Capturing Clinical Reasoning in AI

Semantic annotation and knowledge representation for sepsis detection

Melissa Y. Yana,b,c, Lise Tuset Gustadc,d, Lise Husby Høvikc,e, Øystein Nytrøa,c,f

aDepartment of Computer Science, Norwegian University of Science and Technology
bDepartment of Circulation and Medical Imaging, Norwegian University of Science and Technology
cCentral Norway Sepsis Research Centre
dFaculty of Nursing and Health Sciences, Nord University
eClinic of Anaesthesia and Intensive Care, St. Olavs Hospital, Trondheim University Hospital
fDepartment of Computer Science, UiT The Arctic University of Norway

Poster presented at the Nature Conference: Redefining Healthcare in the Age of AI.

Introduction:

  • AI models achieve strong predictive performance, but AI adoption in clinical practice remains limited [1].
  • The mismatch between model explanations and clinical reasoning results in a lack of clinician trust that is a key barrier to real-world AI adoption [2].
  • Clinical text improves model performance, but annotations are task-specific and corpora are rarely shared which limits reuse [3].
  • Context matters: a combination of documented signs in a specific situation can describe a medical condition never explicitly written in text [4], [5].
  • Capturing this requires clinical expertise encoded as reusable knowledge, not just entity labels [1], [4], [5].
  • Thus, clinical natural language processing needs to shift from solely extracting entities towards reasoning with domain knowledge [5].

Aim: To develop a transparent, ontology-based method that captures clinical reasoning from clinical text for sepsis detection by structuring explicit observations into a terminology and inferring implicit signs via an ontology.

Clinical Problem:

  • Sepsis is a life-threatening syndrome caused by a dysregulated immune response to an infection that leads to multi-organ failure [6].
  • Peripheral intravenous catheters (PIVCs) cause 7.6-35% of S. aureus bloodstream infections [7].
  • Despite routine PIVC usage and high sepsis mortality rates, both are poorly documented [8], [9].

The 5 step approach:

  1. Identify a clinical problem & clinical questions
  2. Curate a semantically annotated corpus
  3. Develop a terminology & an ontology to capture clinicians’ reasoning
  4. Answer clinical questions via SPARQL queries
  5. Automate question answering with LLM annotation and ontology-based clinical reasoning

Results:

  • Curated clinical reasoning AI resources with clinicians:
    • A systematic semantic annotation method [10].
    • An annotated corpus of incident reports for signs and events of catheters and infections [11].
    • A clinician-verified terminology to reflect how catheter- and infection-related signs and events appear in clinical documentation [12].
    • A corresponding clinician-verified ontology to infer the presence of documented catheters and infections [12].
  • The ontology infers implicit signs from explicit observations which results in interpretable clinical reasoning linked to documented evidence.
  • Demonstrated feasibility on Norwegian adverse event reports.
  • Work in progress: electronic health record integration of sepsis AI surveillance for prevention and informed management.

References

  1. [1] K. Sokol, J. Fackler, and J. E. Vogt, “Artificial intelligence should genuinely support clinical reasoning and decision making to bridge the translational gap,” NPJ Digit. Med., vol. 8, no. 1, p. 345, June 2025, doi: 10.1038/s41746-025-01725-9
  2. [2] K. Zhang, D. Wang, F. Lin, J. Xie, and W. Zhou, “A comprehensive review of explainable artificial intelligence in healthcare methods, evaluation, and clinical integration,” iScience, vol. 29, no. 3, p. 115026, Mar. 2026, doi: 10.1016/j.isci.2026.115026
  3. [3] I. Spasic and G. Nenadic, “Clinical text data in machine learning: Systematic review,” JMIR Med. Inform., vol. 8, no. 3, p. e17984, Mar. 2020, doi: 10.2196/17984
  4. [4] S. Sheikhalishahi, R. Miotto, J. T. Dudley, A. Lavelli, F. Rinaldi, and V. Osmani, “Natural language processing of clinical notes on chronic diseases: Systematic review,” JMIR Med. Inform., vol. 7, no. 2, p. e12239, Apr. 2019, doi: 10.2196/12239
  5. [5] S. van Baalen, M. Boon, and P. Verhoef, “From clinical decision support to clinical reasoning support systems,” J. Eval. Clin. Pract., vol. 27, no. 3, pp. 520–528, June 2021, doi: 10.1111/jep.13541
  6. [6] M. Singer et al., “The third international consensus definitions for sepsis and septic shock (sepsis-3),” JAMA, vol. 315, no. 8, pp. 801–810, Feb. 2016, doi: 10.1001/jama.2016.0287
  7. [7] L. A. Mermel, “Short-term peripheral venous catheter-related bloodstream infections: A systematic review,” Clin. Infect. Dis., vol. 65, no. 10, pp. 1757–1762, Oct. 2017, doi: 10.1093/cid/cix562
  8. [8] E. Alexandrou et al., “Use of short peripheral intravenous catheters: Characteristics, management, and outcomes worldwide,” J. Hosp. Med., vol. 13, no. 5, May 2018, doi: 10.12788/jhm.3039
  9. [9] J. M. Rohde et al., “The epidemiology of acute organ system dysfunction from severe sepsis outside of the intensive care unit: Epidemiology of Organ System Dysfunction,” J. Hosp. Med., vol. 8, no. 5, pp. 243–247, May 2013, doi: 10.1002/jhm.2012
  10. [10] M. Y. Yan, L. T. Gustad, L. H. Høvik, and Ø. Nytrø, “Method for designing semantic annotation of sepsis signs in clinical text,” in Proceedings of the 5th Clinical Natural Language Processing Workshop, T. Naumann, A. B. Abacha, S. Bethard, K. Roberts, and A. Rumshisky, Eds., Stroudsburg, PA, USA: Association for Computational Linguistics, 2023, pp. 236–246. doi: 10.18653/v1/2023.clinicalnlp-1.29
  11. [11] M. Y. Yan, L. H. Høvik, A. Pedersen, L. T. Gustad, and Ø. Nytrø, “Preliminary Processing and Analysis of an Adverse Event Dataset for Detecting Sepsis-Related Events,” in IEEE International Conference on Bioinformatics and Biomedicine, BIBM 2021, Houston, TX, USA, December 9-12, 2021, Y. Huang, L. A. Kurgan, F. Luo, X. Hu, Y. Chen, E. R. Dougherty, A. Kloczkowski, and Y. Li, Eds., IEEE, 2021, pp. 1605–1610. doi: 10.1109/BIBM52615.2021.9669410
  12. [12] M. Y. Yan, L. T. Gustad, L. H. Høvik, and Ø. Nytrø, “Terminology and ontology development for semantic annotation: A use case on sepsis and adverse events,” Semant. Web, vol. 14, no. 5, pp. 811–871, May 2023, doi: 10.3233/sw-223226