Author(s): SONG, Y., ZHEN, W., RUI, Z., ZHIMO, J., BINGNAN, G. , Institution: CAPITAL UNIVERSITY OF PHYSICAL EDUCATION AND SPORTS, Country: CHINA, Abstract-ID: 1459

INTRODUCTION:
Energy Expenditure (EE), a key indicator reflecting exercise intensity and volume, has been widely researched in public health and sports training. Accurate monitoring of EE is crucial in physiology, health management, and exercise prescription design. Consequently, accurately predicting or estimating energy expenditure using simple, easy-to-measure indicators combined with artificial intelligence (AI) algorithms has become a critical direction in energy consumption research. Building on previous research, this study integrates heart rate and acceleration data from incremental load exercises, constructing multiple dynamic time series prediction models. The model with the best fit was selected based on performance evaluation metrics, providing a theoretical reference for future energy expenditure prediction studies.
METHODS:
Twenty-five college students participated as subjects, undergoing an incremental load exercise test on a treadmill. Data collection comprised exercise load metrics and heart rate readings captured by accelerometers, with energy expenditure predictions made using dynamic time-series models (e.g., LSTM, BiLSTM, TCN, GRU, CNN+LSTM). Model performance was evaluated using metrics such as RMSE, R², and Bias, with prediction consistency further assessed through Bland-Altman plots.
RESULTS:
The LSTM model achieved the lowest RMSE (0.2423) and Bias (0.1060), indicating its high accuracy and consistency in predicting energy expenditure. It also displayed the highest R² value (0.9255), demonstrating its strong ability to interpret the data.The TCN model also performed well, with an RMSE of 0.2710, an R² of 0.9068, and the lowest Bias (0.0216) among the models. The low Bias value indicates that TCN offers a reduced systematic error in predictions compared to other models. TCNs ability to capture long-term contextual information through its convolutional structure helped it mitigate biases effectively. BiLSTM, GRU, and CNN+LSTM showed moderate performance. Overall, LSTM and TCN were the top-performing models.
CONCLUSION:
In this study, we developed an EE model for incremental loading exercise by utilizing acceleration and heart rate data, comparing the predictive performance of multiple time series models, including LSTM, BiLSTM, GRU, TCN, and CNN+LSTM. Through evaluation metrics such as RMSE, R², and Bias, along with Bland-Altman plots to assess prediction consistency, we identified LSTM and TCN as the most accurate and stable models for EE prediction. Our findings underscore the effectiveness of time series-based algorithms in capturing dynamic patterns in incremental loading exercise. Moreover, incorporating joint internal (heart rate) and external (acceleration) metrics as input features proves effective in reflecting exercise intensity. This approach suggests that the integration of multiple metrics enhances the accuracy of EE estimation, providing a reliable means for quantifying exercise load and calculating exercise intensity.