ML Model Predicts Prematurity Complications in Newborns Using EMR Data

In a study published this week in Science Translational Medicine, researchers from the Stanford School of Medicine revealed that a machine-learning (ML) algorithm could predict prematurity complications before birth using EMR data.

According to a news release discussing the study, premature birth is typically defined as birth occurring at least three weeks early. The phenomenon is linked to a higher risk of complications in infants’ brains, lungs, digestive system, vision, and hearing. While premature birth is known to carry a higher risk of complications, the timing of the birth can only approximately predict an infant’s health outcomes, with some premature births leading to no complications while others result in severe illness or death.

The press release further indicates that many of these complications only emerge in the days or weeks after birth, by which time they have already caused significant damage to an infant’s health. Identifying at-risk newborns before the emergence of complications could help clinicians prevent adverse outcomes.

“Preterm birth is the single largest cause of death in children under age 5 worldwide, and we haven’t had good solutions,” said senior study author Nima Aghaeepour, PhD, an associate professor of anesthesiology, perioperative and pain medicine and of pediatrics at Stanford, in the press release. “By focusing our research on predicting the health of these babies, we can optimize their care.”

The researchers have developed a type of ML algorithm known as a long short-term memory neural network to evaluate whether EMR data from mothers and babies could be used to predict 24 possible health outcomes for infants up to two months after birth.

“There is a computational challenge in using electronic health records because they are longitudinal and contain a large amount of data from each patient,” Aghaeepour explained. “A long short-term memory neural network operates similarly to a person reading a book. When we’re reading, we don’t remember every word, but we remember the key concepts, read the next part, add more key concepts and carry that forward. The algorithm doesn’t memorize the entire electronic health record of every patient, but it can remember key concepts and carry those forward to the point where we make a prediction.”

The researchers leveraged EMR data from mothers and babies seen at Stanford Health Care and Stanford Medicine Children’s Health between 2014 and 2020, resulting in a cohort representing 32,354 live births.

The mothers’ EMR data included information from their pregnancy and health data from before they became pregnant for those who had been patients at Stanford Medicine prior to pregnancy. Data from infants’ records used in the study included information recorded at birth, such as weight, blood tests, and Apgar scores, a measure taken in the delivery room one and five minutes after birth that evaluates factors such as the infant’s pulse, breathing, and muscle tone.

“We look mainly at the baby to make treatment decisions in neonatology, but we are finding that we can get valuable information from the maternal health record, really homing in on how individual babies’ trajectories have been shaped by exposure to their specific maternal environment,” said study co-author David Stevenson, MD, a neonatologist at Lucile Packard Children’s Hospital Stanford, professor of pediatrics and director of the March of Dimes Prematurity Research Center at the Stanford School of Medicine, in the press release. He called the research “a move toward precision medicine for babies.”

Overall, the ML model provided strong predictions at the time of birth for which newborns would develop multiple serious conditions, including bronchopulmonary dysplasia, retinopathy of prematurity, anemia of prematurity, and necrotizing enterocolitis.

The model also achieved significant prediction accuracy in forecasting outcomes such as retinopathy of prematurity, mortality, and 11 other conditions the week before birth. However, other complications, including candidiasis, polycythemia, and meconium aspiration syndrome, were not reliably predicted.

The researchers validated their findings using an independent group of 12,258 mother-baby pairs from the University of California San Francisco. They found that model predictions at birth provided more accurate information than standard risk assessment tools like the National Institute of Child Health and Human Development risk score and Apgar scores, which consider only the baby's condition at birth and do not take into account the mother’s medical history.

These findings indicate connections between various health and social conditions in mothers and the health of their babies, which warrants additional research in larger, more diverse populations, the research team noted.

“We need to explore what linkages explain these relationships at a biological level, as these might offer clues to how certain conditions occur,” Stevenson stated. “That will allow us to intervene better to help those kids.”