ABS 022: Wearable ECG Arrhythmia Monitoring with Embedded Spiking Neural Network Inference

Julia Wong ¹, Jake Yoshimoto ¹, Ian Schiller ¹, Allen Shen ¹, Xinyi Wu ²

¹ University of California, Santa Cruz
² University of Electronic Science and Technology of China

The Van Wickle Journal (2026) Volume 2, ABS022

Introduction: Accurate and timely arrhythmia detection is essential for improving cardiovascular monitoring and enabling earlier clinical insight. Existing electrocardiogram (ECG) monitoring systems often rely on short-duration recordings or cloud-based analysis, limiting their ability to provide continuous, low-latency cardiac monitoring in portable settings. This work presents the development of a wearable ECG platform capable of performing fully local arrhythmia classification using embedded neural network inference.

The device employs a three-lead configuration following the Lead II convention, incorporating analog front-end amplification, right-leg drive for common-mode rejection, anti-aliasing, and digitization prior to digital preprocessing. A real-time signal processing pipeline performs filtering, R-peak detection, and fixed-window beat segmentation to generate standardized heartbeat inputs for machine learning classification. The filtering stage improves signal quality by attenuating baseline wander, muscle noise, and powerline interference commonly present in ambulatory ECG recordings.

For classification, a convolutional neural network (CNN) is trained on the MIT-BIH Arrhythmia Database using an inter-patient paradigm to promote unbiased generalization across individuals and reduce patient-specific overfitting. The trained CNN weights are subsequently transferred to a structurally aligned spiking neural network (SNN) architecture, enabling temporal spike-based inference for embedded deployment. ECG amplitudes are converted into firing rates, and spiking neurons integrate activity across discrete time steps to produce class predictions. Data-driven threshold calibration preserves classification fidelity while introducing event-driven computational dynamics compatible with neuromorphic computing principles.

The complete system is deployed on a Raspberry Pi 5 to demonstrate real-time, on-device classification into five clinically standardized beat categories derived from the MIT-BIH Arrhythmia Database: Normal (N), Supraventricular (S), Ventricular (V), Fusion (F), and Unknown (Q). This work establishes an integrated hardware-to-inference framework that bridges biomedical signal acquisition, embedded machine learning, and neuromorphic-ready deployment for next-generation wearable cardiac monitoring systems.

Methods: The wearable ECG device was developed using a three-electrode Lead II acquisition configuration connected to an analog front-end with amplification, right-leg drive stabilization, anti-alias filtering, and analog-to-digital conversion. Digitized ECG signals were processed on a Raspberry Pi 5 using a real-time digital signal processing pipeline consisting of bandpass filtering, adaptive R-peak detection, and fixed-window beat segmentation.

A convolutional neural network (CNN) was trained using heartbeat samples from the MIT-BIH Arrhythmia Database under an inter-patient training paradigm. The model classified beats into five AAMI-standardized categories: N, S, V, F, and Q. Following training, CNN weights were transferred into a structurally equivalent spiking neural network (SNN). ECG amplitudes were rate-encoded into spike trains, and leaky integrate-and-fire neurons accumulated membrane potentials over 64 discrete time steps. Threshold calibration was performed to preserve inference accuracy during CNN-to-SNN conversion and embedded deployment.

Results: The implemented signal processing pipeline achieved reliable real-time ECG preprocessing suitable for embedded inference, including accurate R-peak detection and heartbeat segmentation. The trained CNN demonstrated strong inter-patient classification performance on the MIT-BIH Arrhythmia Database, and the converted SNN maintained comparable classification behavior following CNN-to-SNN transfer. Real-time inference was successfully deployed on a Raspberry Pi 5, demonstrating fully local arrhythmia classification without external computation or cloud connectivity. The integrated system was capable of classifying ECG beats into five clinically standardized categories while maintaining low-latency operation compatible with wearable cardiac monitoring applications.

Discussion: This work demonstrates the feasibility of combining biomedical signal acquisition, embedded digital signal processing, and spiking neural network inference within a wearable ECG. The successful CNN-to-SNN conversion shows that neuromorphic-inspired computation can preserve classification capability while enabling event-driven inference suitable for on-chip devices. By integrating the complete pipeline into a portable embedded system, this project advances the translation of adaptive ECG classification from offline research environments toward real-time wearable deployment. Future work will focus on power optimization, expanded patient validation, improved robustness to motion artifacts, and deployment onto dedicated neuromorphic hardware for further reductions in energy consumption and computational overhead.