Abstract 4137678: Maximal Oxygen Uptake Prediction From Resting and Submaximal Variables of Cardiopulmonary Exercise Testing by Using Machine Learning Methods

Background: Cardiorespiratory fitness, as measured by maximal oxygen consumption (VO2max) during a cardiopulmonary exercise test (CPET), is a prognostic indicator for longevity and adverse cardiovascular event prevention. This study aims to formulate a regression model utilizing resting and submaximal variables during CPET evaluation in the general population to predict VO2max. Methods: We used 13,535 CPET results on cycle ergometer collected by the UCLA Exercise Physiology Research Laboratory over eight years. Patients were divided into a normal group (NG; n=1,400) and an other group (OG; n=12,135). The inclusion criteria for the NG were: the absence of any known clinical diagnosis, BMI <30, no current use of beta blocker, and VO2max >85% of the predicted value for age and sex (Wasserman equation). OG refers to all other patients not in NG. Models were trained and evaluated for each group using stratified 5-fold cross-validation. We also trained different models using only resting variables (R-VARS) and resting and submaximal exercise variables (R+S-VARS). Feature importance was assessed using Shapely additive explanations (SHAP) values to identify how the feature related to the VO2max prediction. Results: The regression models were trained on the NG, OG, and total (N+OG) groups. The optimal models were Bayesian Ridge for the NG and Light Gradient Boosting Machine for the other two groups. The mean (standard deviation) R2 when using only R-VARS was 0.67(0.037) for the NG, 0.54 (0.014) for the OG, and 0.55 (0.009) for the N+OG. When using R+S-VARS, performance increased to 0.82 (0.014) for the NG, 0.80 (0.010) for the OG, and 0.8 (0.008) for the N+OG. Chronotropic index (ci), body mass index (BMI), VO2 at the first ventilatory threshold (VO2h1kg), minute ventilation at the second ventilatory threshold (VEH2), and forced expiratory volume (FEV1) were important features across the models trained with R+S-VARS. CI, BMI, and VEH2 had a negative effect, while VO2h1kg, FEV1 had positive effect on VO2max prediction. Conclusion: Our VO2max prediction model demonstrated remarkable accuracy in this dataset representing a generally healthy, non-athlete population using R+S-VARS. Particularly noteworthy was the enhanced performance within subgroups exhibiting a lower VO2max. This methodology offers a means to assess VO2max for individuals who might not achieve maximal exhaustion during CPET due to non-cardiopulmonary reasons.