The DOI matches its Crossref record; 99 of the 99 references it cites were found in a registry or an index; 25 of the 36 sources its claims cite were checked for support.
The shaky foundations of large language models and foundation models for electronic health records
Open accessAbstract
The success of foundation models such as ChatGPT and AlphaFold has spurred significant interest in building similar models for electronic medical records (EMRs) to improve patient care and hospital operations. However, recent hype has obscured critical gaps in our understanding of these models' capabilities. In this narrative review, we examine 84 foundation models trained on non-imaging EMR data (i.e., clinical text and/or structured data) and create a taxonomy delineating their architectures, training data, and potential use cases. We find that most models are trained on small, narrowly-scoped clinical datasets (e.g., MIMIC-III) or broad, public biomedical corpora (e.g., PubMed) and are evaluated on tasks that do not provide meaningful insights on their usefulness to health systems. Considering these findings, we propose an improved evaluation framework for measuring the benefits of clinical foundation models that is more closely grounded to metrics that matter in healthcare.
Overview
CoChat Summary
## TL;DR
This narrative review examines 84 clinical foundation models for non-imaging electronic health record data and argues that current evaluations mostly measure predictive accuracy rather than the healthcare value these models are claimed to provide. The authors propose that clinical FMs need new evaluation tasks, metrics, and datasets tied more directly to clinical utility.
## Key findings
- The review identified 84 distinct clinical FMs published before March 1, 2023: 50 CLaMs and 34 FEMRs.
- Nearly all CLaMs were trained on just two data sources: MIMIC-III and PubMed; the paper states these two corpora contain about 18.5 billion words combined and that MIMIC-III includes about 2 million clinical notes.
- Most CLaMs are evaluated on standard NLP tasks such as named entity recognition, relation extraction, and document classification, and roughly half were not validated on clinical text.
- For both CLaMs and FEMRs, the paper says evaluations mostly show better predictive accuracy, while other claimed FM benefits—such as simpler deployment, less labeled data, novel human-AI interfaces, and emergent clinical applications—are largely unstudied.
## Methods
- Design: Narrative review and taxonomy-building paper focused on clinical foundation models for EMR data, excluding images, genetics, and wearables.
- Data/sample: The authors curated 84 models published before March 1, 2023, including 50 clinical language models (CLaMs) and 34 foundation models for EMRs (FEMRs), identified through citations from representative samples and manual article curation.
- Main analysis: They summarized each model’s architecture, training data, accessibility, and evaluation tasks, then reorganized evaluations around six claimed FM benefits to assess whether current benchmarks support those claims.
## Limitations
- The authors explicitly state they do not claim to include every possible model or every recent advancement, and that their search was not systematic; this is a limitation stated in the paper.
- They also note the review excludes imaging data and therefore is scoped only to structured and unstructured EMR data; this is stated in the paper as a scope choice.
- Evident from the text: because the paper is a narrative review with curated examples, its conclusions depend on the selected models and evaluations rather than a systematic search or meta-analysis.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Figures




Verification
Claims the paper makes
12 Valid · 8 Not valid · 7 Not checked · 27 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Clinical FMs lack the shared evaluation frameworks and datasets20 that have underpinned progress in other fields, such as natural language processing (NLP) and computer vision21.”
§ The shaky foundations of large language models and foundation models for electronic health records > Introduction
What the sources say
- Partly supportedcites [20] Blagec, K., Kraiger, J., Frühwirt, W. & Samwald, M. Benchmark datasets driving artificial intelligence development fail to capture the needs of medical professionals. J. Biomed. Inform. 137, 104274 (2023). sourceExperimental — accuracy not yet measured
The Introduction describes public benchmarks as shared, standardized ways to compare AI models, and the Analysis of benchmark datasets finds that few benchmarks have direct clinical relevance. But the paper examines clinical and biomedical AI benchmarks, not clinical foundation models specifically; it also identifies some existing clinical benchmarks rather than a complete lack of them. Page numbers were not available in our copy. (LLM, conf=0.86)
- Not checked — no full text of the sourcecites [21] Donoho, D. 50 years of data science. J. Comput. Graph. Stat. 26, 745–766 (2017). source
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Almost all CLaMs trained on the clinical text used a single database: MIMIC-III, which contains approximately 2 million notes written between 2001–2012 in the ICU of the Beth Israel Deaconess Medical Center56.”
