FU-HU-P5 at #SMM4H-HeaRD 2026: MedSynth Dialogue-to-Note Generation

Paper · 2026 · Jessica Ying En Wong

Summary

This system paper addresses Task 4 of the #SMM4H-HeaRD 2026 shared task: automatically turning a doctor–patient conversation into a clinical note in SOAP (Subjective, Objective, Assessment, Plan) format. The aim is to reduce the documentation workload on clinicians. The team used the synthetic MedSynth corpus of paired diarized dialogues and SOAP notes. After light cleaning, they fine-tuned the compact Qwen3.5-2B model using QLoRA together with Unsloth and SFTTrainer. Only about half a percent of the weights were updated, and training ran on a single consumer-grade T4 GPU. The abstract also mentions semi-supervised learning, but the method section describes only instruction-based supervised fine-tuning. On the test set, the model reached an average score of 0.44 across BLEU, ROUGE and METEOR, close to its validation average of 0.46. This suggests reasonable generalization for its size. The authors also report that some reference notes contain hallucinated details absent from the dialogues, which is a data-quality concern for MedSynth. The work is relevant as a low-resource baseline for clinical note generation.

Key points

Data: MedSynth splits are 8528 train, 1506 eval and 368 test samples. Filtering removed missing, short (<50 characters) and duplicate entries, which dropped a single training sample.
Model setup: Qwen3.5-2B (max sequence length 2048) was fine-tuned with QLoRA (rank 16, alpha 32, all attention and MLP projections). About 10.9M of 2.22B parameters (0.49%) were trained. Training used lr 2e-4, adamw_8bit, effective batch size 32 and 2 epochs.
Compute: Training took roughly 20 hours on one NVIDIA Tesla T4 (fp16). The authors observed frequent loss spikes and recommend gradient clipping and lower learning rates.
Test scores:

Metric Test Validation

BLEU 0.33 0.35

ROUGE-1 0.60 0.61

ROUGE-2 0.35 0.39

ROUGE-L 0.42 0.42

METEOR 0.51 0.54

Average 0.44 0.46
Limitations and authorship: Overly long prompts degraded output coherence, and scaling to Qwen3.5-4B or 9B is planned. The PDF lists Vardhani, Lee and Wong as authors, not Wong alone.

Metric	Test	Validation
BLEU	0.33	0.35
ROUGE-1	0.60	0.61
ROUGE-2	0.35	0.39
ROUGE-L	0.42	0.42
METEOR	0.51	0.54
Average	0.44	0.46

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 4Computer Science · 18,614 works Artificial Intelligence · 17,965 works Arts and Humanities · 17,450 works Language and Linguistics · 17,296 works

🧭 Topics 1Dialogue systems · 654 works

Cite

Cite as Jessica Ying En Wong. 2026. FU-HU-P5 at #SMM4H-HeaRD 2026: MedSynth Dialogue-to-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-1.38

@inproceedings{wong2026fu,
  title = {{FU-HU-P5 at \#SMM4H-HeaRD 2026: MedSynth Dialogue-to-Note Generation}},
  author = {Jessica Ying En Wong},
  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-1.38},
  url = {https://doi.org/10.18653/v1/2026.smm4h-1.38}
}

References

5 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

To verify the generalizability of our model, we conducted an evaluation on unseen test dataset. The results demonstrated that the model maintained consistent performance across all primary metrics. find
Anjanava Biswas and Wrick Talukdar. 2024. Intelligent clinical documentation: Harnessing generative ai for patient-centric clinical note generation. arXiv preprint arXiv:2405.18346. — [arXiv:2405.18346](https://arxiv.org/abs/2405.18346) find
Limitations and Future Work While the current experiments yielded promising results, several technical limitations were identified during the training process that provide clear directions for future improvements. find
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 Conf… find
Qwen Team. 2026. Qwen3.5: Towards native multimodal agents. 239 3 find

Updated 2 November 2026

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wong-2026-fu-hu-p5-smm4h-heard/