Using Supervised Deep Learning for Real-Time Motor Unit Decomposition of High-Density Surface Electromyography Signals

Thumbnail Image

Files

master_Talke_Marcus_2022.pdf (7.69 MB)

URL

Journal Title

Journal ISSN

Volume Title

Perustieteiden korkeakoulu | Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.

Authors

Date

2022-06-13

Department

Major/Subject

Biomedical Engineering

Mcode

SCI3059

Degree programme

Master’s Programme in Life Science Technologies

Language

en

Pages

80

Series

Abstract

High-density surface electromyography (HDsEMG) is a non-invasive neural interface that records the electrical activity of muscles from the skin surface. These electrical signals can be reverse engineered to recover motor neuron activity in the spinal cord, a process known as motor unit (MU) decomposition. The recovered motor neuron firing times represent the neural drive that stimulates muscle contraction and can be used to estimate muscle force production and predict human movements. Current MU decomposition algorithms such as fastICA are too computationally expensive for human-machine interfaces. This thesis combines deep learning with fastICA to provide a faster and more stable approach to MU decomposition with the potential for prosthetic control. In this thesis, a parallel network of 3D convolutional neural networks (3D CNN) was trained on fastICA decomposition output to identify MUs from HDsEMG with minimal latency. The HDsEMG data was recorded from the forearm of thirteen able-bodied subjects during three degrees of freedom (DoF) of wrist motion. Each 3D CNN was trained to track a single MU for a single DoF. The 3D CNNs were then tested on their single training DoF and multiple, unseen DoFs. Decomposition performance metrics (recall and precision) were compared to the reference fastICA algorithm. In addition, normalized cross-correlation of estimated MU action potential (MUAP) waveforms assessed the similarity of MUs between training and testing. The latency between HDsEMG signals and MU prediction was 0.12 ms per 3D CNN, far below the human electromechanical delay. When tracking a single DoF, the mean recall and precision of the 3D CNNs were 0.91±0.05 and 0.74±0.08, respectively. MUAP waveform analysis suggests that low precision is likely attributed to inaccurate fastICA test labels. While it is currently not possible to fully validate 3D CNN decomposition performance during multiple, unseen DoFs, this work demonstrates that 3D CNNs are a fast, alternative approach to online MU decomposition.

Description

Supervisor

Vujaklija, Ivan

Thesis advisor

Taleshi, Mansour

Keywords

motor unit decomposition, motor unit spike train, electromyography, blind source separation, independent component analysis, convolutional neural network

Other note

Citation

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-202206194075

Collections

[dipl] Perustieteiden korkeakoulu / SCI