Classification of spinal cord injured EMG data for locomotion recovery
- Publication Type:
- Thesis
- Issue Date:
- 2024
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
A century ago, neuroscientists observed rhythmic behaviours in decerebrate cats, leading to the discovery of the locomotor central pattern generator, which drives locomotion through peripheral sensory feedback. Since then, significant evidence has shown the spinal cord's capacity for use-dependent motor learning, even without descending brain input. The spinal cord can generate complex motor tasks given an appropriate neural environment and training, playing a role in preparation, execution, and adaptation of coordinated sensorimotor functions.
This study aimed to explore the optimal neural environment for use-dependent motor learning. An experiment on spinally transected rats, housed in enriched environments, analysed the effects of pharmacological and electrical stimulation therapies after six weeks of training. Hindlimb step-like activity was recorded using a rule-based algorithm that relied on sparse electromyogram (EMG) data. The algorithm performed better than previous methods in reducing false positives. Further improvements were made by incorporating deep learning techniques such as multi-label classification, continuous wavelet transform inputs, vision transformers, and domain-invariant adversarial learning.
To enhance the deep learning model, curriculum learning and domain adaptation across subjects were employed. Results from task-relevant EMG datasets were compared to self-supervised contrastive learning, demonstrating improved classification by pretraining feature extraction layers before training the classification layer. This approach allowed the model to integrate long- and short-term information across different subjects.
By analysing the EMG and motor-evoked potential (MEP) activity during step-like events, the study inferred how the spinal cord's neurological state influences spontaneous motor actions, even without sensory input from treadmill activity. Combining pharmacological agents—quipazine (a serotonin agonist), strychnine (a glycine antagonist)—and electrical stimulation effectively elevated locomotor neural network activity, promoting self-training events. MEPs recorded during locomotion displayed spiking activity correlated with the functional state of the spinal cord, specifically in the middle and late responses.
A biologically constrained spiking neural network model was developed to explain the integration of sensory and neuromodulatory inputs within the flexor reflex circuit. The model explored the effects of body-weight-supported locomotion, serotonin agonists, and electrical stimulation in a simulated spinal cord injury environment. Achieving a balance of excitation and inhibition facilitated phasic flexion activation, providing a mechanistic basis for locomotor recovery.
Future work suggests incorporating human pathological data and extending the model to include extensor reflex circuitry to further investigate the role of polysynaptic late responses in MEPs during recovery.
Please use this identifier to cite or link to this item: