RadOncRAG: A Novel Retrieval-Augmented Generation Framework Improves Large Language Model Benchmark Performance in Radiation Oncology.

Abstract

Large language models (LLMs) show promise in assisting knowledge-intensive fields such as oncology, where up-to-date information and multidisciplinary expertise are critical. Traditional LLMs risk hallucinations and reliance on static, possibly outdated data that lack domain-specific context. Retrieval-augmented generation (RAG) has emerged as a strategy to address these issues by incorporating domain-specific information from external knowledge repositories. We evaluated 15 LLMs, including Meta Llama-2/3, generative pretrained transformer (GPT)-3.5/4/4o variants, Claude-3, Gemini-2.0, and DeepSeek-R1. In a zero-shot workflow, each LLM answered 298 scorable questions from the 2021 American College of Radiology in-training examination. We implemented a RAG pipeline (Iridium Model) that transforms user prompts into vector embeddings, queries a specialized radiation oncology database, and merges relevant text with the original prompt to form an augmented query. We compared zero-shot versus RAG-augmented performance. Larger-parameter LLMs had higher zero-shot accuracy, with six models outscoring graduating residents (<i>P</i> < .01). Top scorers were reasoning models GPT-4o1, o3-mini, and DeepSeek-R1, which achieved 91.6%, 86.6%, and 91.6% without RAG, respectively. Gemini-2.0 improved 6.7% (to 79.2%), Llama-3-70b 8.4% (to 75.8%), and GPT-4o 5.7% (to 85.6%) with RAG. Top scoring reasoning models surpassed graduating resident averages by 17.7%-20% (<i>P</i> < .01), but had no improvement or detriment with RAG. Domain-specific gains occurred in clinical, biology, and physics. Majority voting boosted aggregate accuracy when individual model performance exceeded 50%. RAG workflows and reasoning models incurred higher computational costs. Radiation-oncology-specific retrieval-augmented generation pipeline enhances nonreasoning LLM performance in radiation oncology by integrating domain-specific evidence, whereas it does not improve performance of reasoning models. These findings demonstrate that RAG can elevate clinical decision support by enabling simpler, cost-effective nonreasoning models to tackle complex tasks through retrieval capabilities-an efficient alternative to extensive model training that also yields citable, evidence-based explanations.