Intelligent Injury Risk Assessment through Retrospective Analysis of Multimodal Sports Data among Professional Gymnasts

Abstract

Sports injuries still pose a significant problem in professional gymnastics, as a consequence of the combination of repetitive movement patterns, high physical demands, long hours of training and accumulated physiological stress. Knowing early in the process of development when injury is likely to occur is important for athlete safety, continuing training and performance, and making evidence-based decisions. Traditional injury evaluation techniques rely on visual inspection and single physiological measurement, and have only a partial retrospective perspective of injury factors. The comparative analytical framework of this study involves machine learning and deep learning methods, to explore injury risk patterns based on retrospective analysis of multimodal sports injury records. The Multimodal Sports Injury Prediction Dataset is used, which consists of physiological, workload, and athlete condition indicators from professional sports participants. Before developing the models, data preprocessing was done which involved data cleaning, data normalization, class balancing, and statistical exploration. To predict injury risk, four machine learning models (Logistic Regression, Random Forest, XGBoost, LightGBM) and three deep learning models (Artificial Neural Network (ANN), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM)) were tested. The experimental analysis showed that the LightGBM model had the highest accuracy of 95.16%, F1-score of 94.92% and ROC–AUC of 0.981 compared to other machine learning models. In deep learning techniques, CNN showed the highest predictive power with 95.84% of accuracy, 0.986 ROC–AUC and 95.55% F1-score. An explainability analysis identified workload index, heart rate variability, recovery time and respiratory factors as dominant contributors to injury. The results show that multimodal retrospective injury analysis and prediction based on data can be useful for athlete risk monitoring and decision making in professional gymnastics.