Predicting peak cardiorespiratory fitness and its response to cardiac rehabilitation using machine learning
Jungwon Suh, Bo Ryun Kim, Hyo Kyung Lee, Jae Seung Jung, Hee Jung Kim, Ho Sung Son, You Ha Kwon, Kyung Cheon Seo, Cho Rong Bae, Hongbum Kim, Jong Hoon Kim, Sejeong JangBackground
Peak cardiorespiratory fitness (CRF), commonly quantified as peak oxygen consumption (VO 2 peak), is an important prognostic marker in cardiovascular disease (CVD). This study developed machine-learning (ML) models to estimate VO 2 peak from clinical and functional information and to predict its longitudinal change during cardiac rehabilitation (CR).
Methods
This retrospective longitudinal cohort study included 333 visits from 162 patients with CVD. Task 1 estimated VO 2 peak at eligible visits using 29 clinical and functional candidate predictors without CPET-derived predictors. Task 2 predicted the change in VO 2 peak between consecutive visits (ΔVO 2 peak) using two approaches: Task 2-1 combined clinical, functional, and inter-visit exercise information with the VO 2 peak predicted by Task 1, whereas Task 2-2 combined pre-CR CPET-derived and inter-visit exercise information. Four linear regression approaches and six ML algorithms were evaluated using 5-fold patient-grouped cross-validation.
Results
In Task 1, CatBoost yielded an RMSE of 4.17 ± 0.24 mL·kg -1 ·min -1 ; the 6-minute walk distance, age, Korean Activity Scale Index, hand grip strength, and body mass index were among the highest-ranked features. In Task 2, CatBoost yielded similar internal prediction errors in the pathway incorporating the Task 1-predicted VO 2 peak (Task 2-1: RMSE, 3.46 ± 0.78 mL·kg -1 ·min -1 ) and the CPET-based pathway (Task 2-2: RMSE, 3.39 ± 0.75 mL·kg -1 ·min -1 ). Recent cardiac intervention in Task 2-1 and measured baseline VO 2 peak in Task 2-2 were the highest-ranked features for predicting ΔVO 2 peak.
Conclusions
ML models showed promising internal performance for estimating VO 2 peak and predicting its inter-visit change in this single-center cohort. These results are exploratory and do not establish clinical utility or replacement of CPET. Independent multicenter validation and prospective evaluation are required.