P194 - Data-Driven Psychosocial Phenotyping for Precision Behavioral Nutrition

Abstract

Background (Background, Rationale, Prior Research, and/or Theory): People respond differently to behavior change interventions; therefore, behavioral nutrition interventions often have small effect sizes. Precision Behavioral Nutrition aims to leverage information about individuals to make these interventions more effective via personalization. Psychosocial phenotyping describes variability in psychological and social characteristics in a population, identifying salient mediators of behavior change. Previously, we used psychosocial phenotyping to qualitatively explain results of a behavioral nutrition intervention and noted its potential for designing precision interventions. For this purpose, psychosocial phenotyping must be scalable and identify mediators that can serve as intervention targets along with related moderators and behaviors. Mixed membership models, including Latent Dirichlet Allocation (LDA) and UPhenome, are machine-learning methods used for clinical phenotyping. Whereas LDA treats mediators, moderators, and behaviors as a single data type, UPhenome handles them separately, learning phenotypes that represent each equally. These data-driven models have not yet been applied in behavioral nutrition. Objective: We assessed the utility of LDA and UPhenome for automatic learning of psychosocial phenotypes from survey data. Study Design, Setting, Participants, Intervention: We extracted individual's (n = 5,883) psychosocial characteristics from surveys collected during a cohort study of a predominantly Latino New York City community affected by high levels of chronic disease and health disparity. We conducted a secondary analysis of these data using UPhenome and LDA. Outcome Measures and Analysis: Participant responses were tokenized (n = 4,604) and used as input to LDA and UPhenome to identify phenotypes (k = 10). Domain experts (n = 4) assessed internal consistency among phenotypes resulting from LDA; a heatmap (n = 3,233) depicted dietary patterns for each phenotype. Results: Experts found within-phenotype consistency and between-phenotype discrimination. Differences in dietary patterns were visually detected. UPhenome identified more specific intervention targets than LDA. Conclusions and Implications: LDA and UPhenome can identify psychosocial phenotypes in a large cohort of Latino New Yorkers and could advance Precision Behavioral Nutrition. Future research will assess generalizability of psychosocial phenotypes and integrate psychosocial phenotyping into behavioral nutrition interventions. Funding: NIH Sackler Institute for Nutrition Science at the New York Academy of Sciences.