Elamur Research⌕
☰
home / papers / 2026 / URJC-Team at #SMM4H-HeaRD 2026: TNM Stage Extraction with a Regex-LLM Workflow

URJC-Team at #SMM4H-HeaRD 2026: TNM Stage Extraction with a Regex-LLM Workflow

Paper · 2026 · Natalia Madrueño, Jose Walter Hernández Pérez, Rubén R. Fernández et al.

Summary

The paper tackles automatic TNM cancer staging from free-text pathology reports for Task 6 of the #SMM4H-HeaRD 2026 Shared Tasks. Systems must predict the T (T1–T4), N (N0–N3) and M (M0–M1) components independently, following the AJCC 7th edition. The data come from TCGA pathology reports covering about 7,000 patients and 23 cancer types. Labels are strongly imbalanced, and many values are missing.

The URJC-Team built a two-stage pipeline. Hand-written regular expressions first capture explicitly stated staging codes. Any component the rules miss is passed to Qwen3.5-27B, used either zero-shot or after supervised fine-tuning with QLoRA. On validation, the fine-tuned variants scored best, beating several baselines including BB-TEN, logistic regression and a fully fine-tuned BioClinical-ModernBERT-large. The regex layer did not reduce accuracy and lowered inference cost.

The official evaluation set mostly contained explicit stage codes, and all submissions scored perfectly. On the harder post-evaluation set, explicit codes were almost absent, yet the system still clearly beat the baseline. This shows that cheap rules plus an LLM fallback can work both when staging is stated outright and when it must be inferred.

Key points

  • Training split: 6,774 reports, with many labels missing; for example, M is absent in 42% of reports and M1 makes up only 4%.
  • Validation phase: LLM SFT and Regex + LLM SFT both reached 0.82 Macro-F1 Stage and 0.92 Micro-F1. For comparison, LLM ZS scored 0.78, SML FFT 0.76, BB-TEN 0.71 and Regex alone 0.34 (Macro-F1 Stage).
  • Evaluation phase: the regex rules resolved 99.3% of T, 99.4% of N and 96.5% of M cases. All three submissions scored 1.00 on every metric, against a baseline Micro-F1 of 0.95.
  • Post-evaluation phase: only 0.2% of reports carried explicit TNM codes. Regex + LLM SFT achieved 0.86 Macro-F1 Stage and 0.93 Micro-F1 (baseline 0.53 / 0.71), with a perfect 1.00 Macro-F1 on M.
  • Stated limitations: high compute demands of the LLM, dependence on labeled data for SFT, evaluation limited to TCGA reports, and no explanations or confidence flags for predictions.

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 · 18,614 worksArtificial Intelligence · 17,965 worksPsychology · 1,649 worksBiochemistry, Genetics and Molecular Biology · 263 worksMolecular Biology · 260 worksSocial Psychology · 330 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,894 works

Cite

Cite as Natalia Madrueño, Jose Walter Hernández Pérez, Rubén R. Fernández and Soto Montalvo. 2026. URJC-Team at #SMM4H-HeaRD 2026: TNM Stage Extraction with a Regex-LLM Workflow. 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-1.22

@inproceedings{madrueno2026urjc,
  title = {{URJC-Team at \#SMM4H-HeaRD 2026: TNM Stage Extraction with a Regex-LLM Workflow}},
  author = {Natalia Madrueño, Jose Walter Hernández Pérez, Rubén R. Fernández and Soto Montalvo},
  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-1.22},
  url = {https://doi.org/10.18653/v1/2026.smm4h-1.22}
}

References

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

  1. Abdulrahman K. Aalabdulsalam, Jennifer H. Garvin, Andrew Redd, Marjorie E. Carter, Carol Sweeny, and Stephane M. Meystre. 2017. Automated extraction and classification of cancer stage mentions from unstructured text fields in a central cancer registry. In AMIA Joint Summits on Translational Science… find
  2. Selen Bozkurt, Christopher J. Magnani, Martin G. Seneviratne, James D. Brooks, and Tina Hernandez-Boussard. 2022. Expanding the secondary use of prostate cancer real world data: Automated classifiers for clinical and pathological stage. Frontiers in Digital Health, 4. find
  3. Mariia Chizhikova, Pilar López-Úbeda, Teodoro Martín-Noguerol, Manuel C. Díaz-Galiano, L. Alfonso Ureña-López, Antonio Luna, and M. Teresa Martín-Valdivia. 2024. Automatic TNM staging of colorectal cancer radiology reports using pre-trained language models. Computer Methods and Programs in Biomedici… find
  4. Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. QLoRA: Efficient fine-tuning of quantized LLMs. In Advances in Neural Information Processing Systems, volume 36, pages 10088–10115. find
  5. Stephen B. Edge, David R. Byrd, Carolyn C. Compton, April G. Fritz, Frederick L. Greene, and Andrew Trotti, editors. 2010. AJCC Cancer Staging Manual, 7th edition. Springer, New York. find
  6. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations.
  7. Kentaro Ishida, Ryusuke Murakami, Koji Yamanoi, Koki Hasebe Kohei Hamada, Azusa Sakurai, Taito Miyamoto, Rin Mizuno, Mana Taki, Ken Yamaguchi, Junzo Hamanishi, Kenichi Saito, Kazumasa Kishimoto, Goshiro Yamamoto, Tomohiro Kuroda, and Masaki Mandai. 2025. Real-world application of large language mode… find
  8. Ruonan Jin, Chao Ling, Yixuan Hou, Yuhan Sun, Ning Li, Jiefei Han, Jin Sheng, Qizhao Wang, Yuepeng Liu, Shen Zheng, Xingyu Ren, Chiyu Chen, Jue Wang, and Cheng Li. 2026. Augmenting large language model with prompt engineering and supervised fine-tuning in non-small cell lung cancer tumor-node-metast… find
  9. Jenna Kefeli, Jacob Berkowitz, Jose M. Acitores Cortina, Kevin K. Tsang, and Nicholas P. Tatonetti. 2024. Generalizable and automated classification of TNM stage from pathology reports with external validation. Nature Communications, 15(1):8916. find
  10. Yikuan Li, Ramsey M Wehbe, Faraz S Ahmad, Hanyin Wang, and Yuan Luo. 2023. A comparative study of pretrained language models for long clinical text. Journal of the American Medical Informatics Association, 30(2):340–347. find
  11. Guillermo Lopez-Garcia, Jose Miguel Acitores Cortina, Jacob Berkowitz, Joey Chan, Ganesh Chandrasekar, Sumon Kanti Dey, Ivan Flores Amaro, Fernando Gallego, Lauren Gryboski, Ari Z. Klein, Martin Krallinger, Salvador Lima-López, Tomohiro Nishiyama, Lisa Raithel, Ahmad Rezaie Mianroodi, Roland Roller,… find
  12. Yuta Nakamura, Shohei Hanaoka, Shuntaro Yada, Shoko Wakamiya, and Eiji Aramaki. 2023. NTCIR-17 MedNLP-SC radiology report subtask overview: Dataset and solutions for automated lung cancer staging. In NTCIR Conference on Evaluation of Information Access Technologies. find
  13. Hyung Jun Park, Namu Park, Jang Ho Lee, Myeong Geun Choi, Jin-Sook Ryu, Min Song, and Chang-Min Choi. 2022. Automated extraction of information of lung cancer staging from unstructured reports of pet-ct interpretation: natural language processing with deep-learning. BMC medical informatics and decis… find
  14. You are an expert oncology pathologist . find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/madrueno-2026-urjc-team-smm4h-heard-2026/