Elamur Research⌕
☰
home / papers / 2026 / Differentially Private De-identification of Dutch Clinical Notes: A Comparative Evaluation

Differentially Private De-identification of Dutch Clinical Notes: A Comparative Evaluation

Paper · 2026 · Michele Miranda, Xinlan Yan, Nishant Mishra et al.

Summary

Secondary use of hospital notes requires removing personal information, and manual redaction is slow and expensive. This study is the first to compare three families of automatic de-identification on real Dutch clinical text: NER-based redaction, LLM prompting, and differential privacy (DP) methods that formally bound leakage. The authors use the Dutch ADE corpus of ICU notes and build five pipelines. Three apply one method alone: gliner-multi-v2.1 for NER, zero-shot LLM redaction, and token-level DP via Metric-DP or RANTEXT. Two are hybrids that run NER or an LLM before DP. They measure residual PII and also train MedRoberta.nl on each privatized corpus for two downstream tasks: drug/disorder entity classification and adverse-drug-event relation classification. DP alone severely damages utility. Redacting text first, particularly with an LLM, gives a better privacy–utility balance. LLM redaction alone stays close to the original data on downstream tasks but offers no formal guarantee. For clinical NLP in languages other than English, the results show the cost of DP and suggest that pre-redaction reduces it.

Key points

  • Data: The Dutch ADE dataset contains 102 ICU notes (107,110 words). About 0.83% of tokens are PII, covered by 17 placeholder types such as <NAAM> and <ZIEKENHUIS>.
  • DP methods and models:
    • Metric-DP uses BERTje embeddings with an ANN index.
    • RANTEXT uses Dutch GPT-2 (gpt2-small-dutch).
    • The LLMs tested were GPT-4o, DeepSeek-R1 8B/70B, LLaMA-3.1 8B and MedGemma 27B. DeepSeek-70B served as the LLM step in the LLM→DP hybrids.
  • Redaction F1: MedGemma scored highest (≈0.59) and GPT-4o lowest (≈0.21), mostly because GPT-4o produced 4,848 false positives.
  • Downstream results without DP:
    • Entity classification (micro F1): 0.91 on original data, 0.97 after LLM redaction, 0.70 after NER.
    • Relation classification (macro F1): 0.62 on original data, 0.59 after LLM redaction, 0.29 after NER.
  • Downstream results with DP:
    • The best entity classification score was ≈0.51, from NER+RANTEXT at ε=512.
    • The best relation classification score was 0.33, from LLM+RANTEXT at ε=512.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓University of AmsterdamAmsterdam · Netherlands144 works · 173,906 citations🎓Sapienza University of RomeRome · Italy57 works · 9,564 citations🏥Amsterdam University Medical CentersAmsterdam · Netherlands3 works

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksInformation Systems · 1,437 worksPsychology · 1,552 worksSocial Psychology · 306 works

Cite

Cite as Michele Miranda, Xinlan Yan, Nishant Mishra, Rachel Murphy, Ameen Abu Hanna, Sébastien Bratières and Iacer Calixto. 2026. Differentially Private De-identification of Dutch Clinical Notes: A Comparative Evaluation. Proceedings of the 8th Workshop on Clinical Natural Language Processing (Clinical NLP) @ LREC 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/2xrg7fmndfhy

@inproceedings{miranda2026differentially,
  title = {{Differentially Private De-identification of Dutch Clinical Notes: A Comparative Evaluation}},
  author = {Michele Miranda, Xinlan Yan, Nishant Mishra, Rachel Murphy, Ameen Abu Hanna, Sébastien Bratières and Iacer Calixto},
  booktitle = {Proceedings of the 8th Workshop on Clinical Natural Language Processing (Clinical NLP) @ LREC 2026},
  year = {2026},
  doi = {10.63317/2xrg7fmndfhy},
  url = {https://doi.org/10.63317/2xrg7fmndfhy}
}

References

10 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774. ↗
  2. Loick Bourdois, Marta Avalos, Gabrielle Chenais, Frantz Thiessard, Philippe Revel, Cedric Gil-Jardine, and Emmanuel Lagarde. 2021. De-identification of emergency medical records in french: Survey and comparison of state-ofthe-art automated systems. The International FLAIRS Conference Proceedings, 34. find
  3. Konstantinos Chatzikokolakis, Miguel E. Andrés, Nicolás Emilio Bordenabe, and Catuscia Palamidessi. 2013. Broadening the scope of differential privacy using metrics. In International Symposium on Privacy Enhancing Technologies. find
  4. Andrew Kweku Conduah, Sebastian Ofoe, and Dorothy Siaw-Marfo. 2025. Data privacy in healthcare: Global challenges and solutions. Digital Health, 11:20552076251343959. ↗
  5. Wietse de Vries and Malvina Nissim. 2020. As good as new. how to successfully recycle english gpt-2 to make models for other languages.
  6. Wietse De Vries, Andreas van Cranenburgh, Arianna Bisazza, Tommaso Caselli, Gertjan van Noord, and Malvina Nissim. 2019. Bertje: A dutch bert model. arXiv preprint arXiv:1912.09582. ↗
  7. Franck Dernoncourt, Ji Young Lee, Ozlem Uzuner, and Peter Szolovits. 2017. De-identification of patient notes with recurrent neural networks. Journal of the American Medical Informatics Association (JAMIA), 24(3):596–606. ↗
  8. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv e-prints, pages arXiv– 2407.
  9. Cynthia Dwork. 2006. Differential privacy. In Proceedings of the 33rd International Conference on Automata, Languages and Programming - Volume Part II, ICALP’06, page 1–12, Berlin, Heidelberg. Springer-Verlag. find
  10. Cynthia Dwork and Aaron Roth. 2014. The algorithmic foundations of differential privacy. Found. Trends Theor. Comput. Sci., 9(3–4):211–407. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/miranda-2026-differentially-private-de-identification-dutch/