Researchers at the Children’s Hospital of Philadelphia (CHOP) developed machine learning models that can detect the presence of sepsis in infants, hours before physicians. Findings from the study were published in PLOS One.
Sepsis is a major cause of illness and death among infants, and incidence is 200-fold higher in infants who are born prematurely or who are chronically hospitalized. Prematurely born infants experience the highest rates of death and many of them incur major “long-term impairments.”
Rapid diagnosis is necessary to avoid further illness like organ failure and possible deaths. However, impacted children often exhibit ambiguous clinical signs which mimic other diseases. And screening laboratory tests also have limited diagnostic accuracy in this cohort.
Recent research has shown sepsis deaths are likely unpreventable with better hospital care; however, there are AI-based sepsis screening tools that help physicians find patients who are most at risk of developing sepsis.
In this first-of-its-kind study, the researchers, led by Aaron J. Masino, PhD, of the University of Pennsylvania and CHOP, sought to develop a model using electronic health record (EHR) data capable of recognizing infant sepsis at least four hours prior to clinical recognition.
“Because early detection and rapid intervention is essential in cases of sepsis, machine-learning tools like this offer the potential to improve clinical outcomes in these infants,” Masino said in a prepared statement.
Masino et al. evaluated how well eight different machine learning models could analyze EHR data from 618 infants who were admitted to the CHOP neonatal intensive care unit (NICU) between 2014 to 2017. Many of the infants in the cohort were premature and the cohort had an average gestational age of 34 weeks.
They created a list of 36 features that are associated with or could be associated with infant sepsis. The features, which were extracted from EHR notes, were grouped under vital signs, laboratory values, comorbidities and clinical factors.
The researchers found six (AdaBoost, gradient boosting, logistic regression, Naive Bayes, random forest and SVM) of the eight machine learning models were able to predict sepsis up to four hours before clinical recognition.
Because this was a retrospective study, the researchers were able to compare each model’s predictions to subsequent findings, or whether or not a patient developed sepsis.
"Follow-up clinical studies will allow researchers to evaluate how well such systems perform in a hospital setting,” Masino said.
Masino noted their algorithms are “a preliminary step toward developing a real-time clinical tool for hospital practice.” The researchers will continue their studies to improve their models and investigate their algorithms in a clinical study.
“If research validates some of these models, we may develop a tool to support clinical decisions and improve outcomes in critically ill infants,” Masino concluded.