Researchers have developed an AI-powered tool that can analyze nurses’ shift notes to identify—far earlier than traditional methods—when an admitted patient’s health may be deteriorating or on the cusp of “crashing.”
In early trials, the AI-tool, dubbed the CONCERN Early Warning System (CONCERN EWS), helped lower a patient’s risk of death by more than 35% while reducing the average hospital stay by more than half a day.
The clinical trials—which involved more than 60,000 patients from 2020 to 2022—also showed a 7.5% decrease in risk of sepsis among patients admitted to hospitals deploying CONCERN EWS. The results, if replicated widely, could offer hospital systems a reliable way to improve patient outcomes while simultaneously reducing the costs associated with inpatient care.
In a study published in April in Nature Medicine, the team behind the AI—led by researchers from Columbia University and the University of Pennsylvania—outlined how the machine-learning algorithm prioritizes nurses’ keen medical observations. Nurses typically interact with patients frequently and often identify subtle—but important—changes in a patient’s health that might otherwise slip by unnoticed.
CONCERN EWS dives deep into what nurses are seeing to develop accurate and insightful predictions—but in an unexpected way. The AI understands natural language and can read what nurses write in a patient’s electronic health records (EHRs). But its main innovation is connecting the dots of the metadata linked with those notes.
For instance, a nurse may notice a patient’s color has slightly changed, they appear lethargic, or seem cognitively off. As a result, the nurse may increase how frequently they check on the patient, or hold off administering certain medications until the patient appears healthier. CONCERN EWS focuses on these micro-decisions nurses make.
As a proxy for understanding a nurses’ insights, CONCERN EWS analyzes metadata connected with each EHR entry—things like date, time, and location—looking for patterns that suggest trouble. The system notices if nurses assess patients more frequently than normal, and if those visits occur in the middle of the night or at other uncommon times. When the system sees alarming patterns, it alerts care teams that a patient’s health might be taking a turn for the worse.
By scrutinizing patients through the eyes—and records—of nurses, the ML model helped reduce a patient’s hospital stay by an average of 11%, noted the study’s lead author, Sarah Rossetti, an associate professor of biomedical informatics and nursing at Columbia.
The researchers used NVIDIA RTX A2000 12GB Graphics Cards to develop their algorithm.
During the clinical trials, CONCERN EWS was deployed in four hospitals across two hospital systems in Massachusetts and New York. It was able to help care teams spot signs of trouble, on average, 42 hours earlier than traditional methods. This gave teams a much better shot at stepping in before a patient’s health was at risk.
In May, the promising results helped the research team win one of three prestigious “Reimagining Nursing Initiative” grants the American Nurses Foundation awards each year. Each recipient will receive part of $1.5 million in total grants.
According to Rossetti, the team—which is co-led by Kenrick Cato, a professor of informatics at the University of Pennsylvania—will use its portion of the grant to partner with Children’s Hospital Colorado to create a pediatric version of its current model and evaluate it across community hospitals.
Read additional news coverage about CONCERN EWS or watch a video about the technology.
