Rare Disease Diagnosis Assistant

• GIF

  Rare Disease Diagnosis Assistant

Inspiration

Rare diseases affect 300+ million people worldwide, yet the average diagnosis takes 5-7 years. Doctors lack access to consolidated, up-to-date information scattered across medical databases. We built RDDA to bridge this gap using AI.

What it does

RDDA is an explainable, RAG-based AI assistant that:

• Retrieves real-time data from PubMed, Orphanet, and FDA APIs
• Builds an interactive knowledge graph linking diseases, drugs, symptoms, and phenotypes
• Uses Anthropic Claude to generate evidence-based answers with source citations
• Self-assesses confidence (0-1) and auto-generates follow-up questions when uncertain

How we built it

1. Data Collection — PubMed (NCBI E-utilities), Orphanet (REST API), FDA (OpenFDA) fetch medical data in real-time
2. Text Chunking & Embedding — Documents split into overlapping chunks, embedded using Sentence Transformers (all-MiniLM-L6-v2)
3. FAISS Vector Search — Semantic similarity search over 1500+ indexed chunks
4. Claude RAG Generation — Anthropic Claude synthesizes answers from retrieved context with inline citations
5. Knowledge Graph — NetworkX graph with interactive canvas-based visualization
6. Confidence Scoring — Two-stage: FAISS retrieval score + Claude self-assessment

Anthropic Claude Integration (4 uses)

1. Medical Entity Extraction — extracts diseases, drugs, symptoms from free text
2. RAG Answer Generation — generates cited answers from PubMed/Orphanet/FDA context
3. Confidence Self-Assessment — rates answer confidence, triggers follow-up questions if < 0.7
4. Knowledge Graph Unification — merges entities across data sources into unified graph

Challenges

• Rate limiting on Claude API during rapid-fire queries (entity extraction + RAG run simultaneously)
• Building real-time knowledge graphs from 3 different API formats with different schemas
• Ensuring citation URLs point to actual source pages, not just API endpoints

What we learned

• RAG with domain-specific medical data produces significantly better answers than general LLM queries
• Confidence self-assessment adds crucial transparency for medical AI applications
• Multi-source data fusion (PubMed + Orphanet + FDA) provides more comprehensive answers than any single source

Built With

• anthropic-claude
• docker
• faiss
• fastapi
• html5
• javascript
• networkx
• python
• sentence-transformers
• spacy