A Machine Learning Approach to Non-Exercise Prediction of Peak VO2 in Advanced Heart Failure

Introduction: Peak oxygen consumption (VO 2 ), assessed via cardiopulmonary exercise testing (CPET), is central to heart transplant evaluation in advanced heart failure. However, many patients have limitations that prevent them from completing CPET. This study was aimed at developing machine learning (ML) models to estimate peak VO 2 by using non-exercise clinical data. Methods: We retrospectively analyzed 524 heart transplant candidates who underwent comprehensive clinical, laboratory, echocardiographic, and hemodynamic assessment. The Boruta algorithm selected key predictors among 55 variables. Multiple ML algorithms were trained to estimate peak VO 2 , and performance was assessed according to root mean square error (RMSE), mean absolute error (MAE), and R 2 . Feature importance was evaluated with SHAP values. Results: Models were developed with two sets of variables: a full model (17 features) and a simplified model (seven features). Extreme Gradient Boosting performed best overall (RMSE: 2.44, MAE: 1.89, R 2 : 0.689). In the simplified model, gradient boosted decision trees were optimal (RMSE: 3.20, MAE: 2.54, R 2 : 0.481). Key predictors included TAPSE/PASP, mixed venous saturation, proBNP, age, hemoglobin, pulmonary vascular resistance, and BMI. Conclusion: ML-based models with non-exercise data can estimate peak VO 2 with reasonable accuracy, thus offering a practical tool for evaluating transplant candidates unable to undergo CPET.