Food craving (FC) peaks are highly context-dependent and variable. Accurate prediction of FC might help preventing disadvantageous eating behavior. Here, we examine whether data from 2 weeks of ecological momentary assessment (EMA) questionnaires on stress and emotions (active EMA, aEMA) alongside temporal features and smartphone sensor data (passive EMA, pEMA) are able to predict FCs ~2.5 h into the future in N = 46 individuals. A logistic prediction approach with feature dimension reduction via Best Item Scale that is Cross-Validated, Weighted, Informative and Transparent (BISCWIT) was performed. While overall prediction accuracy was acceptable, passive sensing data alone was equally predictive to psychometric data. The frequency of which single predictors were considered for a model was rather balanced, indicating that aEMA and pEMA models were fully idiosyncratic.

Background: Food craving precedes unhealthy eating behaviors such as overeating or binge eating and is thus a promising intervention target. Yet, craving varies rapidly across the day and responds to external and internal context changes. This makes it a candidate for just in time adaptive interventions (JITAI), which, however, requires that it can be predicted ahead of time.Objective: To investigate whether upcoming food cravings could be detected from passive smartphone sensor data (excluding geolocation information) without the need for EMA-questionnaires.Methods: A sample of 56 participants recorded their food craving 6 times a day for 14 days as dependent variable. Predictor variables were environmental noise, light, device movement, screen activity, notifications and time of the day. For each participant we determined the best fitting food craving split and prediction algorithm in 10-fold cross-classification with a 75/25 train/test split.Results: Individual high vs. low craving ratings could be predicted with a mean Area Under the Curve (AUC) of 0.78 which outperformed a baseline model trained on past craving values in 85% of participants. Conclusions: Passive sensing of craving-preceding states seems viable based on person specific baseline data and individualized machine learning approaches. Craving prediction allows implementation of craving-preventive JITAIs. Within subject modeling increases the feasibility of the notoriously challenging task of craving prediction.

BackgroundFood craving relates to unhealthy eating behaviors such as overeating or binge eating and is thus a promising target for digital interventions. Yet, craving varies strongly across the day and is more likely in some contexts (external, internal) than in others. Prediction of food cravings ahead of time would enable preventive interventions.ObjectiveThe objective of this study was to investigate whether upcoming food cravings could be detected and predicted from passive smartphone sensor data (excluding geolocation information) without the need for repeated questionnaires.MethodsMomentary food craving ratings, given six times a day for 14 days by 56 participants, served as the dependent variable. Predictor variables were environmental noise, light, device movement, screen activity, notifications, and time of the day recorded from 150 to 30 min prior to these ratings.ResultsIndividual high vs. low craving ratings could be predicted on the test set with a mean area under the curve (AUC) of 0.78. This outperformed a baseline model trained on past craving values in 85% of participants by 14%. Yet, this AUC value is likely the upper bound and needs to be independently validated with longer data sets that allow a split into training, validation, and test sets.ConclusionsCraving states can be forecast from external and internal circumstances as these can be measured through smartphone sensors or usage patterns in most participants. This would allow for just-in-time adaptive interventions based on passive data collection and hence with minimal participant burden.