Energy Efficient Arrhythmia Classifier Based on Event-Driven Data

Abstract

Electrocardiogram (ECG) monitoring is important for early detection of cardiac arrhythmias, but traditional uniformly sampled ECG acquisition often leads to high power consumption in long-term wearable systems. Level-crossing (LC) sampling provides an event-driven representation that can reduce redundant data. This thesis investigates heartbeat classification using LC ECG signals.

First, a feature-based classification framework is developed. ECG fiducial points are detected from LC signals, and morphological, rhythm, and LC dynamic features are constructed. After feature optimization, a multilayer perceptron classifier achieves an accuracy of 99.4% with a Macro-F1 of 0.949 on the MIT-BIH database.

Second, an event-driven spiking neural network (SNN) classifier is proposed. A time-aware leaky integrate-and-fire neuron model is introduced to incorporate event intervals into membrane dynamics. The proposed SNN achieves 99.02% accuracy and 0.936 Macro-F1, while improving the recognition of fusion beats. The results demonstrate the potential of event-driven ECG classification for low-power wearable monitoring systems.