// the find
healthchainai/HealthChain
Python SDK for healthcare AI — typed, validated FHIR tools for agents, real-time EHR connectivity, production deployment ✨ 🏥
A Python SDK for building healthcare AI on FHIR. It provides typed R4B resources, a gateway that aggregates live data from multiple EHR FHIR APIs, CDS Hooks and CDA support, and a toolkit that exposes validated FHIR operations to LLM agents over MCP or LangChain. It is aimed at ML and AI engineers who need model output to land in a clinical system without first becoming FHIR specialists.
Validation and terminology checks sit inside the agent tools, so a model's write is checked for resource validity and code-system correctness before it reaches the record. Most agent-to-EHR attempts skip that step. The FHIR models come from fhir.resources, so you get typed, validated objects rather than dict-shaped JSON. The 'no invented facts' rule is a real design position: helpers don't stamp timestamps or guess statuses, and the gateway's provenance tagging marks AI-added content rather than silently changing it. There is more here than a README, with a Synthea and MIMIC sandbox, a CDA parser driven by Liquid templates, and a CLI that scaffolds a gateway project.
Most of the README is argument rather than evidence. 'Production-ready' and 'built for NHS and HIPAA environments' come with no threat model, compliance notes, or audit documentation, and the roadmap leaves governance, observability, and deeper EHR connectivity unchecked. Those are the parts a health system would need first. The headline Epic and Cerner example passes client_id in the source URL and doesn't show the auth flow a live EHR requires (SMART Backend Services uses signed JWTs and registered keys), so expect to read clients/auth.py before connecting to anything real. The surface is wide for a project with 228 stars: CDA, CDS Hooks, a SOAP NoteReader gateway, an interop engine, sandboxes, terminology, MCP and LangChain adapters, and a plugin marketplace. Depth in any one of these is hard to judge from the README, and it states no version or stability policy, so treat the API as unsettled until there is a 1.0.