Lightweight element-wise product enhanced neural network for efficient arrhythmia detection on embedded devices

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Cardiovascular Technology

Background:

• Cardiovascular disease necessitates accurate arrhythmia detection for improved patient outcomes.
• Deep learning models show promise for ECG-based arrhythmia classification.
• Computational efficiency is crucial for deploying models on resource-constrained embedded devices.

Purpose of the Study:

• To develop a lightweight, computationally efficient model for real-time arrhythmia detection on embedded platforms.
• To address the challenge of deploying deep learning for cardiac monitoring in resource-limited settings.

Main Methods:

• An elementwise-product enhanced lightweight model (EPLM) with spindle-shaped architectures and depthwise separable convolutions was developed.
• A novel elementwise product fusion mechanism enhanced high-dimensional feature representation.
• The model was evaluated on MIT-BIH, SVDB, INCART, and PTB databases, assessing accuracy, precision, recall, F1-score, computational complexity, inference time, and power consumption.
• Practical validation included deployment on a Raspberry Pi 5 and a smartphone for real-time applicability assessment.

Main Results:

• The EPLM achieved high classification accuracies (99.10% on MIT-BIH, 98.85% on PTB).
• The model requires minimal parameters (25,981) and FLOPs (525,400), with a fast inference time of 2.29 ms.
• Real-time operation was confirmed through deployment on embedded systems, demonstrating suitability for wearable applications.

Conclusions:

• The EPLM framework offers ultra-efficient arrhythmia detection with low computational overhead for continuous cardiac monitoring.
• This advancement facilitates early clinical intervention and improved patient outcomes in resource-limited settings.
• The model's efficiency and accuracy have the potential to transform real-time wearable healthcare technologies.