Improved Surface Electromyogram-Based Hand-Wrist Force Estimation Using Deep Neural Networks and Cross-Joint Transfer...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Rehabilitation Technology

Background:

• Surface electromyogram (sEMG) signals are crucial for non-invasive human movement analysis.
• Deep neural networks (DNNs) and transfer learning (TL) show promise for improving sEMG-based force estimation.
• Existing TL studies are limited by focusing on single joint applications, restricting dataset size and diversity.

Purpose of the Study:

• To investigate the efficacy of cross-joint TL for sEMG-based force estimation between two upper-limb joints.
• To evaluate four different DNN architectures for this cross-joint TL application.
• To determine the optimal sliding window characteristics for capturing sEMG-force dependencies.

Main Methods:

• Four DNN architectures (two feedforward, two recurrent) were employed with sliding windows for sEMG-based force estimation.
• Transfer learning was implemented by pre-training models on elbow joint data and fine-tuning on hand-wrist data.
• Feature engineering and feature learning approaches were utilized within the DNN models.

Main Results:

• sEMG-force dependencies were found to be short-term (<400 ms), indicating sliding windows are sufficient.
• DNNs reduced the necessary sliding window length for accurate force estimation.
• A convolutional neural network (CNN) with TL achieved a 6.03 ± 0.49% maximum voluntary torque error using only 20s of training data, outperforming other models.

Conclusions:

• Cross-joint TL is a viable strategy to enhance sEMG-based force estimation accuracy and reduce training data requirements.
• The findings suggest that complex recurrent structures may not be necessary, and sliding window DNNs are effective.
• Successful cross-joint TL can significantly enrich data diversity for future deep learning research in biomechanics and rehabilitation.