A Cascaded Quantized Spiking Neural Network for Real-Time ECG Arrhythmia Detection on Edge Hardware

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Cardiovascular Technology

Background:

• Wearable ECG monitors offer continuous cardiac surveillance but often depend on cloud analysis, limiting real-time on-device arrhythmia detection.
• Spiking neural networks (SNNs) show promise for low-power edge computing, but their application in directly trained, quantized SNNs for multi-class arrhythmia detection, especially with RR-interval features and FPGA validation, is underexplored.

Purpose of the Study:

• To propose and validate a quantized convolutional spiking neural network (QCSNN) for real-time, multi-class arrhythmia detection on resource-constrained hardware.
• To investigate the interaction between class-imbalance loss functions (Cross-Entropy and Focal Loss) and RR-interval features in directly trained quantized SNNs.
• To evaluate the performance of the proposed QCSNN model on a PYNQ-Z2 FPGA for practical edge deployment.

Main Methods:

• Developed a QCSNN with a dual-head architecture for joint binary and four-class classification, reorganized into a cascaded pipeline for efficient inference.
• Evaluated two loss functions (Cross-Entropy and Focal Loss) across four RR-feature routing strategies to assess their impact on classification performance.
• Deployed the optimized CE:RR→Both configuration on a PYNQ-Z2 FPGA, measuring accuracy, latency, and power consumption.

Main Results:

• RR-interval features significantly improved Stage-2 macro F1 by +0.028 to +0.034, compensating for class imbalance and outperforming the advantage of Focal Loss over Cross-Entropy.
• The FPGA deployment achieved 99.02% cascaded accuracy with 11.54 ms per-beat latency and 0.33 W accelerator power, representing substantial power and energy reductions compared to GPU inference.
• The QCSNN demonstrated a practical solution for real-time edge arrhythmia monitoring, independent of cloud connectivity.

Conclusions:

• Quantized SNNs are a viable solution for real-time, low-power edge arrhythmia detection in wearable devices.
• RR-interval features are crucial for enhancing SNN performance in handling class-imbalanced ECG data.
• The study provides one of the first systematic analyses of loss-function/RR-feature interactions in SNN arrhythmia classification and an early FPGA implementation of a fully quantized, directly trained SNN for this task.