Automatic Detection of Chewing and Swallowing

Area of Science:

• Biomedical Engineering
• Acoustic Signal Processing
• Health Informatics

Background:

• Eating behaviors (chewing, swallowing) are vital for health.
• Current data collection methods are invasive (X-rays, fiberscopes).
• A simpler, non-invasive measurement method is needed for healthcare.

Purpose of the Study:

• To develop an automated method for recognizing eating behaviors from sounds.
• To enable non-invasive monitoring of chewing and swallowing patterns.

Main Methods:

• Utilized a hybrid CTC/attention model for sound analysis.
• Employed 2-channel microphones placed under the ear for sound recording.
• Applied N-gram based data augmentation with weakly labeled eating sound data.

Main Results:

• Successfully detected left chewing, right chewing, front biting, and swallowing.
• The hybrid CTC/attention model improved detection performance by learning context.
• Achieved similar detection accuracy for both open and closed food types.

Conclusions:

• The developed sound-based method accurately recognizes eating behaviors non-invasively.
• This approach provides a simpler alternative to invasive techniques for health monitoring.
• The model's ability to learn context and handle augmented data enhances its practical application.