From promising prototypes to “instructions for use”: embedding LLMs safely in perioperative and intensive care
Open accessThe DOI matches its Crossref record; 6 of the 6 references it cites were found in a registry or an index; 13 of the 13 sources its claims cite were checked for support.
Abstract
Large language models (LLMs) show promise for supporting clinical decision-making in perioperative and intensive care settings... We argue that responsible deployment of LLMs in ICU and perioperative workflows must clarify: (1) intended clinical scope and non-indications; (2) role in the decision-making hierarchy and when clinicians should override model recommendations; and (3) mechanisms for transparency, governance, and staff training.
Overview
CoChat Summary
## TL;DR
This paper argues that large language models (LLMs) in ICU and perioperative care should be deployed only with explicit, operational “instructions for use” covering scope, decision authority, transparency, governance, and training. It uses a preoperative triage study by Xu et al. as a template and cautionary example, emphasizing that combining structured and unstructured data improved performance but also risked reproducing local practice biases.
## Key findings
- Xu et al.’s preoperative triage model used 1,761 patients and 46,922 clinical notes from a single centre to triage surgical patients to either a nursing-led phone visit or an in-person/telehealth evaluation.
- Across four transformer-based models (BERT, RoBERTa, ClinicalBERT, PubMedBERT), performance was best when structured diagnosis codes and unstructured notes were combined.
- With combined data, RoBERTa achieved an AUROC of about 0.70–0.74 and F1-scores around 0.75–0.77.
- A general-purpose model outperformed domain-specific variants, and the model appeared to learn local triage practice standards, highlighting both utility and the risk of embedding local biases.
## Methods
- This is a commentary/perspective that synthesizes lessons from Xu et al. and an AI policy checklist to propose ICU/perioperative “instructions for use” for LLMs.
- The paper’s evidence base is primarily a narrative interpretation of the Xu et al. triage study; it does not present new empirical data of its own.
- It organizes recommendations around intended use/scope, decision-making hierarchy, and transparency of model behavior, and discusses governance, training, and monitoring.
## Limitations
- The paper does not report original empirical evaluation of ICU or perioperative LLM deployment; this is stated in the text by its reliance on a prior study and policy framework rather than new data.
- The authors explicitly note that preoperative triage is retrospective and lower-stakes than real-time ICU decisions, so transferability is limited by the difference in setting and risk.
- Evident from the text: the article is normative and conceptual, so its recommendations are not validated by prospective clinical outcome data in ICU settings.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
12 Valid · 0 Not valid · 0 Not checked · 5 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“The architecture described by Xu et al.—a classifier integrating diagnosis codes and summarised clinical notes—is readily transferable to ICU-related tasks such as admission and transfer triage, alert prioritisation for sepsis or respiratory failure, and routing of complex consults [1].”
§ From promising prototypes to “instructions for use”: embedding LLMs safely in perioperative and intensive care > Lessons from preoperative triage for ICU applications of LLMs
What the sources say
- Partly supportedcites [1] Xu NY, Litake O, Tully JL et al. A pre-trained Language model approach for triaging surgical patients for preoperative anesthesia clinics. J Clin Monit Comput 2025 Dec 24. https://doi.org/10.1007/s10877-025-01401-z sourceExperimental — accuracy not yet measured
The Methods section (Data pre-processing of clinical notes; page unavailable) describes summarising clinical notes, and the Discussion (page unavailable) confirms that the classifiers combine notes with diagnosis-code descriptions for preoperative triage. The paper does not address ICU admission or transfer triage, sepsis or respiratory-failure alerts, or consult routing; its Discussion instead says external validation is needed to assess generalisability. It therefore does not support the claim that the architecture is readily transferable to those tasks. (LLM, conf=0.88)
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
“A general-purpose model outperformed domain-specific variants”
§ From promising prototypes to “instructions for use”: embedding LLMs safely in perioperative and intensive care > Abstract
