JMIR Publications

ABSTRACT

Background:

Pregnancy is a period marked by heightened nutritional vulnerability and elevated risks for adverse maternal and infant outcomes. The Harris Health Food Farmacy project is a cross-sector collaborative effort to improve pregnancy-related outcomes using a medically-tailored groceries (MTG) approach providing home delivery of fresh produce among low-income, Medicaid-eligible high-risk pregnant individuals in Houston, TX receiving care at receiving care at Harris Health, a county safety net hospital system.

Objective:

We conducted a randomized control trial (RCT) to evaluate if on-demand nutrition education using Flora, a conversational agent (AI-based chatbot) improves food security, utilization of MTG provided, and dietary behaviors in a standard home delivery-based MTG model in high-risk pregnant individuals.

Methods:

We used an RCT design, randomizing 1:1 recruited participants to the control arm (biweekly produce deliveries + standard care), and the other half to a Chatbot arm (biweekly produce deliveries + standard care+ facilitated access to Flora). Pregnant individuals (n=212)≤20 weeks gestation and meeting the high-risk pregnancy criteria were consented, enrolled and randomized between June 2023-June 2025. The primary functions of Flora were to provide participants with on-demand advice related to pregnancy health, cooking tips, and culturally tailored recipes to support the use of fruits and vegetables included in the MTG boxes. Outcomes measured using self-report surveys at baseline, 3 months and 6 months post-baseline included food security, nutrition security, utilization of the food (amount of MTG consumed by the participants vs. Others in the household), and consumption of fruits and vegetables. For main outcomes, we estimated study arm-specific slope of change in different utilization and consumption metrics (e.g. linear change in proportion of vegetables consumed by patient and household respectively, and linear change in proportion of fruits consumed by patient and household respectively) using linear mixed effects models specifying time in months as a continuous variable and adjusting for person-specific random effects and co-variates.

Results:

The pilot RCT shows that the addition of an AI-based nutrition education chatbot to an MTG program targeting high risk pregnant participants did not substantially improve food security, utilization and consumption of vegetables. Results suggest that provision of MTG alone may be sufficient to improve diet-related outcomes, and call into question the additional utility of chatbots in providing nutrition education to this population.

Conclusions:

AI-based chatbot exposure had less than desirable uptake in our population of Medicaid pregnant women in Texas resulting in minimal impact on FV intake. Understanding barriers and facilitators to adoption and utilization of nutrition education in an AI-based chatbot format is an important next step to inform future design and delivery. Clinical Trial: ClinicalTrials.gov NCT07165990; https://clinicaltrials.gov/study/NCT07165990