Dynamic prediction of adolescent sports injury probability based on the improved ARIMA model in the background of digital physical education teaching.

This study aims to solve the problem of dynamic prediction of sports injuries in adolescent physical education teaching. By integrating multi-source heterogeneous data and improving time series modeling, a real-time early warning system suitable for digital education scenarios is constructed. In view of the limitations of traditional ARIMA models in capturing the suddenness and multi-frequency characteristics coordination of injury events, this study innovatively integrates Time-Varying Filtering Empirical Mode Decomposition (TVF-EMD) and Mixed-Frequency Data Sampling (MIDAS) techniques to achieve separate modeling of injury trend and fluctuation terms, and establish cross-scale correlation between second-level sports data and quarterly physical fitness assessment. In an empirical study involving 217 middle school student athletes, the model significantly improved prediction accuracy (multi-scenario average RMSE reduced by > 40%), with an accuracy rate of 86.2% in high-intensity basketball training, a false alarm rate of 5.2% for ACL injuries in football, and the critical threshold of fracture risk was determined when the monthly running distance exceeded 280 km in track and field training. According to the research, Key technical breakthroughs include dynamically optimizing model parameters through particle filtering to adapt to sudden changes in teaching load, using LightGBM residual correction mechanism to improve the ability to capture sudden injury events, and verifying the reliability of three-dimensional motion trajectory monitoring in virtual reality teaching (error < 12.3 mm). The research results provide core methodological support for constructing a real-time and personalized campus sports health protection system, and will be further validated through multi-center cohort studies to promote clinical translational applications.