Athlete Injury Prediction: A Time Series Forecasting Method Based on Wearable Sensor Data and Deep Learning

Athlete injuries pose significant challenges in sports, leading to performance declines, financial losses, and long-term health issues. This paper presents a novel approach for predicting injury risks using time series data from wearable sensors, integrating permutation entropy as a feature with an LSTM model enhanced by attention mechanisms. Synthetic data simulating 20 athletes’ physiological metrics (heart rate, accelerometer, gyroscope) were generated to train and evaluate the model, incorporating realistic trends like gradual escalations in heart rate and accelerations to mimic pre-injury states. The dataset includes trends toward injury risks, enabling robust predictions and addressing data scarcity issues in real-world scenarios. Experiments demonstrate high performance: accuracy of 99.87%, precision of 100%, recall of 99.36%, F1-score of 99.68%, and AUC of 99.78%. Ablation studies confirm the value of permutation entropy, improving F1 by 0.13% through better capture of signal complexity. Cross-validation yields a mean accuracy of 99.79% with low variance (std 0.10%), underscoring model stability. Visualizations, including confusion matrices, ROC curves, and feature importances, highlight the model’s effectiveness in capturing injury precursors and provide interpretable insights for practitioners. This method advances early warning systems, potentially reducing injuries through proactive interventions and personalized training adjustments. Future work could incorporate real-world data for enhanced generalizability and explore hybrid models with additional modalities.