When Michael Draugelis, chief data scientist in Penn Medicine’s Predictive Healthcare team, worked at Lockheed Martin, he created algorithms to analyze huge amounts of data to seek out enemy missiles. Now, he and his team at Penn are repurposing this technology for a different cause: better outcomes for patients.
Although its use in patient care is relatively new—Penn Medicine is the first in the region to have a dedicated data science team—predictive analytics is already a big part of modern life for many Americans. When people get a movie from Netflix or books from Amazon, they’ll also get suggestions: “If you liked X, then you might like Y.” This is data science. In medicine, it means “training” a computer to recognize patterns and using that information to predict outcomes.
Draugelis says that “clinical champions” drive the process in healthcare. “Clinicians can identify variables that might increase a specific risk and data scientists can help them connect the dots,” he says. “Success is less dependent on how we can deploy technology but more on being well-integrated with the team and clinically relevant.”
Relevant data—which can include results from lab tests or changes in vital signs—is pulled from the Penn Medicine EMR (electronic medical record) systems and funneled through Penn Signals, a real-time, big-data platform which “harmonizes all pulled clinical data into a common format and orchestrates multiple predictive applications,” Draugelis says, adding that the system can also serve as a platform for research and collaboration.
In scanning healthcare data in real-time, algorithms can pick up previously unrecognized patterns and connections that may indicate a risk that providers may not have considered or even anticipated. For example, the early warning system (EWS) algorithm developed at Penn for early detection of sepsis showed that if a patient’s red blood cells are broken up—fractured in the bloodstream—it correlates more highly with the condition.
“Data analytics is like panning for gold,” says C. William Hanson, Penn Medicine’s chief medical information officer. “The computer sifts through a lot of information to find gold nuggets that you might not otherwise find.”
Data scientists partner with members of both the Information Services and Quality and Safety teams. A steering committee which includes Hanson; PJ Brennan, chief medical officer and senior vice president; Pat Sullivan, chief quality officer; and Kevin Mahoney, executive vice president and chief administrative officer, examines potential projects. To move ahead, each project must align with the Blueprint for Quality and Patient Safety, Penn Medicine’s framework for clinical strategy, and balance the connection between value and patient outcome.
Heading Off a Fierce Enemy
Despite all the advances in modern medicine, including effective antibiotics, sepsis remains a primary cause of preventable in-hospital death. Early diagnosis is difficult because many of its symptoms, such as fever, increased heart rate and difficulty breathing, are similar to those associated with other conditions.
“What makes sepsis deaths preventable is that cases can go undetected or may be detected too late. And we’re talking hours, not days,” says Craig Umscheid, vice chair of quality and safety in the Department of Medicine.
In 2012, Umscheid and his team developed EWS 1.0, a simple predictive algorithm that identified patients with sepsis based on six variables. “We wanted to make sure no one was being missed,” he says. Two years later, they joined forces with Penn Medicine’s data scientists to take this surveillance one step further, predicting patients at high risk before it happened.
EWS 2.0 was more accurate, training on more than 500 variables—expanding beyond blood pressure and heart rate to include lab results, such as measures of renal function and blood urea nitrogen—and sending out alerts to physicians up to 30 hours in advance of patients developing several of the key symptoms of sepsis.
Early results from EWS 2.0 suggest that while more providers may be acting on the alerts—for example, there was a small increase in testing for sepsis in those who were triggered—“it may not have an effect on the prevention of septic shock,” Umscheid says. Still, he feels it is “an important step forward to plot what’s next in this lifesaving effort.
“EWS 1.0 was crawling; EWS 2.0 was walking,” Umscheid says. The next version, he added, will continue to improve our abilities to decrease sepsis. “It’s tough to run before you can even walk.”
A One-Two Punch to Improve Heart Failure Outcomes
Improving outcomes for patients with heart failure—which afflicts nearly five million Americans—was a two-part challenge for data scientists. The first part focused on identifying inpatients who were hospitalized for reasons other than heart failure but who still, according to the predictive algorithm, “looked like” (that is, showed the same pattern as) those who have been diagnosed with the condition.
“It’s hard to find these patients—they can be on any service,” says Joanne Fante-Gallagher, director of quality and safety, heart & vascular service line. Once the new algorithm was deployed, the multidisciplinary heart failure team was alerted to every inpatient the system flagged, enabling them to examine the person’s cardiac history and needs. If necessary, patients were connected with a cardiologist or in some cases, a primary care physician, Fante-Gallagher says.
The second component focused on decreasing readmissions of diagnosed heart failure patients.
“In general, heart failure patients have a higher risk of readmission than other patients but a certain subset has the highest risk,” Hanson explains. Patients identified as “fitting the pattern” of being in this subset received more aggressive post-acute care.
“We looked at what their outpatient care looked like. For example, are they getting the right meds? Can we connect the patient with Penn Care at Home nurses?” Fante-Gallagher says. “This put all information in one bucket, getting them the services they need. ... It’s a matter of getting the right patient the right care at the right time.”
The team is currently targeting heart failure patients at the three downtown campuses but is planning to expand to heart failure teams throughout Penn Medicine.
This model is being used in areas across the Health System. Under the leadership of Barry Fuchs, the medical intensive care unit at HUP developed an alert for when patients could be safely weaned from mechanical ventilation. Corrina Oxford, of Maternal-Fetal Medicine at HUP, is working with data scientists to identify women who could be at risk for infection or hemorrhage after childbirth. On the ambulatory side, data scientists have targeted lung cancer patients who are more likely to end up in the ER and being admitted to the hospital.
Predicting is Just the Beginning
Data science is taking healthcare providers from the ability to detect when a disease or condition has started to predicting it before it happens. But, Sullivan says, “The work doesn’t stop at just predicting who’s at risk. We need systems and processes in place to act on that information.”
With the recent opening of its Oncology Evaluation Center, Penn Medicine has already established one such process for outpatient cancer patients. The Center provides symptom-based care for patients who aren’t critically ill, such an antibiotics or fluids if they’re dehydrated, that will keep them out of the emergency department and the hospital.
“It’s a whole new area of focus,” she says. “It’s about being proactive.”
“Gone is the era where data was reported on paper charts and drawing links between one patient and another was done manually. Today all that information is increasingly being captured in computer-readable format in real-time,” Hanson says. “The bottom line: We want to intervene earlier to prevent something bad from happening or provide the necessary resources that lead to better outcomes.”
This article originally appeared in Penn Medicine’s SystemNews.