An Energy-Efficient and Robust SNN Classifier for LC-ADC Sampled ECG Signals

This paper presents an energy-efficient and robust spiking neural network (SNN) classifier design for level-crossing analog-to-digital converter (LC-ADC) sampled electrocardiogram (ECG) signals. A quantization-aware training method is proposed for the fixed-point quantization stage of the hardware-aware training to achieve better classification accuracy while further compressing the quantization resolution. The SNN classifier is trained with a mixed ECG dataset in which the ECG signals are sampled by LC-ADCs with multiple resolutions. Thus, the robustness of the SNN model is improved, achieving a classification accuracy improvement of up to 2% on various LC resolutions. The classifier is validated on a neuromorphic processor chip and achieves an accuracy of 98.24% with a measured energy consumption of 0.36 $\mu$J/inference. The energy-delay product is 164 pJ$\cdot$s, achieving more than $12 \times$ improvement over other works.