Multivariate fusion modeling for dynamic prediction of oxygen uptake during incremental load exercise in college students

The aim of this study is to construct a model that can easily and accurately predict oxygen uptake (VO2) in college students through multiple machine learning and deep learning methods. We collected data on triaxial acceleration, heart rate, and resting characteristics (e.g., weight and height) from 22 college students during incrementally loaded exercise for prediction using different time-series models, including LSTM, GRU, Transformer, and XGBoost. Model performance was thoroughly evaluated by using metrics such as RMSE, MAE, R² and Bland-Altman plots. The results show that LSTM and XGBoost exhibit high accuracy and stability in predicting oxygen uptake. Specifically, LSTM (RMSE=0.1973, R2=0.9563, MAE=0.1386) was able to effectively capture the time dependence in acceleration and heart rate data and excelled in oxygen uptake estimation by leveraging its ability to model long-term dependence and nonlinear relationships. XGBoost (RMSE=0.2077, R2=0.9515, MAE=0.1421), on the other hand, showed strong prediction stability and low error, especially when facing data with different exercise load intervals, the prediction accuracy was more balanced. In addition, we found that combining temporal features (heart rate and acceleration) and resting features (e.g., body weight, height, resting heart rate) as input features significantly improved the accuracy of oxygen uptake prediction. By integrating data from multiple sources, the model is able to comprehensively capture internal and external load changes during exercise and effectively address the delayed response of physiological metrics to exercise intensity in model predictions. This study provides a method and idea reference for oxygen uptake detection during exercise, personalized exercise plan customization, and exercise rehabilitation.