Top-Down Detection of Eating Episodes by Analyzing Large Windows of Wrist Motion Using a Convolutional Neural Network

Area of Science:

• Wearable sensor technology
• Machine learning applications
• Human activity recognition

Background:

• Accurate detection of eating behavior is crucial for dietary monitoring and health research.
• Previous methods focused on individual hand-to-mouth gestures, often missing broader eating contexts.
• Longer temporal analysis windows can provide richer contextual information for improved detection.

Purpose of the Study:

• To develop and evaluate a novel method for detecting eating periods using wrist motion data.
• To investigate the impact of analyzing longer time windows on eating detection accuracy.
• To improve the precision and recall of automated meal detection systems.

Main Methods:

• Utilized a convolutional neural network (CNN) to analyze wrist motion data.
• Employed extended time windows (0.5-15 minutes) to capture contextual eating-related gestures.
• Tested the method on the public Clemson all-day dataset.

Main Results:

• The CNN approach demonstrated a 15% increase in eating detection accuracy with time windows of 4 minutes or longer compared to shorter windows (≤15 seconds).
• A 6-minute analysis window achieved 89% accuracy in detecting eating episodes.
• The method yielded a favorable false positive to true positive ratio of 1.7.

Conclusions:

• Analyzing longer time windows of wrist motion data with CNNs enhances the accuracy of detecting eating periods.
• This method offers a significant improvement over previous gesture-specific detection techniques.
• The findings represent a state-of-the-art performance for automated eating detection on the tested dataset.