FHIR Experiments and Experiences
In this presentation series, we aim to organize our thoughts, challenges, integration strategies, and lessons learned from working with FHIR.
Overview
We will cover:
- Challenges
- Experiments and tests conducted before, during, and after implementation
- Final thoughts and conclusions from the entire process
FHIR Related Challenges
Different types of challenges when working with FHIR
Structured Data Conversion
While not extremely challenging by itself, this process requires significant time for:
Unstructured Data Conversion
This presents significant challenges due to numerous variables affecting data transformation:
- Multiple file formats and document types
- Varying text structures and information organization
- Diverse expressions for the same concepts in plain text
- Limited, different or partially codification at the same documents
- And many more..
Data Conversion Validation
Due to the sensitive nature of clinical data, FHIR requires robust systems to ensure data integrity throughout the transformation process.
While feasible for small projects, validation becomes increasingly complex with larger system integrations.
Experiences and Experiments
List of conducted experiments
HAPI FHIR JPA Starter Server
This technology provides a quick deployment solution for local FHIR servers
Its advantage over other frameworks includes pre-configured FHIR resources and advanced search capabilities
API Testing Framework
A lightweight testing framework for custom FHIR APIs
Particularly useful for local FHIR servers with non-standard resources
IoT Devices with FHIR
Experiment using ESP32 microcontrollers with C++ libraries for IoT applications
FHIR in Workflow Applications
Integration possibilities with workflow tools like Node-RED, n8n, and Apache Airflow
We explored implementation approaches and potential applications
Synthetic Data Generation
AI-powered approach to create test data using prompt engineering
Valuable when real data is limited by availability or privacy constraints
AI Clinical Assistant
Google Notebook-based solution to address clinical queries about the project recommendations
AI-Powered FHIR Conversion API
Experimental API for FHIR data conversion using generative AI and LLMs
In the future, this kind of AI data conversion could be a good resource for unstructured data conversion
Gemma 2 Fine-Tuning
Easy way of fine-tuned Gemma2 model using unsloth, for more accurate FHIR data parsing (€10 cost)
Anticipate improved results as local LLM technology evolves
Chat with FHIR DATA
Easy and quick agent development using Google adk, mcp server and Hapi FHIR Server
Decoupled development of agents and tools as many experts tells is the future of AI.
Conclusions
Key takeaways from our FHIR implementation experience:
Development Best Practices
| Recommendations |
|---|
| Start simple: Begin with a local FHIR server (HAPI FHIR) and tools like Hoppscotch/Postman to understand FHIR workflows |
| Wait for maturity: Avoid coding until the Implementation Guide (IG) reaches advanced stages |
| Libraries over prototypes: Prefer mature libraries (HAPI FHIR) over quick Python prototypes for production |
| Validation first: JSON schema validation saves more time than traditional testing libraries |
| FHIR Bundle resources: Transaction bundles provide better context than individual resources |
| FHIR Search wisely: Reserve advanced search for complex operations; keep queries simple |
IoT & Workflow Integration Insights
| IoT | Workflow |
|---|---|
| C++ remains ideal for microcontroller compatibility between PIC, AVR (ATMega) and ESP32 | Multiple options to choose node-red, n8n, Apache Airflow, etc., |
| No mature open-source FHIR libraries for C++ | Excellent for heterogeneous EHR systems and multi-database environments |
| FHIR works over mutliple protocols like Bluetooth/MQTT/Zigbee but the payload could be big, so use a lightweight fhir resource or a server as middleware | Some has graphic programming, so it's accessible for developers of all skill levels |
| Not easy access or debugging | Complex debugging |
AI in FHIR: Tradeoffs
| Advantages | Challenges |
|---|---|
| Cost-efficient: Saves development/time costs for data conversion | Hallucination risk: Requires validation tools for model outputs |
| Context-aware: Handles varied clinical expressions and text structures | Limited models: Few quality in open source local models trained on medical data |
| JSON-friendly: Naturally aligns with FHIR's JSON structure | Resource-intensive: High inference costs for typical hospital infrastructure |
| Affordable tuning: Open-source models can be fine-tuned cost-effectively | Expertise gap: Requires specialized AI/ML knowledge |
Members of the IT Team
Marisa, Antonio, Vicente, Jose, Celia, Paco, Lucas