Elamur Research⌕
☰
home / papers / 2026 / NM at CRF Filling 2026: A Two-Stage LLM Pipeline for Clinical CRF Population

NM at CRF Filling 2026: A Two-Stage LLM Pipeline for Clinical CRF Population

Paper · 2026 · Niccolò Morabito

Summary

This system paper reports a submission to the CRF Filling Shared Task 2026 at BioNLP 2026. The task is to fill a fixed Case Report Form (CRF) of 134 items from free-text clinical notes about patients with dyspnea, in English and Italian. The data is highly sparse: about 95% of fields have the value "unknown". The author uses a pipeline with two large language models. First, Gemini 3 Flash receives three gold note–CRF demonstrations in its prompt and returns all fillable fields as a single JSON object. This extractor tends to over-predict. A second, more conservative model therefore checks each proposed field against the note and drops any that lack explicit support. It may not add fields or change values. On the development data, this filtering removed some false positives at little cost to true positives and raised macro F1 slightly. On the official test set, the system reached a macro F1 of 0.56 in both languages. The paper argues that splitting recall-oriented extraction from precision-oriented checking is a useful design for structured extraction from clinical text.

Key points

  • Data used: of the training resources, only the 10 gold-standard Dyspnea CRF pairs (Kaczmarek et al., 2026) were used, as a pool for the 3-shot prompt. The development set has 80 pairs and the test set has 200, with parallel English and Italian versions.
  • Model choice: the verifier differs by language, with LLaMA4 Maverick 17B for English and GPT-OSS 120B for Italian. Each was chosen because it was the most conservative predictor when tried as an extractor. Gemma3 27B and Qwen3 were also tested in preliminary runs.
  • English development results: F1 went from 0.584 (Gemini alone) to 0.612 with verification. False positives fell from 420 to 373, while true positives dropped only from 278 to 274.
  • Italian development results: the gain was smaller, from 0.599 to 0.604. False positives went from 385 to 372 and true positives from 289 to 276.
  • Test results: the final test macro F1 was 0.56 for both English and Italian, lower than on the development set. Only the final score is reported for the test set, without a TP/FP/FN breakdown.

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 Niccolò Morabito. 2026. NM at CRF Filling 2026: A Two-Stage LLM Pipeline for Clinical CRF Population. Proceedings of the BioNLP 2026 (Shared Tasks).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.bionlp-2.1

@inproceedings{morabito2026nm,
  title = {{NM at CRF Filling 2026: A Two-Stage LLM Pipeline for Clinical CRF Population}},
  author = {Niccolò Morabito},
  booktitle = {Proceedings of the BioNLP 2026 (Shared Tasks)},
  year = {2026},
  doi = {10.18653/v1/2026.bionlp-2.1},
  url = {https://doi.org/10.18653/v1/2026.bionlp-2.1}
}

References

6 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Pietro Ferrazzi, Mattia Franzin, Alberto Lavelli, and Bernardo Magnini. 2026a. Small LLMs for Medical NLP: a Systematic Analysis of Few-Shot, Constraint Decoding, Fine-Tuning and Continual Pre-Training in Italian. Preprint, arXiv:2602.17475. ArXiv preprint. find
  2. Pietro Ferrazzi, Soumitra Ghosh, Alberto Lavelli, and Bernardo Magnini. 2026b. Overview of the CRF 2026 Shared Task on Clinical Case Report Forms filling. In Proceedings of the Third Workshop on Patient-Oriented Language Processing (CL4Health), Palma, Mallorca (Spain). ELRA. find
  3. Pietro Ferrazzi, Alberto Lavelli, and Bernardo Magnini. 2025. Converting Annotated Clinical Cases into Structured Case Report Forms. In Proceedings of the 24th Workshop on Biomedical Language Processing, pages 307–318, Vienna, Austria. Association for Computational Linguistics. find
  4. Gabriela Anna Kaczmarek, Pietro Ferrazzi, Lorenzo Porta, Vicky Rubini, and Bernardo Magnini. 2026. Toward automatic filling of case report forms: A case study on data from an italian emergency department. Preprint, arXiv:2602.23062. ArXiv preprint. — [arXiv:2602.23062](https://arxiv.org/abs/2602.23062) find
  5. Ita F1 0.56 *Table 3: Final macro F1 scores of the proposed system on the test set.* OpenAI, Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Altman, Andy Applebaum, Edwin Arbus, Rahul K. Arora, Yu Bai, Bowen Baker, Haiming Bao, Boaz Barak, Ally Bennett, Tyler Bertao, Nivedita Brett, Eugene Brevdo, Greg… — gpt-oss-120b & gpt-oss-20b Model Card
  6. A - Do NOT change values. - Do NOT normalize or reinterpret. - Do NOT infer from medical knowledge. - Judge ONLY based on what is explicitly written in the note. find

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/morabito-2026-nm-crf-filling-2026-two/