Elamur Research⌕
☰
home / papers / 2026 / Uncertainty-Aware Multi-Label Routing of Clinical Text to Surveillance Pathways

Uncertainty-Aware Multi-Label Routing of Clinical Text to Surveillance Pathways

Paper · 2026 · Agathe Zecevic, Sebastian Zeki and Angus Roberts

Summary

Decision-support tools that serve several gastrointestinal (GI) surveillance pathways must first decide which pathways apply to a patient. One report can point to more than one condition at once. The authors treat this routing as a six-label classification problem over paired endoscopy and histopathology reports, and they stress the cost of wrongly dropping a relevant pathway. Using 1,773 report pairs from one NHS trust, they compare four baselines: regex keywords, TF-IDF with logistic regression, frozen BGE-Large-EN-v1.5 embeddings with a linear head, and fine-tuned ClinicalBERT. Each is combined with two ways of handling uncertainty: min-margin threshold abstention and per-label conformal prediction, in marginal and Mondrian variants. Fine-tuned ClinicalBERT performed best overall. Abstention lowered exact-match error on the cases the model accepted. Mondrian conformal prediction kept most true pathways in its candidate sets while making the extra review workload measurable. The paper's main lesson is that aggregate scores hide important clinical failure modes. Set-valued outputs should therefore support human review, not automatically activate pathways.

Key points

  • Data and labels: 1,773 merged endoscopy–histopathology reports from Guy's and St Thomas' NHS Foundation Trust, with separate train, calibration and test splits.
    • The six labels are Normal, Barrett's, Colorectal Polyps, Gastric Atrophy, Other and Incomplete.
    • All labels come from a single annotator.
    • The data is not public; code is available on request.
  • Hard-prediction results: fine-tuned ClinicalBERT reached 0.811 subset accuracy and 0.861 macro-F1. Under min-margin abstention it had the lowest AURC, 0.084.
  • Mondrian conformal prediction at α = 0.10: mean positive-label recall coverage ranged from 0.883 to 0.917 across the learned baselines. For ClinicalBERT:
    • recall coverage was 0.891;
    • average set size was 1.70 labels;
    • candidate-label precision was 0.642;
    • there were 0.61 false-positive labels per report.
  • Comparison with a recall-tuned threshold: a per-label threshold set for at least 90% recall reached similar recall (0.895). However, it produced larger sets (2.03), lower precision (0.540) and more false positives (0.93 per report) than Mondrian conformal prediction.
  • Trade-off on the rare Gastric Atrophy class: Mondrian conformal prediction raised recall from 0.67 to 0.93, but precision fell from 0.80 to 0.18. Marginal conformal prediction recovered too few positives for several labels.

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

🎓King's College LondonLondon · United Kingdom46 works · 1,714 citations

Also on the byline: Guy's and St Thomas' NHS Foundation Trust.

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksHealth Professions · 75 worksHealth Information Management · 56 works

Cite

Cite as Agathe Zecevic, Sebastian Zeki and Angus Roberts. 2026. Uncertainty-Aware Multi-Label Routing of Clinical Text to Surveillance Pathways. BioNLP 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.bionlp-1.16

@inproceedings{zecevic2026uncertainty,
  title = {{Uncertainty-Aware Multi-Label Routing of Clinical Text to Surveillance Pathways}},
  author = {Agathe Zecevic, Sebastian Zeki and Angus Roberts},
  booktitle = {BioNLP 2026},
  year = {2026},
  doi = {10.18653/v1/2026.bionlp-1.16},
  url = {https://doi.org/10.18653/v1/2026.bionlp-1.16}
}

