xmiao/causal-odyssey

Causal Odyssey — Rare Disease Diagnosis

Gemma 4 Good Hackathon submission — Health & Sciences · Main · Ollama tracks.

A locally-hosted diagnostic reasoner that runs entirely through Gemma 4. No cloud calls, no API keys. Patient phenotypes never leave the machine.

What's running here

The demo combines:

• —Structural retrieval — gene-rescue + missing-pathognomonic penalty
• —Per-patient FCI causal edges (iGFCI-style individualized framework) — bidirected HPO pairs from patient-specific FCI
• —Gemma 4 reranking with per-patient FCI edges and Reactome pathway evidence in chain-of-thought
• —Chain-closure audit — deterministic Mondo bridge check; 21.6% tight closure rate on 1,122 patients
• —Verifier — flags top-1 when a pathognomonic HPO is missing; P(correct | confirm) = 32% vs P(correct | refute) = 10% (3.3× rejection-signal discrimination)
• —Interrogator — proposes the most-discriminating next HPO with Gemma-generated clinical rationale
• —Closed loop — answer, re-rank, watch rank trajectory update (novel ΔR@10 metric: +11.2pp pooled / +38.2pp verified)

HF Spaces build — important

HF free CPU tier can't run Gemma 4 26B. This Space ships with pre-computed Gemma outputs for curated showcase patients from our 1,122-patient benchmark evaluation run. Toggle Use Gemma 4 reranker to load the cached reasoning + chain-closure audit for those cases.

For live Gemma runs on arbitrary HPO inputs, clone the repository and run locally:

git clone https://github.com/xueerchen1990/CausalOdyssey
cd CausalOdyssey

# 1. Install Python dependencies (pre-requirements first, then runtime deps)
pip install -r pre-requirements.txt
pip install -r requirements.txt

# 2. Fetch the two large public datasets the demo needs (~115 MB)
bash scripts/download_data.sh

# 3. Start Ollama and pull Gemma 4 26B (~16 GB)
ollama serve &
ollama pull gemma4:26b

# 4. Run the demo
streamlit run app.py

Note on `pre-requirements.txt`: HF Spaces installs this file before requirements.txt automatically. For local installs, run both files in order as shown above — the pre-reqs pin modern pip/setuptools/wheel/numpy so that causal-learn and other scientific packages build cleanly.

Headline result

On 1,122 RareBench patients across four benchmarks (FullMA — full 4-agent stack):

• —Pooled R@1 +8.4pp (McNemar p = 1.4×10⁻¹⁴)
• —RAMEDIS R@1 +13.3pp (in-domain)
• —MME R@5 +12.5pp
• —HMS R@1 +2.3pp

Primary method contribution: individual-patient FCI causal edges supplied as Gemma chain-of-thought context, with a Verifier that calibrates rejection at the top of the differential.

What the showcase cases demonstrate

Disclosures

• —Quechua audio (referenced in the video): best-effort MMS-TTS synthesis from approximate Cusco Quechua, not native-speaker translation.
• —RareBench data (HPO + disease labels) is public; no proprietary biobank or clinical records used.
• —Gemma 4 26B (quantized) is the default model; pipeline is model-agnostic.
• —Not FDA-cleared. Decision support, not decision.

Repro

Full benchmark code and deployment configs used to compute the headline numbers are not released in this repository — a preprint with the evaluation details is forthcoming. This repo and the Space contain the method implementation and cached Gemma outputs for the curated showcase patients.