Food Frequency Questionnaire Personalisation Using Multi-Target Regression

Abstract

Fondazione Bruno Kessler is developing a mobile app prototype for empowering citizens to improve their health conditions through different lifestyle interventions that will be incorporated into a mobile application for lifestyle promotion of the Province of Trento in the context of the Trentino Salute 4.0 Competence Center. The envisioned interventions are based on promoting behaviour change in various domains such as physical activity, mental health and nutrition. In particular, the nutrition component is a self-monitoring module that collects dietary habits to analyse them and recommend healthier eating behaviours. Dietary assessment is completed using a Food Frequency Questionnaire on the Mediterranean diet that is presented to the user as a grid of images. The questionnaire returns feedback on 11 aspects of nutrition. Although the questionnaire used in the application only consists of 24 questions, it still could be a bit overwhelming and a bit crowded when shown on the screen. In this paper, we tried to find a machine-learning-based solution to reduce the number of questions in the questionnaire. We proposed a method that uses the user's previous answers as additional information to find the goals that need more attention. We compared this method with a case where the subset of questions is randomly selected and with a case where the subset is chosen using feature selection. We also explored how large the subset should be to obtain good predictions. All the experiments are conducted as a multi-target regression problem, which means several goals are predicted simultaneously. The proposed method adjusts well to the user in question and has the slightest error when predicting the goals.

Keywords: Food Frequency Questionnaires; dietary assessment; feature selection; machine learning; multi-target regression; self-monitoring.

Figure A1

Complete results. We compare results by the number of activated goals (rows), by the number of selected features (columns) and by the method used for feature selection (yellow, pink and blue colours within one subplot).

Figure 1

Example of FFQ representation as implemented in the LBC application. The FFQ is represented as a grid of images, and the system asks the user about his daily consumption of the 24 food items. The number of button presses on each image indicates the number of consumed portions of the food item presented with the corresponding image/question. (a) The figure on the left represents the FFQ in which 24 questions are represented in a 6 × 4 grid. (b) The figure on the right represents an example of a filled-in questionnaire after the user has marked the consumption of at least one portion of bread, fruit, vegetables, lentils, fish, cheese, olive oil and honey. The chosen food items have a darker (gray) background.

Figure 2

Graphical representation of the pipeline. We conduct three types of experiments: (1) we choose n random features (yellow section); (2) we choose n best static features in the case where all chosen goals are equally important (pink section); (3) we find n most important features based on previous answers gathered from the user (blue section).

Figure 3

Error in predicting consumed portions for the 11 goals. Each row presents the results for a certain number of activated goals. The first column (yellow bar plots) shows results for random features. The second column (pink bar plots) shows results for statically optimised features. Finally, the third column (blue bar plots) shows the results for personalised features.

Figure 4

Error in predicted portions per day averaged over the 11 goals. We compare the errors based on the number of activated goals (x-axes in the subplots), based on the number of features we want to select (different lines within one subplot), and based on the method for feature selection, each of the three subplots corresponding to one of the methods.