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

Energy expenditure (EE) assessment is crucial in both sports science and health management. However, current EE prediction models often overlook individual differences and lack dynamic correlation analysis between multi-modal data and EE. Building upon previous research, this study proposes a personalized dynamic-static feature fusion framework, which integrates two types of information to improve energy expenditure (EE) prediction during incremental load exercise: dynamic signals (physiological signals recorded continuously during exercise, such as tri-axial acceleration and electrocardiography [ECG]) and static physiological metrics (stable individual traits measured at rest, such as BMI, body-fat percentage, resting heart rate, and resting oxygen uptake [VO₂]). These two feature sets were combined through a hybrid convolutional neural network (CNN) and long short-term memory (LSTM) neural network architecture. CNN layers extract local temporal patterns from dynamic signals, and LSTM layers model temporal dependencies over longer intervals. The model prediction performance was evaluated using root mean square error (RMSE), coefficient of determination (R²), mean absolute error (MAE) and Bland-Altman plots, and the results show that the CNN + LSTM model significantly outperforms both the traditional autoregressive (AR) linear model and the LSTM model that uses only a single modality (acceleration or ECG). Analysis of feature values and SHAP values revealed that accelerometer features played a dominant role in EE prediction during moderate-to-high intensity exercise. As exercise intensity increased, the contribution of ECG features gradually increased, with ECG features dominating during high-intensity exercise, demonstrating the complementary effect and dual contribution of these two types of features in EE prediction at different exercise intensities. This study demonstrates that personalized dynamic-static feature fusion can effectively predict EE during incremental exercise tests and analyzes the dynamic changes in the contribution of different features across different intensity ranges, providing a theoretical basis and methodological reference for related research.