Abstract 4357549: Multi-Center Validation of a Non-exercise Machine Learning Prediction Model Equivalent to Cardiopulmonary Exercise Parameters in Cardiovascular Patients

Abstract

Background: A symptom-limited incremental cardiopulmonary exercise testing (CPET) is rigidly applied to patients with cardiovascular disease for assessing exercise capacity and prognosis, yet its widespread use is limited by the need for sophisticated equipment and specialists. We developed a machine learning (ML) prediction model for anaerobic threshold heart rate (AT-HR) and peak oxygen uptake (VO 2peak ) using over 20,000 CPET cases, which provides these measures from non-exercise clinical and demographic variables. This study aimed to validate the ML derived algorithm on multi-center cohorts. Methods: We analyzed CPET data during 2018 and 2024 from patients with cardiovascular disease across three high volume institutions: Sakakibara Heart Institute (n=336), St. Marianna University Hospital (n=300), and Showa Medical University Fujigaoka Hospital (n=101). The testing protocols were the same among three institutes. Gradient boosting regression created the ML model incorporating non-exercise 78 features for the predictions of AT-HR and VO 2peak . Predictive accuracy was assessed by the mean of the absolute error (MAE) and R 2 across institutions and disease subgroups (ischemic heart disease excluding post-coronary artery bypass grafting (CABG), post-CABG, post valvular surgery, heart failure, and aortic disease). Results: The mean age was 67 ± 29 years, with 552 participants (75%) being male. Although the institutions differed in patient demographics and disease profiles, the ML algorithm demonstrated high predictive accuracy across centers and subgroups of each diseases. AT-HR and VO 2peak difference between measured and predicted by ML ranged from 6.37 to 7.32 bpm, and from 2.64 to 1.88 mL/kg/min, respectively across disease categories, with consistent R 2 values indicating strong model performance. Bland-Altman plots confirmed minimal bias. The algorithm effectively and equally applied to each institutions despite differences in equipment, patient characteristics, and clinical staffs. Conclusions: This multi-center validation study confirms the robustness and clinical applicability of a ML algorithm. The model enables practical prediction of key CPET parameters without the need for exercise testing, offering a standardized, accessible tool for risk stratification and management of cardiovascular disease. This approach has the potential to expand the use of CPET-derived insights into diverse clinical situations.