Compression of EMG Signals Using Deep Convolutional Autoencoders.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Efficient storage and transmission of electromyogram (EMG) data are crucial for telemedicine and big data applications.
• Current internet speeds and hardware limitations pose challenges for handling large EMG datasets.
• Advanced compression techniques are needed to overcome these data handling obstacles.

Purpose of the Study:

• To propose and evaluate a novel EMG data compression method using deep convolutional autoencoders (CAE).
• To assess the compression efficiency and reconstruction performance of the CAE for both standard and high-density EMG data.
• To compare the CAE method against state-of-the-art compression techniques.

Main Methods:

• Developed a deep convolutional autoencoder (CAE) architecture for EMG data compression.
• Investigated eight-channel and high-density EMG data from multiple subjects.
• Evaluated compression ratio (CR), percentage RMS difference normalized (PRDN), and classification accuracy (CA).

Main Results:

• The CAE achieved efficient compression (CR=1600) with an average PRDN of 31.5% and a minor ~5% reduction in wrist motion classification accuracy.
• The CAE significantly outperformed high-efficiency video coding and wavelet-thresholding methods.
• Reducing bit resolution to 6 bits yielded an additional 4-fold compression with no significant performance degradation.
• Promising inter-subject performance (PRDN only 2.6% higher than within-subject) demonstrates generalizability.

Conclusions:

• The proposed CAE method offers a powerful and automatic end-to-end solution for EMG data compression and reconstruction.
• The CAE demonstrates high efficiency, remarkable reconstruction quality, and excellent generalizability across subjects.
• This approach has significant potential for advancing telemedicine and big data analysis in EMG applications.