Background

Biomedical large language models (LLMs) allow extraction of medical insights deeply embedded within large repositories of unstructured text data. This study exploits natural language processing (NLP) to systematically analyze clinical narratives and provide data-driven total parenteral nutrition (TPN) practice behavior, serving as a proof-of-concept application of LLM in AI-assisted clinical decision-making in bariatric care. Complete whole text charts from 45 consecutive postoperative bariatric patients who were started on TPN at Brigham and Women's Hospital (Boston, MA) were collated and analyzed using an open source LLM within a secure network (Llama2). Search queries were formulated with good coverage and employed Elasticsearch retrieval system to maximize output relevance which was analyzed in aggregate. Wide collections of both structured and unstructured data were integrated for LLM analysis, as outlined in Figure 1 describing an example prompt used in our study. Iterative responses were quality checked for clinical relevance, accuracy, and hallucination effects. The most common reason for TPN initiation was poor oral intake and inability to tolerate food orally, followed by surgical complications including staple line leak. TPN was most commonly stopped due to resolution of oral intolerance and achievement of nutrition goals set at TPN initiation. Interestingly, analysis of patient narratives objectively reveals previously underappreciated insight into patient preference and perspective surrounding TPN versus tube feeds due to ease of management, comfort, and fear of enteral feeding tubes. Additional optimization of prompt engineering is in pursuit to further elucidate unrealized relationships between postoperative complications, patient satisfaction, and TPN therapy parameters.