Spiking Neural Networks for Efficient Classification of Peripheral Nerve Stimuli

Abstract

Accurate electroneurographic (ENG) signal classification is a key function in biological and neural communication systems, where peripheral nerves act as noisy, bandwidth-limited information channels interfacing with implantable bioelectronic devices. In neural decoding and stimulation (ND&S) systems, ENG signals are transmitted and subsequently classified to extract task-relevant information prior to stimulation. This processing must operate in real-time on resource-constrained implantable platforms, limiting the applicability of conventional deep neural networks in molecular and biological communication settings. To address these challenges, we propose two event-driven spiking neural network (SNN) architectures for energy-efficient ENG decoding. A Parametric Leaky Integrate-and-Fire Spiking Neural Network (PLIF-SNN) employs trainable membrane time constants to model temporal dynamics, while a Parallel Spiking Neural Network (PSNN) uses parallel spiking neurons with learnable temporal kernels and reset-free dynamics for compact temporal filtering. The proposed models were evaluated on a multichannel cuff-electrode dataset recorded from nine Long Evans rats and compared with ESCAPE-NET and MobilESCAPE-NET. For single compound action potential classification, PSNN achieved 87.21% accuracy and an 84.98% macro F1-score using only 25.5k parameters, reducing model complexity by up to 99%. On 500-sample windows, PLIF-SNN reached 90.25% accuracy with 15.4k parameters, matching MobilESCAPE-NET performance with one quarter of its parameter count. These results demonstrate that event-driven SNN-based decoding enables state-of-the-art ENG information extraction with substantially reduced computational and energy costs, supporting low-power, low-latency bioelectronic platforms for bidirectional biological and multiscale communication systems.