CARTE seed projects seminar series is pleased to welcome Eugene Hwang (2nd year MASc student in Biomedical Engineering under the supervision of Professor José Zariffa) and José Sales Filho (4-th year Ph.D. student in Electrical and Computer Engineering under the supervision of Professor Roman Genov).

To attend this seminar, please contact CARTE[AT]utoronto.ca

Abstract:

Background: Closed-loop control of functional electrical stimulation involves using recorded nerve signals to make decisions regarding nerve stimulation in real-time. Surgically implanted devices that can implement this strategy have significant potential to restore natural movement after paralysis and improve the quality of life of individuals living with spinal cord injury. Discriminating between recorded signals from different neural pathways can provide information about the body’s state and external inputs, which can then be used to decide on the most appropriate form of stimulation to deliver. Our previous work demonstrated the use of a convolutional neural network (CNN), ESCAPE-NET, to discriminate spatiotemporal signatures produced by different neural pathways. Despite state-of-the-art performance, this approach required too much data storage, power and computation time for implementation on hardware that is small enough to be surgically implanted in patients and fast enough for real-time decision making.

Objective: This study aimed to minimize resource consumption while maintaining performance accuracy for machine learning techniques that can distinguish between neural pathways in high-density multi-contact nerve cuff electrode recordings.

Methods: Several CNN architectures were evaluated using a dataset of rat sciatic nerve recordings previously collected using 56-channel (7 x 8) spiral nerve cuff electrodes. The CNNs were trained to classify individual, natural compound action potentials (nCAPs) elicited by ankle dorsiflexion, plantarflexion or heel prick. Reduced-complexity CNN architectures have been explored and designed using techniques such as dropout and fully convolutional networks (FCN).

Results: Reduced ESCAPE-NET using dropout was 16x smaller than the original architecture and operated with 2.4x fewer floating-point operations per second (FLOPS), performing with 80.6% accuracy compared to a baseline of 80.7%. Reduced ESCAPE-NET using FCNs was 420x smaller than the original and operated with 3.5x fewer FLOPS, performing with 77.8% accuracy compared to a baseline of 80.7%.

Conclusion: These reduced versions of ESCAPE-NET require significantly fewer resources with minimal accuracy loss, and thus will be more easily incorporated into a surgically implantable device that performs closed-loop control of stimulation in real-time. Beyond the impact on people with paralyzed limbs and spinal cord injury, this selective peripheral nerve recording technology has applications in the control of prosthetic limbs and neuromodulation interventions for diseases such as chronic pain, diabetes, or incontinence.