Squat errors classification based on National Academy of Sports Medicine guidelines using IMU and deep learning...

Accurate squat assessment is essential for injury prevention and performance improvement. Traditional visual assessments are often subjective and inconsistent, highlighting the need for automated classification methods. This study proposes a deep learning-based approach to classify squat movement errors using data from inertial measurement unit (IMU), following the guidelines of the National Academy of Sports Medicine (NASM). Twenty athletes (10 male, 10 female) performed five squat variations: one correct and four incorrect forms, including knees valgus, knee varus, low back arching, and low back rounding. Kinematic data were analyzed using CNN, CNN- TabNet, GRU, GRU- TabNet, CNN-GRU, and CNN-GRU-TabNet architectures. Model performance was evaluated using accuracy, precision, recall, and F1-score. The CNN-GRU model achieved the highest classification accuracy (0.997) when using data from all five IMUs. When limited to a single IMU, the optimal sensor placement was on the right thigh, achieving 0.988 accuracy. Although the CNN-GRU-TabNet model offered slightly lower accuracy, it provided enhanced interpretability through its attention-based feature selection mechanism. CNN-TabNet and GRU-TabNet models compared to CNN and GRU while using all five IMUs data showed higher classification accuracy. In conclusion, hybrid deep learning models, particularly CNN-GRU, demonstrate strong potential for automated squat error classification. Moreover, near-optimal accuracy can be achieved with a single strategically placed IMU, supporting the development of efficient and practical wearable systems for use in sports science and rehabilitation.