§ The shaky foundations of large language models and foundation models for electronic health records > State of published clinical FMs > CLaMs
What the sources say
- Partly supportedcites [56] Johnson, A. E. W. et al. MIMIC-III, a freely accessible critical care database. Sci. Data 3, 160035 (2016). sourceExperimental — accuracy not yet measured
The Background & Summary identifies MIMIC-III as clinical data from Beth Israel Deaconess Medical Center, and the Data Records and Additional information sections describe its notes. Patient characteristics gives 2001–2012 for adult critical-care admissions, but the passages do not establish approximately 2 million notes or that all notes were written in the ICU during that period. The statement about how many CLaMs used MIMIC-III is the citing authors’ framing, not a finding attributed to this database description. (LLM, conf=0.85)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Other public datasets vary greatly in size, from eICU’s 139,000 patients61 to the CPRD’s longitudinal records on 7% of all patients in the UK62.”
§ The shaky foundations of large language models and foundation models for electronic health records > State of published clinical FMs > FEMRs
What the sources say
- Supportedcites [61] Pollard, T. J. et al. The eICU Collaborative Research Database, a freely available multi-center database for critical care research. Sci. Data 5, 180178 (2018). source
The Data Records section reports 139,367 unique patients in eICU, supporting the citing sentence’s rounded figure of 139,000. The CPRD figure is attributed to a different citation and is not assessed here; no page number is provided for the passage. (LLM, conf=0.95)
- Supportedcites [62] Li, Y. et al. Hi-BEHRT: hierarchical transformer-based model for accurate prediction of clinical events using multimodal longitudinal electronic health records. IEEE J. Biomed. Health Inform. 27 (2022). source
The citation to this paper is attached to the CPRD portion of the sentence, not the eICU figure. In Methods, “Data Source and Cohort Selection,” the paper describes CPRD as a longitudinal EHR dataset containing data from approximately 7% of the U.K. population. (LLM, conf=0.94)
plainn = 139,000- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“In reality, clinical workflows are capacity constrained—a nursing team may only be able to act on a handful of model predictions per day79,80.”
§ The shaky foundations of large language models and foundation models for electronic health records > State of published clinical FMs > FEMRs
What the sources say
- Not checked — no full text of the sourcecites [79] Shah, N. Making machine learning models clinically useful. JAMA 322, 1351 (2019). source
- Supportedcites [80] Wornow, M., Gyang Ross, E., Callahan, A. & Shah, N. H. APLUS: a Python library for usefulness simulations of machine learning models in healthcare. J. Biomed. Inform. 139, 104319 (2023). source
The Introduction says that resource constraints affect clinical workflows. In Results (“Simulating unlimited downstream specialist capacity”), the paper describes low-capacity nursing settings where only the top few model predictions can be acted upon, including a simulated capacity of three patients per day. (LLM, conf=0.92)
hedged- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“Currently, the majority of evaluation tasks span one data modality78”
§ The shaky foundations of large language models and foundation models for electronic health records > State of published clinical FMs > FEMRs
What the sources say
- Could not verifycites [78] McDermott, M. et al. A comprehensive EHR timeseries pre-training benchmark. In Proc. Conference on Health, Inference, and Learning 257–278 (Association for Computing Machinery, 2021). sourceExperimental — accuracy not yet measured
The High-Level Overview (Section 4; page unavailable) describes this paper’s ten downstream tasks using EHR time-series data. Neither it nor the task descriptions in Section 5.2 establish that the majority of evaluation tasks generally span one data modality; searches for single-modality and multimodal evaluation tasks found no such statement. (LLM, conf=0.73)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Could not verify
30 more claims · 4 with their cited sources checked
Create a free account to see all 54 claims and every source verdict.
- It may state two separate claims as one. 16
- Probably not a substantial claim. 8
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?
The cited source says what the claim attributes to it.
In the citing paper
“Previous work has demonstrated potential across clinical NLP tasks22,23, adapting to the medical domain by training a new model24,25, fine-tuning an existing model26,27 or supplying task-specific examples in the model prompt27,28.”
From Adapted large language models can outperform medical experts in clinical text summarization (2024) · cites this paper as [22]
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“Third, many state-of-the-art models are inaccessible13, necessitating the development of effective open-source models for biomedicine.”
From A clinically accessible small multimodal radiology model and evaluation metric for chest X-ray findings (2025) · cites this paper as [13]
The check wasn’t confident either way, so it gives no verdict — never disputed.
In the citing paper
“Some LLMs have already been integrated into electronic medical record systems,174 such as the Nuance DAX Copilot, a generative AI-program embedded in an electronic health record software called Epic, which records patient interactions and automatically produces clinical documentation.175,176”
From Large Language Models for Chatbot Health Advice Studies: A Systematic Review (2025) · cites this paper as [174]
Experimental — accuracy not yet measuredThe check wasn’t confident either way, so it gives no verdict — never disputed.
In the citing paper
“Smaller models incur smaller carbon footprint12 and offer reduced serving costs and latency, which is of particular importance in resource-constrained settings outside of data centers13.”
