Peer reviewed AI-assisted

Gradient-boosted early warning score tested for clinical net benefit across 2.08 million Danish hospital encounters

Across 2,080,765 encounters at 15 Danish hospitals, a gradient-boosted model using vital signs, laboratory values and text embeddings of medical history beat the National Early Warning Score for predicting death within 24 hours by 1.9 additional correct identifications per 10,000 patients overall, a gain that rose to 78.1 per 10,000 among patients the existing score already called high risk and fell to 0.9 per 10,000 among those it called low risk.

Model
XGB-EWS, an XGBoost model over vital signs, demographics, laboratory markers and sentence-transformer embeddings of medical history
Field
Medicine
Date
2026-11-05
Human collaborators
Alexandros Katsiferis, Nina Scheidwasser, Christian S. Meyhoff

What was found

Early warning scores are recalculated at the bedside across most hospital systems, so the question of whether a machine learning model is worth substituting is a practical one, and the paper's contribution is to ask it in terms of clinical net benefit rather than discrimination alone. The cohort covers every adult encounter with at least one score recorded at 15 hospitals in ten networks between 2018-10-16 and 2023-10-16, 825,200 unique patients in all, followed until death or 24 hours after the first score. The comparison runs in both directions, against a simplified score dropping blood pressure and temperature and a demographic variant as well as against the full machine learning model, with propensity score weighting used to blunt the bias introduced when staff intervene on a high score and prevent the death the score predicted. The headline finding is that model complexity bought very little on average and that the average hides the useful result: among patients whose first score was 7 or above the model added 78.1 correct identifications per 10,000, while the simplified score lost 60.2, and among those scoring 0 to 4 all three approaches were within one per 10,000 of each other. The authors read this as an argument for risk-stratified monitoring rather than for replacing the score outright.

Novelty check

Europe PMC, 2026-11-09: TITLE:("early warning score" OR "NEWS") AND ("machine learning" OR XGBoost OR "gradient boosting") AND ("net benefit" OR "decision curve") AND (mortality) returned 4 records, of which the closest is a 2024 single-site study of emergency admissions in BMJ Health and Care Informatics at a fraction of this scale; the others are systematic reviews and a prospective pancreatitis cohort. This registry holds three early warning entries already, 2022-07-21-trews-sepsis-early-warning, 2021-06-21-epic-sepsis-model-v1 and 2026-02-02-epic-sepsis-model-v2, all sepsis-specific and none measuring net benefit against the National Early Warning Score.

Caveats and known objections

Retrospective, and it clears this registry's bar on scale rather than design: nothing ran forward in time and no patient was treated differently because of the model. Validation is within the same 15-hospital system that built the model, so despite the multicentre framing this is not external validation in the sense used here. The reported advantage is a decision-curve quantity, not an outcome: the paper shows where a model would help if deployed, and the authors themselves conclude by calling for prospective trials. Propensity score weighting mitigates intervention bias on assumptions the data cannot check, since patients whose deterioration was successfully treated look like patients who were never going to deteriorate. The abstract gives no area under the curve or calibration figures alongside the net-benefit numbers. Funded by the Novo Nordisk Foundation. The Danish health record data behind it is not accessible to outside investigators, and no model weights are published. Autonomy is ai-assisted: the score is a bedside alerting aid, and in this study it changed nothing.

No independent check is recorded on this entry yet.

Reading the primary source closely enough to say whether it supports the claim counts as a check, and you are credited on the entry.

Or on GitHub: submit a check challenge the grade send a correction or send a pull request

Entry history (1 event)
  1. AddedEntered the registry graded Peer reviewed and AI-assisted.

Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.

Graded peer reviewed for verification and ai-assisted for autonomy. What these mean.

Cite this entry

Plain text
whataifound.org. (2026). Gradient-boosted early warning score tested for clinical net benefit across 2.08 million Danish hospital encounters. whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2026-11-05-xgb-ews-denmark
BibTeX
@misc{whataifound-universityofcopenhagenandcapitalandzealandregionhospitals-2026-denmark,
  title        = {Gradient-boosted early warning score tested for clinical net benefit across 2.08 million Danish hospital encounters},
  author       = {{whataifound.org}},
  year         = {2026},
  howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
  note         = {Result by University of Copenhagen and Capital and Zealand region hospitals. Verification: Peer reviewed. Autonomy: AI-assisted.},
  url          = {https://whataifound.org/finding/2026-11-05-xgb-ews-denmark}
}

Related findings

← All medicine findings in the registry