Elamur Research⌕
☰
home / papers / 2026 / VerbaNexAI at ClinicalSkillQA: From Visual Evidence to Procedural Order A Two-Stage Generative Vision-Language Framework for ClinSkillQA

VerbaNexAI at ClinicalSkillQA: From Visual Evidence to Procedural Order A Two-Stage Generative Vision-Language Framework for ClinSkillQA

Paper · 2026 · Andrea Menco Tovar, Jairo E. Serrano, Edwin Puertas et al.

Summary

The paper describes the VerbaNexAI Lab entry to the ClinSkillQA 2026 shared task. In this task, shuffled keyframes from videos of students performing clinical procedures must be put back in the correct order and explained. The authors work only on the Basic Life Support (BLS) subset. They use a zero-shot pipeline built on Qwen2-VL-2B-Instruct with no fine-tuning.

The pipeline has two stages:

  • Stage 1: The model writes a short factual description of each frame. Deterministic lexical rules turn it into a structured record with entities, actions, a coarse procedural state and uncertainty flags.
  • Stage 2: The same model gives each frame a single ordinal score for its BLS stage. Frames are sorted by that score. Ties are broken by projecting averaged hidden-state embeddings onto one PCA axis.

On the test leaderboard the system placed fourth of eight visible teams. Its explanation quality was fairly close to the leader's, but exact sequence reconstruction lagged well behind. The study offers a transparent, reproducible baseline. It also shows that recognising frames one at a time is not enough to separate visually similar neighbouring procedural steps.

Key points

  • Data: The BLS subset of ClinSkillQA has 200 samples, 150 with four frames and 50 with six, loaded from BLS_input.json. A 10-sample DEV subset was used for debugging.

  • Ordinal scale (0–5): approach, prepare, expose_chest, position_hands, compressions, check_or_adjust. If the model's output cannot be parsed, the system retries with a stricter prompt and then falls back to a conservative default score.

  • Setup: Qwen2-VL-2B-Instruct ran in float16 with deterministic decoding (do_sample=False) on a Kaggle Tesla T4 GPU.

  • Test results:

    MetricVerbaNexAI LabBest visible system
    Task Accuracy0.170.63
    Pairwise Micro Accuracy0.600.86
    BERT F10.710.79
    Overall score37.9671.43

    The system ranked 4/8 and produced predictions and rationales for every sample.

  • Main weakness: Temporal discrimination between adjacent, look-alike BLS states. The authors name better temporal modelling and better tie resolution as future work.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

Also on the byline: Universidad Tecnológica de Bolívar.

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksComputer Vision and Pattern Recognition · 2,869 worksMedicine · 241 worksEngineering · 133 worksHealth Informatics · 184 worksControl and Systems Engineering · 45 works

🧭 Topics 1Research subjects and languages the work is about.Vision & language · 657 works

Cite

Cite as Andrea Menco Tovar, Jairo E. Serrano, Edwin Puertas and Juan Carlos Martinez-Santos. 2026. VerbaNexAI at ClinicalSkillQA: From Visual Evidence to Procedural Order A Two-Stage Generative Vision-Language Framework for ClinSkillQA. 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-2.2

@inproceedings{tovar2026verbanexai,
  title = {{VerbaNexAI at ClinicalSkillQA: From Visual Evidence to Procedural Order A Two-Stage Generative Vision-Language Framework for ClinSkillQA}},
  author = {Andrea Menco Tovar, Jairo E. Serrano, Edwin Puertas and Juan Carlos Martinez-Santos},
  booktitle = {Proceedings of the BioNLP 2026 (Shared Tasks)},
  year = {2026},
  doi = {10.18653/v1/2026.bionlp-2.2},
  url = {https://doi.org/10.18653/v1/2026.bionlp-2.2}
}

References

11 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. For explanation quality, the system obtained BERT Precision = 0.74, BERT Recall = 0.68, and BERT F1 = 0.71. In addition, both Predicted Coverage and Rationale Coverage reached 1.0, showing Rawan AlSaad, Ala Abd-Alrazaq, Sabah Boughorbel, Arfan Ahmed, Marie A. Renault, Rami Damseh, and Javaid Sheikh.… find
  2. Lakshmi Seenivasan, Mobarakol Islam, A. K. Krishna, and Hongliang Ren. 2022. Surgical-VQA: Visual question answering in surgical scenes using transformer. In Lecture Notes in Computer Science, volume 13437, pages 33–43. find
  3. Ian Hartsock and Ghulam Rasool. 2024. Vision-language models for medical report generation and visual question answering: a review. Frontiers in Artificial Intelligence, 7:1430984. find
  4. Julie H. Vermylen, Elaine R. Cohen, David A. Cook, William C. McGaghie, S. Barry Issenberg, Jennifer Arnold, Heather Ballard, Mohamed Bayoumi, Michael Beestrum, Ruth Bremner, Sarah Crawford, Naomi Einstein, Christina Mannarino, Anjali Misra, T. M. Tomita, Hannah Waldron, Francesca Yanko, and Deborah… find
  5. Xiyang Huang, Jiawei Lin, Keying Wu, Jiaxin Huang, Kailai Yang, Renxiong Wei, Jiayi Xiang, Ziyan Kuang, Min Peng, Qianqian Xie, and 1 others. 2026. Siming-bench: Evaluating procedural correctness from continuous interactions in clinical skill videos. arXiv preprint arXiv:2604.09037. — [arXiv:2604.09037](https://arxiv.org/abs/2604.09037) find
  6. S. Barry Issenberg, William C. McGaghie, E. Ronald Petrusa, David L. Gordon, and Rosemary J. Scalese. 2005. Features and uses of high-fidelity medical simulations that lead to effective learning: a BEME systematic review. Medical Teacher, 27(1):10–28. find
  7. Peng Wang, Shuai Bai, Shidong Tan, Shunqiang Wang, Zhihao Fan, Jing Bai, Kai Chen, Xian Liu, Jie Wang, Wei Ge, Yu Fan, Kai Dang, Meng Du, Xin Ren, Rui Men, Dong Liu, Chen Zhou, Jingren Zhou, and Ji Lin. 2024. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. arXi…
  8. Mahdieh Ebrahimi Khanghahi and Farahnaz Esmaeili Fomani Azar. 2018. Direct Observation of Procedural Skills (DOPS) evaluation method: Systematic review of evidence. Medical Journal of the Islamic Republic of Iran, 32(1):45. find
  9. X. Zhang, C. Wu, Z. Zhao, W. Lin, Y. Zhang, Y. Wang, and W. Xie. 2024. Development of a large-scale medical visual question-answering dataset. Communications Medicine, 4(1):277. find
  10. Jessica J. Lau, Shibashis Gayen, Asma Ben Abacha, and Dina Demner-Fushman. 2018. A dataset of clinically generated visual questions and answers about radiology images. Scientific Data, 5(1):180251. .1 find
  11. Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Hao Liu, Jian Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. 2023. LLaVA-Med: Training a large language-and-vision assistant for biomedicine in one day. In Advances in Neural Information Processing Systems, volume 36.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tovar-2026-verbanexai-clinicalskillqa-visual-evidence-procedural/