ULECGNet: An Ultra-Lightweight End-to-End ECG Classification Neural Network

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Computer Science

Background:

• Electrocardiogram (ECG) classification is crucial for intelligent health monitoring.
• Traditional machine learning methods (SVM, KNN) show limited accuracy for ECG classification.
• Current end-to-end neural networks offer high accuracy but suffer from significant computational complexity and hardware demands.

Purpose of the Study:

• To propose an ultra-lightweight end-to-end neural network for ECG classification.
• To achieve high classification accuracy with reduced computational complexity.
• To enable deployment on low-cost microcontrollers (MCUs) for real-time applications.

Main Methods:

• Development of a novel, computationally efficient end-to-end neural network architecture.
• Minimization of network parameters and operations (approx. 8.2k parameters, 227k MACs).
• Implementation and evaluation on a low-cost MCU (MSP432) for real-time ECG analysis.

Main Results:

• Achieved an overall ECG classification accuracy of 99.1%, outperforming state-of-the-art methods.
• Demonstrated extremely low computational complexity suitable for resource-constrained devices.
• Real-time implementation on MSP432 consumed minimal energy (0.4 mJ/normal, 3.1 mJ/abnormal heartbeat).

Conclusions:

• The proposed ultra-lightweight neural network offers a viable solution for efficient and accurate real-time ECG classification.
• This approach facilitates the integration of intelligent ECG monitoring into low-cost, portable devices.
• The design significantly reduces power consumption and computational overhead compared to existing neural network models.