Attention-Enhanced CNN-LSTM Model for Exercise Oxygen Consumption Prediction with Multi-Source Temporal Features

Area of Science:

• Physiology
• Biomedical Engineering
• Machine Learning

Background:

• Dynamic oxygen uptake (VO2) is crucial for exercise science and clinical applications.
• Existing methods face challenges in fusing diverse sensor data and modeling temporal patterns.

Purpose of the Study:

• To enhance VO2 prediction by integrating wearable accelerometer and heart-rate data using a CNN-LSTM architecture.
• To evaluate the impact of attention mechanisms on prediction accuracy.

Main Methods:

• Collected physiological signals and VO2 from 21 adults during rest and exercise tests.
• Developed and compared baseline CNN-LSTM, LSTM, and CNN-LSTM models with spatial, temporal, and spatio-temporal attention modules.
• Assessed model performance using R-squared values.

Main Results:

• Pairing accelerometer and heart-rate data improved VO2 prediction compared to heart rate alone.
• The baseline CNN-LSTM (R²=0.946) outperformed a plain LSTM (R²=0.926).
• Spatial attention further boosted accuracy (R²=0.962), while temporal attention slightly decreased it (R²=0.930).

Conclusions:

• CNN-LSTM with sensor fusion offers a robust approach for wearable metabolic monitoring.
• Spatial attention mechanisms can enhance VO2 prediction accuracy in wearable systems.
• Model performance degrades during high-intensity exercise, indicating a need for further research into non-linear physiological responses.