A Study of the Lexicography of Hand Gestures During Eating

TL;DR

This study develops and tests a standardized lexicography for classifying hand gestures during eating, demonstrating high inter-rater reliability and consistency across multiple raters and datasets, to support dietary intake monitoring research.

Contribution

It introduces a clear, objective gesture lexicography for eating actions and validates it through extensive labeling and reliability testing with multiple raters.

Findings

92.5% inter-rater agreement on gesture labeling

Low mistake rates for intake gestures (0.6% bite, 1.9% drink)

High agreement (95.8%) with previous gesture annotations

Abstract

This paper considers the lexicographical challenge of defining actions a person takes while eating. The goal is to establish objective and repeatable gesture definitions based on discernible intent. Such a standard would support the sharing of data and results between researchers working on the problem of automatic monitoring of dietary intake. We define five gestures: taking a bite of food (bite), sipping a drink of liquid (drink), manipulating food for preparation of intake (utensiling), not moving (rest) and a non-eating category (other). To test this lexicography, we used our definitions to label a large data set and tested for inter-rater reliability. The data set consists of a total of 276 participants eating a single meal while wearing a watch-like device to track wrist motion. Video was simultaneously recorded and subsequently reviewed to label gestures. A total of 18 raters manually labeled 51,614 gestures. Every meal was labeled by at least 1 rater, with 95 meals labeled by 2 raters. Inter-rater reliability was calculated in terms of agreement, boundary ambiguity, and mistakes. Results were 92.5% agreement (75% exact agreement, 17.5% boundary ambiguity). Mistakes of intake gestures (0.6% bite and 1.9% drink) occur much less frequently than non-intake gestures (16.5% utensiling and 8.7% rest). Similar rates were found across all 18 raters. Finally, a comparison of gesture segments against single index labels of bites and drinks from a previous effort showed an agreement of 95.8% with 0.6% ambiguity and 3.6% mistakes. Overall, these findings take a step towards developing a consensus lexicography of eating gestures for the research community.