EHR Risk Workbench hub · UI route /train
Train & audit
Configure a task, start training or multi-model compare, and run the leakage audit job against the active artifact.
Controls in detail
| Control | What it does |
|---|---|
| Task preset | Loads YAML from tasks/ (horizon, windows, labels) |
| Data path / format | Defaults to data/demo/ehr_data.csv for the teaching longitudinal preset; accepts longitudinal events or flat tabular CSV |
| Model | logreg · random_forest · xgboost · lightgbm (if installed) |
| Isotonic calibrate | Fits calibration; reports Brier / ECE |
| Patient-level split | Same patient never in train and hold-out |
| Temporal patient split | Tests future performance when timestamps exist |
| Multi-window features | w7d_ / w30d_ / w180d_ aggregates |
| Index strategy | last_event · before_last · column |
Jobs
- Start train — background job →
model.pkl+ metrics - Compare models — ranks by hold-out ROC-AUC; optional promote-best
- Light HPO — optional small research grid; Results shows its best trial and individual trials
- Leakage audit — writes
reports/leakage_audit.json
Live job panel shows status, message, and log tail.