ML models predict sepsis in neonates hours before clinical recognition
Using routine electronic health record data, researchers at the Children’s Hospital of Philadelphia have developed machine learning models able to identify infants with sepsis hours before clinical recognition.
The research team conducted a retrospective case-control study that leveraged EHR data from 618 infants hospitalized in CHOP’s neonatal intensive care unit from 2014 to 2017.
In all, eight machine learning models were evaluated for their ability to analyze patient data to predict which infants had sepsis. Of the eight, six models—AdaBoost, gradient boosting, logistic regression, Naïve Bayes, random forest and SVM—performed well in accurately predicting sepsis as much as four hours before clinical recognition of the life-threatening condition.
“We found that machine learning models that utilize input features derived from data collected in most EHRs can predict sepsis in infants hospitalized in the NICU hours prior to clinical recognition,” state the authors of the a study published last week in the journal PLOS ONE.
Sepsis is a complex syndrome, caused by the body’s response to an infection, and is one of the most frequent causes of hospital deaths. However, there is no single confirmatory diagnostic test, and the condition tends to be under-recognized by clinicians.
“To our knowledge, this was the first study to investigate machine learning to identify sepsis before clinical recognition, using only routinely collected EHR data,” said Aaron J. Masino, who led the research team’s machine learning activities and is an assistant professor in CHOP’s Department of Anesthesiology and Critical Care Medicine, and a member of the Department of Biomedical and Health Informatics.
The researchers noted in their study that the machine learning models “would almost definitely require retraining for non-NICU settings, however the method should still be applicable.”
At the same time, they contend that “further research is warranted to assess potential performance improvements and clinical efficacy in a prospective trial.”
“Follow-up clinical studies will allow researchers to evaluate how well such systems perform in a hospital setting,” added Masino. “If research validates some of these models, we may develop a tool to support clinical decisions and improve outcomes in critically ill infants.”