Winner, 2026 HackRare Genomic Diagnostics track
HackRare winner profile · Devpost submission
CheckMate is an equity-aware decision-support system for rare disease diagnosis. Given a patient's observed Human Phenotype Ontology (HPO) terms, it produces a calibrated disease differential and recommends the next symptom, exam finding, or test that would reduce diagnostic uncertainty the most.
DeepRare tells you what it could be. CheckMate tells you what to do next.
- Build a differential. CheckMate starts from DeepRare output when the service is available and provides local GNN and ontology-based fallback modes.
- Represent uncertainty. A Partial Variational Autoencoder uses frozen SHEPHERD knowledge-graph embeddings to model incomplete patient records.
- Choose the next action. The acquisition engine simulates positive and negative findings and ranks unobserved phenotypes by expected information gain.
- Account for access and documentation. Cost-aware ranking favors accessible observations for sparse records, while group-conditional conformal calibration adjusts diagnostic set sizes by documentation depth.
The submitted model was trained for 45 epochs on 226,000 HPOA-derived synthetic patients across 12,971 rare disease classes.
| Metric | Result |
|---|---|
| Top-1 disease accuracy | 64.2% |
| Top-5 disease recall | 82.5% |
| Expected calibration error | 0.021 |
| Temperature-scaling factor | 1.39 |
In the synthetic evaluation, group-conditional calibration reduced the coverage disparity between sparsely and well-documented cases from roughly 18 percentage points to about 1 percentage point. See the winner profile and Devpost write-up for the full methodology and results.
Requires Python 3.10 or newer.
git clone https://github.com/henrygreenhut/HackRare2026.git
cd HackRare2026
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
make downloadStart the API and UI in separate terminals:
make serve-api
make serve-uiThe repository includes the submitted CheckMate checkpoints. To run with the external DeepRare service, use make setup-deeprare and make serve-deeprare before starting the API.
activeraredx/— model, calibration, acquisition, and pipeline codeapi/andui/— FastAPI backend and Streamlit interfacescripts/— data preparation, training, and evaluation utilitiescheckpoints/— submitted model artifactsMODEL_ARCHITECTURE.md— detailed technical architecture
Built at Brown University by Sanil Desai, Henry Greenhut, and Akshay Vakharia for HackRare 2026.