Interpretable deep learning for personalized energy expenditure prediction using ECG and acceleration signals in...

Area of Science:

• Sports Science
• Health Management
• Biomedical Engineering

Background:

• Accurate energy expenditure (EE) assessment is vital for sports science and health management.
• Existing EE prediction models often fail to account for individual differences and dynamic multi-modal data correlations.
• There's a need for personalized and dynamic approaches to improve EE prediction accuracy.

Purpose of the Study:

• To develop and validate a personalized dynamic-static feature fusion framework for improved EE prediction during incremental exercise.
• To integrate continuous physiological signals (dynamic) with resting physiological metrics (static) for enhanced prediction.
• To analyze the contribution of different features across varying exercise intensities.

Main Methods:

• A hybrid Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) neural network architecture was employed.
• The model fused dynamic signals (tri-axial acceleration, ECG) with static metrics (BMI, body-fat percentage, resting heart rate, resting oxygen uptake).
• Performance was evaluated using RMSE, R², MAE, and Bland-Altman plots, comparing against traditional AR and single-modality LSTM models.

Main Results:

• The proposed CNN+LSTM fusion model significantly outperformed traditional AR and single-modality LSTM models in EE prediction.
• Accelerometer features were dominant in moderate-to-high intensity exercise prediction.
• ECG features increasingly contributed with higher exercise intensities, demonstrating a complementary effect with acceleration data.

Conclusions:

• Personalized dynamic-static feature fusion is an effective method for predicting EE during incremental exercise tests.
• The study highlights the complementary roles of dynamic and static features across different exercise intensities.
• Provides a theoretical basis and methodological reference for future EE prediction research and applications.