Related Experiment Video
Updated: Jul 9, 2025
12:09
Functional and Physiological Methods of Evaluating Median Nerve Regeneration in the Rat
Published on: April 18, 2020
12.2K
Selective peripheral nerve recording using simulated human median nerve activity and convolutional neural networks
Taseen Jawad1,2, Ryan G L Koh1, José Zariffa3,4,5,6
1KITE Research Institute, Toronto Rehabilitation Institute-University Health Network, Toronto, Canada.
Biomedical Engineering Online
|December 8, 2023
Summary
The Extraneural Spatiotemporal Compound Action Potentials Extraction Network (ESCAPE-NET) shows robust performance in classifying neural signals for prosthetic limb control. This convolutional neural network (CNN) model demonstrates potential for human neural interfaces, outperforming other networks in simulations.
Area of Science:
- Computational Neuroscience
- Biomedical Engineering
- Machine Learning for Neural Interfaces
Background:
- Intuitive control of prosthetic limbs remains a challenge, leading to device abandonment.
- Peripheral nerve interfaces translate motor intent into prosthetic commands.
- The Extraneural Spatiotemporal Compound Action Potentials Extraction Network (ESCAPE-NET), a convolutional neural network (CNN), effectively discriminates neural sources in rat sciatic nerves.
Purpose of the Study:
- To assess the applicability of ESCAPE-NET to larger, more complex nerves, specifically the human median nerve.
- To characterize ESCAPE-NET's performance in a computational model of the human median nerve for future human translation.
Main Methods:
- A finite-element model of the human median nerve was generated from immunohistochemistry images.
- Simulated extraneural recordings were used to train and test ESCAPE-NET for classifying naturally evoked compound action potentials (nCAPs) based on source location.
- ESCAPE-NET's performance was compared against ResNet-50 and MobileNet-V2 under varying numbers of sources and noise levels.
Main Results:
- ESCAPE-NET demonstrated high classification accuracy, with performance decreasing as the number of nCAP sources and noise levels increased.
- In low-noise conditions, ESCAPE-NET achieved 97.8% (3-class) and 89.3% (10-class) accuracy.
- In high-noise conditions, ESCAPE-NET achieved 70.3% (3-class) and 52.5% (10-class) accuracy, outperforming ResNet-50 and MobileNet-V2 across tested conditions.
Conclusions:
- All tested networks learned to differentiate nCAPs from different sources, performing significantly above chance.
- ESCAPE-NET exhibited the most robust performance, providing valuable translational guidelines for human neural interface design.
- The study highlights ESCAPE-NET's potential for developing advanced prosthetic limb control systems.