From A clinically accessible small multimodal radiology model and evaluation metric for chest X-ray findings (2025) · cites this paper as [13]
Experimental — accuracy not yet measuredIn the citing paper
“Clinical notes summarization assists healthcare practitioners in identifying potential health risks within patients' electronic health records (Wornow et al. 2023), thereby reducing errors and facilitating informed decision-making.”
From Large language models in medical and healthcare fields: applications, advances, and challenges (2024) · cites this paper as Wornow et al. 2023
Citations
Cites
Highly accurate protein structure prediction with AlphaFold
Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Žídek, A., Potapenko, A., et al. (2021) · Nature · doi:10.1038/s41586-021-03819-2 · Cited by 43,764 · cited 2×
“FMs have received signi fi cant attention given their impressive range of capabilities across multiple domains, from text generation 2 and video editing 3 to protein folding 4 and robotics 5 .”
BackgroundLanguage Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372
Learning Transferable Visual Models From Natural Language Supervision
Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. (2021) · arXiv · doi:10.48550/arxiv.2103.00020 · Cited by 15,293 · cited 1×
“Substantial prior work has shown that these models ’ large parameter counts and dataset sizes enable them to effectively model disparate modalities in the same shared latent space, thereby deriving richer representations for each modality than possible with unimodal models 48 – 51 .”
BackgroundTransformers: State-of-the-Art Natural Language Processing
Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., et al. (2020) · Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations · doi:10.18653/v1/2020.emnlp-demos.6 · Cited by 8,741
MIMIC-III, a freely accessible critical care database
Johnson, A. E. W., Pollard, T. J., Shen, L., Lehman, L.-W. H., Feng, M., Ghassemi, M., Moody, B., Szolovits, P., Anthony Celi, L., & Mark, R. G. (2016) · Scientific Data · doi:10.1038/sdata.2016.35 · Cited by 8,141 · cited 2×
“Again, the most popular public dataset is MIMIC-III, which contains less than 40,000 patients 56 .”
BackgroundMethodChain-Of-Thought Prompting Elicits Reasoning in Large Language Models
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q. V., & Zhou, D. (2022) · Advances in Neural Information Processing Systems 35 · doi:10.52202/068431-1800 · Cited by 4,505
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., & Zhou, D. (2022) · arXiv · doi:10.48550/arxiv.2201.11903 · Cited by 4,263 · cited 1×
“This enables a two-way conversation between humans and machine, and allows for the decomposition of problems into smaller steps via techniques such as “ chain-of-thought ” prompting 54 .”
MethodKung, T. H., Cheatham, M., Medenilla, A., Sillos, C., De Leon, L., Elepaño, C., Madriaga, M., Aggabao, R., Diaz-Candido, G., Maningo, J., et al. (2023) · PLOS Digital Health · doi:10.1371/journal.pdig.0000198 · Cited by 4,115 · cited 1×
“Physicians were quick to apply the model to pass medical licensing exams 8 – 11 , simplify radiology reports 12 , and write research articles 13 .”
MethodKung, T. H., Cheatham, M., ChatGPT, Medenilla, A., Sillos, C., De Leon, L., Elepaño, C., Madriaga, M., Aggabao, R., Diaz-Candido, G., et al. (2022) · doi:10.1101/2022.12.19.22283643 · Cited by 4,115
Universal Language Model Fine-tuning for Text Classification
Howard, J., & Ruder, S. (2018) · Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/p18-1031 · Cited by 4,100
Research works printed but not linked: 13
- Lehman, E. et al. Do we still need clinical language models? Preprint at arXiv [cs.CL] (2023).
- Moradi, M., Blagec, K., Haberl, F. & Samwald, M. GPT-3 models are poor few-shot learners in the biomedical domain. Preprint at arXiv [cs.CL] (2021).
- Steinberg, E., Xu, Y., Fries, J. & Shah, N. Self-supervised time-to-event modeling with structured medical records. Preprint at arXiv [cs.LG] (2023).
- Lee, S., Da Young, L., Im, S., Kim, N. H. & Park, S.-M. Clinical decision transformer: intended treatment recommendation through goal prompting. Preprint at arXiv [cs.AI] (2023).
- Sushil, M., Ludwig, D., Butte, A. J. & Rudrapatna, V. A. Developing a general-purpose clinical language inference model from a large corpus of clinical notes. Preprint at arXiv [cs.CL] (2022).
- Dash, D. et al. Evaluation of GPT-3.5 and GPT-4 for supporting real-world information needs in healthcare delivery. Preprint at arXiv [cs.AI] (2023).
- Thomas, R. & Uminsky, D. The problem with metrics is a fundamental problem for AI. Preprint at arXiv [cs.CY] (2020).
- Lee, P., Goldberg, C. & Kohane, I. The AI Revolution in Medicine: GPT-4 and Beyond. (Pearson, 2023).
