Elamur Research⌕
☰
home / papers / 2026 / Operation-Mechanism Alignment for Reliable Clinical Reasoning over Electronic Health Records

Operation-Mechanism Alignment for Reliable Clinical Reasoning over Electronic Health Records

Paper · 2026 · Guanyu Tao, Siyao Wang, Yong Xue et al.

Summary

LLM pipelines for clinical decisions over electronic health records usually fold very different steps into one generation call: reading free-text notes, calculating lab-derived values, restricting data to time windows, and applying guideline rules. As a result, arithmetic or date mistakes reach the answer unnoticed, and the reasoning cannot be audited. The authors propose "operation-mechanism alignment." Each task becomes a directed acyclic graph whose nodes are typed as EVALUATE, COMPUTE or AGGREGATE. Each type goes to a different mechanism:

  • interpretive judgments to LLM evaluators;
  • numeric, temporal and unit-sensitive work to deterministic tools;
  • combination of criteria to explicit logical operators.

Every intermediate result keeps structured links to its supporting record evidence. Domain experts author the task graphs from clinical guidelines. The method was tested on six clinician-labelled binary tasks built from MIMIC-III, MIMIC-IV and a partner-hospital dataset. It outperforms direct prompting, single-step RAG and chain-of-thought, mostly through higher specificity. The gains are largest where criteria involve multi-step quantities and time filtering. The work argues that matching the execution mechanism to the operation type is a practical route to inspectable clinical reasoning.

Key points

  • Typed DAG nodes (EVALUATE / COMPUTE / AGGREGATE) are routed to LLM evaluators, deterministic tools, or logic operators. Each node's output pairs a result with evidence references. The running example is chronic kidney disease staging via eGFR and its decline rate.
  • The evaluation covers six binary tasks across four disease areas, using MIMIC-III, MIMIC-IV and de-identified Cone Health data. Cohorts range from 50 to 151 patients, and certain tasks have heavily skewed class distributions (the asthma exacerbation risk task, for instance, splits 46/4).
  • Task graphs follow validated guidelines such as KDIGO 2024 and GOLD 2025. The authors present the approach as complementary to ReAct/Toolformer-style agents rather than a replacement for them.
  • The advantage shrinks on simple diagnostic-lookup tasks.
  • The authors list several open issues:
    • templates are built manually;
    • there is no mechanism-level ablation;
    • there is no direct comparison with tool-using agents;
    • compute cost and latency are not quantified.

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

🎓Hong Kong University of Science and TechnologyHong Kong · China142 works · 25,136 citations🎓Imperial College LondonLondon · United Kingdom48 works · 12,181 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksMedicine · 241 worksHealth Informatics · 184 worksHealth Professions · 69 worksHealth Information Management · 50 works

🧭 Topics 1Research subjects and languages the work is about.Reasoning · 1,426 works

Cite

Cite as Guanyu Tao, Siyao Wang, Yong Xue, Ashwani Tanwar, Yuting Ji, Kai Sun, Monica Mok, Marzana Chowdhury, Deepa Gupta, Ashok Gupta, Jingqing Zhang, Vibhor Gupta and Yike Guo. 2026. Operation-Mechanism Alignment for Reliable Clinical Reasoning over Electronic Health Records. BioNLP 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.bionlp-1.48

@inproceedings{tao2026operation,
  title = {{Operation-Mechanism Alignment for Reliable Clinical Reasoning over Electronic Health Records}},
  author = {Guanyu Tao, Siyao Wang, Yong Xue, Ashwani Tanwar, Yuting Ji, Kai Sun, Monica Mok, Marzana Chowdhury, Deepa Gupta, Ashok Gupta, Jingqing Zhang, Vibhor Gupta and Yike Guo},
  booktitle = {BioNLP 2026},
  year = {2026},
  doi = {10.18653/v1/2026.bionlp-1.48},
  url = {https://doi.org/10.18653/v1/2026.bionlp-1.48}
}

