Choose your path¶
MedDeID is a suite, but most users need only a small part of it. Start from the outcome you want.
All Python components install from PyPI. For inference, install meddeid or use
the published GHCR image. The browser applications also have versioned public
GHCR images, so ordinary users do not need Node.js or a source checkout.
| Your goal | Begin with | Add only when needed |
|---|---|---|
| De-identify one note or a batch of notes | meddeid |
meddeid[server] for an internal web service |
| Import hospital data for review | meddeid-data |
meddeid for model pre-annotations |
| Review and correct PII spans with one annotator | meddeid-annotate |
meddeid for optional model pre-annotations |
| Reconcile multiple reviewers | meddeid-curate |
Use only when the study protocol requires it |
| Create a detailed evaluation benchmark | meddeid-subannotate |
Start from completed reviewed annotations |
| Train or adapt a model | meddeid-training[train] |
meddeid-data to organize training and test data |
| Score predictions or test stability | meddeid-eval |
meddeid-eval[plots] for figures |
| Add support for another language | meddeid-core |
Implement a separate meddeid-language-* package and model bundle |
Common paths¶
Continue to local inference.
source notes → meddeid-data → meddeid-annotate
→ optional meddeid-curate → reviewed primary annotations
Continue to prepare and annotate data.
Continue to domain adaptation.
What you do not need¶
- You do not need the grouped suite workspace. Python users install released packages directly from PyPI; browser-application users pull only the GHCR image they need.
- You do not need training, evaluation, or annotation packages for ordinary inference.
- You do not need curation for a completed single-reviewer dataset.
- You do not need detailed character-level benchmark labels for ordinary model training.