Source: Unite.AI
Wiring a widely deployed hospital prediction model to automatic pages for critical-care teams was associated with an 18% reduction in the risk-adjusted odds of in-hospital death, according to a study of 23,132 high-risk patients published in NEJM AI on July 29, 2026. Researchers at Rutgers Robert Wood Johnson Medical School and RWJBarnabas Health tracked the rollout across 11 New Jersey hospitals, from an academic medical center down to community non-teaching facilities.
The model is the Epic Deterioration Index, a proprietary machine-learning score that Epic ships inside its electronic health record. It reads vital signs, lab results, nursing assessments and age already sitting in the chart and returns a risk number on a 0-to-100 scale, which the health system’s own announcement says recalculates every 15 minutes. What RWJBarnabas changed was the routing: once a patient crossed the highest-risk threshold, the record pushed a notification straight to the hospital’s rapid response team.
What the numbers show
The team ran a quasi-experimental, staggered pre- versus post-implementation comparison rather than a randomized trial. It covered adult medical-surgical admissions scoring 60 or above on the index between October 1, 2022 and August 30, 2024: 10,803 patients before the change and 12,329 after, with a mean age of 71.9 years. Roughly 47% of post-implementation encounters generated a page to the response team, though not every page produced an activation.
Rapid response activations among these patients rose from 25.3% of hospital stays to 37.5%. Unadjusted in-hospital mortality fell from 23.1% to 18.6%. After adjusting for age, comorbidities, hospital type, index score and clustering by hospital, the odds of dying in the hospital were 18% lower in the post-implementation group, an adjusted odds ratio of 0.82. Escalations of care held near 1% in both periods, so the extra response-team traffic did not translate into a wave of new intensive-care transfers.
How the model has fared in head-to-head tests
Hospital risk-prediction models are usually judged on how well they separate patients who deteriorate from patients who don’t, and that literature has been rough on this particular score. In the largest published analysis of the Epic index to date, researchers at Yale New Haven Health and the University of Chicago scored six early warning systems against the same 362,926 patient encounters across seven hospitals and placed it near the bottom, with an area under the curve of 0.808. That trailed the National Early Warning Score at 0.829, a points system a clinician can add up by hand, and sat well behind the machine-learning eCART model at 0.895.
Lead time was the sharper gap. The Epic score’s high-risk alerts arrived a median of one hour before deterioration, against 11 hours for eCART and eight for the hand-calculated score, and the authors noted that earlier work suggests moving a patient to intensive care within four to six hours of meeting deterioration criteria is where outcomes improve. Yale New Haven said it moved its seven hospitals onto eCART in January 2024 on the strength of those findings.
That same comparison noted that the FDA treats clinical decision support software that reads chart data to flag deteriorating patients as a regulated medical device, and that only two general early warning scores had been cleared at the time: the Rothman Index and eCART. The Epic index, the most widely available of the group, has not gone through that review, and its own development and validation work has not appeared in a peer-reviewed journal. Epic’s record system covered roughly 48% of US acute-care hospital beds in 2022, by the figure cited in that paper, so the score is computing in a very large number of buildings.
What hospitals running Epic can take from it
The senior author put the reconciliation plainly. “The mortality benefit was not produced by an algorithm but by the partnership around the algorithm,” said Stephen P. O’Mahony, chief medical information officer at RWJBarnabas Health.
The authors credit a package of changes rather than the score alone: clinician education, alert tuning during the initial pilot at Robert Wood Johnson University Hospital, the automated notifications, and a standing response team with critical-care staff to send. Read next to the Yale comparison, that is the operative finding for anyone else running the tool. The mortality difference tracked the response system built around the score, not the score’s ranking against its competitors.
The group is now evaluating alerts triggered by how fast a patient’s risk score is climbing rather than the level it reaches, on the theory that velocity buys more warning than a threshold crossing does. For the many hospitals that already have this model running in the background, the transferable half of the study is operational: who gets paged, how quickly, and who arrives at the bedside.
