According to the CDC, one of every three patients who die in a hospital has sepsis: a result of the body’s natural defenses against infection. Each year 1.7 million adults develop sepsis and 270,000 people die from from the condition annually. The University of California – San Francisco Medical Center focuses on research in sepsis detection in order to treat patients early to increase the chance of survival.
Working on the assumption that machine learning could detect sepsis in a timely manner by using electronic health records (EHR), UCSF researchers pulled data from the UCSF Medical Center and the Beth Israel Deaconess Medical Center in Boston. The records included in the study were of adult patients who did not have sepsis on admission and for whom there was at least one recording each of six vital signs: SpO2, heart rate, respiratory rate, temperature, and systolic and diastolic blood pressure.
The UCSF machine learning algorithm (MLA) analyzed the data from USCF to predict the onset of sepsis. The algorithm’s performance then was compared to the predictions of four systems commonly used to detect sepsis: SIRS, MEWS, SOFA, and qSOFA. The outcome of the research showed the MLA was significantly superior to the standard tools in predicting sepsis up to 48 hours in advance of onset. After the MLA trained on the USCF data, the system was tested using the BIDMC data to compare its predictions against the actual outcomes. The algorithm performed at an even higher level. The USCF team’s study was published in Computers in Biology and Medicine.
The MLA’s success in early sepsis detection has the potential to save many lives through early detection and intervention. The MLA’s performance when trained with one dataset and tested with a different population suggests further-reaching value than if all of the training and testing were done with a single dataset. This example of machine learning algorithm success opens potential discussion and consideration for applications with different diseases and conditions.