Stanford Hospital uses an AI-based model that predicts when a patient will deteriorate and alerts the patient’s doctors and nurses.
The alert system helps doctors connect and intervene more efficiently and effectively to prevent patients’ conditions from worsening and ending up in the intensive care unit, said Ron Li, MD, clinical associate professor of medicine and director of medical informatics in the field digital health. Li is the lead author of the Stanford Medicine study showing the potential of AI to mediate the connection between doctors and nurses. In a news update with Stanford Medicine’s Hanae Armitage, Li shared details of the project and how it fosters connection in a perpetually busy hospital environment. Below are the highlights:
How does the model work?
The algorithm is a predictive model that retrieves data – such as vital signs, information from electronic health records and laboratory results – in near real time to predict whether a patient is at risk of health deterioration while in the hospital.
Doctors are unable to monitor all of these data points for every patient at all times, so the model runs in the background and checks these values approximately every 15 minutes. It then uses artificial intelligence to calculate a risk score for the likelihood that the patient’s condition will worsen. If the patient feels like they are getting worse, the model sends an alert to the care team.
How can hospitals benefit from this?
This model is based on AI, but the action it triggers, the intervention, is essentially a conversation that might not have happened otherwise.
Nurses and doctors have conversations and handoffs when they change shifts, but it is difficult to standardize these communication channels due to busy schedules and other hospital dynamics, Li said. The algorithm can help standardize and focus the doctor’s attention on a patient, who may need additional care. Once the alert is received by the nurse and doctor at the same time, a conversation is initiated about what the patient needs to ensure that they do not decline to the point of requiring transfer to the intensive care unit.
Implementation and evaluation
The model originally sent an alert when the patient’s condition was already deteriorating. So researchers adapted it to focus on predicting ICU transfers and other indicators of worsening health status. The goal was to ensure that the care team was highly involved and felt empowered to initiate conversations with physicians about adjusting a patient’s care.
When the tool, used on nearly 10,000 patients, was evaluated, there were significant improvements in clinical outcomes. In a subset of 963 patients with risk scores within a “regression discontinuity window,” which basically means that, there was a 10.4% decrease in deterioration events, which the researchers defined as ICU transfers, rapid response team events or codes were defined at the threshold of high risk. These are patients whose clinical course may not be as obvious to the medical team. For this patient group, this model was particularly helpful because it encouraged doctors and nurses to work together to determine which patients needed special care.
Feedback from nurses and doctors
Responses have been broadly positive, but there are concerns about alert fatigue because not all alerts indicate a true decline, researchers said. When the model was validated using patient data before implementation, researchers calculated that about 20% of patients reported by the model actually experienced deterioration within six to 18 hours. While it’s not a completely accurate model at this point, it’s accurate enough to warrant conversation. It shows that the algorithm does not have to be perfect to be effective.
Researchers are now working to improve accuracy to increase confidence. The study of Li; Computer science postdoctoral researcher and lead author Robert Gallo, MD; Lisa Shieh, MD, PhD, clinical professor of medicine; Margaret Smith, operations manager for primary care and population health; and Jerri Westphal, director of nursing informatics, was published in JAMA Internal Medicine.
Originally published here.