Widely used sepsis prediction tool is less effective than Michigan doctors thought

A physician reviews a patient's medical record in a hospital.

Sepsis is a severe inflammatory response to infection and affects 1.7 million U.S. adults each year. Common symptoms may include fever, a high heart rate, and a low white blood cell count, which can resemble other conditions. To help identify patients at risk of sepsis early, many doctors use prediction models embedded into a patient’s electronic health record. If a patient shows signs of the inflammatory condition, their medical team receives an alert. However, a study in JAMA Internal Medicine found the Epic Sepsis Model (EMS), a common prediction tool, wasn’t as effective at detecting sepsis as doctors at Michigan Medicine, part of the University of Michigan, originally thought.

To assess the effectiveness of EMS, researchers defined sepsis by using diagnostic criteria from the CDC and Medicare. Their definition varied from the developer’s, which was based on sepsis-related treatment and care. After evaluating the medical records of 27,697 patients during a 10-month period in 2018-2019, which connected to 38,455 hospitalizations, the researchers found 2,552 hospitalizations, 7%, were due to sepsis. The EMS prediction missed 1,709 patients, 67%, who had sepsis even though the algorithm sent out alerts for 18% of hospitalized patients. Overall, the Michigan research team found the EMS was 63% effective at detecting sepsis early compared to the developer’s predictions of 77-83%.

In a corresponding
editorial and podcast interview, physicians discuss how this insight may help the medical community evaluate how prediction models for many conditions can be studied, reviewed, and recalibrated to help doctors detect and treat disease early. The research was supported by the NHLBI.