Machine Learning Tool Outperforms Clinicians in Identifying Sepsis in Infants
The researchers implemented eight machine learning models to attempt to identify infant sepsis at least four hours prior to clinical recognition, and the models were tasked with classifying input data from control and case windows as “sepsis negative” or “sepsis positive.”
Six of the eight models accurately predicted sepsis as many as four hours sooner than clinicians.
>> READ: How Machine Learning Predicted Post-Cardiothoracic Surgery Complications
The team studied infants hospitalized for at least 48 hours in the Neonatal Intensive Care Unit at CHOP between September 2014 and November 2017 who received at least one sepsis evaluation before 12 months of age. From CHOPS’s registry that is automatically populated with data abstracted from the EHR, the team identified 618 unique infants with 1,188 sepsis evaluations that met the criteria.
Of the 618 infants, 92 were culture positive, 199 had clinical sepsis and 492 had negative evaluations.
Through literature and expert review, the team identified 36 features collected in the registry that could be associated with infant sepsis. The features were grouped under vital signs, laboratories values, co-morbidities and clinical factors, and were extracted from EHR entries to provide input data for the machine learning models. Among the features were weight, glucose, temperature and heart rate.
It was also determined that the presence of an underlying chronic medical condition — potentially related to immune dysfunction and impaired resistance to bacterial pathogens — increases the risk for sepsis in children and adults.
Due to the increased risk, the baseline risk for sepsis might also be higher among infants who have experience co-morbid conditions such as necrotizing enterocolitis, prolonged ventilation for chronic lung disease or surgical procedures. Indicator variables for the presence of comorbidities were included in the analyses.
Based on the findings, the team learned that the models that used the input features derived from the data collected in most EHRs can predict sepsis in infants hospitalized in the Neonatal Intensive Care Unit hours prior to clinical suspicion.
Sepsis is a major cause of infant mortality and morbidity nationwide. And those who survive infant sepsis have an increased chance of suffering long-term problems such as lung disease, neurodevelopmental disabilities and prolonged hospital stays.
Delays in recognizing sepsis can lead to delays in intervention, including antibiotic treatment and supportive care.
“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,” said first author Aaron J. Masino, Ph.D., department of biomedical and health informatics at CHOP.
The team suggests that further research is needed to find potential performance improvements and clinical efficacy in another trial, but the researchers believe that their approach is generalizable and has the potential to be adopted at many institutions.
Get the best insights in healthcare analytics directly to your inbox.
This year at HIMSS, I connected with Bridget Duffy, M.D., chief medical officer of health IT communi
A team of researchers at the Children’s Hospital of Philadelphia (CHOP) used electronic health recor