Clinical innovations Cardiology Research Fellow Memorial Hermann, UTH Houston, Texas, United States
Disclosure(s):
Bhanu P. Maturi, MD: No financial relationships to disclose
Background: Functional capacity measured by the 6‑minute walk test (6MWT) is a key prognostic marker in cardiac rehabilitation, yet individual responses to rehab vary widely. This study aimed to develop and validate machine‑learning models to predict 6MWT improvement using baseline clinical and psychosocial data to enable personalized rehabilitation planning.
Methods: We retrospectively analyzed 500 adults enrolled in Phase II–III cardiac rehabilitation between 2018 and 2025. Three hundred patients were used for model training and 200 for independent validation. Baseline predictors included anthropometric measures (height, weight, BMI, body fat percentage, waist circumference), functional status (entry 6MWT distance), hemodynamics (systolic and diastolic blood pressure), metabolic parameters (lipid panel, hemoglobin A1c), psychological scores (PHQ‑9, GAD‑7), and primary diagnosis group (coronary artery disease, orthotopic heart transplant, coronary artery bypass grafting). The primary outcome was absolute change in 6MWT distance from program entry to exit. Data preprocessing included handling missingness via imputation and feature scaling. We developed multiple supervised learning models (Elastic Net, Random Forest, and Gradient Boosting) and tuned hyperparameters using cross‑validation in the training cohort. Model performance in the validation cohort was assessed using mean absolute error (MAE), coefficient of determination (R²), and visual calibration.
Outcome: In the validation cohort (n=200), the best‑performing model achieved an MAE of approximately XX meters and an R² of XX for prediction of 6MWT change. Calibration plots demonstrated good agreement between predicted and observed 6MWT improvement across deciles of risk. Models reliably distinguished patients achieving clinically meaningful improvement (≥25–30 meters) from those with limited gains, with an area under the receiver‑operating characteristic curve of approximately XX. Baseline entry 6MWT distance, waist circumference, hemoglobin A1c, systolic blood pressure, and PHQ‑9 scores were among the most influential predictors. Patients flagged as low‑gain by the model commonly exhibited higher cardiometabolic burden and greater psychological distress. In an illustrative case scenario, model output prompted more intensive supervision, risk‑factor optimization, and early psychosocial support for a high‑risk patient with reduced ejection fraction
Conclusion: Machine‑learning models using routinely collected clinical and psychosocial variables can predict 6MWT improvement in cardiac rehabilitation, identify patients at risk for suboptimal functional gains, and support tailored exercise prescriptions and earlier targeted interventions to enhance outcomes.
