Evaluating large language model workflows in clinical decision support for triage and referral and diagnosis
Open accessPublished version of a preprint
Abstract
Accurate medical decision-making is critical for both patients and clinicians. Patients often struggle to interpret their symptoms, determine their severity, and select the right specialist. Simultaneously, clinicians face challenges in integrating complex patient data to make timely, accurate diagnoses. Recent advances in large language models (LLMs) offer the potential to bridge this gap by supporting decision-making for both patients and healthcare providers. In this study, we benchmark multiple LLM versions and an LLM-based workflow incorporating retrieval-augmented generation (RAG) on a curated dataset of 2000 medical cases derived from the Medical Information Mart for Intensive Care database. Our findings show that these LLMs are capable of providing personalized insights into likely diagnoses, suggesting appropriate specialists, and assessing urgent care needs. These models may also support clinicians in refining diagnoses and decision-making, offering a promising approach to improving patient outcomes and streamlining healthcare delivery.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
Key risks of monitoring agents include alert fatigue from overreporting, critical changes missed due to improperly calibrated thresholds, or failure to report emergencies.
“Key risks include alert fatigue from overreporting, critical changes missed due to improperly calibrated thresholds, or failure to report emergencies.”
From Medical AI Agents for Clinical Decision Support: Viewpoint Using the Planning, Action, Reflection, and Memory (PARM) Analytical Lens (2026) · cites this paper as [22]
The Introduction (page unavailable) says under-triage can delay treatment for patients needing immediate care, while over-triage wastes resources and increases waiting times. These support the general risks of missing urgent cases and excessive reporting, but the paper evaluates clinical triage, not monitoring agents; it does not address alert fatigue, threshold calibration, or failures to report emergencies. (LLM, conf=0.78)
Experimental — accuracy not yet measured
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
The citation lists haven’t been fetched yet
References 36 · Cited by 165
Research works printed but not linked: 28
- [1]Sutriningsih, A., Wahyuni, C. U. & Haksama, S. Factors affecting emergency nurses’ perceptions of the triage systems. J. Pub. Health Res. 9, 85–87 (2020). PMC7376454 32728554
- [2]Ma, M. D. et al. CliBench: Multifaceted evaluation of Large Language Models in clinical decisions on diagnoses, procedures, lab tests orders and prescriptions. arXiv10.48550/arXiv.2406.09923 (2024).
- [4]Abbas, A., Rehman, M. S. & Rehman, S. S. Comparing the performance of popular large language models on the National Board of Medical Examiners sample questions. Cureus 16, e55991 (2024). 38606229 PMC11007479
- [5]Brin, D. et al. How large language models perform on the United States medical licensing examination: A systematic review. medRxiv10.1101/2023.09.03.23294842 (2023)
- [7]Lupton, J. R. et al. Under-triage and over-triage using the Field Triage Guidelines for injured patients: A systematic review. Prehosp. Emerg. Care 27, 38–45 (2023). 35191799 · DOI
- [9]Mori, Y., Izumiyama, T., Kanabuchi, R., Mori, N. & Aizawa, T. Large language model may assist diagnosis of SAPHO syndrome by bone scintigraphy. Mod. Rheumatol. 34, 1043–1046 (2024). 38153762 · DOI
- [10]Kwon, T. et al. Large language models are clinical reasoners: Reasoning-aware diagnosis framework with prompt-generated rationales. ACM AAAI Conf. AI 38, 18417–18425 (2023).
- [11]Daher, M. et al. Breaking barriers: can ChatGPT compete with a shoulder and elbow specialist in diagnosis and management? JSES Int. 7, 2534–2541 (2023). 37969495 PMC10638599 · DOI
- [12]Madadi, Y. et al. ChatGPT assisting diagnosis of neuro-ophthalmology diseases based on case reports. JNO10.1097/WNO.000000000000227 (2023). PMC12341750 40830998
- [13]Delsoz, M. et al. Performance of ChatGPT in diagnosis of corneal eye diseases. Cornea 43, 664–670 (2023). 38391243 · DOI
- [14]Sorin, V. et al. GPT-4 multimodal analysis on ophthalmology clinical cases including text and images. medRxiv10.1101/2023.11.24.23298953 (2023)
- [16]Ueda, D. et al. Evaluating GPT-4-based ChatGPT’s clinical potential on the NEJM quiz. BMC Digit Health 2, 4 (2024). · DOI
- [17]Han, T. et al. Comparative analysis of GPT-4Vision, GPT-4 and open source LLMs in clinical diagnostic accuracy: A benchmark against human expertise. medRxiv10.1101/2023.11.03.23297957 (2023)
- [20]Johnson, A. et al.MIMIC-IV. PhysioNet 10.13026/HXP0-HG59 (2024). · DOI
- [27]Lewis, P. et al. Retrieval-augmented generation for knowledge-intensive NLP tasks. Adv. Neural Inf. Process. Syst.33, 9459–9474 (2020).
