Identification of Spared and Proportionally Controllable Hand Motor Dimensions in Motor Complete Spinal Cord Injuries Using Latent Manifold Analysis.

The loss of bilateral hand function is a debilitating challenge for millions of individuals that suffered a motor-complete spinal cord injury (SCI). We have recently demonstrated in eight tetraplegic individuals the presence of highly functional spared spinal motor neurons in the extrinsic muscles of the hand that are still capable of generating proportional flexion and extension signals. In this work, we hypothesized that an artificial intelligence (AI) system could automatically learn the spared electromyographic (EMG) patterns that encode the attempted movements of the paralyzed digits. We constrained the AI to continuously output the attempted movements in the form of a digital hand so that this signal could be used to control any assistive system (e.g. exoskeletons, electrical stimulation). We trained a convolutional neural network using data from 13 uninjured (control) participants and 8 tetraplegic participants (7 motor-complete, 1 incomplete) to study the latent space learned by the AI. Our model can automatically differentiate between eight different hand movements, including individual finger flexions, grasps, and pinches, achieving a mean accuracy of 98.3% within the SCI group. Analysis of the latent space of the model revealed that proportionally controllable movements exhibited an elliptical path, while movements lacking proportional control followed a chaotic trajectory. We found that proportional control of a movement can only be correctly estimated if the latent space embedding of the movement follows an elliptical path (correlation =0.73; p <0.001). These findings emphasize the reliability of the proposed system for closed-loop applications that require an accurate estimate of spinal cord motor output.