Early Signals

Catching gastric cancer before stage IV, using the clinic records Rwanda already collects.
Surveillance, risk flags, care coordination and an AI agent, built on a synthetic OpenMRS-style EMR.

▶ Watch the 39-second demo reel

Important

All data in this project is synthetic. No real patients, facilities or district statistics appear anywhere. The risk models are proofs of concept and are not for clinical use.

Background

Gastric cancer cases are rising in Rwanda, and most patients are diagnosed late, when treatment options are limited and survival is poor. Hospitals already record diagnoses, lab results, endoscopy reports and demographics in their EMRs, but that data is rarely analysed to see how the disease is moving through the population or who is most at risk. Key risk factors such as H. pylori infection are often not tested or not recorded.

Early Signals shows what routine EMR data can do when it is put to work end to end. A synthetic 1.5M-person Rwandan EMR feeds an incremental analytics pipeline and three tiers of risk models. Ministry analysts see where the burden is growing. District doctors see who needs an endoscopy, and why. And once a doctor approves a care plan, the patient gets a plain-language nudge on their phone, with no diagnosis words, only advice to visit.

Features

National picture. Age-standardised rates, joinpoint trends, 3D district hotspots, survival and H. pylori testing for the ministry.	Case analysis in 3D. A patient's symptoms, labs and vitals light up on an anatomical body, with a month-by-month replay.
Who needs a scope, and why. Risk-ranked patients for each facility, with the reasons behind every flag.	Care plans in the patient's pocket. Doctor-approved plans, reminders by app, SMS and community health worker, closed by EMR evidence.
Outlook to 2031. Forecasts with 80/95% bands, the drivers behind the change and what-if scenarios.	AI agent. Ask in plain language and get charts with numbers checked against the data. Also available over MCP.

• Surveillance: age-standardised rates with confidence intervals, joinpoint regression, spatial clustering (LISA, Gi*, SIR), Kaplan–Meier survival and care-quality funnel plots.
• Early warning: pre-diagnostic signal curves, missed alarm symptoms, H. pylori testing gaps and diagnostic delays.
• Three tiers of risk models: a points score, XGBoost with SHAP reasons and a JAX sequence model, combined into an ensemble (AUROC 0.970 and a median lead time of 6.2 months on the synthetic test set).
• Care coordination: six care pathways, a reminder ladder, a patient app and a simulation clock that shows whether patients come back.
• Learning loop: a challenger model trained on verified outcomes, with gates and a human Promote button.
• Role-scoped by design: the ministry sees only aggregates (counts below 5 suppressed), doctors see only their facility, and patients see only their own record.
• Data videos and daily briefs: Remotion videos of a case or the national picture, and a one-page PDF brief each morning.

See Architecture for every screen and component.

Tech stack

Layer	Tools
Synthetic EMR	Python 3.11, NumPy, Polars, PyArrow, OpenMRS-style MySQL schema
Pipeline and analytics	DuckDB (blue/green publish), statsmodels, lifelines, PySAL (libpysal, esda), SciPy
Machine learning	scikit-learn, XGBoost, SHAP, JAX (GRU/Transformer sequence model), APC + ETS forecasting
API	FastAPI, Pydantic, sqlglot (guarded NL→SQL), WebSockets
Dashboard	React 18, TypeScript, Vite, HeroUI, Framer Motion, ECharts, deck.gl, react-three-fiber
AI agent	Node 22, Hono, Vercel AI SDK, MCP (Streamable HTTP), Claude / OpenRouter / OpenAI-compatible models, Python sandbox, libSQL
Videos	Remotion
Quality	pytest, Playwright, DeepEval
Tooling	uv, pnpm, Make, Docker Compose, Claude Code

Quick start

cp .env.example .env
uv sync --extra dev --extra report
make dev-data     # small synthetic dataset, about 15 min, no MySQL needed
make serve        # dashboard http://localhost:5173, API http://localhost:8000

The full 1.5M-person build, demo mode and the live MySQL loop are covered in Getting started and Setup.

Documentation

Getting started	Run it locally
Architecture	Components, roles, screens
Care loop	Care plans, patient app, forecasts, 10-step demo
AI agent and MCP	Agent settings and MCP connection
Built with Claude Code	How we built it and the repo's Claude automations
Tests and results	Test suites and verified numbers
Data, privacy and licences	Synthetic data rules and third-party licences
Specification and decisions	The full spec and agreed deviations

Built at the Anthropic AI Week Hackathon

Early Signals was built at the Anthropic AI Week Hackathon.

Team: Benjamin, Remy, Andy and Irere.

Built with Claude. The code was written almost entirely by Claude Opus 5.5 in Claude Code, with the team setting direction and reviewing the work. Along the way we used:

• /goal for long autonomous runs against a clear finish line;
• /loop to repeat tasks such as /validate-risk on an interval;
• Claude Code on the web cloud sessions and Routines (a daily 07:00 Kigali report);
• project skills (/validate-risk, /daily-report), hooks and CLAUDE.md;
• subagents, multi-agent workflows, git worktrees, /fork and Monitor for parallel work;
• MCP, to plug the Early Signals agent into Claude Code;
• /brag, which made the demo reel from real captures of the app.

Details are in Built with Claude Code.

Licence

Code is released under the MIT Licence. The 3D anatomy (Z-Anatomy / BodyParts3D, CC BY-SA) and district boundaries (geoBoundaries, CC BY 4.0) keep their own licences; see Data, privacy and licences.