300 Section - Personalizing Exercise Capacity and Prescription in Cardiac Rehabilitation Through a Data-Driven Approach

Accurate exercise capacity assessment is vital for personalized cardiac rehabilitation (CR), but CPET is often limited by cost and feasibility. This study aimed to develop and validate a machine learning model to predict key exercise capacity parameters without CPET, using routinely collected clinical, functional, and behavioral data during CR.

Design:

We retrospectively analyzed data from 221 patients who underwent CR between March 2020 and May 2025. Data from visits 1 to 6 (every 2–3 months) were used to predict VO₂ max, MET max, Fat max, and recommended MET levels. Predictors included baseline clinical characteristics (demographics, comorbidities, procedures, vitals, lifestyle factors) and repeated measures: patient-reported outcomes (EQ-5D, KASI), functional tests (handgrip strength, 6MWD), body composition, and exercise behavior (frequency/duration of strength, walking, cycling). Longitudinal trends were captured via differencing and rolling averages. Missing data were imputed with MICE, and features normalized. Feature selection and hyperparameter tuning were performed via Bayesian optimization. An XGBoost model trained on top-ranked features was evaluated via 5-fold stratified group cross-validation. Performance was assessed via RMSE, MAE, and MAPE.

Results:

The model demonstrated strong performance:

• VO₂ max – RMSE 3.91, MAE 3.08, MAPE 14.17%

• MET max – RMSE 1.11, MAE 0.86, MAPE 13.99%

• Fat max – RMSE 0.71, MAE 0.56, MAPE 16.82%

• Recommended MET (lower) – RMSE 1.20, MAE 0.84, MAPE 20.60%

• Recommended MET (upper) – RMSE 1.41, MAE 0.99, MAPE 19.97%

SHAP analysis highlighted both static (e.g., age, body fat %, SMI, resting BP) and dynamic variables (e.g., 6MWD, KASI, exercise frequency/duration) as key predictors. KASI and 6MWD, at baseline and over time, consistently ranked highest. Changes in exercise behavior further improved predictions. The model performed consistently across subgroups, supporting clinical applicability.

 

Conclusion:

This model accurately estimated exercise capacity from routine CR data, offering a practical alternative to CPET for personalized exercise prescriptions.