JMIR Publications

ABSTRACT

Background:

Prevention of binge eating through just-in-time mobile interventions requires prediction of high-risk times, e.g., by using affective states and associated contextual factors. Yet, these factors and states are highly idiosyncratic, and thus prediction models that are based on averages across individuals often fail.

Objective:

We thus developed an idiographic (within-individual) binge-eating prediction approach based on Ecological Momentary Assessment (EMA).

Methods:

We first derived a novel EMA-item set that covers a particularly broad set of potential idiographic binge-eating antecedents from literature research and an eating disorder focus group (N=11). Female patients with Bulimia Nervosa or Binge-Eating Disorder (N=13) then answered the final EMA-item set (14 days with 6 daily prompts). We used the correlation-based machine learning approach ‘BISCUIT’ (Best Items Scale that is Cross-validated, Unit-weighted, Informative and Transparent) to select parsimonious, idiographic item subsets and predict binge-eating occurrence from said EMA data (32-EMA items assessing antecedent contextual and affective states, and 12 predictors indicating district times of the day and time cycles).

Results:

The derived item subsets predicted binge-eating episodes on average with high accuracy (Area Under the Curve=.80, specificity=.87, sensitivity=.85, maximum reliability of rD=.52 and rCV=.17). Across participants, highly heterogeneous predictor sets of varying sizes (between 5 and 9 predictors) were chosen for prediction.

Conclusions:

Predicting binge-eating episodes from psychological and contextual states seems feasible and accurate, but predictor sets are highly variable and idiographic. This has practical implications for mHealth and Just-in-Time Adaptive Intervention approaches. Further, current theorizing around binge eating needs to account for this high between-person variability and broaden the scope of potential antecedent factors. Ultimately, a radical shift from purely nomothetic models towards idiographic prediction models and theories seems warranted.