Available at: https://digitalcommons.calpoly.edu/theses/3274
Date of Award
6-2026
Degree Name
MS in Electrical Engineering
Department/Program
Electrical Engineering
College
College of Engineering
Advisor
Nishith Chakraborty
Advisor Department
Electrical Engineering
Advisor College
College of Engineering
Abstract
Accurate diagnosis of pathological conditions from biomedical signals, such as electrocardiograms (ECGs) is often performed offline, making it time-consuming, costly, and inefficient, especially when abnormal patterns are rare and long-term monitoring generates large amounts of data. To address this, this work proposes a compact, scalable, and programmable neuromorphic system designed for real-time preliminary arrhythmia detection and classification using ECG signals, that can be extended to other biomedical signals. The proposed design processes ECG signals using a delta modulation-based spike encoder, followed by classification with a dot-product engine (DPE) based spiking neural network (SNN) processor and winner-take-all (WTA) circuit. The architecture supports online learning via spike-timing-dependent-plasticity (STDP) and includes a scan-chain programming circuit that allows processor parameters to be reconfigured during runtime.
The proposed system has been implemented on four FPGA platforms, demonstrating promising results in terms of resource utilization, power consumption, and operating frequency. The optimized design uses only 2.63% of available LUTs and 1.04% of available FFs and features a maximum operating frequency of 192.8 MHz with 20 mW of dynamic power consumption at 100 MHz, which linearly scales, on the low cost, entry level Basys-3 board. This design was subsequently implemented in a 130nm CMOS process, taking up a total die area of 95,792.5 µm2 and dynamic power consumption as low as 26.2 µW for a 1MHz clock. Finally, a software model was built utilizing snnTorch to train a spiking neural network with surrogate gradient backpropagation and evaluation on a hardware-faithful simulation of the proposed system. Tested on the MIT-BIH Arrhythmia Database, the system achieves classification across the four main AAMI heartbeat classes, with significant potential for improving multi-class arrhythmia sensitivity and detection. These results, alongside the compact hardware footprint, position this initial design – an efficient, scalable solution for real-time arrhythmia detection, adaptable to diverse patient profiles and other biomedical signals – as a foundation for future enhancements including increasing synapse weight bit-widths and accumulator precision, scaling the DPE array size, cascading multiple arrays to enable multi-layer feature extraction, and implementing per-neuron programmable thresholds while maintaining ultra-low-power operation.