Elamur Research⌕
☰
home / papers / 2026 / Persona-Conditioned Generation of Patient Self-Reports from EHRs

Persona-Conditioned Generation of Patient Self-Reports from EHRs

Summary

Large English collections of patient-written symptom descriptions are rare, which hinders diagnosis and triage models meant for real first-contact settings. The authors address this by synthesizing first-person self-reports from MIMIC-IV Note. Only patient-stated content is used, and each case is paired with a persona built from structured MIMIC-IV fields such as age, gender, insurance, language and visit urgency, plus an education level that was assigned by simulation. A rule-based prompt controls which denials appear, what history is included, and how the persona shapes register. Two generators, GPT-4o-mini and GPT-5-mini, were run on 10k stratified cases. Two LLM graders, GPT-4.1 and Claude Sonnet 3.7, then judged the outputs using a weighted four-dimension rubric and a separate rubric-free pairwise preference test. GPT-5-mini scored higher on the factual dimensions. In the free-form comparison, however, the graders disagreed, largely because of stylistic biases such as favoring longer texts. The study offers a planned public resource and an evaluation design that separates factual quality from grader taste.

Key points

  • Personas combine eight MIMIC-IV attributes with a synthetic education category ranging from primary to PhD. Race, outcomes and clinician-side findings were excluded to prevent leakage.
  • The rubric covers Clinical Correctness, Faithfulness, Persona alignment and Realism. The first two carry the most weight, and on them GPT-5-mini outperformed GPT-4o-mini.
  • In the rubric-free test, the graders preferred opposite systems (both p < 0.001), and their agreement was very low (κ ≈ 0.07).
  • Claude Sonnet 3.7 favored reports with more sentences and longer sentences (OR/SD ≈ 2.46 and ≈ 2.03). GPT-4.1 showed almost no sensitivity to length.
  • Flesch–Kincaid grade rose steadily with assigned education for both generators, and GPT-5-mini wrote at a higher level throughout. The final dataset will therefore use GPT-4o-mini for personas up to high school and GPT-5-mini for college and above.

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: Linguistics & languages

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/34prj6o9qfrk

@inproceedings{wu2026persona,
  title = {{Persona-Conditioned Generation of Patient Self-Reports from EHRs}},
  author = {Wu, Yuexin and Wei, Jianming and Rus, Vasile},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
  year = {2026},
  doi = {10.63317/34prj6o9qfrk},
  url = {https://doi.org/10.63317/34prj6o9qfrk}
}

