Classification of 41 Hand and Wrist Movements via Surface Electromyogram Using Deep Neural Network

Area of Science:

• Biomedical Engineering
• Neuroscience
• Robotics

Background:

• Surface electromyography (sEMG) offers non-invasive control for prosthetic limbs.
• Previous sEMG-based movement classification studies show significant variability.
• Accurate classification of numerous hand and wrist movements is crucial for advanced prosthetics.

Purpose of the Study:

• To investigate deep neural networks for classifying 41 hand and wrist movements using sEMG signals.
• To evaluate model performance on two distinct sEMG datasets (DB5 and DB7) from the Ninapro project.
• To assess performance improvements when focusing on a subset of clinically relevant movements (SHAP).

Main Methods:

• Utilized deep neural networks for sEMG signal classification.
• Trained and evaluated models on Ninapro datasets DB5 (16 channels, 200 Hz) and DB7 (12 channels, 2 kHz).
• Compared classification accuracy for 41 movements versus a subset of six Southampton Hand Assessment Procedure (SHAP) movements.

Main Results:

• Achieved high overall accuracies: 93.87% (DB5) and 91.69% (DB7) for 41 movements.
• Attained superior accuracies for SHAP movements: 98.82% (DB5) and 99.00% (DB7 amputee data).
• Demonstrated improved balanced accuracies, particularly for the SHAP subset.

Conclusions:

• Deep neural networks provide a robust method for classifying a wide range of hand and wrist movements from sEMG.
• Focusing on clinically validated movement subsets like SHAP significantly enhances classification performance.
• The proposed approach holds potential for controlling versatile prosthetic hands, especially with further data from amputee subjects.