Feasibility of Data-driven EMG Signal Generation using a Deep Generative Model.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Rehabilitation Technology

Background:

• Myoelectric control systems rely on pattern recognition for prosthetic limb control.
• High-quality electromyography (EMG) data collection for training is a significant challenge.
• Limited training data hinders the widespread adoption of advanced myoelectric control.

Purpose of the Study:

• To propose a framework for augmenting limited training data using synthetic EMG data.
• To generate subject-specific synthetic EMG data with a deep generative network (SinGAN).
• To evaluate if synthetic data can improve classification accuracy in myoelectric control.

Main Methods:

• Utilized SinGAN, a deep generative network, to create synthetic EMG data.
• Generated 1000 synthetic EMG data windows from a single window of six different motions.
• Assessed synthetic data quality through qualitative (visual inspection) and quantitative (Lepage tests) methods.
• Evaluated the impact of synthetic data on classification accuracy in a cross-validation scheme.

Main Results:

• Qualitative assessment showed comparable feature space distributions between real and synthetic data.
• Quantitative analysis indicated similarities in location and scale for 11 out of 32 synthetic features.
• Multivariate distributions of synthetic data were statistically different from real data (p <0.05).
• The addition of synthetic data to a limited training set significantly improved classification accuracy by 5.4% (p <0.001).

Conclusions:

• Subject-specific synthetic EMG data generated by SinGAN can augment limited real training data.
• This data augmentation strategy shows potential to enhance classification accuracy in myoelectric control.
• The proposed framework addresses the challenge of high-quality data collection for myoelectric control adoption.