The Restrain Food Database: validation of an open-source database of foods that should be eaten more or less as part of a healthy diet

Abstract

Studies of food-related behaviours often involve measuring responses to pictorial stimuli of foods. Creating these can be burdensome, requiring a significant commitment of time, and with sharing of images for future research constrained by legal copyright restrictions. The Restrain Food Database is an open-source database of 626 images of foods that are categorized as those people could eat more or less of as part of a healthy diet. This paper describes the database and details how to navigate it using our purpose-built R Shiny tool and a pre-registered online validation of a sample of images. A total of 2150 participants provided appetitive ratings, perceptions of nutritional content and ratings of image quality for images from the database. We found support for differences between Food Category on appetitive ratings which were also moderated by state hunger ratings. Findings relating to individual differences in appetite ratings as well as differences between BMI weight categories are also reported. Our findings validate the food categorization in the Restrain Food Database and provide descriptive information for individual images within this investigation. This database should ease the burden of selecting and creating appropriate images for future studies.

Keywords: food; food stimuli; healthy diet; open-source image database; personalization.

Figure 1.

Image showing the homepage of the OSF project for the Restrain Food Database. The wiki section on the left-hand side provides a description of the project and the various components which comprise the project. The components can be found on the right-hand side of the page. Each of these have a web page similar to the format above with the wiki page detailing its contents. Relevant files and folders can be found in the Files section at the bottom left-hand corner of the web page.

Figure 2.

Image showing the homepage of the R Shiny search tool. The left-hand panel lists various filters that can be applied to produce a subset of the database. All filter options will be applied upon opening the web page. To filter the dataset, either untick the categorical filters, or drag slider filters to where appropriate. The database in the centre of the page displays the output of the current filters using an icon of the image as well as descriptive information. The subset of images can be downloaded by pressing ‘Download Images’ and the descriptive information (e.g. taste, craving ratings, nutritional categorization, etc.) can be downloaded in CSV format by pressing the ‘Download Data’ button in the bottom left of the screen.

Figure 3.

A schematic diagram of the study design. Upon opening the survey, participants provided consent followed by demographic details and completed a state hunger measure. Participants were randomly allocated into 1 of 52 between-group image sets. Each set consisted of eight ‘Eat Less’ and two ‘Eat More’ foods which were created by randomly assigning all images from the food groups to a set so that each image appeared only in one set. The rating task used a randomized block design for Block A and B with Block C being presented at the end. Participants rated all 10 food images in the set before they moved on to the next block. The presentation of images and order of rating measures were randomized within each block. Following Block C, a catch-trial was implemented as an attention-check with the order of response options being randomized. After completing the check, participants were provided debrief information. The study took approximately 12 min for participants to complete.

Figure 4.

Panel showing an example of a trial from each block and the catch-trial. For the rating tasks, the image is presented in the middle of the screen with a description presented above. Rating measures are presented below the image in a randomized order. Panels (a–c) show the respectively named blocks; (d) shows the catch-trial used.

Figure 5.

Interval plots showing mean responses for cravings, taste and healthiness ratings, and their associated 95% confidence intervals. (a) shows ratings at the Food Category level and (b) shows ratings at the subcategory level.

Figure 6.

Bar chart describing responses for frequency of weekly consumption as the proportion of the responses within a subcategory. Ratings from each individual item in the study were counted according to their class interval (response categories for frequency of consumption) and subcategory. These frequencies were presented as the proportion of responses for a given class interval of the total amount of responses within a subcategory (i.e. approx. 20% of fruit items in this study were reported as being consumed 1–2 times a week across the sample). Error bars are 95% Wilson score intervals.

Figure 7.

Interaction plots of hunger and food category for primary outcomes. Shading around lines indicates 95% confidence intervals. Note that hunger score has been mean-centred.

Figure 8.

Interaction plots of sex and food category for cravings on the left hand panel with lines denoting 95% intervals. The right hand panel is an interaction plot of age and food category with shading around lines indicating 95% confidence intervals. Note that age has been mean-centred.