Fast Inference of Spinal Neuromodulation for Motor Control using Amortized Neural Networks

This research explores using machine learning to improve motor function after spinal cord injury (SCI) through epidural electrical stimulation (EES). EES involves stimulating the spinal cord with electricity to help restore movement, but finding the right stimulation parameters is difficult. The researchers used deep neural networks to create models that can predict motor outputs based on EES parameters and, conversely, suggest EES parameters to achieve specific motor outputs. This approach aims to automate and accelerate the process of finding effective EES settings. Data was collected from sheep implanted with EES electrodes, and the neural networks were trained to learn the relationship between EES and muscle activity. The models were then tested in vivo to see if they could accurately identify EES parameters that would produce desired muscle activation patterns.

• 1
  Neural networks can effectively approximate the complex computations of the spinal cord related to sensorimotor control, accurately predicting muscle activity based on EES parameters.
• 2
  The developed inverse model can identify novel EES parameters capable of producing desired muscle recruitment in vivo in under 20 minutes, significantly faster than manual methods.
• 3
  The study discovered functional redundancies within the spinal sensorimotor networks, identifying different EES parameters that can achieve similar motor outcomes, suggesting potential for personalized and optimized EES strategies.