Hybrid AlexNet–Tabnet with Lyrebird Feature Selection for Injury Prevention and Athlete Health Monitoring

Sports healthcare increasingly relies on intelligent motion analysis to monitor athlete performance, identify risky movements, and prevent injuries. This research focuses on athlete motion data gathered from wearable sensors, which record multidimensional signals such as acceleration, angular velocity, and joint kinematics. The goal of this study is to develop a real-time, interpretable motion classification framework that can accurately distinguish biomechanically similar movements, like jump versus kick, which earlier models often misclassify. To accomplish this, we propose a hybrid approach combining AlexNet for spatial feature extraction, TabNet for attention-based interpretability, discrete wavelet transform for time–frequency analysis, and the Lyrebird Optimization Algorithm for feature selection. Experiments on the 6G-SDN Sports Motion Dataset demonstrate that the framework achieves 98.85% accuracy, 98.60% precision, 98.61% sensitivity, and 98.65% F1-score, outperforming CNN-only, TabNet-only, and LSTM baselines by 2.7–4.6%. Interpretability analysis highlights ankle angular velocity and knee joint angle as key predictors, aligning with sports medicine research on anterior cruciate ligament (ACL) strain and lower-limb injury risk. Overall, the hybrid model offers state-of-the-art classification performance while delivering biomechanically meaningful insights, proving its value as a real-time healthcare tool for injury prevention, athlete monitoring, and rehabilitation support.