The patient, 56, had undergone heart surgery and was recovering in a UCHealth medical-surgical unit when his nurse got an alert from the UCHealth Virtual Health Center. The nurse had been periodically checking in on the patient, and he had seemed to be doing fine – certainly compared to another patient who had been her focus all morning.
The Virtual Health Center had noted an onset of rapid breathing and a drop in oxygen saturation to just 86% (below 90% is a warning sign). The nurse checked the patient’s heart rate and found an abnormal beat. The patient was deteriorating – a medical term for an often abrupt plunge in status that can land someone in an intensive care unit (ICU) or, too often, worse.
The nurse ordered tests, upped the patient’s oxygen, and supplemented low potassium and magnesium levels. Rather than deteriorating into an ICU visit, the patient recovered quickly. He was well enough to go home the next day.
It was a success story with roots in a three-year UCHealth-wide effort to apply artificial intelligence (AI) to the problem of detecting patients who are poised to abruptly decline. The approach is saving more than 200 lives a year.
First, monitoring for sepsis
The initial focus was sepsis, a systemic blood infection that strikes 1.7 million people in the United States each year (that’s twice the number of annual U.S. heart attacks), killing 270,000 of them, according to the U.S. Centers for Disease Control and Prevention.
Time is of the essence with sepsis treatment. The typically subdued CDC resorts to all-caps in its patient explainer: “ACT FAST. Get medical care IMMEDIATELY…” The same applies to those already hospitalized, and the problem’s magnitude is such that it has attracted plenty of attention among technologists. Epic, whose electronic health record powers MyHealthConnection, offers an AI-powered algorithm with dozens of indicators that could point to incipient sepsis – high heart rate, low blood pressure, fever, shortness of breath, and lab results, among others. But Dr. CT Lin, a University of Colorado School of Medicine internal medicine physician and UCHealth’s chief medical information officer, says the technology, however sophisticated, isn’t enough.
“Most programs around the country are very happy to draw the finish line at saying, ‘Look, we built a sepsis algorithm,’” Lin said. “And we’re like, ‘That’s negligible work. It doesn’t affect patient care. You haven’t actually put it into place to change the quality of care. We’ve actually done that.’”
It wasn’t easy. Using the Epic Sepsis Model and the Epic Deterioration Index as starting points, UCHealth data scientists and machine learning experts had to choose and customize algorithms to improve the models’ predictive accuracy, Lin says, the aim being to foresee a patient’s deterioration up to 12 hours in advance. They fed the system three years of historical patient data, with variables including dozens of vital signs, lab values, medications being administered, procedures, and more. There was also the matter of deciding how to determine when a patient should be considered to have deteriorated. The team settled on unplanned ICU transfers, unexpected mortality, and code blues (the patient’s heart or breathing has stopped).
Virtual Health Center provides continued monitoring
A key decision was how to handle the familiar tradeoff between a diagnostic’s sensitivity and specificity. Sensitivity is about detecting what could be a problem. Specificity is about nailing down an actual problem. The more sensitive, the more false alarms and the less specific. UCHealth leadership made the decision to go for sensitivity, Lin says. That way, they would be more certain to catch deteriorations.
“Our leadership team said, ‘Don’t miss anything,’” Lin recalled. “But when you tune the system to be super sensitive, guess what? The noise is going to be terrible. It’s not going to be specific.”
The result was a system that would flag 61 patients a day for every one or two actual deteriorations. That’s not a problem in itself. But for an alert system to really work, you need four things, says Dr. Diana Breyer, UCHealth’s Northern Colorado medical director and the health system’s chief quality analytics officer.
“You need the data. You need the algorithms. Somebody has to notice. Then you have to do something,” Breyer said.
The team had accomplished the first two steps. The second two presented the entirely different challenges of noting and acting. Initially, the team fed the alerts directly to bedside nurses. But those nurses were already busy helping patients deteriorating now, not at risk of doing so 12 hours from now. The 61 alerts per day were more than they could digest.
The answer came in the form of the UCHealth Virtual Health Center, which is staffed around the clock with a critical-care physician and four critical-care nurses who serve all 12 UCHealth hospitals.
Virtual presence combined with AI
The Virtual Health Center team monitored the deterioration status of patients the AI system flagged and called bedside nurses only with regard to patients, who their experience and expertise told them were at risk of deteriorating. Their ability to do so had to do with more than just having the time to dedicate to monitoring deterioration, says Amy Hassell, a critical-care nurse and the UCHealth’s Virtual Health Center’s nursing director.
“We have sepsis-response teams through the system, but one team may only see a deterioration every 15 to 20 days, so it’s not something that frontline nurses – especially novice nurses – have seen,” Hassell said. “So it was a matter of making it scalable.”
That step improved the accuracy of the overall deterioration-detection system by a factor of 30. Now, someone flagged by a machine and filtered by a trained professional poses a greater than 50-50 risk of deteriorating in the next 12 hours.
The result has been a sharply improved speed of administering fluids (from 96 minutes to 19 minutes) and IV antibiotics (from 94 minutes to 45 minutes). Those and other measures have saved an estimated 211 lives a year, Lin says.
In March, UCHealth’s hospitals in northern Colorado rolled out an additional monitoring tool: rollaway patient-observing cameras that the Virtual Health team monitors for six hours after a likely deterioration call is made. That system is slated to go live in UCHealth’s southern Colorado hospitals in August and at UCHealth University of Colorado Hospital on the Anschutz Medical Campus in October, Breyer says.
The team is also expanding beyond sepsis to deterioration of other sorts as well as applying its lessons to hospital-acquired pressure injuries (HAPI, a.k.a. bedsores) and to patients at high risk of falls, Lin says.
He likens UCHealth’s deterioration-monitoring program to the centaur of Greek mythology, though rather than half-man, half-horse, it’s half-person, half-machine – a distributed cyborg. Regardless, the program’s key lessons – that success with AI depends on what human intelligence can do with the machine’s outputs – will endure.