As hospitalists at SHM Converge 2025 in Las Vegas, we may have wished for a crystal ball to win big at the tables. However, our professional aspirations focus on predicting which patients may deteriorate on the ward. Jessica Nave, MD, an assistant professor and academic hospitalist at Emory University in Atlanta, started her presentation by describing the different rounding styles of her fellow hospitalists, from alphabetical and geographic, to prioritizing new patients or discharges. She highlighted a scenario where the wrong prioritization could lead to a decompensating patient being overlooked for the first hour and a half of a shift due to the system chosen. She posited that there might be a better way to prioritize rounding.
The ability to discern which patients are sick and which aren’t is fundamental to our training and a focus of what we teach our trainees. Early Warning Scores use the objective data that we gather on the wards to identify which of our patients are at risk of needing escalation in care. The COVID-19 pandemic put pressure on many hospitals to make sure resource utilization, including intensive care unit (ICU) beds, was optimized. Emory University hospitals directly faced this issue, requiring the development of a system to intervene and prevent ICU transfers through the use of their medical emergency teams.
Over the past three decades, several predictive models for clinical deterioration have emerged, aimed at forecasting ICU transfers, cardiac arrest, and mortality. The first Early Warning Scores included data from bedside assessments, including common variables like heart rate, blood pressure, respiratory rate, temperature, and mental status. These scores have been adapted, including the Modified Early Warning Score (MEWS) and National Early Warning Score (NEWS and NEWS2), which apply different weights to the variables and have additional variables added, including urine output, oxygenation, and mental status assessments.
With electronic health records (EHR), additional variables can be integrated as well as evaluation over time. The Rothman index (RI), Epic’s Deterioration Index (DI), and Electronic Cardiac Arrest Risk Triage (eCART) are three of these models. In a recent head-to-head comparison, eCART and NEWS outperformed the other models.1 Recently, eCART and RI became U.S. Food and Drug Administration-approved for use in the hospital. Two machine learning systems, eCART and DI, are proprietary. Depending on the EHR, these scores can be directly integrated into patient lists to be displayed to practitioners or specialized teams. Newer systems incorporating artificial intelligence to include non-discrete fields are being developed, including the recently published CONCERN tool, which incorporates nursing notes and level of concern as part of the notification system.2
After reviewing their internal database of Medical Emergency Team (MET) activations, Emory found that 19.7% were preventable. When developing their intervention, Dr. Nave discussed four pivotal variables: 1) choosing the patient population, 2) available scores and tools, 3) determining the end user, and 4) defining the intervention. These variables had to fit with their overall goals of reducing mortality, ICU transfers, cardiac arrest, and MET activations.
Choosing the patient population
When choosing where the early warning score should fire, the patient population can be anywhere in the hospital, from the emergency department to the ICUs. As the end goal was to prevent ICU transfers and target a population where the nursing and physician ratios require additional outreach, hospital floor patients were chosen for intervention. This is a common area targeted by rapid response systems and a prime target for the use of early warning scores.
Additional tools
At the time of the implementation of their system, proprietary machine learning systems were not available. MEWS data was available to be gathered bedside and could be incorporated into the EHR. Additional variables were considered, including the use of the Glasgow Coma Scale, continuous pulse oximetry, and telemetry.
The end-user
Determining whom to display the results to can come with challenges if the end user does not understand the data. Emory chose to focus on their medical emergency teams, which specialize in intervention and prevention of decompensation. Their familiarity with common decompensation scenarios helped them develop active protocols and connect to acute utilization of resources. Individual clinicians have access to the data, but do not yet have targeted education on how to put the scores to use.
The intervention
After the activation from the early warning score, the team responds with proactive rounding and communicating with the primary teams. Interventions include ordering labs, medications, and additional imaging or studies, and activating teams and protocols such as stroke or sepsis activation.
After implementation, Emory changed EHRs, and the proprietary DI score was included as part of this change. In comparing the two scores, MEWS kept the census of their rapid response team lower and focused on those most likely to decompensate during the shift of the team. DI often predicted with a longer time frame who may decompensate, with many with elevated DI scores developing a higher MEWS score on the following shift.
The Emory intervention led to decreased mortality, fewer transfers to the ICU, and decreased cardiac arrests, but did result in more activations of the rapid response team. The successful intervention provides a model for other hospitals to implement an integrated system. Following the stepwise fashion Dr. Nave provided and addressing the key variables of implementing a system can improve care and prevent decompensation.
Key Takeaways
- Predictive models have a wide range of complexity and can provide different insights towards decompensation.
- Be thoughtful about implementation by knowing which patient population to target, what tools you will need to gather data, who you want to see the data, and what intervention should be used depending on the data.
- Use of predictive models can be implemented at a system level to result in earlier intervention and decrease mortality and ICU transfers.
- Individual provider knowledge of these scores can influence rounding behaviors to see the sickest patients first.
Dr. Molitch-Hou is an assistant professor, the director of hospital medicine sub-internship, core faculty for the internal medicine residency program, and co-director of the care transition clinic at the University of Chicago Medical Center in Chicago.
References
- Green M, et al. Comparison of the between the flags calling criteria to the MEWS, NEWS, and the electronic cardiac arrest risk triage (eCART) score for the identification of deteriorating ward patients. Resuscitation. 2018;123:86-91. doi:10.1016/j.resuscitation.2017.10.028
- Rossetti SC, et al. Real-time surveillance system for patient deterioration: a pragmatic cluster-randomized controlled trial. Nature Medicine website. https://www.nature.com/articles/s41591-025-03609-7. Published April 2, 2025. Accessed May 25, 2025. doi:10.1038/s41591-025-03609-7