EdgeECG: a lightweight edge-oriented network with dual criterion pruning for real-time ECG arrhythmia classification.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Embedded Systems

Background:

• Miniature electrocardiogram (ECG) devices offer real-time cardiac signal acquisition for timely heart disease warnings.
• Accurate arrhythmia classification on resource-constrained edge devices is crucial for effective patient monitoring.

Purpose of the Study:

• To propose EdgeECG, an ultra-lightweight neural network for accurate arrhythmia classification on low-power microcontrollers.
• To enable efficient deployment of ECG analysis on embedded platforms.

Main Methods:

• Designed a compact convolutional architecture for EdgeECG to ensure compatibility with resource-limited embedded platforms.
• Implemented a dual criterion pruning (DCP) strategy for precise model compression by evaluating weight importance.
• Applied quantization to reduce storage cost and enhance deployment efficiency on an STM32F103 microcontroller.

Main Results:

• EdgeECG achieved 98.07% overall classification accuracy on the MIT-BIH arrhythmia dataset, outperforming existing methods.
• The model, with 2680 parameters, demonstrated low inference latency (0.025 s) and energy consumption (0.156 mJ) on an STM32F103.
• DCP reduced non-zero parameters by nearly 50% while maintaining high classification performance.

Conclusions:

• EdgeECG offers an effective solution for five-class ECG arrhythmia classification on resource-constrained edge devices.
• Its compact design, pruning strategy, and successful deployment highlight its potential for low-power edge-based ECG analysis.
• The study demonstrates the feasibility of EdgeECG for wearable cardiac monitoring applications.