Can we predict snacking behaviour just from previous instances? (Preprint)

BACKGROUND Consuming too much food or drink with high levels of saturated fats, salt or sugar can be harmful for health. Many snack foods fall into this category (HFSS snacks). However, the palatability of these snacks means that people can sometimes struggle to reduce their intake. Machine learning algorithms could help by predicting the likely occurrence of HFSS snacking, so that just-in-time adaptive (JITAI) interventions can be deployed. However, HFSS snacking data has characteristics (such as sparseness and incompleteness), which make snacking prediction a challenging machine learning problem. Previous attempts have employed several potential predictor variables, achieving considerable success. Nevertheless, collecting information along several dimensions requires several potentially burdensome user questionnaires per day, so that this approach may be less acceptable among the general public. OBJECTIVE Our aim is to consider the capacity of machine learning algorithms to predict HFSS snacking based on minimal data that can be collected in a mostly automated way: day of week, time of day, and location (coarsened as work, home, other). METHODS A sample of 111 participants in the UK were asked to record HFSS snack occurrences and location category, over a period of 28 days, leading to a new dataset on HFSS snacks. Data collection was facilitated by a purpose-specific app. Additionally, we use a similar dataset from the Netherlands. For both datasets, we employ machine learning methods (random forests and neural networks). RESULTS We report results concerning the ability of machine learning methods to predict the time of the next HFSS snack. The quality of the prediction depended on both the dataset and temporal resolution employed. In some cases, predictions were accurate to as few as 17 minutes on average. CONCLUSIONS We demonstrated that prediction of HFSS snacks on sparse data is possible to reasonable accuracy. We consider the type of prediction problem which may be most suitable for putative interventions in relation to HFSS snacking. While we think it was important to employ standard machine learning algorithms in this work, we also discuss ways to tailor both the machine learning algorithms and the prediction problem to better align with the unique characteristics of the problem.