n=818). Machine learning prediction model (LightGBM) was evaluated on Prediction of early postoperative pulmonary complications (PPCs) within 7 days (AUC 0.746, 95% CI 0.670-0.823). The LightGBM machine learning model integrating CPET-derived cardiopulmonary reserve parameters and perioperative clinical variables predicted early postoperative pulmonary complications with an AUC of 0.746.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: