📄 A Dataset of Brazilian Portuguese Clinical Notes for Anaphylaxis Detection

Summary

Automatically spotting anaphylaxis in free-text medical records could support large-scale retrospective research. Supervised methods, however, need labeled data, and Portuguese clinical corpora with document-level condition labels are very rare. SemClinBr, for instance, is annotated only for entities and relations.

The authors build and document a public Brazilian Portuguese resource of 969 clinical narratives. Allergists tagged each one as anaphylaxis present or absent, applying the NIAID/FAAN and WAO 2020 diagnostic criteria. The texts come from three origins:

• fictional notes written by hand by specialists, with no language models involved;
• published case reports recast in a clinical-note style;
• de-identified notes drawn from SemClinBr.

Positive cases make up roughly 5% of the data. This mirrors real-world prevalence and turns the task into a difficult, imbalanced rare-event problem.

The paper is deliberately dataset-centric and reports no new classifiers. Instead, it details provenance, annotation procedure, composition, limitations and ethics. The result is a reusable benchmark for document classification, information extraction and LLM evaluation in Portuguese clinical NLP, a field that otherwise depends heavily on private or restricted data.

Key points

• Size and labels: 969 Brazilian Portuguese clinical narratives with a binary label only. Severity, triggers, time course and outcomes are not annotated.
• Class balance: about 5% positive cases, consistent with prevalence estimates cited from Cardona et al. (2020).
• Sources: clinician-written synthetic notes (manual, not LLM-generated), literature case reports adapted into note form, and a de-identified subset of SemClinBr (a corpus of 1,000 notes with 65,117 entities).
• Annotation and privacy: allergists of the University of São Paulo / Hospital Sírio-Libanês team labeled the texts using NIAID/FAAN and WAO criteria. The SemClinBr portion complies with Brazil's General Data Protection Law.
• Availability: released on Hugging Face as matos1012/brazilian-portuguese-anaphylaxis. The authors flag the modest size, possible stylistic bias in synthetic notes, and the dataset's research-only scope.