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

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Closed-loop functional electrical stimulation (FES) systems utilize recorded nerve signals for real-time decision-making.
• Previous convolutional neural network (CNN) models achieved high accuracy in discriminating neural pathways but were too resource-intensive for implantable devices.

Purpose of the Study:

• To develop resource-efficient neural network (NN) architectures for implantable FES systems.
• To minimize data storage and power consumption for CNNs used in multi-contact nerve cuff electrode recordings with minimal performance degradation.

Main Methods:

• Evaluated various neural network architectures, including ESCAPE-NET, fully convolutional networks, and recurrent neural networks, using rat sciatic nerve recordings.
• Trained NNs to classify natural compound action potentials (nCAPs) from 56-channel cuff electrode data.
• Assessed NN performance based on F1-score, number of weights, and floating-point operations (FLOPs).

Main Results:

• Identified NN variations requiring 1,132-1,787x fewer weights and 389-995x less memory than ESCAPE-NET.
• Achieved macro F1-scores of 0.70-0.71, comparable to the 0.75 baseline, with significantly reduced computational load (6-11,073x fewer FLOPs).
• Memory requirements ranged from 22.69 KB to 58.11 KB, suitable for on-chip integration in ASIC-based deep learning accelerators.

Conclusions:

• Reduced ESCAPE-NET versions offer substantial resource savings without compromising accuracy.
• These optimized NNs are suitable for integration into surgically implantable devices for responsive closed-loop neural stimulation.