Scalable, Clinician-Supervised Generative-AI Food-Chaining Assistant for Pediatric ARFID: A Pilot Randomized Controlled Trial
Scalable, Clinician-Supervised Generative-AI Food-Chaining Assistant for Pediatric ARFID: A Pilot Randomized Controlled Trial
Children with Avoidant/Restrictive Food Intake Disorder (ARFID) often lack access to specialty dietitians, and scalable nutritional guidance/food chaining tools are currently not available. The investigators will evaluate a web-based, clinician-supervised, generative-AI assistant that produces individualized food-chaining plans.
Develop an AI assistant that generates ≥15 allergy-safe, evidence-based chaining steps per participant and meets ≥90 % expert agreement for safety/appropriateness.
Validate the assistant against gold-standard clinician recommendations (Cohen's κ ≥ 0.80).
Test clinical impact in a three-month pilot RCT (n = 96) by comparing change in Nine-Item ARFID Screen (NIAS) scores between intervention and usual-care groups.
Hypothesis: AI-generated plans will reduce NIAS scores by ≥3 points relative to controls.
ARFID is an eating disorder marked by highly selective eating that can lead to significant nutrient and energy deficiencies. Patients may show low interest in food because of certain tastes, textures, or smells; have chronically low appetite; and/or avoid eating out of fear (e.g., choking, vomiting), resulting in inadequate oral intake. A recent systematic review estimates the prevalence of ARFID in children and adolescents at 0.3-15.5%. Unlike anorexia nervosa or bulimia nervosa, body image concerns do not drive ARFID. Comorbidities-including autism spectrum disorder, attention-deficit/hyperactivity disorder, anxiety, and depression-are common.
Care is provided either by frontline clinicians (e.g., pediatricians, community dietitians) or multidisciplinary specialty teams. Outpatient management may combine cognitive-behavioral therapy adapted for ARFID (CBT-AR), family-based or exposure therapy, and dietitian-guided food-repertoire expansion via food chaining. Food chaining gradually introduces foods similar to an individual's "safe" items, creating an actionable, stepwise plan that accumulates meaningful dietary change.
Large-language models such as ChatGPT are increasingly used in health care. By constraining model output via an application programming interface (API), these models can power domain-specific chatbots. While ChatGPT excels at general queries, it is not inherently expert in ARFID or food chaining; however, with appropriate fine-tuning, LLMs could deliver scalable, personalized, high-quality chaining plans.
AI tools still pose risks. In 2023, the National Eating Disorders Association released "Tessa," an AI chatbot intended for eating-disorder support. Although a clinical trial showed modest reductions in weight/shape concerns at three and six months (d ≈ -0.2; p ≈ 0.03-0.04), the bot later provided harmful weight-loss advice, recommending calorie restriction and other disordered behaviors, leading to its shutdown. This case underscores the need for rigorous clinician oversight during early deployment of medical chatbots.
This research group has developed the only generative-AI tool that produces allergy-safe food-chaining recommendations, but it has not yet been clinically tested. This proposal builds on that proof of concept to evaluate its effectiveness in a broader pediatric ARFID population.
For this study, the investigators will be using two validated assessments of eating disorders:
EDYQ: The Eating Disorder in Youth Questionnaire (EDY-Q) is a brief, 14-item self-report screening tool designed for children and adolescents (roughly ages 8-13) to identify behaviors and attitudes consistent with DSM-5 feeding and eating disorders, especially ARFID. Each item is rated on a 5-point Likert scale (0 = "never" to 4 = "always"), yielding subscores for restrictive/avoidant eating and for weight- or shape-related concerns. Higher scores flag youths who may require a full clinical assessment for an eating disorder.
NIAS: The Nine-Item ARFID Screen (NIAS) is a concise, self-report questionnaire developed to detect symptoms of avoidant/restrictive food intake disorder across three DSM-5-aligned domains: sensory-based avoidance ("picky eating"), low appetite/lack of interest, and fear of aversive consequences. Respondents rate each item on a 0-5 Likert scale, producing domain scores and a total score that collectively gauge ARFID severity. Higher scores signal greater clinical concern and help clinicians identify individuals who warrant a full diagnostic evaluation and intervention.