What the sources say
- Supportedcites [1] Xu NY, Litake O, Tully JL et al. A pre-trained Language model approach for triaging surgical patients for preoperative anesthesia clinics. J Clin Monit Comput 2025 Dec 24. https://doi.org/10.1007/s10877-025-01401-z source
The Methods identify RoBERTa as a general-purpose model and ClinicalBERT and PubMedBERT as domain-specific models. The Results (Performance of pre-trained language model-based classifiers; page not provided) report that RoBERTa had the highest AUC of all four models for each input combination. (LLM, conf=0.92)
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.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“The Xu et al. classifier reproduces an institutional triage standard [1]”
§ From promising prototypes to “instructions for use”: embedding LLMs safely in perioperative and intensive care > Lessons from preoperative triage for ICU applications of LLMs
What the sources say
- Partly supportedcites [1] Xu NY, Litake O, Tully JL et al. A pre-trained Language model approach for triaging surgical patients for preoperative anesthesia clinics. J Clin Monit Comput 2025 Dec 24. https://doi.org/10.1007/s10877-025-01401-z sourceExperimental — accuracy not yet measured
The Methods (Study population and outcomes assignment; page unavailable) describe institution-specific triage criteria and committee-assigned outcomes, and the Discussion says the models may learn institutional practice patterns. But the Results (Performance of pre-trained language model-based classifiers; page unavailable) report imperfect classification, with a best AUC of 0.74. The paper therefore supports approximating the institution’s triage decisions, not fully reproducing its standard. (LLM, conf=0.82)
plainstates a cause- 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
“models that combine structured and unstructured information outperform those based on diagnosis codes alone”
§ From promising prototypes to “instructions for use”: embedding LLMs safely in perioperative and intensive care > Lessons from preoperative triage for ICU applications of LLMs
What the sources say
- Supportedcites [1] Xu NY, Litake O, Tully JL et al. A pre-trained Language model approach for triaging surgical patients for preoperative anesthesia clinics. J Clin Monit Comput 2025 Dec 24. https://doi.org/10.1007/s10877-025-01401-z source
The Results report higher AUC and F1-score ranges for models using diagnosis codes plus clinical notes than for models using diagnosis codes alone. The Discussion explicitly describes improved classification with the combined inputs in this study of preoperative triaging. (LLM, conf=0.93)
- Could not verifycites [2] Thissen A, Salgado-Baez E, Fürstenau D, et al. Barriers and facilitators of the use of computerized critical care information systems in the intensive care unit: qualitative interview study. J Med Internet Res. 2025;27:e49254. https://doi.org/10.2196/49254. sourceExperimental — accuracy not yet measured
The Results (Suitability for the Task; page unavailable) discuss free-text entry and synchronizing information across systems, not predictive-model performance. Searches for structured and unstructured information, diagnosis codes, and model comparisons found no passage supporting the claimed outperformance. (LLM, conf=0.76)
plainstates a cause- 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.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“with potential to entrench inequities if not carefully governed [4]”
§ From promising prototypes to “instructions for use”: embedding LLMs safely in perioperative and intensive care > Lessons from preoperative triage for ICU applications of LLMs
What the sources say
- Supportedcites [4] Shi T, Ma J, Yu Z, et al. Large Language models in critical care medicine: scoping review. JMIR Med Inf. 2025;13:e76326. https://doi.org/10.2196/76326. source
The Discussion’s “Bias and Alignment Challenges” section says LLM biases in critical care can disproportionately affect patient populations and lead to disparities in treatment and care. Its “Mitigating Bias and Ensuring Fairness” section calls for measures to detect and reduce those biases. Together, these support the paraphrase that ICU tools could entrench inequities without careful oversight. (LLM, conf=0.92)
hedgedstates a possible cause- Asserted: passes — The authors put it forward as true.
- Checkable: unclear — It is unclear whether it could be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
- Supported
5 more claims · 1 with their cited sources checked
Create a free account to see all 17 claims and every source verdict.
