Neurophysiological Signals Analysis with Nanoelectronic System Platforms

Abstract

To address the issue of high costs in physiological signal processing, this paper combines memristor technology with brain-inspired computing methods to explore a low-power, high-performance on-chip learning and inference system to improve the efficiency and accuracy of Electroencephalogram (EEG) and Electrocardiogram (ECG) signal processing.

This study proposes two novel neural network architectures, including the Kolmogorov-Arnold Network (KAN) based on learnable activation functions and the Spiking Neural Network (SNN) combined with dendrites based on neuromorphic devices. By introducing memristors as hardware implementations for neurons and synapses, an efficient memristor array was constructed, improving the ability to integrate algorithms with hardware systems. In addition, based on memristor technology, we designed an innovative neuron model based on dendritic Leaky Integrated and Fire (LIF) neurons that improve network performance without relying on the traditional delay layer approach. This not only enhances the spatiotemporal processing capability of the network for physiological signals but also improves the system's biological plausibility and computational efficiency.

Experimental results show that the proposed networks achieve strong performance in EEG seizure prediction and ECG abnormality detection. They offer effective physiological signal processing with significantly reduced power consumption, delivering comparable or superior accuracy at lower computational cost than conventional neural networks.

This work presents a novel solution for low-power physiological signal processing, supporting applications in intelligent medical devices and brain-computer interfaces. Future research will focus on optimising memristor dynamics, extending to more signal modalities, and facilitating large-scale hardware integration.