An End-to-End Energy-Efficient Approach for Intake Detection With Low Inference Time Using Wrist-Worn Sensor

Area of Science:

• Human-Computer Interaction
• Wearable Technology
• Biomedical Engineering
• Digital Health

Background:

• Automated detection of intake gestures using wearable sensors is vital for understanding and intervening in eating behaviors.
• Existing algorithms often prioritize accuracy over efficiency, limiting real-world, on-device deployment for continuous dietary monitoring.
• Energy inefficiency of current methods impedes long-term, real-time tracking of dietary intake.

Purpose of the Study:

• To develop an optimized, template-based multicenter classifier for accurate intake gesture detection.
• To ensure low-inference time and minimal energy consumption for on-device deployment.
• To enable real-time, continuous dietary monitoring using wrist-worn sensors.

Main Methods:

• Utilized a wrist-worn accelerometer and gyroscope for data collection.
• Developed the Intake Gesture Counter smartphone application (CountING).
• Validated the algorithm against seven state-of-the-art approaches on three public datasets (In-lab FIC, Clemson, OREBA).

Main Results:

• Achieved optimal accuracy (81.60% F1 score) and low inference time (15.97 msec) on the Clemson dataset.
• Demonstrated comparable accuracy with superior inference times (13.8x faster on In-lab FIC, 33.9x faster on OREBA) on other datasets.
• Resulted in an average 25-hour battery lifetime, a 44%-52% improvement over existing methods on a commercial smartwatch.

Conclusions:

• The proposed template-based optimized multicenter classifier offers an effective and efficient solution for intake gesture detection.
• The algorithm enables accurate, real-time monitoring with significantly reduced energy consumption.
• This approach facilitates longitudinal studies and practical, on-device dietary tracking using wrist-worn devices.