Doxiverse Staff - July 2026
A study published in NEJM AI found that pairing an AI-based patient deterioration score with automated alerts to hospital Rapid Response Teams (RRTs) was associated with a drop in risk-adjusted in-hospital mortality among high-risk patients, from 23.1% to 18.6%, across 11 hospitals in the RWJBarnabas Health system.
Why it matters: This is one of the larger, peer-reviewed, outcomes-based evaluations of an AI early-warning tool that measures an actual mortality difference. The underlying tool, Epic's Deterioration Index, is already present in many hospitals running Epic, meaning the relevant variable for other systems may be less about acquiring new software and more about how alerts are routed and acted on.
What Happened
Researchers from RWJBarnabas Health and Rutgers Robert Wood Johnson Medical School evaluated a systemwide rollout of the Epic Deterioration Index (EDI), an AI tool that continuously analyzes existing electronic health record data including vital signs, lab results, nursing assessments, and age to generate a deterioration risk score. The study examined adult medical-surgical admissions with an EDI score of 60 or higher, across 11 acute care hospitals (a mix of academic, community teaching, and nonteaching facilities) between October 1, 2022, and August 30, 2024.
The design was what the authors describe as a "quasi-experimental, staggered cohort study" using a pre- versus post-implementation comparison, not a randomized controlled trial, but a design that compares outcomes at the same hospitals before and after the intervention was rolled out, with different hospitals implementing at different times. Among 23,132 high-risk patients included in the analysis, in-hospital mortality fell from 23.1% before implementation to 18.6% after, which the authors report as an 18% reduction in the risk-adjusted odds of death. Lead author Dr. Thomas Nahass is joined by co-authors including Dr. Joseph Hanna, Dr. Jason Roy, and Dr. Andy Anderson, RWJBarnabas Health's Chief Medical and Quality Officer.
According to coverage of the study, the intervention's key operational change was routing high-risk ("red") alerts as automated push notifications directly to mobile devices carried by RRT staff, rather than relying solely on passive EHR dashboards that a clinician would need to actively check. Built-in suppression rules were reportedly used to reduce alert fatigue blocking redundant notifications for patients already receiving ICU-level care, comfort care, or who had triggered a rapid response or sepsis alert within the prior six hours.
Notably, RRT activations among high-risk patients rose substantially after implementation, from 25.3% of hospital stays to 37.5% without a corresponding significant increase in ICU transfers, suggesting the added activations reflected earlier bedside evaluation rather than simple overtriage.
Study co-author Dr. Thomas Nahass, VP of Health Informatics and an intensive care physician at RWJBarnabas Health, said the team's goal was to identify patients earlier, before intervention becomes much more difficult. "The deterioration index gives us an earlier point in time," Nahass said. "If we can get a critical care eye on the patient sooner, we can change the course of their outcome."
What This Means for Hospitals and Health System Leaders
For hospitals already using Epic and its built-in Deterioration Index, this study suggests the tool's clinical value may depend heavily on how alerts are operationalized, specifically, whether high-risk scores trigger an automated, direct notification to a response team, versus sitting in a dashboard that depends on a clinician noticing it. That's a workflow and alerting-configuration question as much as a technology-adoption question, and it may be a lower-cost lever for systems that already have EDI available but haven't built an automated RRT-alerting pathway around it.
The alert-fatigue suppression logic described in coverage of the study is also a relevant detail for any team designing or tuning a similar alerting system: unfiltered high-risk alerts risk overwhelming response teams or triggering unnecessary escalations, and the suppression rules used here, excluding patients already in ICU-level or comfort care, or recently flagged are a specific, reportable design choice other systems could evaluate adopting. The fact that RRT activations rose by roughly 12 percentage points without a corresponding jump in ICU transfers may offer some reassurance that more aggressive alerting doesn't necessarily translate into unnecessary escalations, though this should be evaluated in each system's own context.
What's Still Unsettled
This was a pre- versus post-implementation observational study at a single regional health system, not a randomized controlled trial and it establishes a strong association, but the study design is more susceptible to confounding from other concurrent changes in care (staffing, protocols, case mix) than an RCT would be, even after risk-adjustment.
It's also worth noting this evaluates one specific configuration (EDI plus automated RRT push alerts plus suppression logic) at one health system over a roughly two-year period; results at a different hospital, EHR configuration, or with a different RRT staffing model could differ.
Sources:
Nahass, T.A., et al. "Implementation of an AI-Triggered Rapid Response — Association with Mortality." NEJM AI (2026). DOI: 10.1056/AIoa2500973. Published July 29, 2026. https://ai.nejm.org/doi/10.1056/AIoa2500973
HIT Consultant, "RWJBarnabas Links Real-Time Epic EDI Alerts Directly to Rapid Response Teams," July 29, 2026 (secondary reporting, includes attributed quote from Dr. Thomas Nahass).
Rutgers University / HCI Innovation Group / Medical Xpress, coverage of the same study, July 29, 2026 (secondary reporting, used to corroborate figures and quotes).
This article summarizes a single-system observational study. It is not clinical guidance; hospitals considering similar alerting configurations should review the full primary study and consult their own clinical informatics and quality teams.