An automatic consult reply system for therapeutic plasma exchange using retrieval‐augmented generation

Abstract

BACKGROUND AND OBJECTIVES: Large language models (LLMs) show promise for clinical decision support but remain vulnerable to factual errors. Retrieval-augmented generation (RAG) mitigates this limitation by grounding outputs in authoritative domain knowledge. Therapeutic plasma exchange (TPE) requires consistent, guideline-driven decisions based on the 2023 American Society for Apheresis (ASFA) recommendations. This study aimed to evaluate whether an RAG-based framework could improve accuracy, reliability and standardization of decision support for TPE, compared to conventional LLMs. MATERIALS AND METHODS: We built a hybrid RAG pipeline combining BAAI/bge-base-en-v1.5 embeddings with Chroma and BM25, coupled with structured prompts that encode ASFA categories and grades, Health Insurance Review and Assessment (HIRA) service criteria, and plasma volume computation rules. Thirty de-identified real-world consultation cases were converted into standardized queries. Across six RAG and three non-RAG generative pre-trained transformer (GPT)-series model configurations, each case was answered five times (1,350 outputs). Performance was assessed by item-level accuracy for six elements (diagnosis, ASFA category, grade, insurance applicability, plasma volume, and replacement fluid) and reproducibility on 14 disease-name prompts. Response time and output length were also analyzed. RESULTS: RAG configurations consistently outperformed non-RAG baselines across items, with the largest gains in plasma-volume calculation and ASFA classification. Reproducibility was markedly higher with RAG across repeated runs. Among all configurations, RAG GPT-4.1-mini showed the most balanced and superior performance, delivering high accuracy with low latency. CONCLUSION: A guideline-grounded RAG approach substantially enhances the accuracy, stability and standardization of TPE consultation compared with conventional LLMs. This RAG-TPE framework demonstrates the feasibility of reliable, clinically oriented decision support in transfusion medicine, warranting further evaluation in prospective clinical workflows.

Ähnliche Arbeiten