EHR Risk Data guide
Bundled datasets
| ID | Path | Use |
|---|---|---|
| ehr_data | data/demo/ehr_data.csv | Tiny longitudinal teaching / smoke-test fixture (no index_time; use custom / last_event) |
| paper_synthetic | data/raw/paper_synthetic_cohort.csv | Larger synthetic cohort with index_time for diabetes / horizon / readmission tasks |
| sample_ehr | data/demo/sample_ehr.csv | Legacy tabular teaching sample |
Workspace catalog
Datasets lists bundled demos by default. Use Show bundled demo datasets to hide them, select a row to inspect it, or check multiple rows and delete them together. Repeating a delete for an already removed allowed file is reported as already_absent.
Longitudinal columns (typical)
patient_id, timestamp, clinical fields (glucose, blood_pressure, cholesterol, lab_value, vital_signs, icd_code), age, label. Optional but required for horizon-safe tasks: index_time. Also optional: sex, age_band.
Multi-window model features look like w7d_glucose, not raw column names — see schema on Predict / /v1/model/schema.
BYO CSV tips
- Upload via Datasets or map columns with the map-preview API.
- Run dataset health; fix blockers before train.
- Choose index strategy and horizon to match your question.
- Keep real extracts local / gitignored; follow IRB and PhysioNet rules for MIMIC.
Longer notes: repository DATA_GUIDE.md and docs/data_sources_and_schema.md.
Temporal integrity (must-know)
Features must use only information available at or before the prediction index time. Labels belong in the post-index horizon. The leakage audit job checks common failure modes — run it before you trust a metric.