Chewing detection from an in-ear microphone using convolutional neural networks.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Accurate detection of chewing sounds for eating behavior monitoring is challenging due to sound variability and interference from non-chewing noises.
• Existing methods often rely on feature extraction and traditional classifiers, which can struggle with the complexity of audio data.

Purpose of the Study:

• To develop and evaluate a novel approach for chewing sound detection using feature learning in the time domain.
• To improve the accuracy and robustness of in-ear audio-based eating behavior monitoring systems.

Main Methods:

• Utilized 1-dimensional convolutional neural networks (CNNs) for direct feature learning on raw audio samples.
• Applied a CNN architecture comprising convolutional and fully connected layers to process audio windows.
• Aggregated detected individual chews into complete eating events.

Main Results:

• Achieved high effectiveness in chewing detection on a large, semi-free living dataset.
• Reported an accuracy of 0.980 and an F1 score of 0.883 for the proposed method.
• Demonstrated the capability to differentiate chewing sounds from speech and movement-related noises.

Conclusions:

• 1D CNNs offer a powerful approach for feature learning directly from time-domain audio for chewing detection.
• The proposed method significantly enhances the accuracy of in-ear audio-based eating behavior monitoring.
• This technique holds promise for unobtrusive and objective assessment of dietary habits.