Multiday EMG-Based Classification of Hand Motions with Deep Learning Techniques.

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Myoelectric control for upper limb prostheses relies on electromyography (EMG) signal pattern recognition.
• Current clinical methods face challenges in feature selection and long-term performance robustness.
• Deep learning offers intrinsic feature extraction, potentially improving myoelectric control.

Purpose of the Study:

• To evaluate the efficacy of using raw EMG signals as direct inputs to deep neural networks for prosthetic control.
• To compare the performance of a convolutional neural network (CNN) against traditional methods like linear discriminant analysis (LDA) and stacked sparse autoencoders (SSAE).
• To assess the robustness of the proposed method over multiple days of data acquisition.

Main Methods:

• Raw bipolar EMG signals were recorded from seven able-bodied subjects performing six active motions plus rest over 15 consecutive days using a MYO armband.
• A CNN was trained using raw EMG samples as direct input.
• Performance was evaluated against LDA and SSAE (with features and raw samples) across within-session, between-session, and leave-one-day-out analyses.

Main Results:

• CNN demonstrated significantly lower classification error compared to LDA and SSAE-r across all evaluation metrics (p < 0.001).
• No significant performance difference was observed between CNN and SSAE-f (autoencoder with features).
• CNN significantly enhanced performance and robustness over time compared to LDA with handcrafted features.

Conclusions:

• CNNs utilizing raw EMG signals offer a robust and data-driven approach for myoelectric control.
• This method overcomes the limitations of manual feature calibration and selection in traditional systems.
• The findings suggest a promising direction for improving the long-term performance and usability of upper limb prostheses.