Enabling Eating Detection in a Free-living Environment: Integrative Engineering and Machine Learning Study.

Area of Science:

• Engineering and Machine Learning
• Wearable Technology
• Health Monitoring

Background:

• Accurate eating monitoring is crucial for managing chronic conditions like diabetes and eating disorders.
• Current automatic eating tracking in free-living settings is hindered by a lack of mature systems and large, reliable datasets.

Purpose of the Study:

• To address the gap in automatic eating detection by developing a large-scale, wearable-based system.
• To leverage engineering and machine learning for robust eating behavior monitoring.

Main Methods:

• A prospective, longitudinal study collected 3828 hours of data from Apple Watches (diary, accelerometer, gyroscope) streamed to iPhones and the cloud.
• Deep learning models were developed using spatial and temporal augmentation on the collected data.
• Personalized models were created to improve eating detection accuracy over time.

Main Results:

• The general population model achieved an area under the curve (AUC) of 0.825 within 5 minutes.
• Personalized models reached an AUC of 0.872, and meal-level detection achieved an AUC of 0.951.
• An independent validation cohort confirmed model robustness with a meal-level AUC of 0.941.

Conclusions:

• The developed deep learning models and data streaming platform demonstrate high accuracy in detecting eating behavior.
• The system's accuracy and platform robustness suggest immediate applicability for eating monitoring, particularly in diabetic integrative care.