勘误:大语言模型作为强大的电子健康记录编码器研究中的数据与表述更正
Author Correction: Large language models are powerful electronic health record encoders
npj Digital Medicine 发表勘误,更正原刊《大语言模型作为强大的电子健康记录编码器》中的数值、排版及格式错误。修正包括更新 UK Biobank 访问量为 20,354,931 次、事件数为 72,910,130 次,修正 CLMBR-T-Base 模型名称及 DeBERTa v3 large 的注意力头数。
- Author Correction
- Open access
- Published: 14 October 2026
- Stefan Hegselmann1,2,3 na1,
- Georg von Arnim1,4 na1,
- Tillmann Rheude1,4,
- Noel Kronenberg1,
- David Sontag5,6,
- Gerhard Hindricks3,
- Roland Eils1,7 &
- …
- Benjamin Wild1,8
npj Digital Medicine volume 9, Article number: 790 (2026) Cite this article
Correction to: npj Digital Medicine https://doi.org/10.1038/s41746-026-02915-9, published online 06 July 2026
In the original version of this manuscript, there occur some errors consisting of numbers, typographical errors and formatting errors. The complete list of corrections is mentioned below. The original paper has now been corrected.
-
1.
Outdated UK Biobank numbers in Figure 1 c) and Table 1. Table 1 should contain the following revised entries for the UK Biobank:
UK Biobank No. of visits: 20,354,931 (instead of outdated 19,484,777)
UK Biobank No. of events: 72,910,130 (instead of outdated 72,265,684)
UK Biobank Black: 5715 (instead of wrong 5751)
-
2.
Wrong model name in Figure 3, legend. The legend entry reads CLIMBR-T-Base; it should read CLMBR-T-Base.
-
3.
Outdated UK Biobank number in Results, “Experimental setup”, p. 2.
Published: “The processed UKB subset used in our study comprised 387,464 patients, approximately 19.5 million visits, [...]”
Corrected: “The processed UKB subset used in our study comprised 387,464 patients, approximately 20.4 million visits, [...]”
-
4.
Wrong table row referred to in Results, “LLM embeddings achieve strong performance in low-data regimes”, p. 4. The sentence quotes the UK Biobank win–loss counts in a paragraph about EHRSHOT.
Published: “[...] outperformed CLMBR-T-Base on two versus zero tasks in the 8-shot setting and on four versus zero tasks in the 64-shot setting, while most tasks remained indistinguishable (Table 4).”
Corrected: “[...] outperformed CLMBR-T-Base on four versus one tasks in the 8-shot setting and on three versus two tasks in the 64-shot setting, while most tasks remained indistinguishable (Table 4).”
-
5.
Former result referred to in Results, “General-purpose LLM embeddings rival domain-specific EHR models”, p. 4. The sentence refers to the old BioClinicalBERT MEME results instead of the updated BioClinicalBERT.
Published: “[...] BioClinicalBERT with mean embeddings over 512-token chunks performed best at 0.702 (0.671–0.732) (Tables 3 and S4).”
Corrected: “[...] BioClinicalBERT with mean embeddings over 512-token chunks performed best at 0.705 (0.680–0.730) (Tables 3 and S4).”
-
6.
Transposed values in Supplementary Table S6. The chest X-ray values for the two “Counts + TB” rows are transposed between the GBM and LR blocks.
-
7.
Imprecise wording in Discussion, limitations, p. 9. Specified hyperparameter tuning and replaced wrong usage of “two-shot settings” where “two shot sizes” was meant.
Published: “In addition, our fine-tuning experiments were limited to two task groups and two-shot settings due [...]”
Corrected: “In addition, the hyperparameter tuning for our fine-tuning experiments was limited to two task groups and two shot sizes due [...]”
-
8.
Wrong number of attention heads stated for DeBERTa v3 large model in Methods, “LLM embedding models and baselines”, p. 11.
Published: “[...] large variant (24 layers, 12 attention heads, 1024 hidden size), with [...]”
Corrected: “[...] large variant (24 layers, 16 attention heads, 1024 hidden size), with [...]”
-
9.
Wrong use of “an encoder” instead of “a decoder” in Methods, “Existing methods using language models for EHR prediction”, p. 12.
Published: “They also evaluate an encoder-based approach producing Yes and No tokens [...]”
Corrected: “They also evaluate a decoder-based approach producing Yes and No tokens [...]”
-
10.
Outdated UK Biobank numbers in Methods, “External validation on UK Biobank” p.13.
Published: “[...] increasing the number of unique codes from 51,677 to 69,850, and [...]”
Corrected: “[...] increasing the number of unique codes from 50,702 to 71,393, and [...]”
-
11.
Wrong task name in Methods, “Statistical testing”, p. 13. Used task name “CheXpert” while task was introduced as “Chest X-ray findings”.
Published: “For the CheXpert task, which [...]”
Corrected: “For the chest X-ray findings task, which [...]”
-
12.
Missing space in the “Author contributions”, p. 16.
Published: “[...] and B.W.S.H. drafted the initial manuscript, [...]”
Corrected: “[...] and B.W. S.H. drafted the initial manuscript, [...]”
Author information
Author notes
These authors contributed equally: Stefan Hegselmann, Georg von Arnim.
Authors and Affiliations
Berlin Institute of Health at Charité – Universitätsmedizin Berlin, Center of Digital Health, Berlin, Germany
Stefan Hegselmann, Georg von Arnim, Tillmann Rheude, Noel Kronenberg, Roland Eils & Benjamin Wild
Berlin Institute of Health at Charité – Universitätsmedizin Berlin, BIH Biomedical Innovation Academy, BIH Charité Digital Clinician Scientist Program, Berlin, Germany
Stefan Hegselmann
Deutsches Herzzentrum der Charité (DHZC) – Universitätsmedizin Berlin, Berlin, Germany
Stefan Hegselmann & Gerhard Hindricks
Freie Universität Berlin, Berlin, Germany
Georg von Arnim & Tillmann Rheude
Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology (MIT), Cambridge, MA, USA
David Sontag
Layer Health Inc., Boston, MA, USA
David Sontag
Intelligent Medicine Institute, Fudan University, Shanghai, China
Roland Eils
Institute of Medical Informatics (IMI) – Universitätsmedizin Berlin, Berlin, Germany
Benjamin Wild
Authors
- Stefan Hegselmann
- Georg von Arnim
- Tillmann Rheude
- Noel Kronenberg
- David Sontag
- Gerhard Hindricks
- Roland Eils
- Benjamin Wild
Corresponding author
Correspondence to Stefan Hegselmann.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Hegselmann, S., von Arnim, G., Rheude, T. et al. Author Correction: Large language models are powerful electronic health record encoders. npj Digit. Med. 9, 790 (2026). https://doi.org/10.1038/s41746-026-03220-1
Published: 14 October 2026
Version of record: 14 October 2026
DOI: https://doi.org/10.1038/s41746-026-03220-1
来源:Nature 健康护理主题 · nature.com