Classifying Food Items During an Eating Occasion: A Machine Learning Approach with Slope Dynamics for Windowed Kinetic Data.

BACKGROUND

Wearable devices equipped with a range of sensors have emerged as promising tools for monitoring and improving individuals' health and lifestyle.

OBJECTIVES

Contribute to the investigation and development of effective and reliable methods for dietary monitoring based on raw kinetic data generated by wearable devices.

METHODS

This study uses resources from the NOTION study. A total of 20 healthy subjects (9 women and 11 men, aged 20-31 years) were equipped with two commercial smartwatches during four eating occasions under semi-naturalistic conditions. All meals were video-recorded, and acceleration data were extracted and analyzed. Food recognition on these features was performed using random forest (RF) models with 5-fold cross-validation. The performance of the classifiers was expressed in out-of-bag sensitivity and specificity.

RESULTS

Acceleration along the x-axis and power show the highest and lowest rates of median variable importance, respectively. Increasing the window size from 1 to 5 s leads to a gain in performance for almost all food items. The RF classifier reaches the highest performance in identifying meatballs (89.4% sensitivity and 81.6% specificity) and the lowest in identifying sandwiches (74.6% sensitivity and 72.5% specificity).

CONCLUSIONS

Monitoring food items using simple wristband-mounted wearable devices is feasible and accurate for some foods while unsatisfactory for others. Machine learning tools are necessary to deal with the complexity of signals gathered by the devices, and research is ongoing to improve accuracy further and work on large-scale and real-time implementation and testing.