Venturus/ptt5-v2-AnonyMED-BR

ptt5-v2 for Medical Data Anonymization

Model description

ptt5-v2-AnonyMED-BR is a Brazilian Portuguese adaptation of the T5 architecture fine-tuned and evaluated on the AnonyMED-BR dataset for the task of anonymizing sensitive medical records. Unlike extractive models, ptt5-v2 can generate rewritten medical records, embedding tags around sensitive entities.

• Architecture: ptt5-v2-large
• Pretraining corpus: Large Brazilian Portuguese web text
• Evaluation domain: Medical anonymization with AnonyMED-BR
• Language: Brazilian Portuguese

---

Intended uses & limitations

Intended uses

• Text-to-text anonymization of Brazilian Portuguese medical records.
• Identifies and tags sensitive entities such as names, dates, IDs, and hospitals.
• Suitable for local deployment, preserving privacy compared to API-based models.

---

How to use

Example usage with Hugging Face Transformers:

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_name = "Venturus/ptt5-v2-AnonyMED-BR"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

text = "O paciente João da Silva foi internado no Hospital das Clínicas em 12/05/2023."
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_length=128)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Example output:

O paciente <PATIENT>João da Silva</PATIENT/> foi internado no <HOSPITAL>Hospital das Clínicas</HOSPITAL/> em <DATE>12/05/2023</DATE/>.

Training procedure

• Base architecture: ptt5-v2 Large.

Fine-tuning for text-to-text anonymization using AnonyMED-BR.

• Hyperparameters:

  • Learning rate: 5e-5

  • Batch size: 4

  • Epochs: 2

  • Precision: FP32

• Hardware: NVIDIA Tesla T4 (16 GB).

Evaluation results

The model was evaluated on the AnonyMED-BR test set.

Overall results

F1-score: 0.8748

Precision: 0.8918

Recall: 0.8645

Entities covered

• Personal data:

  • <PATIENT>
  • <DOCTOR>
  • <AGE>
  • <PROFESSION>

• Identifiers:

  • <IDNUM>
  • <MEDICAL_RECORD>
  • <HEALTH_PLAN>

• Locations:

  • <CITY>
  • <STATE>
  • <COUNTRY>
  • <STREET>
  • <HOSPITAL>
  • <LOCATION_OTHER>
  • <ZIP>

• Other sensitive data:

  • <DATE>
  • <EMAIL>
  • <PHONE>
  • <ORGANIZATION>
  • <OTHER>

Citation

If you use this model, please cite:

@article{schiezzaro2025guardians,
  title={Guardians of the Data: NER and LLMs for Effective Medical Record Anonymization in Brazilian Portuguese},
  author={Schiezzaro, Mauricio and Rosa, Guilherme and Pedrini, Helio and Campos, Bruno Augusto Goulart},
  journal={Frontiers in Public Health},
  year={2025},
  publisher={Frontiers},
  url={https://github.com/venturusbr/AnonyMED-BR}
}