A Comprehensive Evaluation of LLM Phenotyping Using Retrieval-Augmented Generation (RAG): Insights for RAG Optimization

Objective: ICD codes are commonly used to filter patient cohorts but may not accurately reflect disease presence. Furthermore, many health problems are recorded in unstructured clinical notes, complicating cohort discovery from EHR data. Existing computed phenotyping methods have limitations in identifying evolving disease patterns and incomplete modeling. This study explores the potential of LLMs, by evaluating GPT-4o type II diabetes mellitus (T2DM) phenotyping ability using Retrieval-Augmented Generation (RAG). Methods: A RAG system was built, leveraging 275 patients entire notes. We performed total 336 experiments to study the sensitivity of RAG to various chunk sizes, the number of chunks, and prompts across seven embedding models. Then the effectiveness of GPT-4o in T2DM phenotyping was assessed using optimized RAG configurations, comparing with ICD code and PheNorm phenotype performance. Token usage was also evaluated. Results: The results show that GPT-4o with optimized RAG significantly outperformed ICD-10 and PheNorm in sensitivity, NPV, and F1, although PPV and specificity need improvement. When used with general embedding models or a zero-shot prompt, the results showed better sensitivity, NPV, and F1-scores, while domain-specific models and a few-shot prompt excelled in specificity and PPV. Furthermore, RAG optimization allowed lower-ranked embedding models achieve reliable performance. Gte-Qwen2-1.5B-instruct and GatorTronS provided the highest performance in specific evaluation metrics at a substantially lower cost. Conclusion: Optimized RAG configurations significantly enhanced key performance metrics compared to existing methods. This study provides valuable insights into optimal configurations and cost-effective embedding model choices, while identifying limitations such as ranking issues and contextual misinterpretation by LLM.