S742 Natural Language Processing to Assess Patient Sentiments Toward Different Diets for Inflammatory Bowel Disease

Introduction: Randomized controlled trials and prospective observational studies have demonstrated the efficacy of several medical nutrition therapies for inflammatory bowel disease (IBD). However, unlike pharmacologic agents, diets are heterogeneous, and corresponding nutrition research has been scarce. A cost-effective approach to explore the benefits and detriments of different diets—and discover new dietary strategies—may be to crowdsource patients’ experiences. The aim of this study was to develop a natural language processing (NLP)-aided system to efficiently identify diets and analyze sentiment in a large volume of patient narratives. Methods: We retrieved all public posts between January 2011 and April 2025 from the nutrition section of an IBD-focused Internet forum. At least 2 investigators independently reviewed each post to identify the diet discussed and the poster’s sentiment (positive, neutral, negative). An NLP pipeline was constructed, including tokenization, word/phrase parsing, and integration with a local implementation of the Meta Llama 2 13-billion parameter large language model. The NLP system was instructed to review each post, identify the relevant diet discussed, and assess sentiment, mirroring the investigators. Using investigator-adjudicated labels as ground truth, diet and sentiment concordance (accuracy) were calculated. Results: Of 2,410 posts reviewed, 1,254 (52.0%) did not include enough information to identify a particular diet. Mean context length was 570.2 characters with mean processing time of 50.9 seconds. Overall concordance between the NLP system and investigators were 85.7% for diet identification and 76.3% for sentiment assessment. Excluding posts without diet information, concordances were 96.1% for diet and 77.0% for sentiment. The most discussed dietary interventions were the Specific Carbohydrate Diet (SCD), plant-based diets, and low gluten diet with favorability ratings at 70.8%, 58.5%, and 65.7%, respectively. Diets with the highest favorability ratings were the “organic diet,” “healthy diet,” and the autoimmune protocol at 93.8%, 86.5%, and 86.2%, respectively. Conclusion: The NLP-aided system can effectively identify diet interventions and assess sentiment in a large volume of patient narratives. This approach can help assess the relative benefits—or detriments—of diverse dietary strategies in IBD and identify novel diet candidates for systematic investigation and clinical trials in the future.