DiagRAG: AI-Powered Rare Disease Diagnostic

Rare disease diagnosis is uniquely difficult: each disease has very few documented cases, symptoms overlap heavily, and diagnosis is treated as static classification — when in reality it is a sequential, uncertainty-heavy decision process.

A three-layer AI pipeline: phenotype ingestion → probabilistic inference → retrieval-augmented LLM reasoning. A Partial VAE generates a disease posterior with calibrated uncertainty, an information-gain module recommends the most informative next phenotype (cost-aware), and a RAG + LLM layer retrieves biomedical evidence to produce transparent clinical reasoning.

• Clinician enters observed HPO symptoms
• Probabilistic engine generates calibrated disease probabilities with uncertainty
• System suggests the most informative and cost-aware next phenotype or test
• RAG + LLM produces a transparent, evidence-grounded clinical explanation

• RAG-based clinical reasoning — evidence-grounded explanations for ranked diagnoses
• Uncertainty-aware reasoning — surfaces calibrated probabilities, not just top picks

• LLM: OpenAI GPT / Claude
• Embeddings & vector store: Hugging Face + FAISS / Qdrant
• ML: PyTorch
• Domain knowledge: SHEPHERD knowledge graph embeddings
• Stack: FastAPI + React
• Vibe coding: Antigravity