Regression convolutional neural network for improved simultaneous EMG control.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Conventional regression models for electromyography (EMG) control often rely on handcrafted features.
• Deep learning models can automatically learn relevant data representations, potentially improving prediction without manual feature engineering.

Purpose of the Study:

• To propose and validate a regression convolutional neural network (CNN) as a substitute for conventional regression models in EMG control.
• To evaluate the usability of the regression CNN model for controlling individual and simultaneous wrist motions.

Main Methods:

• An online Fitts' law style test was employed to validate the regression CNN model.
• The performance of the CNN-based system was compared against a support vector regression (SVR) scheme utilizing established extracted features.

Main Results:

• The CNN-based EMG control system demonstrated superior performance over the SVR scheme in terms of throughput.
• Higher regression accuracies, particularly with high EMG amplitudes, contributed to the enhanced performance of the CNN model.

Conclusions:

• The regression CNN model effectively extracts underlying motor control information from EMG signals for single and multiple degrees-of-freedom (DoF) tasks.
• Regression CNNs offer an advantage over classification CNNs by enabling independent and simultaneous control of movements.