EHR Risk Blog · EHR risk prediction
How to build an EHR risk prediction model
A short tutorial for clinical machine learning education: train a binary risk model on teaching EHR-style data with the open-source EHR Risk Framework—then review leakage, calibration, and SHAP.
1. Start the workbench
docker compose up --build
# UI http://localhost:8080 · API http://localhost:80002. Use the teaching dataset
Prefer data/demo/ehr_data.csv (bundled synthetic fixture). Hide demos in Datasets if you only want uploads.
See the data guide and Datasets UI.
3. Train with a leakage-aware task
Open Train, select a task (e.g. diabetes / custom), start training, then open Results for metrics, optional light HPO, fairness, and figures.
4. Audit and explain
- Run the leakage audit job against the trained artifact.
- Review calibration plots and SHAP when available.
- Download the results ZIP for a reproducible methods pack.
Deep dive: Quickstart · Fine-tuning · Leakage guide