Detecting Eating Episodes From Wrist Motion Using Daily Pattern Analysis.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Accurate detection of eating episodes is crucial for health monitoring and dietary studies.
• Previous methods often rely on short-term data, missing broader contextual information.

Purpose of the Study:

• To develop and validate a novel framework for detecting eating episodes using full-day wrist motion data.
• To leverage diurnal context for improved accuracy in meal detection.

Main Methods:

• A two-stage framework analyzing local and daily probabilities of eating.
• Incorporation of an iterative retraining augmentation technique for day-length samples.
• Testing on the Clemson All-Day (CAD) and FreeFIC datasets.

Main Results:

• Achieved an eating episode true positive rate (TPR) of 89% and 84% time-weighted accuracy on the CAD dataset.
• Demonstrated substantial improvement in detecting eating episodes by including day-length analysis.
• Reduced transient false detections compared to shorter-window methods.

Conclusions:

• Full-day wrist motion analysis provides significant diurnal context, enhancing eating detection accuracy.
• The proposed two-stage framework with augmentation is effective for analyzing long-term sensor data.
• This method sets a new benchmark for accuracy in detecting eating episodes from wearable sensor data.