An innovative use of artificial intelligence that essentially monitors hospital patients through the eyes of the nurses involved in care resulted in a stunning 36% drop in deaths. However, funding to extend the research with adults to sick children was abruptly cancelled by the Trump administration.
In a one-year trial of an AI system operating in real time at four hospitals in two highly regarded organizations – NewYork-Presbyterian and Mass General Brigham – in-hospital deaths plunged 35.6% compared to usual care, while the length of stay decreased 11.2%, according to a study in Nature Medicine. Put differently, the study estimated that 84 people who might previously have died did not, and some 31,000 individuals on average went home almost 15 hours earlier.
“Our vision is that this system is put into every hospital in this country and beyond,” said Sarah Rosetti, the co-principal investigator of the study, in an interview. Rosetti is both a veteran critical care nurse and an associate professor at Columbia University’s Department of Biomedical Informatics.
It was Rosetti’s academic affiliation that triggered a sudden cancellation of the study’s National Institute of Nursing Research grant meant to enable continued research and expansion to pediatric patients. The grant was a casualty of the Trump administration withdrawing $400 million in funding from Columbia over its alleged weak response to antisemitic campus demonstrations. Although Rosetti’s co-PI, Kenrick Cato, is a clinical informatics researcher at the University of Pennsylvania, and although Columbia eventually acceded to the administration’s demands, there’s been no indication if or when the grant will be restored.
An Idea From the Front Lines
The idea for the AI system goes back more than 20 years, said Rosetti, to a time when she was a hospital nurse working with critically ill patients and using thick paper flow sheets to document everything from vital signs to personal observations. When experienced nurses were worried, she noticed, they returned to the patient’s room more often and wrote more detailed descriptions. “You can be critically stable or critically unstable,” Rosetti noted, “and there’s a very different approach to care depending on which of those two buckets you’re in.”
After becoming an informatics researcher, Rosetti assembled a team from several institutions to build an AI system that would transform what nurses observed “into a new data point that everyone could see and focus on.”
The result was a system with an easy-to-remember acronym and a cumbersome full name: The Communicating Narrative Concerns Entered by RNs (CONCERN) Early Warning System (EWS). CONCERN uses more than 1,200 “ensemble-based models” that take into account not only what nurses document in the electronic health record, but factors such as the number of days the patient has been hospitalized, whether it’s a weekday or weekend, the time of year, the time of day or night, the nurse shift and many more. The resulting prediction model identified patients at risk of deterioration two days earlier than the standard Modified Early Warning Score, which relies on vital signs and lab results.
Data Drives Action
A key to an actual death-rate reduction was that CONCERN provided a concrete data point to replace a nurse’s more nebulous “I’m worried” instinct that sometimes gets expressed, sometimes is repressed due to a reluctance to bother a physician and sometimes, research shows, even when conveyed promptly fails to get promptly acted upon. The researchers worked extensively with both doctors and nurses in CONCERN’s design, integrating it with the existing clinical workflow and explaining how different inputs influenced the predictions. That, in turn, built the trust that led to a swifter response when the system’s green light turned to a warning yellow or urgent red.
As a result, more patients were rushed to the intensive care unit – unexpected ICU transfers rose – but there was a 7.5% reduced risk of sepsis, a hard-to-detect infection that can cause fatal organ failure without immediate treatment.
The hospitals involved included one academic medical center and one community institution from each of the two health systems. But Rosetti told me that a less-advanced version of CONCERN was independently validated by another research team using retrospective data at over 200 hospitals. What that indicates, she said, is that nurse documentation behavior is consistent and that the system should be effective at all types of institutions in all sorts of situations.
Despite the extraordinary drop in patient deaths demonstrated so far, the team is scrambling to make up the $500,000 grant that was to enable it to test the system at a children’s hospital. Nonetheless, the researchers hope to be able to proceed with the work at the pediatric hospital and also plan to expand the research to two more adult hospitals. They are also preparing a limited number of academic research licenses that come with a suite of tools enabling a simple linking of any EHR to the CONCERN system using the FHIR interoperability standard.
Separately, Columbia has applied for a patent that would enable the kind of revenue flow needed to aggressively roll out the system nationwide, including in low-resource areas.
What the researchers have not yet determined is whether saving lives saves money, although they believe they have reduced costs while making care safer. Still, a lessened length of stay with more short bouts of intensive care might or might not reduce total health care costs and might or might not improve a hospital’s own revenue.
The hospitals involved in the study continue to use CONCERN, Rosetti said, and the two health systems are considering how to expand its use to all of their hospitals.