Related Experiment Videos
Machine learning in control of functional electrical stimulation systems for locomotion
A Kostov1, B J Andrews, D B Popović
1Division of Neuroscience, University of Alberta, Edmonton, Canada.
IEEE Transactions on Bio-Medical Engineering
|June 1, 1995
Summary
Machine learning algorithms, adaptive logic networks (ALN) and inductive learning (IL), were tested for functional electrical stimulation (FES) control in spinal cord injured humans. Both methods predicted stimulation events, with IL learning faster and ALN showing better generalization.
Area of Science:
- Biomedical Engineering
- Rehabilitation Technology
- Machine Learning Applications
Background:
- Spinal cord injury (SCI) often impairs locomotion, necessitating assistive technologies.
- Functional Electrical Stimulation (FES) is a promising approach to restore movement in individuals with SCI.
- Developing effective, automated control systems for FES is crucial for improving gait rehabilitation.
Purpose of the Study:
- To evaluate two machine learning techniques, Adaptive Logic Network (ALN) and Inductive Learning (IL), for the automated design of FES control systems.
- To assess the ability of these algorithms to learn the relationship between sensory input and FES control signals for human locomotion.
- To compare the performance, learning speed, and generalization capabilities of ALN and IL in the context of FES-assisted ambulation.
Main Methods:
- Off-line supervised training was employed to teach the machine learning models.
- Sensory data, including foot pressure and joint angles, were recorded from participants.
- FES control signals were based on subject-initiated stimulation via a manual push-button during FES-assisted walking.
Main Results:
- Inductive Learning (IL) demonstrated faster learning compared to Adaptive Logic Network (ALN) with identical training data.
- Both ALN and IL rapidly performed tests and could predict future stimulation events.
- Adaptive Logic Network (ALN) exhibited superior generalization, particularly when incorporating temporal data, while IL provided interpretable decision trees and quantified sensory importance.
Conclusions:
- Both ALN and IL are viable techniques for developing rule-based FES control systems for SCI locomotion.
- ALN offers advantages in continuous learning without knowledge loss, whereas IL provides transparent models and feature importance insights.
- Further research can leverage these findings to enhance FES control systems for improved FES-assisted mobility in individuals with SCI.