Banner Health algorithm unintentionally forecasts which patients are most likely to die

In 2009, a team of researchers at Phoenix-based Banner Health set out to create an algorithm that would warn providers when patients were most at risk for contracting sepsis. The algorithm didn't work, but the failure had an unexpected result, reports BuzzFeed News.

To create the algorithm, the team, led by Banner Health's senior director of health management Hargobind Khurana, MD, who joined in 2011, developed a list of common symptoms of sepsis and organ dysfunction, including high heart rates, unusually high or low body temperature and off-balance chemical levels in a patient's bloodstream. Relying on EHR and medical device sensor data, the algorithm would initiate an alert whenever, within eight hours of each other, a patient would show at least two out of four main symptoms of sepsis and one out of 14 symptoms of organ dysfunction.

Many Banner Health providers doubted the reliability of the algorithm. "There was certainly a lot of skepticism, especially from those who actually answer the alerts, because of the extra workload it would bring on our shoulders," said Nidhi Nikhanj, MD, a physician at Banner Health who helped create the alert, according to the report.

Nevertheless, between April 2011 and June 2013, the algorithm scrutinized the patterns of more than 312,000 patients in Banner Health's 24 hospitals.

The results? Only about 25 percent of the patients the algorithm identified actually had sepsis. But all the flagged patients had one thing in common: They were generally sicker — in fact, much sicker — than the non-flagged patients.

The patients identified by the algorithm were more likely to have chronic medical conditions and to remain in the hospital twice as long as non-identified patients. The algorithm also specifically pinpointed a group of patients who made up almost 90 percent of deaths across Banner Health's hospitals. Patients who set off the alert were four times more likely to die the next day compared to those who didn't set it off.

The results were so striking to Dr. Khurana and his team that Banner Health has continued to use the algorithm. Instead of determining which patients had sepsis, the alert now prompted physicians to ask themselves whether they should be analyzing a patient's care more closely.

Dr. Khurana and his team published their research and results in the American Journal of Medicine this past May. Although they still want to make a sepsis-identifying algorithm, they tweaked their original algorithm to better pinpoint the patients who have a higher chance of dying.

More articles on health IT:
The best, worst states for EHR adoption, HIE, interoperability: 4 findings from ONC
Study finds high variance of errors in pediatric CPOE systems
Russian hackers leak more Olympians' medical records

© Copyright ASC COMMUNICATIONS 2017. Interested in LINKING to or REPRINTING this content? View our policies by clicking here.

 

Top 40 Articles from the Past 6 Months