Estimating oxygen uptake in simulated team sports using machine learning models and wearable sensor data: A pilot...

Area of Science:

• Sports Science
• Biotechnology
• Machine Learning

Background:

• Optimizing training status and reducing injury risk in team sports requires accurate physiological monitoring.
• Current methods for assessing physiological demand can be invasive or impractical during training.

Purpose of the Study:

• To investigate the feasibility of using machine learning (ML) models with wearable sensors to estimate oxygen uptake (VO2) in team sports athletes.
• To compare the performance of various ML models for VO2 estimation using different sensor data and features.

Main Methods:

• Six male team sports athletes underwent incremental fitness tests.
• Data collected included inertial measurement units (IMU), heart rate, and breathing rate.
• Multiple ML models (MLR, XGBoost, LSTM, CNN, MLP) were trained and evaluated using raw and engineered IMU features.

Main Results:

• Long Short-Term Memory (LSTM) models using raw IMU data achieved the highest accuracy in VO2 prediction (RMSE: 4.976, MAE: 3.698).
• Multiple Linear Regression (MLR) models showed competitive performance, especially with engineered features.
• Multi-sensor setups, including torso and limb placements, improved prediction accuracy.

Conclusions:

• ML models show significant potential for non-invasive, real-time VO2 monitoring in team sports.
• Wearable sensor data combined with ML can provide valuable insights into athletes' physiological demands.
• This approach could aid in optimizing training and mitigating injury risk.