Analog Gated Recurrent Unit Neural Network for Detecting Chewing Events.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Accurate detection of eating behaviors, such as chewing, is crucial for various health applications, including dietary monitoring and clinical diagnostics.
• Existing methods for detecting chewing events often rely on complex algorithms or power-intensive hardware, limiting their applicability in real-world scenarios.

Purpose of the Study:

• To develop and evaluate a novel analog integrated circuit implementing a gated recurrent neural network for efficient and accurate chewing detection.
• To assess the performance and power consumption of the analog neural network for real-time chewing event identification.

Main Methods:

• A custom analog integrated circuit was designed and fabricated using 0.18 μm CMOS technology, incorporating a gated recurrent neural network architecture.
• The neural network was trained on 6.4 hours of chewing data collected via a contact microphone placed on volunteers' mastoid bones.
• Performance was evaluated on 1.6 hours of unseen data, measuring recall, F1-score, and power consumption.

Main Results:

• The analog neural network successfully identified chewing events with a 24-second time resolution.
• High performance was achieved, with a recall of 91% and an F1-score of 94%.
• The chewing detection system demonstrated exceptionally low power consumption, operating at 1.1 μW for chewing event detection and an estimated 18.8 μW for whole eating episode detection.

Conclusions:

• The developed analog neural network circuit offers a highly accurate and power-efficient solution for detecting chewing events.
• This technology holds significant potential for unobtrusive, real-time dietary monitoring and other health-related applications through wearable devices.
• The low power footprint of the system makes it suitable for integration into long-term wearable health monitoring solutions.