Low-Power Flexible Classifier Chip for Atrial Fibrillation Detection.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Wearable Technology

Background:

• Atrial fibrillation (AFib) detection is crucial for preventing stroke.
• Existing methods often require bulky equipment or complex processing.
• Need for efficient, real-time, and portable AFib detection solutions.

Purpose of the Study:

• To develop a fully integrated on-chip classifier for real-time AFib detection.
• To utilize a flexible substrate for wearable applications.
• To achieve high classification accuracy with low power consumption.

Main Methods:

• Digitization of electrocardiogram (ECG) signals using a 14-bit analog-to-digital converter (ADC).
• Time-domain feature extraction from ECG signals.
• Classification using a switched-capacitor compute-in-memory, analog, three-layer artificial neural network (ANN).
• Fabrication in 65nm technology.

Main Results:

• Achieved >99.5% accuracy on the Physionet dataset.
• Demonstrated 80% accuracy on a prospective human study.
• Reached 99.9% accuracy on the Rochester dataset.
• Low power consumption of 58.3 µJ per inference.

Conclusions:

• The developed on-chip classifier offers a promising solution for real-time AFib detection.
• The flexible substrate and low power consumption make it suitable for wearable devices.
• High accuracy across multiple datasets validates the system's robustness.