IRIS FHIR SQL Builder + Vector Search

This repository demonstrates a Proof of Concept (PoC) application leveraging InterSystems IRIS For Health to perform semantic search on FHIR data. It combines the capabilities of the FHIR SQL Builder for relational data projection and Vector Search for meaning-based retrieval.

Overview

This project enables semantic search on FHIR data using InterSystems IRIS. By combining the FHIR SQL Builder (for relational projection) with Vector Search (for meaning-based retrieval), it allows users to query clinical records using natural language, effectively bridging the gap between complex FHIR structures and human-readable search.

Features

• FHIR Repository Integration: Automated setup and loading of sample FHIR datasets.
• SQL Projection: Usage of InterSystems FHIR SQL Builder to expose FHIR data as relational tables without moving data.
• Vector Database: Utilization of InterSystems IRIS to store and query high-dimensional embeddings.
• Semantic Search: Find clinical records based on meaning (e.g., “heart issues” matches “cardiac arrest”).
• Containerized: Fully Dockerized setup for easy deployment.

Getting Started

Installation

1. Clone the repository

   git clone [https://github.com/fanji-isc/IRIS-FHIR-sqlbuilder-vectorsearch.git](https://github.com/fanji-isc/IRIS-FHIR-sqlbuilder-vectorsearch.git)
   cd IRIS-FHIR-sqlbuilder-vectorsearch
   

2. Build and Start IRIS For Health

   docker-compose up -d --build iris4health
   

3. Configure FHIR SQL Builder

   • IRIS Portal: http://localhost:8080/csp/sys/UtilHome.csp
     • Username: _SYSTEM
     • Password: ISCDEMO

   In IRIS Portal, navigate to Health–> FHIR Server Management–>FHIR SQL Builder(left side menu), fill out each section with below information.

   • Analysis

     • Name:FHIRSERVER
     • Host:localhost
     • Port:52773
     • Credentials:New
       • Name:Admin
       • Username:_SYSTEM
       • Password:ISCDEMO
     • FHIR Repository URL: /csp/healthshare/demo/fhir/r4

     In the New FHIR Analysis

     • FHIR Repository: FHIRSERVER
     • Selectivity Percentage:100

   • Transformation Specifications

     • Click Import and select the transform1.json file located in the fhirdata/ folder of this repository.

   • Projections

     • FHIR Repository:FHIRSERVER
     • Transformation Specification:transform1
     • Package Name:sql1

4. Accessing FHIR Data via SQL

   In the Management Portal SQL interface, select DEMO as Namespace you will see the SQL table generated from your QuestionnaireResponse data, ready for querying and analysis.

   SELECT * FROM sql1.QuestionnaireResponse
   

   

5. Application Setup & Vectorization

   Now that the data is projected to SQL, start the Python application to generate embeddings and perform searches.

   After running

   docker-compose up -d --build flask
   

   you should be able to access your application’s user interface at:

   http://localhost:8000

6. Useage & Demo

   Vector Search

   Example: Searching for “diabetes medication”
   Searching for “diabetes medication” returns not only patients taking drugs like metformin, insulin glargine, or acarbose, but also those whose family history includes Type 1 or Type 2 diabetes. This demonstrates how IRIS Vector Search connects related medical concepts across different fields, such as medications and family history.

   RAG

   Example: Multi-Condition Patient Query
   When a user asks, “Which patients have a family history of heart disease and are currently prescribed cholesterol-lowering medication?”, the system semantically retrieves relevant data from family history and medication fields — identifying links between terms like cardiac conditions, heart failure, or coronary disease with drugs such as propranolol, spironolactone, or atorvastatin.

   The LLM then summarizes these findings in natural language, giving clinicians a quick overview of at-risk patients without manual filtering or query writing.