Elamur Research⌕
☰
home / papers / 2026 / NU_DeepHealthNLP at #SMM4H-HeaRD 2026: Entity-Conditioned Generation and a Four-Stage Pipeline for Automated SOAP Note Generation

NU_DeepHealthNLP at #SMM4H-HeaRD 2026: Entity-Conditioned Generation and a Four-Stage Pipeline for Automated SOAP Note Generation

Paper · 2026 · Thanya Mysore Santhosh and Deahan Yu

Summary

The paper addresses SMM4H-HeaRD 2026 Task 4, which asks systems to turn transcribed doctor–patient conversations into structured SOAP notes (Subjective, Objective, Assessment, Plan). The motivation is the heavy documentation workload that physicians carry. The authors use the synthetic MedSynth corpus, with 8,529 training dialogues, 1,506 validation dialogues and a blind 368-dialogue test set. On this data they fine-tune Mistral-7B-Instruct-v0.1 with QLoRA.

Their central idea is entity conditioning. A zero-shot Mistral-7B-Instruct-v0.3 extractor identifies ten kinds of clinical entities and sorts them by SOAP section. These entities are added to the prompt in exactly the same format during training and at inference. One submission uses only extraction and generation. The second adds a hybrid BM25/dense retriever that supplies the closest training note, plus a rule-based verifier that checks and repairs the output.

Both systems outperform the published MedSynth baselines on validation. Each reached a 0.54 average on the test set, which placed first on the leaderboard. The results suggest that giving a compact open model explicit, section-sorted clinical facts improves note quality at modest compute cost. Caveats include synthetic-only data and reliance on lexical-overlap metrics.

Key points

  • Generator: Mistral-7B-Instruct-v0.1 fine-tuned with QLoRA.
    • LoRA rank 16, α 16, applied to all seven projection modules.
    • About 13.6M trainable parameters (0.19%).
    • Trained for 1 epoch with Unsloth; inference runs on vLLM.
  • Stage 1 extractor: zero-shot Mistral-7B-Instruct-v0.3 producing JSON output.
    • Covers 10 entity types, grouped into the four SOAP sections.
    • Averages 18.83 entities per validation dialogue.
  • Four-stage pipeline additions:
    • BM25 + FAISS dense retrieval (all-MiniLM-L6-v2), merged with Reciprocal Rank Fusion, in a RAFT-style setup.
    • A rule-based verifier checking faithfulness, negation, completeness, structure and output appropriateness, with no extra model calls.
  • Validation BLEU (1,506 dialogues):
    • Entity-conditioned model: 0.5464.
    • Four-stage pipeline: 0.5423.
    • Best MedSynth baseline (Mistral-7B-v0.3): 0.5346.
  • Test results (368 dialogues):
    • Both submissions averaged 0.54, ranking first among 7 teams (team dy_nu); the runner-up scored 0.53.
    • Test BLEU was 0.4415 and 0.4427, about 0.10 below validation, which the authors attribute to distribution shift.
    • Error analysis found three recurring problems: missed secondary symptoms, flattened timelines of symptom progression, and occasional assessments not grounded in the dialogue.

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 4The 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 works

Cite

Cite as Thanya Mysore Santhosh and Deahan Yu. 2026. NU_DeepHealthNLP at #SMM4H-HeaRD 2026: Entity-Conditioned Generation and a Four-Stage Pipeline for Automated SOAP Note Generation. Proceedings of the 11th Social Media Mining for Health Research and Applications (SMM4H-HeaRD 2026) Workshop and Shared Tasks.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.smm4h-1.17

@inproceedings{santhosh2026nu,
  title = {{NU\_DeepHealthNLP at \#SMM4H-HeaRD 2026: Entity-Conditioned Generation and a Four-Stage Pipeline for Automated SOAP Note Generation}},
  author = {Thanya Mysore Santhosh and Deahan Yu},
  booktitle = {Proceedings of the 11th Social Media Mining for Health Research and Applications (SMM4H-HeaRD 2026) Workshop and Shared Tasks},
  year = {2026},
  doi = {10.18653/v1/2026.smm4h-1.17},
  url = {https://doi.org/10.18653/v1/2026.smm4h-1.17}
}

References

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

  1. Asma Ben Abacha, Wen-wai Yim, Yadan Fan, and Thomas Lin. 2023. An empirical study of clinical note generation from doctor-patient encounters. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics (EACL), pages 2291–2302. find
  2. Satanjeev Banerjee and Alon Lavie. 2005. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Ex-Daniel Han, Michael Han, and Unsloth Team. 2023. Unsloth. https://github.com/unslothai/unsloth. ↗
  3. Edward J. Hu and 1 others. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR).
  4. Albert Q. Jiang and 1 others. 2023. Mistral 7B. arXiv preprint arXiv:2310.06825. find
  5. Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2021. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data, 7(3):535–547. ArXiv preprint arXiv:1702.08734 (2017).
  6. Kundan Krishna, Sopan Khosla, Jeffrey P. Bigham, and Zachary C. Lipton. 2021. Generating SOAP notes from doctor-patient conversations using modular summarization techniques. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Co… find
  7. Woosuk Kwon and 1 others. 2023. Efficient memory management for large language model serving with PagedAttention. In Proceedings of the 29th Symposium on Operating Systems Principles (SOSP), pages 611–626.
  8. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Proc…
  9. Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out.
  10. Ahmad Rezaie Mianroodi, Amirali Rezaie, Niko Grisel Todorov, Cyril Rakovski, and Frank Rudzicz. 2025. MedSynth: Realistic, synthetic medical dialogue-note pairs. Preprint, arXiv:2508.01401. — [arXiv:2508.01401](https://arxiv.org/abs/2508.01401) find
  11. Francesco Moramarco and 1 others. 2022. Human evaluation and correlation with automatic metrics in consultation note generation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL), pages 5739–5754. find
  12. SMM4H Organizers. 2026. \#SMM4H-HeaRD 2026 shared task 4: Dialogue-to-note generation. https://healthlanguageprocessing.org/ smm4h-2026/. ↗
  13. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. BLEU: A method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL), pages 311–318.
  14. Vivek Podder, Valerie Lew, and Sassan Ghassemzadeh. 2023. SOAP notes. StatPearls [Internet]. Treasure Island (FL): StatPearls Publishing. find
  15. Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence embeddings using siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3982–3992.
  16. Stephen E. Robertson, Steve Walker, Susan Jones, Micheline M. Hancock-Beaulieu, and Mike Gatford. 1995. Okapi at TREC-3. In Proceedings of the Third Text Retrieval Conference (TREC-3), NIST Special Publication 500-226, pages 109–126.
  17. Christine Sinsky and 1 others. 2016. Allocation of physician time in ambulatory practice: A time and motion study in 4 specialties. Annals of Internal Medicine, 165(11):753–760. find
  18. Wen-wai Yim, Yujuan Fu, Asma Ben Abacha, Neal Snider, Thomas Lin, and Meliha Yetisgen. 2023. Aci-Bench: A novel ambient clinical intelligence dataset for benchmarking automatic visit note generation. Scientific Data, 10(1):586. find
  19. Tianjun Zhang, Shishir G. Patil, Naman Jain, Sheng Shen, Matei Zaharia, Ion Stoica, and Joseph E. Gonzalez. 2024. RAFT: Adapting language model to domain specific RAG. arXiv preprint arXiv:2403.10131.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/santhosh-2026-nu-deephealthnlp-smm4h-heard-2026/