Skip to content

Workflow overview

MedDeID consists of focused tools that work together. You can use the complete workflow or choose only the steps your project needs.

Clinical inference

note.txt or documents.jsonl
  → meddeid
  → de-identified text or predictions.jsonl

Use this path when a trained model already meets the needs of the target setting. Run local inference.

Dataset preparation and review

hospital TXT / CSV / TSV / Parquet
  → meddeid-data project
  → optional model pre-annotations
  → meddeid-annotate
  → completed reviewed annotations

One completed review is sufficient for training. Add another reviewer and meddeid-curate only when your study requires independent review. Prepare and annotate data.

Detailed evaluation

completed reviewed annotations
  → optional meddeid-curate
  → meddeid-subannotate
  → detailed evaluation benchmark
  → meddeid-eval

This optional step marks which parts of each identifier are essential to be detected. It supports detailed evaluation and is not required for training.

Model adaptation

reviewed development data + separate test data
  → meddeid-data prepare-training
  → meddeid-training
  → exported model bundle
  → meddeid batch
  → meddeid-eval

For published research, MedDeID can first determine how long to train and then train a fresh model on all development data. A simpler one-time training run is available for ordinary experiments. Train and evaluate.

What stays consistent

Across the workflow:

  • one tool's output can be used directly by the next;
  • source files are not silently overwritten;
  • MedDeID records which data, model, and settings produced a result;
  • the independent test set stays separate while training decisions are made.

Implementation details are available in the data contract and artifact lineage pages.