Integrating CPET data in ML models predicted mortality in 1,531 VHD patients with AUCs of 0.974 (Random Forest) and 0.979 (XGBoost); peak VO2/kg <14 indicated highest risk.
Can a machine learning framework integrating cardiopulmonary exercise testing (CPET) parameters with clinical data accurately predict mortality risk and MACE in patients with valvular heart disease?
An interpretable machine learning framework integrating CPET data, specifically peak VO2/kg, demonstrates strong predictive performance for mortality risk stratification in patients with valvular heart disease.
Tasa de eventos absoluta: 0% vs 0%
Abstract Background Valvular heart disease (VHD) is a leading cause of mortality worldwide, particularly among aging populations. Despite advancements in clinical risk assessment, predictive models integrating cardiopulmonary exercise testing (CPET) parameters with clinical data remain limited. This study aims to develop an interpretable machine learning (ML) framework to predict mortality risk in VHD patients and identify key prognostic drivers to improve clinical decision-making. Methods We analyzed electronic health records from 1,531 VHD patients, incorporating demographic, clinical, and CPET data. Model-based clustering was employed to stratify patients into subgroups based on prognosis. SHapley Additive exPlanations (SHAP) analysis was used to identify the most influential predictors of major adverse cardiovascular events (MACE). Random Forest and XGBoost models were developed and optimized to predict mortality risk, with performance evaluated using area under the curve (AUC). Decision tree analysis was applied to establish risk stratification thresholds based on peak VO2/kg. Kaplan-Meier survival analysis was conducted to compare outcomes across risk groups. Results The cohort was stratified into three risk subgroups (low, medium, high) with distinct survival profiles (p0.001). SHAP analysis identified peak VO2/kg and age as the most significant predictors of MACE. The Random Forest and XGBoost models achieved AUCs of 0.974 and 0.979, respectively, demonstrating strong predictive performance. Decision tree analysis validated peak VO2/kg thresholds (14, 14–23, ≥23 mL/kg/min) for risk stratification, with high-risk patients (peak VO2/kg 14) exhibiting the poorest survival outcomes. Kaplan-Meier curves confirmed the prognostic value of these thresholds, with low-risk patients (peak VO2/kg ≥23) showing optimal survival. Over a median follow-up duration of 45.6 months, these risk subgroups maintained their distinct survival trajectories. Conclusions This study highlights the clinical utility of integrating CPET data with ML for mortality risk prediction in VHD patients. The interpretable ML framework, combined with SHAP analysis, provides actionable insights into key prognostic factors, enabling personalized risk assessment and intervention strategies. Future research should focus on multi-center validation and the incorporation of deep learning techniques to further enhance predictive accuracy and generalizability.
Ma et al. (Sat,) reported a other. Integrating CPET data in ML models predicted mortality in 1,531 VHD patients with AUCs of 0.974 (Random Forest) and 0.979 (XGBoost); peak VO2/kg <14 indicated highest risk.