Home - Safety - Algorithm Predicts Onset of Sepsis
Research News

Algorithm Predicts Onset of Sepsis

Researchers develop algorithm that uses machine learning techniques to identify hospitalized patients at risk of severe sepsis

Meeri Kim, Contributor
Mon, 06/19/2017


Machine learning, a type of artificial intelligence that allows computers to learn complex patterns in data without explicit programming, is already embedded in our online lives. Facebook uses machine learning algorithms to rank feeds, ads, and search results. Algorithms from Netflix take input from watched and rated videos to come up with personalized recommendations.

But a different line of machine learning research aims to take these powerful programs out of the online world and into the hospital. Instead of predicting your next favorite movie, scientists in this field are creating algorithms that can flag those patients most at risk for serious conditions.

Sepsis, for instance, is a leading cause of in-hospital mortality that can be difficult to diagnose early. A new study from researchers at the Hospital of the University of Pennsylvania in Philadelphia describes an algorithm that identifies hospitalized patients at risk for severe sepsis and septic shock. The researchers presented their findings on May 24 at the American Thoracic Society 2017 Conference in New York.

The algorithm uses more than 500 variables from patients' electronic health records to make its predictions. The researchers trained it using data from more than 160,000 patients.

“Early intervention to treat these patients saves lives, so creating an algorithm that can identify subtle signs in the electronic health record that point to sepsis may allow us to intervene even faster,” said lead author Heather Giannini, a resident physician at the Hospital of the University of Pennsylvania.

Giannini and her colleagues first provided a training set of electronic health record data such as labs, vitals and demographics from 162,212 patients discharged between July 2011 and June 2014 at three acute care hospitals in the University of Pennsylvania Health System. Out of that initial group, 943 patients met the lab or physiological criteria for severe sepsis or septic shock.

They used a data sorting technique to sift through the large amounts of information in the training set, eventually resulting in an algorithm to predict the onset of sepsis. Variables important to the algorithm included white cell count, heart rate, respiratory rate, temperature, and systolic blood pressure.

To put their program to the test, the researchers then deployed it in real-time “silent mode” at the three hospitals between October and December 2015. During that time, a total of 10,448 patients led to 314 system alerts.

“During the 'silent' period, we knew which patients were screened positive by the algorithm and when, however we did not notify our provider team,” said senior author Craig Umscheid, Associate Professor of Medicine and Epidemiology at the University of Pennsylvania School of Medicine. “We wanted to see how the algorithm performed in real-time prior to going live with an intervention.”

The program correctly identified 98 percent of the people who were later diagnosed with sepsis. It also picked out many people who turned out not to have sepsis, resulting in a sensitivity rate of 26 percent. Next, the researchers will compare a six-month silent period to a six-month period during which the care providers receive alerts from the algorithm.

Samuel Volchenboum, Associate Professor of Pediatrics and Director of the Center for Research Informatics at the University of Chicago, believes the work has the potential to have a high impact in the care of patients. He wasn't involved in the sepsis project, but with his colleagues, he built a predictive model called eCART that identifies patients most likely to suffer cardiac arrest in hospital wards.

“This is an interesting abstract. It was powered by a very large data set over three years,” said Volchenboum. “With the caveat that I have not seen their actual results with confidence intervals and their fully adjusted model, this does seem like an intriguing study.”