Researchers propose RAG-Coding to boost medical LLM accuracy
Adding more AI agents doesn't always mean more accuracy—but when each agent has a distinct role, the whole becomes smarter than the sum of its parts.
A team led by Yidong Gan from Monash University has developed RAG-Coding, a novel approach that leverages retrieval-augmented generation (RAG) to improve large language model (LLM) performance in medical coding. The system orchestrates four specialized LLM agents that coordinate to cross-reference official coding guidelines and tabular lists, grounding their decisions in authoritative external knowledge.
On the MDACE dataset, RAG-Coding demonstrated significant improvements over existing LLM baselines, achieving 8-13% higher micro-F1 scores and 2-8% better macro-F1 scores across multiple LLM backbones. The method also showed superior recall (+11%) compared to the state-of-the-art PLM-ICD system, though PLM-ICD maintained slightly better precision (+6%). The researchers released an updated MDACE-2025 dataset with expert-annotated labels aligned to 2025 ICD-10-CM guidelines, enabling more accurate evaluation against current clinical standards.
- RAG-Coding demonstrates that multi-agent architectures with explicit retrieval steps can improve ICD-10 coding accuracy by 8-13%, but the four-LLM pipeline introduces latency and cost trade-offs that may hinder real-time clinical deployment.
- Commercial competitors like CodaMetrix, Nuance DAX, and Apixio rely on proprietary single-model approaches with human oversight, while RAG-Coding's open-source modularity offers a path to regulatory transparency and easier auditability.
- The hidden risks—fine-grained code granularity, cascading errors, and reliance on static benchmarks—mean that real-world validation is essential before the approach can be adopted in production medical coding workflows.
Why It Matters
Multi-agent RAG could redefine how clinical AI systems balance accuracy, interpretability, and cost—shaping the next phase of healthcare automation.