Classification of Spinal Cord Injured EMG Data for Locomotion Recovery

Abstract

A century ago, neuroscientists first noticed reflexively driven rhythmic behaviour in decerebrate cats. These observations led to the discovery of refex and feedforward locomotion activity driven only by peripheral sensory feedback, namely the locomotor central pattern generator (CPG). Since then, even with limited descending brain input, we have seen extensive evidence of the spinal cord’s capacity for use-dependence motor learning. Indeed, the spinal cord, without supraspinal descending input, can generate complex coordinated motor tasks with the appropriate neural environment and training. More and more evidence supports the idea of the spinal cord participating in preparation, execution, and adaptation events to express a responsive and coordinated sensorimotor system.

The following question was asked: “What is the appropriate neural environment in the spinal cord for use-dependent motor learning?”. An extensive data set from an experiment performed on spinally transacted rats in an enriched cage environment was analysed. The study looked at different combinations of pharmacological and electrical stimulation therapies after 6 weeks of training. Firstly, hindlimb step-like activity was logged during 6-hour recording intervals with a rule-based algorithm using only sparse electromyogram (EMG) ankle fexor and extensor activity. The algorithm results performed better at false positive rejection compared to existing methods in the literature.

The classification algorithm was improved by implementing multi-label deep learning network methodologies and trained on pre-processed continuous wavelet transform (CWT) inputs. Complete deep learning classification analysis included implementing state-of-the-art vision transformers and domain invariant adversarial learning with depth wise separable convolutional neural networks. Results suggested greater spatiotemporal representations in vision transformers compared to depth wise separable convolutional neural networks.

Deep learning results were improved by incorporating curriculum learning with domain adaptation across subjects. Results were compared against self-supervised contrastive learning after pretraining from task-relevant EMG open-source datasets. Pretraining feature extraction layers before linearly training the classification layer with temporally aware domain adversarial strategy successfully bridged long- and short-term information across subjects.

From these captured step-like events, analysis of the EMG and motor-evoked potential (MEP) activity infer how the neurological state of the spinal cord affects spontaneous step-like events without any sensory input from treadmill activity. Combining quipazine (serotonin agonist), strychnine (glycine antagonist), and electrical stimulation most effectively elevated the locomotor neural networks towards a functional state, promoting a more significant number of ‘self-training’ events. MEPs, captured during detected hindlimb locomotion, contained spiking activity in middle and late responses strongly correlated to the functional state of the spinal cord.

A biologically constrained spiking neural network (SNN) model was developed to explain the mechanisms of sensory and neuromodulation integration in the fexor refex circuit. The effects of body-weight-supported (BWS) locomotion, serotonin agonists, and electrical stimulation were investigated in a simulated inhibition-dominant SCI neural environment. Modulating the neural environment to reach a balanced excitation and inhibition state enabled the propagation of phasic fexion activation. The model explains the mechanistic basis of BWS for locomotion recovery. 

Finally, future works are suggested, such as collecting pathological locomotion data from human subjects and incorporating extensor refex circuity in the SNN. The proposed design allows interrogation of the origins and reasons for the emergence of polysynaptic late responses in MEPs during locomotor recovery.