Smartwatch-Based Eating Detection: Data Selection for Machine Learning from Imbalanced Data with Imperfect Labels.

Area of Science:

• Human-Computer Interaction
• Wearable Technology
• Machine Learning

Background:

• Understanding eating habits is vital for healthy lifestyle interventions.
• Objective measurement of meal timing, duration, and intake is needed.
• Wrist-worn devices offer a practical, unobtrusive method for real-time eating monitoring.

Purpose of the Study:

• To develop a novel approach for detecting eating segments using wrist-worn devices.
• To fuse deep and classical machine learning techniques for improved accuracy.
• To address challenges of data selection and imbalanced datasets in wearable sensor data.

Main Methods:

• Developed a novel data selection method for training datasets.
• Implemented a fusion of deep and classical machine learning algorithms.
• Utilized raw and virtual sensor modalities for training with imbalanced data.
• Evaluated the method using accelerometer and gyroscope data from 12 subjects in real-world conditions.

Main Results:

• Achieved high performance in person-independent eating segment detection.
• Reported precision of 0.85, recall of 0.81, and F1-score of 0.82.
• Demonstrated effectiveness across various meal types, cutlery, and locations.

Conclusions:

• Reliable eating detection is feasible using wearable sensors on the wrist.
• The proposed method shows promise for unobtrusive, real-time dietary monitoring.
• This technology can support interventions promoting healthier eating habits.