Preventing Hallucinations in Clinical Coding AI: Utilizing FHIR Terminology Services and MCP
Clinical coding AI: Preventing hallucinations with FHIR
Summary
This article explains how to prevent 'hallucinations'—plausible but incorrect outputs—from Large Language Models (LLMs). It introduces a method using FHIR Terminology Services and the Model Context Protocol (MCP) to ground AI output in internationally standardized vocabularies, thereby enhancing the accuracy and reliability of clinical coding.
Details
While AI excels at natural language understanding, it faces limitations in highly precise fields like clinical coding. Since accurate identifiers must be retrieved from controlled international vocabularies (e.g., SNOMED CT, LOINC), general LLMs often generate plausible but incorrect codes. To solve this, FHIR (Fast Healthcare Interoperability Resources) Terminology Services are utilized. These services provide structured vocabularies like CodeSystems and ValueSets via a standardized REST API, acting as the verifiable 'source of truth' for AI output. Furthermore, the Model Context Protocol (MCP) provides a standard way to reliably connect external authoritative systems (the FHIR Terminology Service) to AI agents. This ensures that the LLM does not rely on internal knowledge and instead generates answers based on deterministic information retrieved from the server. This shift allows for transparent failure reporting—such as stating 'no valid code was found'—which is significantly safer than silent fabrication. This structural solution dramatically improves the safety of AI-driven clinical coding.
Technology Note
FHIRターミノロジーサービス+MCPの組み合わせはAI医療システム開発の新しい品質管理手法。LLM活用時の医療コーディング精度向上に有効。
© Better Care. All rights reserved.