Preprint · 2026 · Open access

De-identification of Brazilian Portuguese Clinical Records Using a Hybrid Pipeline with Local Language Models

Caio Galvão Aragão, Marcelo Costa Oliveira, Thales Miranda de Almeira Vieira

Research Square · Published 3 August 2026 · DOI 10.21203/rs.3.rs-10425027/v1

Open PDFView at the publisherCite this paper

In brief

“De-identification of Brazilian Portuguese Clinical Records Using a Hybrid Pipeline with Local Language Models” is a 2026 preprint by Caio Galvão Aragão and 2 co-authors, published in Research Square. OpenAlex counts 0 citations as of 19 November 2026. It is open access under a CC BY licence: the full text is free to read at Research Square.

Citations
0
References
35
Authors
3

Abstract

Secondary use of electronic health record narratives supports clinical research but requires removing personally identifiable information (PII). Processing identifiable records with external large language model (LLM) services raises critical privacy, governance, and data sovereignty concerns, necessitating locally deployable solutions. We propose and evaluate an offline, hybrid LLM pipeline for de-identifying Brazilian Portuguese clinical narratives on modest hardware. The architecture integrates regular expressions for canonical identifiers, recurrent n -gram detection to withhold boilerplate text from neural inference, and prompted open-weight LLMs for contextual PII extraction. To ensure system stability, a four-detector cascade manages repetitive-generation loops and resumes inference from valid prefixes. Evaluating 28 models and model committees on the synthetic AnonyMed-BR dataset, the best individual model (gemma4:26b) achieved an F 2 score of 0.9595 on target PII categories, outperforming all evaluated committee approaches. Evaluated end-to-end under a broader comparison set, the full pipeline achieved F 2 = 0.9341 , recall of 0.9537 , and precision of 0.8634 . The detector cascade successfully interrupted 99% of observed generation loops, while a statistical scan on an unannotated corpus of 12,674 real hospital notes classified 11.58% of bytes as boilerplate, establishing a practical upper bound for input withholding. These findings demonstrate the technical feasibility and high recall of institutionally controlled, local de-identification without requiring task-specific model fine-tuning.

Cite this paper

Formatted from the catalogue record. Check names, capitals and page numbers against the paper before you submit.

APA 7

Aragão, C. G., Oliveira, M. C., & Vieira, T. M. D. A. (2026). De-identification of Brazilian Portuguese Clinical Records Using a Hybrid Pipeline with Local Language Models. Research Square. https://doi.org/10.21203/rs.3.rs-10425027/v1

MLA 9

Aragão, Caio Galvão, et al. “De-identification of Brazilian Portuguese Clinical Records Using a Hybrid Pipeline with Local Language Models.” Research Square, 2026. https://doi.org/10.21203/rs.3.rs-10425027/v1.

Chicago

Aragão, Caio Galvão, Marcelo Costa Oliveira, and Thales Miranda de Almeira Vieira. 2026. “De-identification of Brazilian Portuguese Clinical Records Using a Hybrid Pipeline with Local Language Models.” Research Square. https://doi.org/10.21203/rs.3.rs-10425027/v1.

Harvard

Aragão, C. G., Oliveira, M. C. and Vieira, T. M. D. A. (2026) ‘De-identification of Brazilian Portuguese Clinical Records Using a Hybrid Pipeline with Local Language Models’, Research Square. doi:10.21203/rs.3.rs-10425027/v1.

IEEE

C. G. Aragão, M. C. Oliveira, and T. M. D. A. Vieira, “De-identification of Brazilian Portuguese Clinical Records Using a Hybrid Pipeline with Local Language Models,” Research Square, 2026, doi: 10.21203/rs.3.rs-10425027/v1.

BibTeX

@misc{aragao2026identification,
  title     = {{De-identification of Brazilian Portuguese Clinical Records Using a Hybrid Pipeline with Local Language Models}},
  author    = {Aragão, Caio Galvão and Oliveira, Marcelo Costa and Vieira, Thales Miranda de Almeira},
  howpublished = {Research Square},
  year      = {2026},
  doi       = {10.21203/rs.3.rs-10425027/v1},
  url       = {https://doi.org/10.21203/rs.3.rs-10425027/v1}
}

Download for a reference manager: BibTeX (.bib) · RIS for Zotero, Mendeley and EndNote (.ris)

Topic and keywords

Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data

  • de-identification
  • Brazilian Portuguese
  • clinical text
  • electronic health records
  • personally identifiable information
  • local large language models
  • prompt engineering
  • open-weight LLMs

Work with this paper

Most-cited papers in Computer Science

All highly cited Computer Science papers

Questions about this paper

How do I cite this paper in APA style?

Aragão, C. G., Oliveira, M. C., & Vieira, T. M. D. A. (2026). De-identification of Brazilian Portuguese Clinical Records Using a Hybrid Pipeline with Local Language Models. Research Square. https://doi.org/10.21203/rs.3.rs-10425027/v1

How many times has this paper been cited?

0 times, according to OpenAlex on 19 November 2026. Google Scholar and Scopus count different sets of sources, so their numbers differ.

Is this paper free to read?

Yes. It is open access at Research Square: sign in and open the PDF from this page.

Who wrote this paper?

Caio Galvão Aragão, Marcelo Costa Oliveira, Thales Miranda de Almeira Vieira.

What is the DOI of this paper?

10.21203/rs.3.rs-10425027/v1, permanently at https://doi.org/10.21203/rs.3.rs-10425027/v1. Sign in and open it from this page.

Bibliographic data from OpenAlex (CC0), last checked 19 November 2026. ResearchGuru does not host this paper; once you are signed in, the links open it at the publisher or the repository that does.