AI Tool Can Identify Sepsis Within 12 Hours of Hospital Admission

A new study published in JAMA Network Open assesses an artificial intelligence (AI) tool that can predict the likelihood of patients developing sepsis and the severity of the infection as quickly as 12 hours after hospital admission.

According to the Centers for Disease Control and Prevention (CDC), 1.7 million adults in the US develop sepsis in a typical year, and 270,000 die. One in three patients who die in the hospital has sepsis, but the condition, or the infection that causes it, starts outside the hospital in 87 percent of cases. Some patients who survive sepsis also develop permanent organ damage and, as a result, have a higher risk of other medical conditions for several years after recovering.

Identifying the risk of sepsis quickly is key to reducing serious adverse outcomes such as septic shock, which can result in multiple organ failures and death. But recognizing and managing sepsis early remains a challenge.

“There is no consistent way of recognizing and triaging critically ill patients when they’re admitted to the ICU,” said Lyle L. Moldawer, PhD, director of the University of Florida’s Sepsis and Critical Illness Research Center and emeritus director of the Laboratory of Inflammation Biology and Surgical Science, in the press release. “While this may not pose a problem at large academic institutions with dedicated specialists, it can be harder for places where tertiary care is less developed.”

Clinicians who treat critically ill patients must contend with two considerations, Moldawer continued. The first is if the patient is likely to have a challenging clinical trajectory, requiring more aggressive interventions and closer monitoring. The second is how the clinician, under these circumstances, can determine the best type of treatment for the patient.

“Sepsis is a very heterogeneous disease,” said Scott Brakenridge, MD, first author of the study and trauma surgeon at the University of Washington, in the press release. “People’s immune systems react in different ways to infection and display different levels of illness. In fact, one of the main reasons that finding effective therapeutics to treat sepsis has been so challenging is due to this variation among patients.”

To combat these challenges, the researchers developed an AI algorithm that could identify sepsis risk and potential severity. They did this by utilizing the health data of 200 critically ill adult patients admitted with suspected sepsis (cohort A) or those at high risk for developing sepsis (cohort B) admitted to the ICU between July 1, 2020, and July 30, 2021.

The diagnostic tool relied on neural networks, which had been previously developed to use the results from whole-blood RNA transcriptomic metric tests, to generate scores representing the likelihood of bacterial infection or 30-day mortality. In this case, 30-day mortality was used as a metric of sepsis severity.

The tool achieved significant accuracy in estimating both the presence of bacterial infection and the risk for 30-day mortality. Compared to several alternative diagnostic and prognostic metrics, the tool achieved equivalent or better performance when measured at hospital admission. It also provided additional information over time that the other methods did not.