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

Area of Science:

• Cardiovascular Physiology
• Biomedical Engineering
• Machine Learning in Healthcare

Background:

• Cardiopulmonary exercise testing (CPET) is the gold standard for VO2max assessment but is resource-intensive.
• Limitations of CPET necessitate alternative, accessible methods for VO2max estimation.
• Machine learning models using physiological data show promise for estimating VO2max.

Purpose of the Study:

• To directly compare multiple machine learning modeling approaches for VO2max estimation.
• To evaluate these models using wearable sensor data in a clinical population.
• To determine the most effective modeling strategy for VO2max prediction from non-exercise data.

Main Methods:

• Wearable ECG and accelerometer data were pre-processed, including signal quality assessment and activity classification.
• Relevant physiological features were extracted based on existing literature.
• Five machine learning models (MLR, SVR, Random Forest, XGBoost, MLP) were compared using 5-fold cross-validation.

Main Results:

• Multiple-Linear Regression (MLR) demonstrated superior performance in predicting VO2max (R = 0.68±0.09, RMSE = 3.35 ± 0.32).
• Performance in this clinical cohort was lower than in studies using exercise-derived features in healthy individuals.
• Wearable sensor data, including heart rate variability, provided meaningful VO2max estimation insights.

Conclusions:

• A linear model (MLR) can effectively estimate VO2max from ECG and accelerometer data in a clinical setting.
• MLR offers superior interpretability compared to more complex machine learning models without sacrificing performance.
• This approach provides a cost-effective and accessible alternative for VO2max assessment in clinical populations.