Applying machine learning to ecological momentary assessment data to identify predictors of loss-of-control eating and overeating severity in adolescents: A preliminary investigation

Abstract

Objective

Several factors (e.g., interpersonal stress, affect) predict loss-of-control (LOC) eating and overeating in adolescents, but most past research has tested predictors separately. We applied machine learning to simultaneously evaluate multiple possible predictors of LOC-eating and overeating severity in pooled and person-specific models.

Method

Twenty-eight adolescents (78.57% female, age = 15.87 ± 1.59 years, BMI %ile = 92.71 ± 8.86) who endorsed ≥ two past-month LOC-eating episodes completed a week-long ecological momentary assessment protocol. Pooled models were fit to the aggregated data with elastic-net regularized regression and evaluated using nested cross-validation. Person-specific models were fit and evaluated as proof-of-concept.

Results

Across adolescents, the median out-of-sample R² of the pooled LOC-eating severity model was .33. The top predictors were between-subjects food craving, sadness, interpersonal conflict, shame, distress, stress (inverse association), and anger (inverse association), and within- and between-subjects wishing relationships were better. The median out-of-sample R² for pooled overeating severity model was .20. The top predictors were between-person food craving, loneliness, mixed race, and feeling rejected (inverse association), and within-subjects guilt, nervousness, wishing for more friends (inverse association), and feeling scared, annoyed, and rejected (all inverse associations). Person-specific models demonstrated poor fit (median LOC-eating severity R² = .003, median overeating R² = −.009); 61% and 36% of adolescents’ models performed better than chance for LOC-eating and overeating severity, respectively.

Discussion

Altogether, group-level models may hold utility in predicting LOC-eating and overeating severity, but model performance for person-specific models is variable, and additional research with larger samples over an extended assessment period is needed. Ultimately, a mix of these approaches may improve the identification of momentary predictors of LOC eating and overeating, providing novel and personalized opportunities for intervention.