Elamur Research⌕
☰
home / papers / 2026 / NormaTex-MapSNOMED: Bridging the Gap Between Brazilian Portuguese Clinical Narratives and SNOMED CT

NormaTex-MapSNOMED: Bridging the Gap Between Brazilian Portuguese Clinical Narratives and SNOMED CT

Summary

The paper addresses the normalization of clinical terms from Brazilian Portuguese free-text medical records to the SNOMED CT terminology. Outside English, there are almost no ready-made solutions for this task. The authors describe NormaTex-MapSNOMED, a module of the broader NormaTex framework. It receives already extracted terms and assigns them to two categories aligned with SNOMED CT: Disease or Syndrome and Sign or Symptom. To do so, it uses an LLM with a structured prompt. The prompt combines step-by-step instructions in the spirit of chain-of-thought, few-shot examples, and HTML-like markup. The model is instructed to skip medications, dosages, and test results. The authors compared four models (LLaMA 3 8B, Gemini 2.5 Pro, GPT-3.5-Turbo, GPT-4.1) on ten cardiology texts from TempClinBr. Quality was assessed in three ways: by Levenshtein distance, by BERT-based similarity, and by verification by the LLM itself. All results were additionally checked manually. The best rates of correct mappings are approximately 62–64%. The model maps diseases more accurately than symptoms. The work matters as one of the few steps toward interoperability of clinical data in Portuguese. However, assigning the actual SNOMED CT concept codes has not yet been implemented.

Key points

  • Data: SemClinBr (1000 narratives, 89 semantic categories) and TempClinBr. TempClinBr is a subset of 126 cardiology texts that serves as the gold standard and has a broader category, Problema.
  • Test set: 10 narratives ranging from 33 to 304 tokens in length (about 140 on average) and 111 reference spans. The metric is span-level precision. Recall was not computed because the models mapped different numbers of spans.
  • Results: by BERT-based evaluation, GPT-4.1 achieves 63.7% correct mappings. By LLM-based evaluation, GPT-3.5-Turbo achieves 62.0%, and LLaMA-3-8B 57.4%. For Gemini 2.5 Pro, 59.6% are incorrect by Levenshtein. LLM-based evaluation was not performed for Gemini due to API limitations.
  • The LLM evaluator was the same model used for mapping, but in a separate run without shared context. The authors state that LLM-based evaluation is consistently better. Yet for GPT-4.1 it yielded 50.5% correct, whereas BERT yielded 63.7%.
  • Limitations: negation is not handled, and there are no comparisons with SapBERT or dictionary-based methods. The SemClinBr annotations contain errors, and the sample is very small.

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

Fields: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{araujo2026normatex,
  title = {{NormaTex-MapSNOMED: Bridging the Gap Between Brazilian Portuguese Clinical Narratives and SNOMED CT}},
  author = {Araujo, Isabela and Moro, Claudia and Martinez, Layslla},
  booktitle = {Proceedings of the 17th International Conference on Computational Processing of Portuguese (PROPOR 2026) - Vol. 1},
  year = {2026},
  url = {https://aclanthology.org/2026.propor-1.115/}
}

References

25 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Afia Z. Khan. 2024. Extraction of normalized symptom mentions from clinical narratives using large language models. Journal of the American Medical Informatics Association. find
  2. Akhila Abdulnazar, Markus Kreuzthaler, Roland Roller, and Stefan Schulz. 2023. Sapbert-based medical concept normalization using snomed ct. In Studies in Health Technology and Informatics, volume 302, pages 825–826. find
  3. Kamyar Arzideh, Henning Schäfer, and 1 others. 2025. From bert to generative ai: Comparing encoder-only vs. large language models in a cohort of lung cancer patients for named entity recognition in unstructured medical reports. Computers in Biology and Medicine, page 110665. find
  4. Yunbing Bai, Wanting Cui, and Joseph Finkelstein. 2025. Performance of open-source large language models to extract symptoms from clinical notes. Studies in Health Technology and Informatics, 329:663– 667. find
  5. Navya Bhagat, Olivia Mackey, and Adam Wilcox. 2024. Large language models for efficient medical information extraction. AMIA Summits on Translational Science Proceedings, 2024:509. find
  6. R. Carneiro. 2017. Sintomas e sinais: uma abordagem científica do exame clínico. Revista Portuguesa de Medicina Geral e Familiar. find
  7. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,… find
  8. Abhimanyu Dubey and 1 others. 2024. The Llama 3 herd of models. arXiv preprint arXiv:2407.21783. find
  9. S. A. Hasan and 1 others. 2015. Ontology-driven semantic search for brazilian portuguese clinical notes. In MEDINFO 2015: eHealth-enabled Health – Proceedings of the 15th World Congress on Health and Biomedical Informatics, volume 216 of Studies in Health Technology and Informatics, São Paulo, Brazi… find
  10. Health Level Seven International. 2024a. Fhir: Fast healthcare interoperability resources. Accessed: Sep 30, 2024. find
  11. Health Level Seven International. 2024b. Hl7 international. Accessed: Sep 30, 2024. find
  12. Yan Hu and 1 others. 2024. Improving large language models for clinical named entity recognition via prompt engineering. Journal of the American Medical Informatics Association, 31(9):1812–1820. find
  13. Florian Leiser, Richard Guse, and Ali Sunyaev. 2025. Large language model architectures in health care: Scoping review of research perspectives. Journal of Medical Internet Research. find
  14. Vladimir I. Levenshtein. 1966. Binary codes capable of correcting deletions, insertions, and reversals. Soviet Physics Doklady, 10(8):707–710. find
  15. Pengfei Liu and 1 others. 2021. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1–35. find our card
  16. Maria Clara Saad Menezes, Alexander F. Hoffmann, Amelia L. M. Tan, and 1 others. 2025. The potential of generative pre-trained transformer 4 (gpt-4) to analyse medical notes in three different languages: a retrospective model-evaluation study. The Lancet Digital Health, 7(1):e35–e43. find
  17. Jingwen Nan and Li-Qun Xu. 2023. Designing interoperable health care services based on fast healthcare interoperability resources: literature review. JMIR Medical Informatics, 11(1):e44842. find
  18. NHS Digital. 2024. Snomed ct browser. Accessed: Jul 21, 2024. find
  19. João Oliveira and 1 others. 2022. Semclinbr - a multi-institutional and multi-specialty semantically annotated corpus for portuguese clinical nlp tasks. Journal of Biomedical Semantics. find
  20. OpenAI. 2023. GPT-4 technical report. arXiv preprint arXiv:2303.08774. find our card
  21. João Figueira Silva, Rui Antunes, João Rafael Almeida, and Sérgio Matos. 2020. Clinical concept normalization on medical records using word embeddings and heuristics. In Studies in Health Technology and Informatics, volume 270, pages 93–97. find
  22. Sonish Sivarajkumar and 1 others. 2024. An empirical evaluation of prompting strategies for large language models in zero-shot clinical natural language processing. JMIR Medical Informatics. find
  23. Dongfang Xu and Timothy Miller. 2022. A simple neural vector space model for medical concept normalization using concept embeddings. Journal of Biomedical Informatics, 130:104080. find
  24. A. Zaghir and 1 others. 2024. Prompt engineering paradigms for medical applications: Scoping review. JMIR. find
  25. Kai Zhang, Tongtong Huang, Bradley A. Malin, and 1 others. 2025. Introducing mcodegpt as a zero-shot information extraction from clinical free text data tool for cancer research. Communications Medicine. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/papers/araujo-2026-normatex-mapsnomed-bridging-gap-between/