Elamur Research⌕
☰
home / papers / 2026 / MIDAS_SYNUR at MEDIQA-SYNUR 2026: A Prompting Study for Clinical Observation Extraction from Nurse Dictation Transcriptions

MIDAS_SYNUR at MEDIQA-SYNUR 2026: A Prompting Study for Clinical Observation Extraction from Nurse Dictation Transcriptions

Paper · 2026 · Swetha Krishna Sriram and Akshitaa Sahoo

Summary

Nurse dictations record useful clinical observations in loose, conversational speech. Turning them into structured records is hard because wording varies, mentions are often implicit, and negations are common. For the MEDIQA-SYNUR 2026 shared task, the authors tested how best to prompt a general-purpose LLM to fill a large, fixed schema of clinical fields. No fine-tuning was used. They compared two strategies with GPT-5.2, both with deterministic decoding:

  • Single-prompt holistic extraction: one call per transcript predicts every field together, guided by few-shot examples grouped by value type.
  • Field-wise decomposed prompting: each field is queried separately, with one positive and one NULL demonstration.

Zero-shot versions of both approaches and a GPT-5.1 comparison isolate the effect of the examples and of the model. All prompts share modular blocks: strict evidence-only rules, field notes written from development-set errors, the full schema, and the demonstrations.

Few-shot GPT-5.2 performed best. The field-wise approach scored only slightly higher in F1, but it needed roughly 190 times more calls. This makes the holistic prompt the more practical option and shows that inference-time prompting is a competitive, reproducible baseline.

Key points

  • Dataset: the MEDIQA-SYNUR data follow the SYNUR framework (Corbeil et al., 2025).
    • The schema has 193 fields: 130 SINGLE_SELECT, 12 MULTI_SELECT, plus STRING and NUMERIC types.
    • The splits contain 122 training, 101 development and 199 test transcripts, each averaging about 180–185 words.
  • Sparse annotations: a transcript fills only 12–14 fields on average, which is under 8% of the schema.
  • Setup: the OpenAI API with temperature 0 and a 16,384-token output cap; the holistic prompt uses up to five training examples per value type.
  • Main result: few-shot F1 is 0.83 for field-wise and 0.82 for single-prompt, but the cost is about 190 calls versus 1 per transcript. GPT-5.2 beat GPT-5.1 on every metric.
  • Main weakness: STRING fields scored only 0.025–0.029 F1, due to verbose outputs and strict exact matching. The authors suggest semantic or relaxed matching as a remedy.

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, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksBiochemistry, Genetics and Molecular Biology · 234 worksMolecular Biology · 231 works

Cite

Cite as Swetha Krishna Sriram and Akshitaa Sahoo. 2026. MIDAS_SYNUR at MEDIQA-SYNUR 2026: A Prompting Study for Clinical Observation Extraction from Nurse Dictation Transcriptions. 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 10.63317/4h7qk3596z83

@inproceedings{sriram2026midas,
  title = {{MIDAS\_SYNUR at MEDIQA-SYNUR 2026: A Prompting Study for Clinical Observation Extraction from Nurse Dictation Transcriptions}},
  author = {Swetha Krishna Sriram and Akshitaa Sahoo},
  booktitle = {Proceedings of the 8th Workshop on Clinical Natural Language Processing (Clinical NLP) @ LREC 2026},
  year = {2026},
  doi = {10.63317/4h7qk3596z83},
  url = {https://doi.org/10.63317/4h7qk3596z83}
}

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. Emily Alsentzer, John Murphy, William Boag, Wei-Hung Weng, Di Jindi, Tristan Naumann, and Matthew McDermott. 2019. Publicly available clinical BERT embeddings. In Proceedings of the 2nd Clinical Natural Language Processing Workshop, pages 72–78, Minneapolis, Minnesota, USA. Association for Computati…
  2. Jean-Philippe Corbeil, Asma Ben Abacha, George Michalopoulos, Phillip Swazinna, Miguel Del-Agua, Jerome Tremblay, Akila Jeeson Daniel, Cari Bader, Kevin Cho, Pooja Krishnan, Nathan Bodenstab, Thomas Lin, Wenxuan Teng, Francois Beaulieu, and Paul Vozila. 2025. Empowering healthcare practitioners with… find
  3. A. H. M. Rezaul Karim and Özlem Uzuner. 2025. Masonnlp at mediqa-oe 2025: Assessing large language models for structured medical order extraction. In Proceedings of the 7th Clinical Natural Language Processing Workshop, pages 57–67. Association for Computational Linguistics. find
  4. Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. 2019. Biobert: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics, 36(4):1234–1240.
  5. Parth Mehta. 2025. Pnlp at mediqa-oe 2025: A zero-shot prompting strategy with gemini for medical order extraction. In Proceedings of the 7th Clinical Natural Language Processing Workshop, pages 75–83. Association for Computational Linguistics. find
  6. George Michalopoulos, Jean-Philippe Corbeil, Cari Bader, Nate Bodenstab, and Asma Ben Abacha. 2026. Overview of the mediqa-synur 2026 shared task on observation extraction from nurse dictations. In Proceedings of the 8th Clinical Natural Language Processing Workshop, ClinicalNLP@LREC 2026, Palma, Ma…
  7. Yifan Peng, Shankai Yan, and Zhiyong Lu. 2019. Transfer learning in biomedical natural language processing: An evaluation of BERT and ELMo on ten benchmarking datasets. In Proceedings of the 18th BioNLP Workshop and Shared Task, pages 58–65, Florence, Italy. Association for Computational Linguistics…
  8. Alexander Turchin, Stanislav Masharsky, and Marinka Zitnik. 2023. Comparison of bert implementations for natural language processing of narrative medical documents. Informatics in Medicine Unlocked, 36:101139. find
  9. Yanshan Wang, Liwei Wang, Majid Rastegar-Mojarad, Sungrim Moon, Feichen Shen, Naveed Afzal, Sijia Liu, Yuqun Zeng, Saeed Mehrabi, Sunghwan Sohn, and Hongfang Liu. 2018. Clinical information extraction applications: A literature review. Journal of Biomedical Informatics, 77:34–49. find
  10. Wu Y., Denny J.C., Rosenbloom S.T., Miller R.A., Giuse D.A., and Xu H. 2012. A comparative study of current clinical natural language processing systems on handling abbreviations in discharge summaries. In AMIA Annual Symposium Proceedings, pages 997–1003. PMID: 23304386. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sriram-2026-midas-synur-mediqa-synur-2026/