AN INTELLIGENT FHIR-INTEGRATED CLINICAL DECISION SUPPORT SYSTEM USING RETRIEVAL-AUGMENTED GENERATION AND LARGE LANGUAGE MODELS
Keywords:
Clinical decision support, FHIR R4, large language models, retrieval-augmented generation, SOAP notes, LangChain, FAISS, EHR integration, clinical NLP, healthcare AI.Abstract
The integration of large language models with clinical data standards has opened new avenues for automated medical documentation and decision support. This paper presents the design, implementation, and evaluation of an intelligent clinical decision support system that connects the HL7 Fast Healthcare Interoperability Resources (FHIR) R4 standard with a LangChain-orchestrated multi-LLM pipeline and a FAISS-backed retrieval-augmented generation knowledge base. The system exposes a Flask REST API through which clinicians can interrogate patient records using natural language, receiving automatically generated SOAP medical notes compiled from live FHIR resource bundles and evidence-grounded treatment recommendations derived from ingested clinical guideline PDFs. The architecture employs Meta Llama-3.1-405B-Instruct via NVIDIA AI Endpoints for inference, BAAI/BGE-M3 for dense semantic embedding, and MongoDB for session persistence, with a custom rotating audit logger. A per-condition RAG retrieval strategy prevents relevance dilution in comorbid patients. Evaluation against the HAPI FHIR R4 public reference server across 50 trials demonstrates 100% retrieval accuracy, SOAP note quality Likert scores above 4.6, zero hallucinated treatment citations, and a mean end-to-end latency of 4.2 seconds. The modular, provider-agnostic configuration design supports on-premises deployment for data-residency-constrained environments.
