A machine learning approach predicts improvement of physical exercise capacity based on pulse wave analysis in coronary artery disease patients.

Abstract

Background: Patients with stable coronary artery disease (CAD) have a residual risk of adverse events and all-cause mortality. Enhancing exercise capacity by exercise training (ET) during cardiac rehabilitation (CR) is a Class 1A guideline recommendation. However, there is a high number of ET non-responders (NR) in CR. We aimed to develop a machine learning (ML) prediction model for the early identification of NR using baseline cardiopulmonary exercise testing (CPET) and pulse wave analysis (PWA) data.

Methods: Participants included 393 CAD patients after myocardial infarction and/or percutaneous coronary intervention and/or coronary artery bypass graft who underwent 3-4 weeks of CR; CPET was conducted at the beginning and end of CR. Responders (R) were defined as participants who demonstrated an increase in exercise capacity (peak oxygen uptake (V̇O₂_peak)) greater than 1 typical error away from 0, all other patients were defined as NR. Only baseline data including diagnosis, medication, PWA, and CPET data were used for modeling, and ML models included 10 different supervised algorithms. The dataset was split into training and test sets, and 10-fold cross-validation was used. Recursive Feature Elimination was used for feature selection to reduce dimensionality and improve generalizability. Independent external validation was performed in a dataset of CAD patients (n = 120) enrolled at 2 different centers (Germany and Spain). Predictions were explained using the model-agnostic SHapley Additive exPlanation methodology.

Results: After data cleaning, 353 patients (20.4% women) with age of 55.8 ± 7.1 years (mean ± SD) were included for analysis, and 225 patients (63.7%) were classified as NR (22% women; age: 56.2 ± 7.1 years). ET participation rates were similar (R: 93.6% ± 7.5%; NR: 92.6% ± 9.3%; p = 0.76). For the prediction model, the Random Forest classifier provided the best mean balanced accuracy of 77.0%. The most influential features were breathing reserve/frequency, oxygen uptake combined with pulse wave velocity, cardiac output, and augmentation time. Of note, primary diagnosis and disease severity had only limited influence on the model. External validation of the Random Forest model showed 82.8% accuracy, with high specificity and moderate sensitivity in long-term outcome prediction.

Conclusion: The developed ML-based model enables an early identification of ET NR, allowing individual patient-centered ET adaptations to improve CR.