- Probably not a substantial claim. 3
- It may state two separate claims as one. 5
Citations
Cites
de Hond, A. A. H., Leeuwenberg, A. M., Hooft, L., Kant, I. M. J., Nijman, S. W. J., van Os, H. J. A., Aardoom, J. J., Debray, T. P. A., Schuit, E., van Smeden, M., et al. (2022) · npj Digital Medicine · doi:10.1038/s41746-021-00549-7 · Cited by 499
AI policy in healthcare: a checklist-based methodology for structured implementation
Bignami, E., Darhour, L. J., Franco, G., Guarnieri, M., & Bellini, V. (2025) · Journal of Anesthesia, Analgesia and Critical Care · doi:10.1186/s44158-025-00278-3 · Cited by 23
Self-citationLarge Language Models in Critical Care Medicine: Scoping Review
Shi, T., Ma, J., Yu, Z., Xu, H., Yang, R., Xiong, M., Xiao, M., Li, Y., Zhao, H., & Kong, G. (2025) · JMIR Medical Informatics · doi:10.2196/76326 · Cited by 21 · cited 1×
“The Xu et al. classifier reproduces an institutional triage standard [1], and similar tools in ICU will learn from local thresholds for escalation, end-of-life decisions, and resource constraints, with potential to entrench inequities if not carefully gov - erned [4].”
Żerdziński, K., Janiec, J., Jóźwik, K., Łajczak, P., & Krzych, Ł. (2025) · Journal of Clinical Medicine · doi:10.3390/jcm15010185 · Cited by 11
Thissen, A., Salgado-Baez, E., Fürstenau, D., Delucchi Danhier, R., Meske, C., Kuss, P., Angermair, S., Spies, C., & Balzer, F. (2025) · Journal of Medical Internet Research · doi:10.2196/49254 · Cited by 7
Xu, N. Y., Litake, O., Tully, J. L., Meineke, M. N., Sinha, A., Meyer, M., & Gabriel, R. A. (2025) · Journal of Clinical Monitoring and Computing · doi:10.1007/s10877-025-01401-z · Cited by 2 · cited 10×
“In Xu et al., preoperative triage is per - formed by nursing staff and the model is intended to reduce workload while preserving oversight [1].”
Influential
Metadata
- Authors
- Bignami, E. G., & Russo, M.
- Affiliations
- University of Parma (IT)
- Corresponding author
- Bignami, Elena Giovanna
- Venue
- Journal of Clinical Monitoring and Computing
- Published
- 2026-02-04
- Volume
- 40
- Issue
- 2
- Pages
- 301-304
- ISSN
- 1387-1307 (ISSN-L)
- 1573-2614
- PMC
- PMC13053538
- OpenAlex
- W7127645522
- Cited by
- 0
- References
- 6
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by-nc-nd
- creativecommons.org/licenses/by-nc-nd/4.0 (tdm)
- creativecommons.org/licenses/by-nc-nd/4.0 (vor)
- Locations
- Journal of Clinical Monitoring and Computing (publishedVersion, cc-by-nc-nd)
- PubMed Central (publishedVersion, cc-by-nc-nd)
- PubMed (publishedVersion)
- Topics
- Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
- Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Keywords
- large language models, clinical decision support, perioperative care, intensive care, instructions for use, AI governance, clinical workflow integration, transparency, employee training, human-in-the-loop decision making, responsible AI deployment
- MeSH terms
- Clinical Decision-Making, Large Language Models, Intelligent Systems, Artificial Intelligence, Critical Care, Humans, Intensive Care Units, Triage, Perioperative Care, Decision Support Systems, Clinical, Workflow
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: From promising prototypes to “instructions for use”: embedding LLMs safely in perioperative and intensive care. https://app.cochat.ai/papers/10.1007/s10877-026-01418-y
- The paper
- Bignami, E. G., & Russo, M. (2026). From promising prototypes to “instructions for use”: embedding LLMs safely in perioperative and intensive care. Journal of Clinical Monitoring and Computing. https://doi.org/10.1007/s10877-026-01418-y
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.