LAUSR

ABSTRACT Background: Maximal oxygen uptake (VO2 max), an indicator of cardiorespiratory fitness (CRF), requires exercise testing and, as a result, is rarely ascertained in large-scale population-based studies. Non-exercise algorithms are cost-effective methods to estimate VO2 max, but the existing models have limitations in generalizability and predictive power. This study aims to improve the non-exercise algorithms using machine learning (ML) methods and data from U.S. national population surveys. Methods: We used the 1999-2004 data from the National Health and Nutrition Examination Survey (NHANES), in which a submaximal exercise test produced an estimate of the VO2max. We applied multiple supervised ML algorithms to build two models: a parsimonious model that used variables readily available in clinical practice, and an extended model that additionally included more complex variables from more Dual-Energy X-ray Absorptiometry (DEXA) and standard laboratory tests. We used Shapley additive explanation (SHAP) to interpret the new model and identify the key predictors. For comparison, existing non-exercise algorithms were applied unmodified to the testing set. Results: Among the 5,668 NHANES participants included in the final study population, the mean age was 32.5 years and 49.9% were women. Light Gradient Boosting Machine (LightGBM) had the best performance across multiple types of supervised ML algorithms. Compared with the best existing non-exercise algorithms that could be applied in NHANES, the parsimonious LightGBM model (RMSE: 8.51 ml/kg/min [95% CI: 7.73 -9.33]) and the extended model (RMSE: 8.26 ml/kg/min [95% CI: 7.44 -9.09]) significantly reducing the error by 15% (P <0.01) and 12% (P<0.01 for both), respectively. Conclusion: Our non-exercise ML model provides a more accurate prediction of VO2 max for NHANES participants than existing non-exercise algorithms. Keywords: Machine learning, GBDTs, Cardiorespiratory fitness, VO2max, NHANES