Attention-Enhanced Convolutional BiLSTM Model for Predicting Recovery Outcomes in Sports Injuries.

Background: Managing sports injuries and ensuring effective rehabilitation are crucial for maintaining optimal athletic performance, particularly among young athletes. Lower limb injuries, due to their frequency and impact on mobility, require precise and personalized rehabilitation approaches. Traditional biomechanical analysis methods often fall short in capturing complex temporal and spatial dependencies inherent in human movement data. To address these limitations, this study proposes a novel hybrid deep learning model aimed at improving injury assessment and recovery prediction. Methods: The proposed framework introduces an Attention-based Random Forest Optimized Convolutional Bidirectional Long Short-Term Memory (A-RF-CBiLSTM) model. Data were collected from four comprehensive datasets: an electromyography dataset, EMG data of limb muscles, a kinematics and EMG dataset during gait-related activities, and the WebAtlas-Human lower limb dataset. Preprocessing involved bandpass filtering, high-frequency noise removal, normalization, and artifact elimination to ensure signal clarity. Feature selection was performed using the Random Forest (RF) algorithm to identify the most relevant inputs for rehabilitation prediction. The selected features were then processed through a Convolutional Block Attention Module (CBAM) to assign adaptive weights based on significance. A depthwise separable Convolutional Neural Network (CNN) was applied for efficient feature extraction. Subsequently, a Bidirectional Long Short-Term Memory (BiLSTM) network captured temporal dependencies in both forward and backward directions. Outputs from the CNN and BiLSTM layers were fused to enhance predictive robustness. Model training was optimized using the Adam optimizer with hyperparameter tuning. Various model performance evaluation tests were conducted to ensure reliability. Results: The A-RF-CBiLSTM model achieved high performance with 98.87% accuracy, 98.60% precision, 97.34% recall, 97.34% F1 score, and 97.88% specificity. These results confirm its effectiveness in predicting rehabilitation outcomes and identifying lower limb injury patterns. Conclusion: This study presents a robust and efficient hybrid model that significantly advances the state-of-the-art in biomechanical analysis for sports injury management.