References

26 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Emily Alsentzer, John R. Murphy, Willie Boag, Wei-Hung Weng, Di Jin, Tristan Naumann, and Matthew B. A. McDermott. 2019. Publicly Available Clinical BERT Embeddings.
  2. Anastasios N. Angelopoulos and Stephen Bates. 2022. A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification.
  3. Matthew Banks, David Graham, Marnix Jansen, Takuji Gotoda, Sergio Coda, Massimiliano di Pietro, Noriya Uedo, Pradeep Bhandari, D Mark Pritchard, Ernst J Kuipers, Manuel Rodriguez-Justo, Marco R Novelli, Krish Ragunath, Neil Shepherd, and Mario Dinis-Ribeiro. 2019. British Society of Gastroenterology… find
  4. Margarida Campos, António Farinhas, Chrysoula Zerva, Mário A. T. Figueiredo, and André F. T. Martins. 2024. Conformal Prediction for Natural Language Processing: A Survey. Transactions of the Association for Computational Linguistics, 12:1497–1516. find
  5. Don Chamil Codipilly, Apoorva Krishna Chandar, Siddharth Singh, Sachin Wani, Nicholas J. Shaheen, John M. Inadomi, Amitabh Chak, and Prasad G. Iyer. 2018. The Effect of Endoscopic Surveillance in Patients With Barrett’s Esophagus: A Systematic Review and Meta-analysis. Gastroenterology, 154(8):2068–… find
  6. Ran El-Yaniv and Yair Wiener. 2010. On the Foundations of Noise-free Selective Classification. Journal of Machine Learning Research, 11(53):1605–1641. find
  7. Adam Fisch, Tal Schuster, Tommi Jaakkola, and Regina Barzilay. 2022. Conformal Prediction Sets with Limited False Positives. find
  8. Rebecca C Fitzgerald, Massimiliano di Pietro, Krish Ragunath, Yeng Ang, Jin-Yong Kang, Peter Watson, Nigel Trudgill, Praful Patel, Philip V Kaye, Scott Sanders, Maria O’Donovan, Elizabeth Bird-Lieberman, Pradeep Bhandari, Janusz A Jankowski, Stephen Attwood, Simon L Parsons, Duncan Loft, Jesper Lage… find
  9. Yonatan Geifman, Guy Uziel, and Ran El-Yaniv. 2019. Bias-Reduced Uncertainty Estimation for Deep Neural Classifiers. find
  10. Juliano Genari and Guilherme Tegoni Goedert. 2025. Mining Unstructured Medical Texts With Conformal Active Learning. find
  11. Vianney Taquet, Vincent Blot, Thomas Morzadec, Louis Lacombe, and Nicolas Brunel. 2022. MAPIE: an open-source library for distribution-free uncertainty quantification. find
  12. Glenn Shafer Vladimir Vovk, Alexander Gammerman. 2005. Algorithmic Learning in a Random World. Springer-Verlag, New York. find
  13. Bingbing Wen, Jihan Yao, Shangbin Feng, Chenjun Xu, Yulia Tsvetkov, Bill Howe, and Lucy Lu Wang. 2025. Know Your Limits: A Survey of Abstention in Large Language Models. find
  14. Yusong Ke, Hongru Lin, Yuting Ruan, Junya Tang, and Li Li. 2025. Correctness Coverage Evaluation for Medical Multiple-Choice Question Answering Based on the Enhanced Conformal Prediction Framework. find
  15. Emma Peterson, Folasade P. May, Odet Kachikian, Camille Soroudi, Bita Naini, Yuna Kang, Anthony Myint, Gordon Guyant, Joann Elmore, Roshan Bastani, Cleo Maehara, and William Hsu. 2021. Automated identification and assignment of colonoscopy surveillance recommendations for individuals with colorectal… find
  16. Matthew D Rutter, James East, Colin J Rees, Neil Cripps, James Docherty, Sunil Dolwani, Philip V Kaye, Kevin J Monahan, Marco R Novelli, Andrew Plumb, Brian P Saunders, Siwan Thomas-Gibson, Damian J M Tolan, Sophie Whyte, Stewart Bonnington, Alison Scope, Ruth Wong, Barbara Hibbert, John Marsh, and… find
  17. Manil Shrestha and Edward Kim. 2026. Conformal Prediction for Risk-Controlled Medical Entity Extraction Across Clinical Domains. find
  18. Gyuseon Song, Su Jin Chung, Ji Yeon Seo, Sun Young Yang, Eun Hyo Jin, Goh Eun Chung, Sung Ryul Shim, Soonok Sa, Moongi Simon Hong, Kang Hyun Kim, Eunchan Jang, Chae Won Lee, Jung Ho Bae, and Hyun Wook Han. 2022. Natural Language Processing for Information Extraction of Gastric Diseases and Its Appli… find
  19. Irena Spasic and Kate Button. 2020. Patient Triage by Topic Modeling of Referral Letters: Feasibility Study. JMIR Medical Informatics, 8(11):e21252. find
  20. Jonathon Stewart, Juan Lu, Adrian Goudie, Glenn Arendts, Shiv Akarsh Meka, Sam Freeman, Katie Walker, Peter Sprivulis, Frank Sanfilippo, Mohammed Bennamoun, and Girish Dwivedi. 2023. Applications of natural language processing at emergency department triage: A narrative review. PLOS ONE, 18(12):e027… find
  21. Reed T. Sutton, David Pincock, Daniel C. Baumgart, Daniel C. Sadowski, Richard N. Fedorak, and Karen I. Kroeker. 2020. An overview of clinical decision support systems: benefits, risks, and strategies for success. npj Digital Medicine, 3(1):17. find
  22. Jonathan R. White and Matthew Banks. 2022. Identifying the pre-malignant stomach: from guidelines to practice. Translational Gastroenterology and Hepatology, 7:8–8. find
  23. Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu Lian, and Jian-Yun Nie. 2024. C-Pack: Packed Resources For General Chinese Embeddings.
  24. Agathe Zecevic, Laurence Jackson, Xinyue Zhang, Polychronis Pavlidis, Jason Dunn, Nigel Trudgill, Shahd Ahmed, Pierfrancesco Visaggi, Zanil YoonusNizar, Angus Roberts, and Sebastian S. Zeki. 2024. Automated decision making in Barrett’s oesophagus: development and deployment of a natural language pro… find
  25. Xin Zhang, Qiyu Wei, Yingjie Zhu, Fanyi Wu, and Sophia Ananiadou. 2025a. THCM-CAL: Temporal-Hierarchical Causal Modelling with Conformal Calibration for Clinical Risk Prediction. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 916–928, Stroudsburg, PA, USA. Associatio… find
  26. Xinyue Zhang, Agathe Zecevic, Sebastian Zeki, and Angus Roberts. 2025b. Improving Barrett’s Oesophagus Surveillance Scheduling with Large Language Models: A Structured Extraction Approach. In Proceedings of the 24th Workshop on Biomedical Language Processing, pages 176–189, Stroudsburg, PA, USA. Ass… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zecevic-2026-uncertainty-aware-multi-label-routing/