Transfer learning-enhanced CNN-GRU-attention model for knee joint torque prediction

Area of Science:

• Biomechanics
• Sports Science
• Machine Learning

Background:

• Accurate joint torque prediction is crucial for injury prevention, offering insights into forces during activities.
• Traditional methods using surface EMG (sEMG) and kinematics struggle with complex movements and individual variability.
• Generalizing joint torque estimation models across subjects is challenging due to declining feature transferability.

Purpose of the Study:

• To develop a more accurate knee joint torque prediction model.
• To address the limitations of traditional methods in capturing non-linear muscle activation-joint motion relationships.
• To enhance model generalizability across different individuals using transfer learning.

Main Methods:

• Proposed a hybrid CNN-GRU-Attention neural network model informed by a neuromusculoskeletal (NMS) solver.
• Integrated transfer learning to improve inter-subject prediction accuracy.
• Trained the model using EMG signals, joint angles, and muscle forces to predict knee joint torque.

Main Results:

• The hybrid-CNN model achieved significantly low error (RMSE ≤0.16 Nm/kg) for within-subject knee joint torque prediction.
• Transfer learning significantly improved inter-subject prediction generalizability, reducing error (RMSE ≤0.14 Nm/kg).

Conclusions:

• The developed transfer learning-enhanced CNN-GRU-Attention model shows significant potential for accurate knee joint torque prediction.
• This approach offers a promising tool for enhancing sports performance and injury prevention strategies.
• The study highlights the effectiveness of hybrid models and transfer learning in complex biomechanical analyses.