Inspiration
The average rare disease patient sees 7 doctors over 4 years before getting a correct diagnosis. The clues are almost always in the chart. No one connects them. We built ATLAS to do exactly that.
What it does
ATLAS is an MCP server that performs end-to-end rare disease diagnostic analysis. It reads a patient's complete longitudinal FHIR history across 8 resource types, extracts HPO phenotype signals via Claude AI, cross-references 6 biomedical databases (Monarch Initiative, ClinVar, PubMed, NCBI Gene, HPO ontology, Epic FHIR), performs two-pass AI reasoning, and returns a structured diagnostic report with candidate diagnoses, missed red flags, causal genes, genetic panels to order, an urgency triage score, a ready-to-send genetics referral letter, and a plain-language patient summary.
How we built it
- MCP server built with FastMCP, streamable HTTP transport, 13 tools, stateless
- FHIR R4 reads and writes 8 resource types via Epic
- Two-pass AI reasoning where pass 1 marshals evidence for/against each candidate and pass 2 synthesizes the final report
- 6 external APIs all free: Monarch Initiative, NCBI ClinVar, PubMed, NCBI Gene, HPO ontology, Anthropic Claude
- Pure computation tools for lab trend analysis and urgency scoring with no external dependencies
- Deployed on Render with a live demo endpoint at /demo requiring no authentication
Challenges we ran into
Connecting everything end-to-end under time pressure. The Anthropic SDK had connection issues on Render's free tier so we replaced it with direct httpx calls to the API. Biomedical APIs have inconsistent response formats across Monarch, ClinVar, and PubMed which required careful fallback handling.
Accomplishments that we're proud of
A fully working rare disease diagnostic pipeline that correctly identifies hEDS and POTS in a synthetic patient with a 9-year diagnostic odyssey, misdiagnosed as fibromyalgia, anxiety, and IBS, in under 60 seconds. The system found 5 persistent multi-year lab abnormalities, 4 worsening trends, and generated a complete genetics referral letter automatically.
What we learned
Rare disease diagnosis is fundamentally a data integration problem. The signal is always there: elevated CRP for 3 years, recurrent subluxations, positive Beighton score, family history of hypermobility. But no single specialist sees the whole picture. AI agents that read longitudinally are uniquely suited to this problem.
What's next for ATLAS - Rare Disease Diagnostic Agent
Real Epic FHIR integration with live patient data, a clinician-facing dashboard, expanding beyond rare diseases to complex multi-morbidity cases, and partnership with rare disease patient advocacy organizations to validate diagnostic accuracy at scale.
Log in or sign up for Devpost to join the conversation.