A Tiny Transformer for Low-Power Arrhythmia Classification on Microcontrollers.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Cardiovascular Health

Background:

• Wearable systems are crucial for continuous cardiovascular disease monitoring.
• Transformer models show promise for real-time electrocardiographic (ECG) signal analysis and arrhythmia detection.
• Efficient implementation of these models on low-power wearable devices presents significant challenges.

Purpose of the Study:

• To develop a compact transformer model for ECG analysis.
• To achieve high accuracy in detecting common arrhythmia classes.
• To ensure efficient deployment on low-power microcontrollers for wearable applications.

Main Methods:

• A tiny transformer model with 6k parameters was designed for ECG signal analysis.
• The model was trained using an augmentation-based approach to enhance robustness against motion artifacts.
• Performance was evaluated using 8-bit integer inference on the MIT-BIH Arrhythmia database and deployed on a GAP9 processor.

Main Results:

• The model achieved 98.97% accuracy in recognizing 5 common arrhythmia classes.
• Post-deployment accuracy, considering noise, was 98.36% in the worst-case scenario.
• Inference on the GAP9 processor took 4.28ms and consumed 0.09mJ, demonstrating high efficiency.

Conclusions:

• The proposed tiny transformer model offers a highly accurate and efficient solution for real-time arrhythmia detection in wearable systems.
• The augmentation-based training improves model robustness, crucial for practical wearable device deployment.
• The model's low complexity and power consumption make it ideal for integration into ultra-low-power microcontrollers for continuous cardiovascular monitoring.