Range Encoding: A Low-Power Spiking Neural Network Encoding Scheme for ECG Classification

As artificial intelligence algorithms become widely embedded in wearable electrocardiogram (ECG) devices for signal processing and analysis, there is a growing demand for highly efficient, low-power neural network implementation methodology. Spiking Neural Networks (SNNs), known for their ability to process event-based information, are particularly well-suited for resource-constrained edge devices in wearable applications. In this paper, we introduce an efficient SNN encoding method called Range Encoding, which captures dynamic variations in ECG signals more efficiently and accurately. Range Encoding achieves comparable or superior results as compared to other encoding schemes while significantly reducing hardware overheads and computational complexity. Experiments on the MIT-BIH Arrhythmia and MIT-BIH Supraventricular Arrhythmia datasets, using only 2 time steps for temporal resolution, yielded accuracy rates of 98.15% and 93.76%, respectively. Synthesis results using the 55nm CMOS process technology achieve an energy consumption per inference of 0.07µJ. Compared to other SNN encoding methods, Range Encoding enables shorter computation latency while delivering equal or even better performance, making it highly practical for SNNs in real-time, low-power medical applications.