Detection and Treatment of Patient Deterioration

The Problem

The current practice of identifying deteriorating patients is reactive, rather than proactive. Patients often show early signs of deterioration hours before a rapid response team (RRT) or code blue is activated. Delays in care for such patients can have a detrimental impact on clinical outcomes, (1, 2) as patients transferred unexpectedly to the intensive care unit (ICU) often have worse outcomes and increased mortality compared with patients with planned admission to an ICU (3–6). Furthermore, survival to discharge following in-hospital cardiac arrest is estimated to be only 22% (4, 7). Unfortunately, predicting clinical deterioration remains intrinsically difficult because critical information such as preceding abnormal vital signs (8–11) are often missed by providers (1, 12).

The current RRT system at Duke Hospital does not capitalize on data available via the electronic health record (EHR) and is dependent on frontline staff with varying levels of knowledge and experience to identify patients who are clinically deteriorating. An initial attempt at implementation of an automated electronic warning system (the NEWS) within MaestroCare at Duke failed to effectively reduce the rate of unanticipated ICU transfer or death, and unfortunately resulted in significant EHR alarm fatigue for frontline nursing staff. (13) In response to the poor performance of the NEWS score we have recently developed and built a new model for patient deterioration into the current EPIC system. While this model performs better than the NEWS score, due to constraints of the EPIC environment, the model is still fairly simplified, consisting of a basic logistic regression. There remains a large opportunity to improve the modeling of patient deterioration. (14)

In addition to not capitalizing on data captured within the EHR to predict deterioration, the care provided when a patient has a deterioration event is not uniform and is often both inefficient and uncoordinated. There is no standardized hand-off between the primary team, RRT team, or ICU. There are no protocols for initial management or evaluation. There is no standard documentation of care provided during the event. Perhaps most importantly, the RRT team does not train as a cohesive emergency response unit and therefore team efficiency and skill are not optimized.

Our Solution

In collaboration with the Duke Institute of Health Innovation (DIHI) and the Department of Statistical Science, our team has separately developed a novel machine-learning model capable of predicting sepsis within 4 hours of onset at Duke University Hospital. This novel sepsis risk model, Sepsis Watch™, has recently been implemented within the MaestroCare production environment, and is currently undergoing prospective validation. We planned to utilize a similar machine-learning based approach applied to EHR data to predict more global early clinical deterioration.

Update in April 2026

The Adult Deterioration prediction solution went live in May 2025 and is currently scaling across the Duke Health System. This go live followed a successful prospective validation of the solution output over the calendar year 2024, which showed strong performance (Table 1, Table 2). 

For each hospitalized patient on an intermediate/stepdown unit, the solution applies a machine learning model to real-time data from Maestro Care, generating a risk prediction of the patient’s clinical deterioration within the subsequent 12 hours. Every 20 minutes, the solution pulls new data, refreshes the risk prediction, and sends an automated page notification if the patient’s risk of deterioration is above a pre-determined threshold. All notifications for that patient are then snoozed for the subsequent six hours to mitigate alarm fatigue and allow for new data to be collected to inform an updated prediction. Figure 1 shows a technical overview of the solution.  

Now, at each of our hospitals, we are optimizing the solution in partnership with the care teams responsible for identifying and intervening on decompensating patients. We aim to reduce the rates of rapid response events, ICU transfers and ICU length of stay, and in-hospital mortality.  

Here is a brief overview of the Adult Deterioration Solution’s status at each hospital, including select news-worthy updates as well as how we are implementing them with our frontline care teams: 

   Duke University Hospital 

• What is the status? The Adult Deterioration solution went live on May 28th, 2025 on all adult patient intermediate and stepdown units at Duke University Hospital. As of mid-April, the solution has generated 3,966 notifications on 1,437 unique patient hospitalizations. The solution sends 8.7 notifications per day to the Patient Response Team (PRT) Charge Nurse. 

• How is the solution applied in the clinical workflow? Figure 2 shows the clinical workflow wherein the PRT nurse receives an automated notification. The notification is sent to the PRT charge nurse’s mobile device as a page, using the Spok mobile application. The content of each notification reads as:  

“Adult Deterioration: ABC123 in 7110-01 is at HIGH risk of deterioration in the next 12 hours (calculated at 06:52AM 04/18)”, where ABC123 is the patient’s MRN, 7110-01 is bed, HIGH is the risk category, and 06:52AM is the time of prediction. 

• Where are we headed? The PRT nursing team has helped us optimize the solution since its go live, including lengthening the snooze window from 4 to 6 hours, and excluding patients on comfort care. In partnership with this team and the Duke Heart Center, we are evaluating the added benefit of including clinical phenotype flags (Figure 3) for additional clinical context on why a patient may be deteriorating. We plan to include both the prediction model output and recent phenotype statuses in the content of each patient’s page. We also plan to apply logic to risk thresholds and phenotype counts, optimizing for sensitivity and specificity of the notifications. 

 Duke Raleigh Hospital 

• What is the status? The solution went live at Duke Raleigh on January 6, 2026 on all adult patient intermediate and stepdown units at Duke Raleigh Hospital. Since go live, the solution has generated 190 notifications on 73 unique patient hospitalizations. It sends 2.3 notifications per day to the Rounding Nurse Team.  

• How is the solution applied in the clinical workflow? The Rounding Nurse Team follows a similar workflow to the PRT workflow at DUH, beginning with a notification formatted the same way as for DUH. Upon receiving a notification in Spok, the Rounding Nurse performs a chart review, then a brief bedside assessment. If the patient is actively deteriorating, the nurse calls an RRT event and huddles with the attending physician, looping in critical care as needed. Otherwise, if there is a deterioration concern, the nurse notifies the First Call physician and bedside nurse. In either case, the nurse adds the patient to their rounding list for the upcoming 24 hours and documents a note in Epic. 

• Where are we headed? The Rounding Nurse Team Leads are collecting feedback from the nursing team and tracking paging volume versus note creation for the impacted patients. The goal is to understand workload and actionability of multiple pages on a given patient over time. The team is likewise interested in the inclusion of clinical phenotypes recently met for the page content, for context and improved performance related to identifying deterioration early.   

 Duke Regional Hospital 

• What is the status? The Early Nurse Intervention Team (ENIT) leads are working with us to finalize the clinical workflow and assess impact on deteriorating patients in the intermediate/stepdown units at Duke Regional. From January through March 2026, the solution logged 119 notifications total across the five in-scope departments. We are comparing these notifications against ENIT consult orders placed, RRT events, Code Blues, ICU transfers, and in-hospital deaths. Based on this performance, and the workload capacity of the ENIT nurses, we will establish thresholds for CRITICAL, HIGH, and MEDIUM that are bespoke to Duke Regional. 

• How is the solution applied in the clinical workflow? The ENIT Nurse workflow is similar to the workflows at DUH and DRAH: the ENIT nurse receives a page, performs a chart review and adds the patient to their prioritized rounding list, and performs a bedside assessment. If there is concern for deterioration, they engage the First Call Provider and Bedside 

• Where are we headed? The Adult Deterioration Prediction solution is planned for go live at Duke Regional in Spring 2026. We will also plan to incorporate the clinical phenotypes in the notifications for additional clinical context and actionability. 

We are collaborating with the Department of Population Health Sciences to study the impact of the solution on deterioration prevention, with the goal of publishing our results and further optimizing the predictive model performance. We are engaging with Duke Health Lake Norman now to understand their deterioration prevention workflows and investigate incorporation of the solution. We are also expanding into the Intensive Care Units across Duke Health to identify patients who are deteriorating within those care settings.