Date of Award
Summer 7-2-2026
Document Type
Dissertation
Publication Status
Version of Record
Submission Date
July 2026
Department
Computer and Electrical Engineering and Computer Science
College Granting Degree
College of Engineering and Computer Science
Department Granting Degree
Electrical Engineering and Computer Science
Degree Name
Doctor of Philosophy (PhD)
Thesis/Dissertation Advisor [Chair]
Behnaz Ghoraani
Abstract
Wearable ECG monitors enable continuous cardiac surveillance, yet most remain limited to basic heart rate metrics or coarse atrial fibrillation detection, relying on cloud-based analysis that introduces latency, connectivity dependence, and battery drain. Deploying advanced multi-class arrhythmia classification directly on-device is constrained by the tight memory, power, and computational budgets of wearable hardware. This dissertation presents a Quantized Convolutional Spiking Neural Network (QCSNN) for real-time ECG arrhythmia detection on edge hardware, developed across three progressive phases.
Phase 1 introduces a separately trained two-stage QCSNN architecture — a binary classifier cascaded with a four-class classifier — trained directly via surrogate gradient descent and 8-bit quantization-aware training using Brevitas and snnTorch. Compared to ANN-to-SNN conversion approaches, direct training substantially improves minority-class detection (Fusion sensitivity: 83.23% vs. 27.00%), while quantization reduces memory requirements by 76.2% (from 285.60 kB to 67.97 kB) with only 1–4.3% AUC degradation.
Phase 2 demonstrates the first FPGA deployment of the separately trained two-stage QCSNN on a PYNQ-Z2 platform, achieving cascaded accuracy within 0.08% of the GPU baseline at approximately 17× lower power consumption (2.15 W vs. 37.09 W).
Phase 3 advances the framework through a jointly trained dual-head architecture and a systematic ablation across eight configurations — two loss functions (Cross-Entropy and Focal Loss) crossed with four RR-interval feature routing strategies. The central finding is that RR-interval features substantially diminish the performance advantage of Focal Loss over Cross-Entropy (from +2.5% to near-zero or negative in Stage 2), suggesting that RR features compensate for class imbalance at the input level. The selected configuration (CE:RR→Both) achieves 99.02% cascaded accuracy. Deployed on the PYNQ-Z2 with a unified single-IP design featuring hardware-resident cascade routing, the system achieves 11.552 ms per-beat latency, 0.33 W accelerator power, and 23.32 mJ energy per inference — a 31.66× power reduction and 4.01× energy reduction versus GPU inference. To the best of our knowledge, this represents the first FPGA deployment of a fully quantized, directly trained SNN for multi-class ECG arrhythmia detection with systematic GPU-FPGA validation.
Recommended Citation
Banjo, Olamilekan, "DESIGN AND FPGA DEPLOYMENT OF QUANTIZED CONVOLUTIONAL SPIKING NEURAL NETWORKS FOR ECG ARRHYTHMIA CLASSIFICATION" (2026). Electronic Theses and Dissertations. 395.
https://digitalcommons.fau.edu/etd_general/395