Background: Strict, lifelong gluten-free diet(GFD) is essential in celiac disease(CeD), however poor availability of trained dietitians and sub-optimal knowledge limits its administration. With increasing availability of digital health infrastructure, a conversational agent powered by augmented intelligence could address the unmet need for available, accessible, affordable and contextually-relevant dietary counselling.
Methods: We developed a physician-supervised conversational agent powered by augmented intelligence, deployed over WhatsApp, to support GFD adherence. The system communicates with patients in multiple vernacular languages, accepts multimodal input (text, voice messages, and photographs of food/labels for gluten-safety assessment) and provides dietary advice. Responses are grounded using retrieval-augmented generation (RAG) over a curated knowledge-base comprising India-specific gluten-free recipes, transcripts of dietitian interviews of 30 unique patients with CeD, general GFD facts, and a structured food ontology database with cost and nutritional information. All patient conversations are logged, enabling physician review and ongoing monitoring of response accuracy and safety. All study procedures were performed with appropriate Institutional ethical clearance (AIIMSA3790/02.05.2025). Full source code is available at https://github.com/samagra-agarwal/virtual-dietitian-v0.1.
Results: In preliminary testing, 5 trained dietitians role-played as simulated patients, posing representative queries spanning dietary safety, meal-planning requests, and food-image assessment, including edge-case scenarios such as hidden gluten sources. Performance metrics related to safety, accuracy, user satisfaction, and cultural relevance were qualitatively assessed to further refine and adapt the model, ensuring user-centred functionality. Responses were consistently rated as accurate and contextually relevant, with appropriate deference to physician follow-up when information was incomplete.
Conclusion: A physician-supervised, RAG-grounded, multimodal conversational AI agent can function as a “virtual celiac dietitian”, and could be a low-cost, scalable tool for supporting GFD adherence in celiac disease across linguistically diverse populations. Prospective evaluation with actual patients, alongside formal safety and usability metrics, is warranted before wider clinical deployment.