- [28]Kresevic, S. et al. Optimization of hepatological clinical guidelines interpretation by large language models: a retrieval augmented generation-based framework. NPJ Digit. Med. 7, 102 (2024). 38654102 PMC11039454 · DOI · Paper page
- [29]Shuster, K., Poff, S., Chen, M., Kiela, D. & Weston, J. Retrieval Augmentation Reduces Hallucination in Conversation. ACL Find. Assoc. Comput. Linguist.: EMNLP 2021, 3784–3803 (2021).
- [30]Ayala, O. & Bechard, P. Reducing hallucination in structured outputs via Retrieval-Augmented Generation. ACL Proc. Conf. North Am. Chapter Assoc. Comput. Linguist.: Hum. Lang. Technol. 6, 228–238 (2024).
- [31]Wang, D. et al. Enhancement of the performance of large language models in diabetes education through retrieval-augmented generation: Comparative study. J. Med. Internet Res. 26, e58041 (2024). 39046096 PMC11584532 · DOI · Paper page
- [32]Park, Y. J., Jerng, S. E., Yoon, S. & Li, J. 1.5 million materials narratives generated by chatbots. Sci. Data 11, 1060 (2024). 39341807 PMC11439064 · DOI
- [33]Khaliq, M. A., Chang, P. Y.-C., Ma, M., Pflugfelder, B. & Miletić, F. RAGAR, your falsehood radar: RAG-augmented reasoning for political fact-checking using multimodal large language models. ACL Proc. FEVER Workshop 7, 280–296 (2024).
- [37]Giray, L. Prompt engineering with ChatGPT: A guide for academic writers. Ann. Biomed. Eng. 51, 2629–2633 (2023). 37284994 · DOI · Paper page
- [38]Bansal, P. Prompt engineering importance and applicability with generative AI. J. Comput. Commun. 12, 14–23 (2024). · DOI · Paper page
- [39]Kojima, T., Gu, S. S., Reid, M., Matsuo, Y. & Iwasawa, Y. Large Language Models are Zero-Shot Reasoners. arXiv Acm. NIPS'22 38, 22199–22213 (2022).
- [40]Wang, L. et al. Prompt engineering in consistency and reliability with the evidence-based guideline for LLMs. NPJ Digit. Med. 7, 41 (2024). 38378899 PMC10879172 · DOI
- [43]Gao, M. et al. Human-like Summarization Evaluation with ChatGPT. arXiv10.48550/arXiv.2304.02554 (2023).
- [45]Zheng, L. et al. Judging LLM-as-a-judge with MT-bench and Chatbot Arena. arXiv Acm. NIPS'23 37, 46595–46623 (2023).
- [46]Davidson, T. R. et al. Self-Recognition in Language Models. ACL Find. Assoc. Comput. Linguist.: EMNLP 2024, 12032–12059 (2024).
- Authors
- Gaber, F., Shaik, M., Allega, F., Bilecz, A. J., Busch, F., Goon, K., Franke, V., & Akalin, A.
- Affiliations
- Max Delbrück Center (DE)
- Agostino Gemelli University Polyclinic (IT)
- University of Chicago (US)
- TUM Klinikum (DE)
- Corresponding author
- Akalin, Altuna
- Venue
- npj Digital Medicine
- Published
- 2025-05-09
- Volume
- 8
- Issue
- 1
- Pages
- 263-263
- ISSN
- 2398-6352 (ISSN-L)
- PMC
- PMC12064692
- OpenAlex
- W4410232646
- Cited by
- 165
- References
- 36
- Citations by year
- 2026: 125 · 2025: 40
- FWCI
- 39.72
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Locations
- npj Digital Medicine (publishedVersion, cc-by)
- medRxiv (acceptedVersion, cc-by-nc-nd)
- PubMed Central (publishedVersion, cc-by)
- PubMed (publishedVersion)
- MDC Repository (Max-Delbrueck-Center for Molecular Medicine) (submittedVersion)
- MDC Repository (Max-Delbrueck-Center for Molecular Medicine) (submittedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- mediaTUM – the media and publications repository of the Technical University Munich (Technical University Munich) (submittedVersion)
- Funders
- Helmholtz-Gemeinschaft (KA-TVP-60)
- Topics
- Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
- Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
- Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
- Keywords
- large language models, clinical decision support, triage, referral, diagnosis, retrieval-augmented generation, LLM benchmarking, MIMIC-III, medical information retrieval, expert selection
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