The first phase of this project will be tool refinement, including GPT prompt refinement and testing, rule-based allergy filter creation, output of sample artificial participant food chaining recommendation sets, and refinement of recommendation set structure.
The second phase of this project will be validation of the tool, wherein the investigating RD and MD will independently assess the artificial recommendation sets for safety and appropriateness, with determination of Cohen's kappa, performed in a double-blinded manner.
The third phase of this project will be a pilot randomized controlled trial (with a subsequent crossover step to provide intervention for the control cohort). This phase involves publicizing the trial on websites and social media sites frequented by patients and families of children with ARFID, the primary means of recruitment. Prospective study participants will complete a comprehensive survey, including baseline demographic data, comorbidity data, and dietary preferences/dislikes, along with completion of the NIAS and EDY-Q questionnaires. A subset of participants who meet inclusion criteria (see below) will be randomized to either Arm 1 (Intervention Cohort) or Arm 2 (Control Cohort).
Arm 1 participants will be provided with 15 food chaining recommendations based on individual child dietary preferences, and will be assessed at the one month point to identify the number of food chaining interventions that have been undertaken and the number of recommendations that have been successful, along with a follow-up NIAS survey to determine differences in pre- and post-intervention ARFID severity. Arm 1 participants will exit the study at this point.
Arm 2 participants at the time of enrollment will not be provided food chaining recommendations, but will be assessed at the one month point to identify the number of food chaining interventions/new foods that have been introduced, along with a follow-up NIAS survey to determine baseline ARFID severity. At the one month point, Arm 2 participants will be provided with 15 food chaining recommendations based on individual child dietary preferences, and will be assessed after one month to identify the number of food chaining interventions that have been undertaken and the number of recommendations that have been successful, along with a follow-up NIAS survey to determine differences in pre- and post-intervention ARFID severity. Arm 2 participants will exit the study at this point.
Inclusion criteria: Children aged 3-17 years, with caregiver- or participant-reported DSM-5 ARFID diagnosis and/or EDYQ-screen-positive ARFID. English proficiency.
Exclusion criteria: As there is no validated non-English version of the NIAS, the investigators must exclude caregivers who do not have English proficiency. Participants must not have been previously treated at Boston Children's for ARFID.
Exposure: Generative AI-developed, respondent-specific set of food chaining recommendations Predictors: Age, gender, comorbid conditions, NIAS score, EDY-Q score, presence/absence of food allergies, presence/absence of celiac disease, subtypes of food preferences and aversions Primary: Change in NIAS total score (NIASΔ) at 4 weeks, number of interventions successfully trialed, number of new foods Secondary: Caregiver-reported trust-in-AI scale Phase 1: Preliminary Work Prototype (React + Python) ingests 15 peer-reviewed chaining papers; generates 15-item plans.
Expert rating (n = 40 plans) showed κ = 0.72, <1% unsafe items. Phase 2: Refinement (Months 0-1) Expand training set to 50 articles, add structured allergen ontology, and embed caloric-adequacy rules.
Guardrail testing on 1,000 adversarial prompts; target false-negative rate < 1 %.
Phase 2: Validation (Months 2-3) Design. Double-blind rating of 100 artificial patient cases; each plan scored by an ARFID MD and RD.
Analysis. Cohen's κ with 95 % CI; success threshold κ ≥ 0.80. Unsafe-item rate monitored.
Phase 3: Pilot RCT (Months 3-5) Trial. Recruitment and screening of participants, assignment of participants to intervention or control arms, evaluation by MD or RD of proposed plans, post-intervention survey; data analysis
Inclusion Criteria:
Exclusion Criteria:
paul.crowley@childrens.harvard.edu16177106519