Sports injury prediction method based on improved LSTM

Sports injury prediction is the core link to ensure athletes' competitive life and scientific training, and its accuracy directly affects the optimization of training plan and the allocation of rehabilitation resources. Traditional models, such as logistic regression and decision tree, are difficult to capture the nonlinear time series characteristics of high-dimensional biomechanical data, especially for the prediction of sudden injury events. In this paper, a hybrid architecture of Bidirectional Long Short-Term Memory (BiLSTM) and Particle Swarm Optimization (PSO) is proposed. The hyperparameters of BiLSTM (the number of hidden layer units, the number of iterations, etc.) are adaptively optimized by PSO to avoid the randomness of manual parameter adjustment, and the causal chain of load accumulation and compensation feedback is modeled by combining the bidirectional information flow of BiLSTM. Based on the multimodal motion database of 1000 athletes (including 10-dimensional biomechanical indicators), the empirical results show that the accuracy of the model in the test set is 93.7%, the F1 value is 0.928, and the acute injury warning deviation is only 3.2%, which is significantly better than comparative model, and provides a robust solution for the dynamic assessment of injury risk in high-noise sports environment.