Estimation of Food Intake Quantity Using Inertial Signals from Smartwatches.

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Behavioral Science

Background:

• Accurate monitoring of eating behavior is vital for obesity and eating disorder management.
• Current methods often require multiple or specialized sensors, leading to poor adherence and data quality.
• There is a need for accessible, non-invasive dietary monitoring solutions.

Purpose of the Study:

• To develop a novel approach for estimating bite weight using data from a commercial smartwatch.
• To establish the feasibility of using inertial sensors alone for dietary intake estimation.
• To lay the groundwork for accessible, non-invasive dietary monitoring systems.

Main Methods:

• Utilized a publicly available dataset of smartwatch inertial data from ten participants.
• Combined extracted behavioral features (e.g., utensil loading time) with inertial signal statistical features.
• Employed a Support Vector Regression model for bite weight estimation, validated using a leave-one-subject-out cross-validation scheme.

Main Results:

• Achieved a mean absolute error (MAE) of 3.99 grams per bite.
• Demonstrated a 17.41% improvement in the MAE compared to a baseline model using the proposed 'improvement metric'.
• Outperformed an adapted state-of-the-art method, which showed a -28.89% performance against the same baseline.

Conclusions:

• Commercial smartwatch inertial sensors can accurately estimate bite weight.
• The proposed method offers a feasible and accessible approach for dietary monitoring.
• This research paves the way for future non-invasive and user-friendly dietary tracking systems.