Machine Learning for VO₂max Predictions: A Comparison of Methods using Wearable Sensor Data

Cardiopulmonary exercise testing is the gold standard for assessing VO₂max, but it is costly in terms of time and personnel. Its limitations drive the need for alternative methods of assessment. Using physiological measurements such as heart rate, several machine learning prediction models have been developed to estimate VO₂max. This paper provides the first direct comparison of multiple modelling approaches in a clinical population using wearable sensor data. Wearable ECG and accelerometer data were first pre-processed. We used a signal quality index for ECG data and ML-based physical activity classification. We then extracted known useful features, based on previous literature. Five models (Multiple-Linear Regression (MLR), Support Vector Regression, Random Forest, XGBoost, Multi-layer Perceptron) were compared using 5-fold cross-validation, with performance evaluated via RMSE, R², correlation, and SEE. MLR outperformed other models in predicting VO₂max (R = 0.68±0.09, RMSE = 3.35 ± 0.32). Overall performance in this clinical population was lower than in studies using exercise-derived features in a healthy individuals, but shows that wearable sensor data, including heart rate variability features, can still provide meaningful insight for VO₂max estimations.

Clinical relevance— This study shows how a linear model can estimate VO₂max from ECG and accelerometer data. This model offers better interpretability to more sophisticated machine learning approaches with no cost in performance in this case.