Originally published by our sister publication Anesthesiology News
By Ethan Covey
A novel, integer-based machine learning model may help in early diagnosis of in-hospital sepsis, thereby reducing patient morbidity and mortality, according to abstract P-81.
“Sepsis has long resulted in high mortality, and the earlier you can start antibiotics the more you can decrease mortality of the patient,” said lead study author Kelvin Kan, MD, a resident in the Department of Anesthesiology at the University of Utah, in Salt Lake City.
Due to the detrimental effects of sepsis, the CDC launched the Hospital Sepsis Program Core Elements, which are intended to aid clinicians, hospitals and health systems in managing the condition. Point-based risk scores, such as the Sequential Organ Failure Assessment and for the systemic inflammatory response syndrome (SIRS), can be used to identify sepsis and can be incorporated into live dashboards that track metrics and provide continuous information to nursing, physicians and unit-based leadership.
“Machine learning algorithms have been shown to outperform these traditional integer risk scores in the task of sepsis prediction,” the study authors wrote. “However, machine learning [is] still underutilized compared to point-based risk scores, partly because clinicians find point-based scores easier to use and understand.”
“Our angle is to design a model to identify sepsis on the floor so we can start antibiotics early and transfer the patient to a higher level of care, if needed,” Kan added.
The temporal logistic regression model was trained by using the MIMIC-III data set of 48,000 ICU patient encounters. The model was used to assign mathematical functions to the patient data and extract the parameters of these functions to be used as prediction features. Results were compared with those from the National Early Warning Score, Modified Early Warning Score, quick Sequential Organ Failure Assessment and SIRS.
The team found that the machine learning model had a significantly higher area under the receiver operating characteristic curve than all others when tasked with identifying sepsis at six hours.
“The second part of our study, which we are working on now, will focus on having attending physicians use the model and integrate it into a patient’s electronic medical record in order to see how well it actually functions in real-world scenarios,” Kan said. “We want to see both how well the model detects sepsis and how attending physicians then react to the alerts that the model gives them.”
Kan reported no relevant financial disclosures.