Skip to content

Choose a workflow

Start with the outcome you need. Each workflow explains which tools to use, what files you will create, and where those files go next. If you have not run MedDeID before, begin with Install and run.

Planning a multi-step study?

Run meddeid start when you want MedDeID to guide you through a longer workflow and keep its decisions and progress together. It asks what you want to accomplish, helps you choose the appropriate workflow, and creates a workspace for continuing it later.

meddeid start

The assistant does not choose your study design for you. See the workflow CLI reference when you are ready to create and use a workspace.

De-identify text

Process one note or a batch with the command line, or add local inference to a Python application.

Follow the inference steps →

Run MedDeID as an internal service

Choose CPU, NVIDIA GPU, or native Apple silicon for the HTTP API, then place the service inside your institution's security and operational controls.

Prepare a deployment →

Prepare and annotate data

Import exported notes, review the identifiers in each document, and reconcile multiple reviewers only when your project requires it.

Prepare and annotate data →

Train a model

Train from reviewed development data and measure the resulting model on an independent test set.

Train and evaluate →

Adapt a model to your setting

Test whether additional reviewed data from a hospital, speciality, or document type improves an existing model fairly.

Plan domain adaptation →

Evaluate or compare systems

Score predictions against reviewed test data and compare systems under the same evaluation conditions.

Evaluate predictions →

Add another language

Understand the language package, model bundle, validation, documentation, and collaboration work involved.

See how to add a language →

How the workflows fit together

The workflows can be used separately or connected. A one-time de-identification run does not require an annotation or training project. A model-development project usually moves through preparation, review, training, and evaluation before the exported model is used for inference or deployment.

flowchart LR
    S["Source notes"] --> R["Prepare and review"]
    R --> T["Train or adapt"]
    R --> E["Evaluate or compare"]
    T --> I["Run or integrate"]
    T --> E
    M["Existing model"] --> I
    I --> D["Deploy internally"]

Whichever route you choose:

  • keep source notes and the private ID mapping inside the approved data boundary;
  • use separate writable files for reviewer work instead of changing the imported source artifact;
  • preserve the manifests and settings that connect data, models, predictions, and results; and
  • keep test answers out of training decisions when reporting model quality.

For component responsibilities and package boundaries, see the suite architecture and component reference.