A Neuromorphic Processing System for Low-Power Wearable ECG Classification

Abstract

This paper proposes a neuromorphic processing system and its classifier design for always-on wearable electrocardiogram (ECG) classification. The ECG signal is captured by level crossing (LC) sampling yielding single-bit temporal coding that can be natively fed into a spiking neural network (SNN) in an event-driven manner. Such an architecture simplifies the quantization of analog-to-digital converter (ADC) and bypasses the coding processing for SNN. Thus, the system power can be reduced by simplified data conversion architecture, single-bit data representation for input data reduction, and spare processing of SNN. Spatio-temporal backpropagation (STBP) training is optimized to adapt to the LC-based data representation and mitigate the firing rate, thus increase network sparsity. The system-level design of the hardware architecture consisting of an LC-ADC and an SNN processor is evaluated by Simulink-ModelSim co-simulation. Trained with the MIT-BIH database, the proposed system achieves 95.34% in classification accuracy with an average of 79 sampling points and 24.6 kFLOPs per inference, corresponding to 55.9 × and 42.4 x reduction on sampling data and FLOPs per inference respectively, compared with conventional ADC and artificial neural network (ANN) approaches.