City Research Online

Abstract

Unhealthy snacking is considered as a risk factor for many diseases including cancer, obesity, cardiovascular disease, and diabetes and many would like to reduce their snacking. However, changing behaviour is challenging and many attempts achieve limited success. Interventions providing suggestions or feedback, often called nudges, can help, especially if they are delivered at the right time, i.e. just before the potential snacking. Machine learning models can be used to predict behaviour and potentially drive the timing of the nudges. Accordingly, the main goals of this project are to a) develop an app to collect data on snacking, b) collect and analyse snacking data, and c) build predictive models to support the timing of interventions.

We developed an app, called ‘SnackTracker’ which enabled users to track their snacking behaviour. We used this app to run a longitudinal study, collecting unhealthy snacking data in the UK between Dec 2019 and April 2020. 184 participants completed the initial survey and 111 participants completed the app-based data collection over a period of approximately 4 weeks. The data about time and location of snacks forms our ‘UK dataset’. For comparison, we analyse another dataset collected with different methodology in the Netherlands (the ‘Dutch dataset’), covering general eating behaviour including unhealthy snacks.

The UK and Dutch datasets were analysed for trends and patterns. We observed demographic and other differences. In both datasets, we found that unhealthy snacking takes place most frequently in the afternoon and at home. The snacking frequencies were similar between weekends and weekdays. We found generally more unhealthy snacking in the UK dataset, and the daily and weekly number of unhealthy snacks tended to reduce during the experiment in both datasets.

Bearing in mind support for timed interventions, we built models to predict unhealthy snacks in different ways. We applied a number of different models, including linear and nonlinear regressions, as well as fixed-context models, requiring fixed input size, and recurrent models. We evaluated the predictive models with standard classification and regression metrics. We also evaluated the trade-off between timing close to the event and the risk of intervening too late. We found that some of our models are usable for timing interventions under reasonable assumptions.

Based on the feedback from users and the observed reduction in snacking, even without interventions delivered by the app, the use of our tracking app can have positive effects. Given our modelling results, the implementation and evaluation of nudges timed by predictive models is now a practical next step for further development and research.

Publication Type:	Thesis (Doctoral)
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science Q Science > QA Mathematics > QA76 Computer software
Departments:	School of Science & Technology > Computer Science School of Science & Technology > School of Science & Technology Doctoral Theses Doctoral Theses

Export

Downloads