A study of 23,132 high-risk patients across 11 hospitals in the RWJBarnabas Health system found that an AI-driven early warning tool helped identify patients heading toward rapid clinical decline earlier — and was associated with in-hospital death rates among those high-risk patients falling from 23.1% to 18.6%, an 18% reduction in risk-adjusted odds of death. The research, conducted with Rutgers Robert Wood Johnson Medical School and published in NEJM AI, a New England Journal of Medicine group publication, is among the largest health-system-wide studies yet to link a deterioration-prediction AI tool directly to fewer patient deaths.
The tool at the center of the study is the Epic Deterioration Index, a proprietary machine-learning model built into the Epic electronic health record system used by RWJBarnabas and thousands of other U.S. hospitals. Unlike a bedside monitor tracking a single vital sign, the index continuously pulls in vital signs, lab results, nursing assessments and patient age already sitting in the EHR, recalculating a composite risk score every 15 minutes for every admitted patient.
The Problem It’s Designed to Catch
Clinical deterioration on a general hospital ward is a notoriously hard thing to catch early. Patients who are going to crash — developing sepsis, respiratory failure, or cardiac instability — often show subtle warning signs hours before an obvious, unmistakable crisis: a slightly elevated heart rate here, a slightly low blood pressure there, none of which individually triggers alarm from a nurse checking vitals on a normal rounding schedule. Traditional early warning scores, like the widely used National Early Warning Score, rely on simple point-based rules applied to a handful of vital signs and have long been criticized for both missing genuine deterioration and generating so many false alarms that staff become desensitized to them — a problem researchers call alarm fatigue.
What Changed at RWJBarnabas
What the RWJBarnabas and Rutgers researchers describe isn’t simply installing software — it’s building an operational loop around it. When a patient’s risk score crosses into the highest-risk category, the system automatically triggers an alert that is routed directly to a rapid response team, rather than depending on a bedside nurse to notice a pattern and decide independently whether to escalate. The study found that this alerting structure increased rapid response team activations without significantly increasing ICU transfers — suggesting the tool was catching genuine at-risk patients and getting them extra attention earlier in their decline, rather than simply funneling more patients into intensive care unnecessarily.
The Researchers’ Own Caveat
Notably, the study’s authors are cautious about attributing the full mortality improvement to the algorithm alone. They describe the benefit as likely resulting from a combination of factors working together as a coordinated systemwide approach: staff education campaigns run alongside the rollout, heightened general clinical awareness of deterioration risk, the EHR alerts themselves, and the automated rapid response notifications. That’s an important nuance — it means the AI model’s predictive scoring was necessary but not sufficient on its own; the surrounding clinical workflow redesign appears to have mattered just as much as the underlying algorithm.
How This Fits Into a Broader Body of Research
The Epic Deterioration Index has been studied at other health systems with more mixed results in some published literature, and academic reviews of AI-based deterioration prediction tools generally note that performance depends heavily on how a hospital operationalizes the alerts, not just the accuracy of the underlying model in isolation — a factor cited in research published in the Journal of the American Medical Informatics Association examining why similar tools succeed at some hospitals and underperform at others. Separately, cardiovascular-specific early warning models have shown promise in ICU settings this year, part of a broader trend of health systems building layered, unit-specific deterioration prediction on top of general hospital-wide tools.
What’s Next
RWJBarnabas Health and Rutgers researchers say the results support continued systemwide use of the Epic Deterioration Index paired with its rapid-response alerting structure, and the study is likely to be closely read by other large health systems considering similar rollouts, given the size of the patient cohort and the multi-hospital scope. The open question for the field is whether the specific combination of technology and workflow redesign that worked at RWJBarnabas — not just the algorithm by itself — can be replicated at other health systems with different staffing models, patient populations and existing alert fatigue levels.
