An Active Learning Algorithm for Control of Epidural Electrostimulation

IEEE Trans Biomed Eng, 2015 · DOI: 10.1109/TBME.2015.2431911 · Published: October 1, 2015

Spinal Cord Injury Bioinformatics Biomedical

Simple Explanation

Epidural electrostimulation shows promise for spinal cord injury therapy, but finding effective stimuli is laborious. An autonomous algorithm could simultaneously deliver therapy and explore the vast stimulus space. This paper proposes a method based on GP-BUCB, a Gaussian process bandit algorithm, to automate stimulus selection. The algorithm was tested in spinally transected rats with implanted epidural electrode arrays. GP-BUCB's performance in selecting stimuli to elicit muscle responses was compared to selections by a human expert. The algorithm consistently discovered effective stimulus patterns, even without anatomical information. GP-BUCB was able to extrapolate from previous sessions to predict performance in new sessions, while remaining flexible enough to capture temporal variability. This validates automated stimulus selection for spinal cord injury therapy.

Study Duration

6 Weeks

Participants

n = 4 spinally transected rats

Evidence Level

Animal study

Key Findings

1
GP-BUCB successfully controlled spinal electrostimulation in 37 testing sessions, selecting 670 stimuli.
2
Delivered performance with respect to the specified metric was as good as or better than that of the human expert.
3
GP-BUCB also consistently discovered such a pattern, despite receiving no information as to anatomically likely locations of effective stimuli.

Research Summary

This paper introduces a novel algorithm which can automatically optimize multi-electrode array stimulation parameters. The methodology is based on a novel Gaussian Process Batch Upper Confidence Bound (GP-BUCB) active machine learning technique. Our experiments showed that an automated algorithm can find high-performing stimulus parameters efficiently, competitively with a human expert. This should lead to a better therapeutic recovery experience, and shows that an automated algorithm can be a competitive option for determining effective therapeutic strategies. Our experiments show that an automated algorithm can handle the difficult task of maximizing the performance of a relatively complicated neural interface. These results provide a strong indication that a structured GP model can effectively capture spinal responses over large sets of possible stimuli (tens to hundreds) and substantial periods of time (weeks).

Practical Implications

Automated Stimulus Selection

Automated algorithms can efficiently determine effective therapeutic strategies for spinal cord injury, reducing the burden on clinicians.

Personalized Therapy

Algorithms can adapt to individual patient needs and spinal cord plasticity, leading to more effective and customized treatment.

Wider Access to Therapy

Automated methods can facilitate widespread, cost-effective distribution of epidural stimulation therapy, addressing the shortage of trained therapists.

Study Limitations

1
The study was conducted on rats, and results may not directly translate to humans.
2
The reward function was based on EMG activity, a low-level function that may not fully reflect complex motor behaviors.
3
The algorithm was tested in a controlled experimental setting, and its performance in real-world clinical scenarios may vary.

Your Feedback

Was this summary helpful?

Back to Spinal Cord Injury