meddeid-english-synth

meddeid-english-synth is a clinical-text de-identification model for explicit en-GB and en-US runtime profiles. It predicts character-offset PII spans and is packaged with its tokenizer, configuration, labels, and MedDeID bundle contract. No real patient notes were used for training.

The MedDeID documentation is the central reference for installation, local deployment, annotation workflows, validation, and security guidance.

Intended use

The model is a research and local-deployment starting point for detecting identifying information in English clinical text. It does not make text anonymous by itself, cannot guarantee that every identifier is removed, and requires validation for the target institution and use case. Sensitive text should be processed inside the user's governance boundary, not in a public demonstration service.

Use with MedDeID

pip install meddeid

Select the relevant locale explicitly:

from meddeid import Deidentifier

engine = Deidentifier.from_pretrained(
    "stighellemans/meddeid-english-synth",
    language_profile="en-GB",  # or en-US
    device="cpu",
)
result = engine("Patient Alex Example attended on 27 August 2026.")
print(result.spans)
print(result.deid_text)

For an air-gapped installation, pre-stage the immutable repository revision:

hf download stighellemans/meddeid-english-synth \
  --revision YOUR_PINNED_COMMIT \
  --local-dir ./meddeid-english-synth

Architecture and training

The model uses FacebookAI/roberta-base pinned at revision e2da8e2f811d1448a5b465c236feacd80ffbac7b, with separate BIO and 14-way entity classification heads. It processes overlapping 512-token windows with a 64-token overlap.

Epoch selection used a held-out partition of the synthetic development corpus. The deliverable model was reinitialised from the pinned base encoder and refit for four epochs on all 6,700 development documents. The fixed 300-document benchmark was evaluated once after selection and refitting.

Benchmark results

On the human-validated English synthetic benchmark, the model-native typed entity evaluation produced precision 99.19%, recall 99.48%, and F1 99.33% (1,717 gold spans). The strict exact-boundary-and-label slice produced precision 99.01%, recall 99.30%, and F1 99.16%. Exact-slice F1 was 99.07% for en-GB and 99.24% for en-US.

Internal synthetic benchmark comparison

The comparison on the released 300-note English synthetic benchmark is shown below. Recall is calculated over the identifying annotation characters. The metadata-enabled condition applies the same optional patient/caregiver-name recovery step to every system. Values in parentheses are document-clustered 95% bootstrap confidence intervals. Higher recall and lower non-PII redaction are better.

These are synthetic-benchmark point estimates. They are not estimates of performance on real clinical notes or another institution.

External portability benchmarks (manuscript Table 3)

The manuscript also evaluates the same systems on the independent Technetium-I and ASQ-PHI synthetic datasets. Values are percentages. Recall is label-agnostic annotation-character recall; higher recall and lower non-PII redaction are better.

These external results assess portability to other synthetic corpora; they are not estimates of performance on real clinical notes.

Predicted labels

The model predicts Address_Location:Caregiver, Address_Location:Other, Address_Location:Patient, Age_Birthdate, Contactdetails, Date, ID:Caregiver, ID:Patient, Name:Caregiver, Name:Other, Name:Patient, Organization:Healthcare, Organization:Other, and Profession.

Limitations

• Training and public evaluation data are synthetic.
• Coverage of specialties, institutions, dialects, styles, and identifier formats is incomplete.
• Locale selection does not replace institution-specific validation.
• The model can miss identifiers and remove clinically meaningful text.
• Threshold, decoder, software, and post-processing changes can change results.

Reproducibility and files

• model.safetensors: weights;
• bundle.json: labels, windowing, base revision, and locale profiles;
• config.json and tokenizer files: self-contained encoder assets;
• train_metrics.json and eval.json: selection/refit and benchmark metrics;
• MODEL_PROVENANCE.json and CHECKSUMS.sha256: provenance and integrity.

Machine-specific absolute path prefixes in the run metadata are replaced with the portable ${MEDDEID_SUITE} marker. Numeric results and configuration values are unchanged. Reproducible use should pin the immutable Hub commit.

Project, citation, and licence

Developed by Stig Hellemans, Tom Stroobants, Elyne Scheurwegs, Pieter Meysman, Philippe Jorens, and Kris Laukens at the University of Antwerp and Antwerp University Hospital (UZA), with support from Research Foundation Flanders (FWO), grant 1SA3226N.

We thank the participating general practice and Co-Medic for enabling the primary-care validation reported in the accompanying MedDeID study.

Model

@software{hellemans_2026_meddeid_english_synth,
  author    = {Hellemans, Stig and Stroobants, Tom and Scheurwegs, Elyne and Meysman, Pieter and Jorens, Philippe and Laukens, Kris},
  title     = {meddeid-english-synth},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/stighellemans/meddeid-english-synth}
}

Benchmark dataset

Hellemans, S., Stroobants, T., Scheurwegs, E., Meysman, P., Jorens, P., and Laukens, K. (2026). MedDeID English synthetic clinical corpus, benchmark and annotation guideline (v3) [Dataset]. Zenodo. https://doi.org/10.5281/zenodo.22689857

MedDeID paper

Hellemans, S., Stroobants, T., Scheurwegs, E., Meysman, P., Jorens, P. G., and Laukens, K. (2026). MedDeID enables locally governed clinical-text de-identification from real or synthetic training data. arXiv:2609.10049.

The model bundle is licensed under AGPL-3.0-only with the MedDeID Private Fine-Tuning Exception, version 1.0. The Exception expressly applies to these model weights and bundled software. It permits private training data and resulting private fine-tuned weights to remain confidential; MedDeID code modifications remain subject to AGPL-3.0-only. The upstream RoBERTa base model remains subject to its own terms and attribution requirements.