References

20 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Suhana Bedi, Hejie Cui, Miguel Fuentes, Alyssa Unell, Michael Wornow, Juan M Banda, Nikesh Kotecha, Timothy Keyes, Yifan Mai, Mert Oez, and 1 others. 2025. Medhelm: Holistic evaluation of large language models for medical tasks. arXiv preprint arXiv:2505.23802. find
  2. Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, and 1 others. 2024. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI conference on a… find
  3. Lang Cao, Qingyu Chen, and Yue Guo. 2026. Ehr-rag: Bridging long-horizon structured electronic health records and large language models via enhanced retrieval-augmented generation. arXiv preprint arXiv:2601.21340. — [arXiv:2601.21340](https://arxiv.org/abs/2601.21340) find
  4. Philip Chung, Akshay Swaminathan, Alex J Goodell, Yeasul Kim, S Momsen Reincke, Lichy Han, Ben Deverett, Mohammad Amin Sadeghi, Abdel-Badih Ariss, Marc Ghanem, and 1 others. 2025. Verifying facts in patient care documents generated by large language models using electronic health records. NEJM AI, 3… find
  5. Emma Croxford, Yanjun Gao, Elliot First, Nicholas Pellegrino, Miranda Schnier, John Caskey, Madeline Oguss, Graham Wills, Guanhua Chen, Dmitriy Dligach, and 1 others. 2025. Evaluating clinical ai summaries with large language models as judges. npj Digital Medicine, 8(1):640. find
  6. Hejie Cui, Alyssa Unell, Bowen Chen, Jason Alan Fries, Emily Alsentzer, Sanmi Koyejo, and Nigam H Shah. 2025. Timer: Temporal instruction modeling and evaluation for longitudinal clinical records. npj Digital Medicine, 8(1):577. find
  7. William J Evans, John E Morley, Josep Argilés, Connie Bales, Vickie Baracos, Denis Guttridge, Aminah Jatoi, Kamyar Kalantar-Zadeh, Herbert Lochs, Giovanni Mantovani, and 1 others. 2008. Cachexia: a new definition. Clinical nutrition, 27(6):793–799. find
  8. Kenneth Fearon, Florian Strasser, Stefan D Anker, Ingvar Bosaeus, Eduardo Bruera, Robin L Fainsinger, Aminah Jatoi, Charles Loprinzi, Neil MacDonald, Giovanni Mantovani, and 1 others. 2011. Definition and classification of cancer cachexia: an international consensus. The lancet oncology, 12(5):489–4… find
  9. Yanjun Gao, Dmitriy Dligach, Timothy Miller, John Caskey, Brihat Sharma, Matthew M Churpek, and Majid Afshar. 2023. Dr. bench: Diagnostic reasoning benchmark for clinical natural language processing. Journal of biomedical informatics, 138:104286. find
  10. Global Initiative for Asthma. 2024. Global strategy for asthma management and prevention. 2024 update. Accessed: 2026-04-17. find
  11. Yixing Jiang, Kameron C Black, Gloria Geng, Danny Park, James Zou, Andrew Y Ng, and Jonathan H Chen. Medagentbench: A realistic virtual ehr environment to benchmark medical llm agents, 2025. URL https://arxiv.org/abs/2501.14654. — [arXiv:2501.14654](https://arxiv.org/abs/2501.14654) ↗
  12. Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T Joshi, Hanna Moazam, and 1 others. 2023. Dspy: Compiling declarative language model calls into self-improving pipelines. arXiv preprint arXiv:2310.03714.
  13. Maya Kruse, Shiyue Hu, Nicholas Derby, Yifu Wu, Samantha Stonbraker, Bingsheng Yao, Dakuo Wang, Elizabeth Goldberg, and Yanjun Gao. 2025. Large language models with temporal reasoning for longitudinal clinical summarization and prediction. In Findings of ACL. EMNLP. Conference on Empirical Methods i… find
  14. Sunjun Kweon, Jiyoun Kim, Heeyoung Kwak, Dongchul Cha, Hangyul Yoon, Kwanghyun Kim, Jeewon Yang, Seunghyun Won, and Edward Choi. 2024. Ehrnoteqa: An llm benchmark for real-world clinical practice using discharge summaries. Advances in Neural Information Processing Systems, 37:124575– 124611. find
  15. Nikita Mehandru, Niloufar Golchini, David Bamman, Travis Zack, Melanie F Molina, and Ahmed Alaa. 2025. Er-reason: A benchmark dataset for llm-based clinical reasoning in the emergency room. arXiv preprint arXiv:2505.22919. — [arXiv:2505.22919](https://arxiv.org/abs/2505.22919) find
  16. Monica Munnangi, Akshay Swaminathan, Jason Alan Fries, Jenelle Jindal, Sanjana Narayanan, Ivan Lopez, Lucia Tu, Philip Chung, Jesutofunmi A Omiye, Mehr Kashyap, and 1 others. 2024. Factehr: A dataset for evaluating factuality in clinical notes using llms. arXiv preprint arXiv:2412.12422. — [arXiv:2412.12422](https://arxiv.org/abs/2412.12422) find
  17. Justice Ou, Tinglin Huang, Yilun Zhao, Ziyang Yu, Peiqing Lu, and Rex Ying. 2025. Experience retrieval-augmentation with electronic health records enables accurate discharge qa. arXiv preprint arXiv:2503.17933.
  18. Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools. Advances in neural information processing systems, 36:68539–68551.
  19. G Xiong, Q Jin, X Wang, M Zhang, Z Lu, and A Zhang. 2024a. Improving retrieval-augmented generation in medicine with iterative follow-up questions. arxiv; 2024. arXiv preprint arXiv:2310.19988. — [arXiv:2310.19988](https://arxiv.org/abs/2310.19988) find
  20. Guangzhi Xiong, Qiao Jin, Zhiyong Lu, and Aidong Zhang. 2024b. Benchmarking retrieval-augmented generation for medicine. In Findings of the Association for Computational Linguistics: ACL 2024, pages 6233–6251.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tao-2026-operation-mechanism-alignment-reliable-clinical/