- Husmann, S., Yèche, H., Rätsch, G. & Kuznetsova, R. On the importance of clinical notes in multi-modal learning for EHR data. Preprint at arXiv [cs.LG] (2022).
- Peng, S., Kalliamvakou, E., Cihon, P. & Demirer, M. The impact of AI on developer productivity: evidence from GitHub copilot. Preprint at arXiv [cs.SE] (2023).
- Zhang, X., Zhou, Z., Chen, D. & Wang, Y. E. AutoDistill: an end-to-end framework to explore and distill hardware-efficient language models. Preprint at arXiv [cs.LG] (2022).
- El-Mhamdi, E.-M. et al. SoK: on the impossible security of very large foundation models. Preprint at arXiv [cs.LG] (2022).
- Sharir, O., Peleg, B. & Shoham, Y. The cost of training NLP models: a concise overview. Preprint at arXiv [cs.CL] (2020).
Cited by
Adapted large language models can outperform medical experts in clinical text summarization
Van Veen, D., Van Uden, C., Blankemeier, L., Delbrouck, J.-B., Aali, A., Bluethgen, C., Pareek, A., Polacin, M., Reis, E. P., Seehofnerová, A., et al. (2024) · Nature Medicine · doi:10.1038/s41591-024-02855-5 · Cited by 891
“Previous work has demonstrated potential across clinical NLP tasks 22,23 , adapting to the medical domain by training a new model 24,25 , fine-tuning an existing model 26,27 or supplying task-specific examples in the model prompt 27,28 .”
Bedi, S., Liu, Y., Orr-Ewing, L., Dash, D., Koyejo, S., Callahan, A., Fries, J. A., Wornow, M., Swaminathan, A., Lehmann, L. S., et al. (2024) · doi:10.1101/2024.04.15.24305869 · Cited by 738
Self-citationTesting and Evaluation of Health Care Applications of Large Language Models: A Systematic Review
Bedi, S., Liu, Y., Orr-Ewing, L., Dash, D., Koyejo, S., Callahan, A., Fries, J. A., Wornow, M., Swaminathan, A., Lehmann, L. S., et al. (2025) · JAMA · doi:10.1001/jama.2024.21700 · Cited by 662
Self-citationThe TRIPOD-LLM reporting guideline for studies using large language models
Gallifant, J., Afshar, M., Ameen, S., Aphinyanaphongs, Y., Chen, S., Cacciamani, G., Demner-Fushman, D., Dligach, D., Daneshjou, R., Fernandes, C., et al. (2025) · Nature Medicine · doi:10.1038/s41591-024-03425-5 · Cited by 615
Creation and Adoption of Large Language Models in Medicine
Shah, N. H., Entwistle, D., & Pfeffer, M. A. (2023) · JAMA · doi:10.1001/jama.2023.14217 · Cited by 435
Self-citation
Metadata
- Authors
- Wornow, M., Xu, Y., Thapa, R., Patel, B., Steinberg, E., Fleming, S., Pfeffer, M. A., Fries, J., & Shah, N. H.
- Affiliations
- Stanford University (US)
- Stanford Medicine (US)
- Stanford Health Care (US)
- Corresponding author
- Wornow, Michael
- ORCID
- Venue
- npj Digital Medicine
- Published
- 2023-07-29
- Volume
- 6
- Issue
- 1
- Pages
- 135-135
- ISSN
- 2398-6352 (ISSN-L)
- PMC
- PMC10387101
- OpenAlex
- W4385381606
- Cited by
- 324
- References
- 87
- Citations by year
- 2026: 97 · 2025: 110 · 2024: 103 · 2023: 14
- FWCI
- 6.77
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0 (tdm)
- creativecommons.org/licenses/by/4.0 (vor)
- Locations
- npj Digital Medicine (publishedVersion, cc-by)
- PubMed Central (publishedVersion, cc-by)
- PubMed (publishedVersion)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Funders
- National Science Foundation
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
- Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
- Health Sciences › Health Professions › Health Information Management › Electronic Health Records Systems
- Keywords
- foundation models, large language models, electronic health records, model evaluation framework, clinical text, structured clinical data, MIMIC-III, PubMed, artificial intelligence in healthcare, model taxonomy
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.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
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
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: The shaky foundations of large language models and foundation models for electronic health records. https://app.cochat.ai/papers/10.1038/s41746-023-00879-8
- The paper
- Wornow, M., Xu, Y., Thapa, R., Patel, B., Steinberg, E., Fleming, S., Pfeffer, M. A., Fries, J., & Shah, N. H. (2023). The shaky foundations of large language models and foundation models for electronic health records. npj Digital Medicine. https://doi.org/10.1038/s41746-023-00879-8
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.