Interoperable Web Platform Based on Large Language Models for Medical Data Analysis: A Protocol for Clinical Decision Support

Summary

This protocol describes the implementation of an interoperable web platform that integrates FHIR-based clinical data with Retrieval-Augmented Generation and multi-agent large language models for clinical decision support. The workflow enables reproducible deployment, standardized data integration, and evaluation of artificial intelligence (AI)-assisted medical data analysis.

Abstract

Clinical decision-making is frequently hindered by fragmented electronic health records and limited interoperability among heterogeneous healthcare information systems. This article presents a step-by-step protocol for implementing an interoperable web platform that integrates clinical data via the Fast Healthcare Interoperability Resources (FHIR) standard, along with Retrieval-Augmented Generation (RAG), Large Language Models (LLMs), and a multi-agent clinical reasoning framework to support medical data analysis and clinical decision support. The protocol describes the complete workflow, including computational environment configuration, clinical dataset preprocessing, FHIR-based data integration, vector database construction, retrieval configuration, prompt engineering, multi-agent orchestration, and system evaluation. Representative results demonstrate the platform's ability to generate clinically relevant and contextually consistent responses while improving semantic interoperability across heterogeneous data sources. System performance was evaluated using complementary quantitative and semantic metrics, including BLEU, ROUGE, BERTScore, and cosine similarity. A qualitative assessment was conducted using publicly available, fully anonymized benchmark healthcare datasets to evaluate the reproducibility of the proposed methodological workflow. The proposed architecture combines standardized healthcare interoperability with retrieval-enhanced language models to improve contextual reasoning, reduce hallucination, and support reproducible AI-assisted clinical workflows. This protocol provides a scalable and reproducible framework for researchers and developers seeking to implement interoperable, privacy-aware, and intelligent healthcare systems for clinical decision support, medical data analysis, and future translational research.

Introduction

Health information is routinely generated across hospitals, diagnostic centers, laboratories, and outpatient clinics, often remaining distributed among heterogeneous information systems. This fragmentation limits interoperability and restricts timely access to comprehensive patient records, creating persistent challenges for clinical care, data integration, and healthcare management^1,2,3,4.

The consequences of this fragmentation become particularly evident in clinical workflows that depend on timely access to pati....

Protocol

This study did not involve the recruitment of human participants, access to identifiable patient records, or experiments involving animals. The protocol was developed and evaluated exclusively using publicly available and fully anonymized datasets for methodological validation. No personal health information was accessed or processed. Therefore, approval from an Institutional Review Board (IRB) or Research Ethics Committee was not required. The protocol was developed in accordance with applicable data protection principles, including the Brazilian General Data Protection Law (LGPD), to support future applications involving clinical data.

Results

System performance was assessed using complementary quantitative metrics together with qualitative analysis to examine both linguistic accuracy and semantic consistency of the generated outputs. This evaluation strategy combines lexical and semantic measures to provide a broader characterization of model behavior in healthcare-related language generation tasks.

Table 1 presents the quantitative results obtained with BLEU, ROUGE, and BERTScore. These metrics were selected becau.......

Discussion

The evaluation presented in this study represents a methodological validation performed using publicly available and fully anonymized benchmark healthcare datasets. No prospective clinical validation involving healthcare professionals or patient recruitment was conducted, as the objective of this work is to demonstrate a reproducible implementation protocol rather than clinical effectiveness.

The findings of this study suggest that combining healthcare interoperability standards with large lan.......

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
FAISS	Meta Platforms Inc.	Version 1.8.0	Vector database used for similarity search in the Retrieval-Augmented Generation (RAG) pipeline.
FHIR Standard	HL7 International	Release 4 (R4)	Healthcare interoperability standard adopted for structured clinical data exchange.
Flask	Pallets Projects	Version 3.0.3	Python web framework used to implement the RESTful backend API.
HL7 Standard	HL7 International	Version 2.x	Legacy messaging protocol used for interoperability with hospital information systems.
HTTP REST API	Custom Implementation	N/A	RESTful communication layer between frontend, backend, database and AI services.
LangChain	LangChain Inc.	Version 0.3.x	Framework used to integrate LLMs, prompt engineering and Retrieval-Augmented Generation workflows.
LangGraph	LangChain Inc.	Version 0.2.x	Framework used for orchestrating the multi-agent clinical reasoning workflow.
LLaMA 3.1 8B Instruct	Meta Platforms Inc.	Version 3.1	Large Language Model employed for clinical reasoning and natural language generation.
MySQL Community Server	Oracle Corporation	Version 8.0	Relational database used to store structured clinical data.
Ollama	Ollama Inc.	Version 0.9.x	Local inference engine used for executing Large Language Models while preserving patient privacy.
Python	Python Software Foundation	Version 3.11	Programming language used for implementing the complete platform.
PyTorch	Linux Foundation	Version 2.4	Deep learning framework used for Large Language Model inference.
Retrieval-Augmented Generation (RAG) Pipeline	Custom Implementation	N/A	AI framework combining semantic retrieval with LLM inference for context-aware response generation.
Sentence Transformers	Hugging Face Inc.	all-MiniLM-L6-v2	Embedding model used to generate dense vector representations for semantic document retrieval.
Ubuntu Linux	Canonical Ltd.	Version 24.04 LTS	Operating system used for development and experimental evaluation.
Visual Studio Code	Microsoft Corporation	Version 1.100	Integrated development environment used for software implementation and debugging.