Resource-Efficient Neural Network Architectures for Classifying Nerve Cuff Recordings on Implantable Devices

Background: Closed-loop functional electrical stimulation can use recorded nerve signals to create implantable systems that make decisions regarding nerve stimulation in real-time. Previous work demonstrated convolutional neural network (CNN) discrimination of activity from different neural pathways recorded by a high-density multi-contact nerve cuff electrode, achieving state-of-the-art performance but requiring too much data storage and power for a practical implementation on surgically implanted hardware. Objective: To reduce resource utilization for an implantable implementation, with minimal performance loss for CNNs that can discriminate between neural pathways in multi-contact cuff electrode recordings. Methods: Neural networks (NNs) were evaluated using rat sciatic nerve recordings previously collected using 56-channel cuff electrodes to capture spatiotemporal neural activity patterns. NNs were trained to classify individual, natural compound action potentials (nCAPs) elicited by sensory stimuli. Three architectures were explored: the previously reported ESCAPE-NET, a fully convolutional network, and a recurrent neural network. Variations of each architecture were evaluated based on F1-score, number of weights, and floating-point operations (FLOPs). Results: NNs were identified that, when compared to ESCAPE-NET, require 1,132–1,787x fewer weights, 389–995x less memory, and 6–11,073x fewer FLOPs, while maintaining macro F1-scores of 0.70–0.71 compared to a baseline of 0.75. Memory requirements range from 22.69 KB to 58.11 KB, falling within on-chip memory sizes from published deep learning accelerators fabricated in ASIC technology. Conclusion: Reduced versions of ESCAPE-NET require significantly fewer resources without significant accuracy loss, thus can be more easily incorporated into a surgically implantable device that performs closed-loop responsive neural stimulation.