References

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

  1. Nancy D. Berkman, Stacey L. Sheridan, Katrina E. Donahue, David J. Halpern, and Karen Crotty. 2011a. Low health literacy and health outcomes: An updated systematic review. Annals of Internal Medicine, 155(2):97–107. find
  2. Nancy D. Berkman, Stacey L. Sheridan, Katrina E. Donahue, David J. Halpern, and Karin Crotty. 2011b. Low health literacy and health outcomes: An updated systematic review. Annals of Internal Medicine, 155(2):97–107. find
  3. John G. Canto, William J. Rogers, Robert J. Goldberg, Eric D. Peterson, Nanette K. Wenger, Viola Vaccarino, Catarina I. Kiefe, Peter D. Frederick, George Sopko, Zhiying Zheng, et al. 2012. Association of age and sex with myocardial infarction symptom presentation and in-hospital mortality. JAMA, 307… find
  4. John G. Canto, Michael G. Shlipak, William J. Rogers, Judith A. Malmgren, Catarina I. Kiefe, Harold V. Barron, Peter D. Frederick, Joseph P. Ornato, Christine M. Gibson, and William D. Weaver. 2000. Prevalence, clinical characteristics, and mortality among patients with myocardial infarction present… find
  5. Juan G. Canto, Emily A. Canto, and Robert J. Goldberg. 2007. Prevalence, clinical characteristics, and mortality among patients with myocardial infarction presenting without chest pain. JAMA, 297(7):813–816. find
  6. Ruihui Hou, Shencheng Chen, Yongqi Fan, Lifeng Zhu, Jing Sun, Jingping Liu, and Tong Ruan. 2024. Msdiagnosis: An emr-based dataset for clinical multi-step diagnosis. arXiv e-prints, pages arXiv–2408. find
  7. Mingyi Jia, Junwen Duan, Yan Song, and Jianxin Wang. 2025. medIKAL: Integrating knowledge graphs as assistants of LLMs for enhanced clinical diagnosis on EMRs. In Proceedings of the 31st International Conference on Computational Linguistics, pages 9278–9298, Abu Dhabi, UAE. Association for Computati… find
  8. Alistair E. W. Johnson, Lucas Bulgarelli, Lu Shen, Aaron Gayles, A. Shammout, Steven Horng, Tom J. Pollard, Leo A. Celi, and Roger G. Mark. 2023. Mimic-iv, a freely accessible electronic health record dataset. Scientific Data, 10(1):1– 10. find our card
  9. Alistair E. W. Johnson, Tom J. Pollard, Lu Shen, Li-Wei H. Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and Roger G. Mark. 2016. Mimic-iii, a freely accessible critical care database. Scientific Data, 3(1):160035. find our card
  10. Stefan Lautenbacher, Jan H. Peters, Michael Heesen, Jennifer Scheel, and Miriam Kunz. 2017a. Age changes in pain perception: A systematic review and meta-analysis of age effects on pain and tolerance thresholds. Neuroscience & Biobehavioral Reviews, 75:104–113. find
  11. Stefan Lautenbacher, Mette L Peters, Maarten Heesen, Julia Scheel, and Miriam Kunz. 2017b. Age changes in pain perception: A systematic-review and meta-analysis of age differences in pain thresholds and tolerance. Neuroscience & Biobehavioral Reviews, 75:104–113. find
  12. Hannah L Semigran, Jeffrey A Linder, Courtney Gidengil, and Ateev Mehrotra. 2015. Evaluation of symptom checkers for self diagnosis and triage: audit study. BMJ, 351:h3480. find
  13. Mary E Short, Ron Z Goetzel, Xue Pei, Mohannad J Tabrizi, Ronald J Ozminkowski, Teresa B Gibson, David M DeJoy, and Mark G Wilson. 2009a. How accurate are self-reports? an analysis of self-reported health care utilization and absenteeism. Journal of Occupational and Environmental Medicine, 51(7):786… find
  14. Mary Eastwood Short, Ron Z. Goetzel, Xiaofei Pei, Mohammad J. Tabrizi, and Ronald J. Ozminkowski. 2009b. How accurate are self-reports? analysis of self-reported health care utilization and absenteeism. Journal of Occupational and Environmental Medicine, 51(7):786– 796. find
  15. Zhixiang Su, Yinan Zhang, Jiazheng Jing, Jie Xiao, and Zhiqi Shen. 2024. Enabling patient-side disease prediction via the integration of patient narratives. In Companion Proceedings of the ACM Web Conference 2024, WWW ’24, page 581–584, New York, NY, USA. Association for Computing Machinery. find
  16. R. E. M. van Oosterhout, A. R. de Boer, A. H. E. M. Maas, F. H. Rutten, M. L. Bots, and S. A. E. Peters. 2020a. Sex differences in symptom presentation in acute coronary syndromes: A systematic review and meta-analysis. Journal of the American Heart Association, 9(9):e014733. find
  17. René E. M. van Oosterhout, Ariënne R. de Boer, Angela H. E. M. Maas, et al. 2020b. Sex differences in symptom presentation in acute coronary syndromes: A systematic review and meta-analysis. Journal of the American Heart Association, 9(13):e014733. find
  18. Evelyn Verlinde, Nele De Laender, Stéphanie De Maesschalck, Myriam Deveugele, and Sara Willems. 2012. The social gradient in doctor–patient communication. International Journal for Equity in Health, 11:12. find
  19. Matthew K Wynia and Chandra Y Osborn. 2010a. Health literacy and communication quality in health care organizations. Journal of Health Communication, 15(Suppl 2):102–115. find
  20. Matthew K. Wynia and Chandra Y. Osborn. 2010b. Health literacy and communication quality in health care organizations. Journal of Health Communication, 15(Suppl 2):102–115. find
  21. Guangtao Zeng, Wenmian Yang, Zeqian Ju, Yue Yang, Sicheng Wang, Ruisi Zhang, Meng Zhou, Jiaqi Zeng, Xiangyu Dong, Ruoyu Zhang, Hongchao Fang, Penghui Zhu, Shu Chen, and Pengtao Xie. 2020a. Meddialog: Large-scale medical dialogue datasets. In Proceedings of the 2020 Conference on Empirical Methods in… find
  22. Guangtao Zeng, Wenmian Yang, Pengtao Xie, et al. 2020b. Meddialog: Large-scale medical dialogue datasets. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9241–9250, Online. Association for Computational Linguistics. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wu-2026-persona-conditioned-generation-patient-self/