Machine learning beats NEWS only for high-risk patients in Danish study of 2.08m stays

A machine learning early warning model added little over the standard National Early Warning Score (NEWS) for most hospital patients in Denmark, but identified substantially more deaths among those NEWS already rated high risk, according to a multicentre retrospective study of 2.08 million hospital encounters published in The Lancet Digital Health in November 2026¹.

What the study compared

The researchers used electronic health record data from 15 hospitals in ten hospital networks across Denmark's Capital and Zealand regions, covering adults with at least one NEWS assessment between 16 October 2018 and 16 October 2023¹. The data set held 2,080,765 encounters from 825,200 unique patients¹. The outcome was death from any cause within 24 hours of the first NEWS recording¹.

Four approaches were tested against each other: NEWS itself; a simplified NEWS that drops blood pressure and temperature; a Demographic Early Warning Score (DEWS) that adds age and sex to the simplified version; and XGB-EWS, an extreme gradient boosting model that combines vital signs, demographics and laboratory markers with embeddings of medical history extracted by sentence transformers¹. To deal with the bias created when clinicians act on a warning and so change the outcome, the team applied propensity score weighting, and judged the models on net benefit, the area under the receiver operating characteristic curve and calibration¹.

What it found

Across all patients the gains were small. At its maximum, XGB-EWS correctly identified 1.9 more deaths per 10,000 patients than NEWS, and NEWS beat the simplified score by 1.5 per 10,000¹.

The picture changed sharply by risk group. Among patients whose first NEWS was 7 or above, XGB-EWS delivered a maximum gain of 78.1 correct identifications per 10,000 patients over NEWS, while the simplified score lost up to 60.2 per 10,000¹. In the low-risk group, with NEWS of 0 to 4, the differences almost vanished: a gain of 0.9 per 10,000 for XGB-EWS and a loss of 0.1 for the simplified score¹.

What it means for the market

For hospitals weighing a machine learning deterioration tool against the score they already use, the study argues against a blanket replacement. Its authors conclude that model complexity brought small average gains, concentrated in patients NEWS already flags as high risk¹. That points to a tiered design: simpler scoring for lower-risk patients, with full NEWS or a model built on the electronic health record kept for those at higher risk¹.

For vendors of deterioration prediction software, the result narrows the case for their products to the sickest patients, where the gain per 10,000 was largest¹. Health systems that use NEWS as their common language for escalation now have large-scale European data on where an added model earns its place and where it does not.

Caveats

The study was retrospective and drew on two regions of one country¹. Weighting reduces but cannot remove the bias created by treatment given after a warning. The findings also rest on a single outcome, death within 24 hours¹. The Novo Nordisk Foundation funded the work¹. The authors call for prospective trials of monitoring stratified by the first NEWS assessment, testing simplified approaches in lower-risk patients, before any change in practice¹.