An artificial intelligence-powered early warning system that automatically notifies rapid response teams about deteriorating patients was associated with a significant reduction in in-hospital deaths, according to a new study published in NEJM AI. The research, conducted across 11 hospitals within the RWJBarnabas Health system, found that implementing automated alerts resulted in an 18% reduction in the risk-adjusted odds of death for high-risk inpatients.
The study, led by researchers from RWJBarnabas Health and Rutgers, evaluated outcomes for 23,132 high-risk adult patients before and after the system was fully activated. The findings demonstrate that integrating predictive AI tools into established clinical response workflows can lead to earlier, more effective interventions and save lives. The unadjusted mortality rate for the high-risk group fell from 23.1% before the system's implementation to 18.6% afterward.
At the core of the intervention is the Epic Deterioration Index (EDI), a predictive tool already integrated into the widely used Epic electronic health record (EHR) system. The EDI continuously analyzes data within a patient's record to generate a risk score, helping clinicians recognize early signs of a patient's decline. While the tool was available to clinicians prior to the study's intervention period, its use was passive. The pivotal change was the creation of an automated alert system. When a patient's EDI score reached the highest-risk threshold, the system sent an automatic notification directly to the hospital's specialized rapid response team, prompting an immediate bedside assessment by critical care specialists.
Comparing Patient Outcomes and Workflow: Pre- and Post-AI System Implementation
To measure the impact of the automated alerts, researchers compared a pre-implementation cohort of 10,803 high-risk patients with a post-implementation cohort of 12,329 patients. The study design allowed for a direct comparison of key clinical outcomes and response metrics, isolating the effect of shifting from a passive risk-scoring system to an active, automated notification workflow. The results showed a marked change not only in patient mortality but also in how clinical teams responded to at-risk individuals.
The table below details the specific changes observed between the two periods for high-risk patients across the 11 participating hospitals, which included academic medical centers as well as community hospitals.
| Metric | Pre-Implementation Phase | Post-Implementation Phase |
|---|---|---|
| In-hospital Mortality (Unadjusted) | 23.1% | 18.6% |
| Risk-Adjusted Odds of Death | Baseline | 18% reduction |
| Rapid Response Team (RRT) Activations | 25.3% of hospitalizations | 37.5% of hospitalizations |
| RRT Notification Method | Clinicians manually checked risk scores and initiated calls | Automated notifications sent directly to RRTs for highest-risk patients |
| Escalation to Higher Care (e.g., ICU) | Baseline | No significant increase |
One of the most telling findings was the dramatic increase in rapid response team activations, which rose from 25.3% to 37.5% of high-risk hospital stays. This indicates that the AI system was successful in identifying at-risk patients who might have otherwise been missed. Crucially, this surge in RRT calls did not lead to a significant increase in patients being escalated to higher levels of care, such as an intensive care unit. This suggests that the earlier interventions prompted by the alerts were effective, allowing clinical teams to stabilize patients at the bedside and prevent further deterioration that would necessitate more intensive and costly care.
Beyond Technology: A Systemwide Approach to Success
While the AI-driven alerts were the technological centerpiece of the intervention, the study's researchers emphasize that the mortality benefit likely resulted from a "coordinated systemwide approach." The success was not attributed to the algorithm alone but to a combination of factors working in concert. This comprehensive strategy included extensive staff education to build trust and understanding of the new system, which in turn fostered enhanced clinical awareness among care teams.
The automated notifications served as a reliable trigger, but it was the prepared and coordinated human response that ultimately improved patient outcomes. By integrating the electronic alerts directly into the rapid response team's workflow, the system ensured that predictive insights were translated into timely clinical action. This socio-technical approach—blending advanced technology with well-designed human processes and training—was critical to achieving the observed reduction in mortality. The study underscores that simply making a predictive score available is not enough; the information must be delivered to the right people at the right time in a way that compels action.
Next Steps: Refining AI for Even Earlier Intervention
Because the study evaluated the Epic Deterioration Index, a tool already built into one of the nation's most prevalent electronic health record systems, the findings may have significant implications for hospitals nationwide. The successful implementation at RWJBarnabas Health provides a potential model for other healthcare systems looking to leverage predictive analytics to improve patient safety and outcomes.
The key takeaway for other institutions is that technology is a powerful enabler but not a standalone solution. The study's success highlights the importance of integrating AI tools within a comprehensive strategy that includes robust staff training, redesigned clinical workflows, and a culture of trust in the technology. For hospitals seeking to replicate these results, the measurable indicator of success will not just be the adoption of an AI tool, but the subsequent increase in timely interventions and a corresponding, evidence-based reduction in mortality rates among their most vulnerable patients.
