Babies born too early often struggle to survive. But doctors can have a hard time telling which preemies are going to develop serious health problems, such as respiratory failure, and which ones are going to be fine. Now, researchers have developed a model that can predict preemie outcomes with greater than 90% accuracy—an advance that could help doctors identify the sickest babies and save billions of dollars in health care costs.
Fifty years ago, physician Virginia Apgar of Columbia University developed a scoring system to rate a newborn’s health. The Apgar score—still the standard method—takes into account factors such as whether a baby flexes its arms and legs or lies still, breathes well or not at all, and whether its skin is healthy and pink or bluish. When it comes to predicting serious illness such as bleeding in the lungs, however, the Apgar score is right only about 70% of the time. New models that factor in white blood cell count and blood pH do better, but “they require a lot of invasive testing,” says Anna Penn, a neonatologist at Lucile Packard Children's Hospital (LPCH) in Palo Alto, California.
The researchers, including Penn and co-senior author Daphne Koller, a computer scientist at Stanford University, set out to develop a more accurate yet noninvasive tool for predicting major complications in the tiniest newborns. The researchers selected 138 infants born at LPCH who spent less than 35 weeks in the womb and weighed less than 2 kg. The team classified the preemies as high or low risk based on the illnesses they developed. Babies in the high-risk group died or developed serious complications such as infections, bleeding, and lung and heart problems. Babies in the low-risk group suffered only minor ailments, such as slight respiratory distress.
Next, the researchers examined physiological data routinely collected in the first 3 hours of life by bedside monitors, such as heart rate, respiratory rate, and the amount of oxygen in the blood. When they modeled this data, they observed signatures in sick babies that were different from the ones they observed in healthy ones. They used these differences to develop a mathematical algorithm that incorporates physiological data from the monitors, birth weight, and length of time spent in the womb to predict the probability that a preemie will develop serious illness. "These are very simple measures," says Penn. "But when combined using the sophisticated tools that come from computer science, we can actually make sense of these in a way that physicians don't normally."
The output of the model is a number between 0 and 1, which the researchers call "PhysiScore." A higher score indicates a greater risk of complications. For example, an infant with a score of 0.8 would have an 80% chance of developing a serious illness.
PhysiScore outperformed not only the Apgar scale but also three models that rely on invasive laboratory tests, the team reports online today in Science Translational Medicine. Using PhysiScore, the researchers were able to predict serious complications with an accuracy of between 91% and 98%. The Apgar score's accuracy ranged from 70% to 74%, and the other models had accuracies from 82% to 91%.
The researchers envision monitors that could compute and display a baby's PhysiScore automatically 3 hours after birth. This number could help physicians decide whether the baby should receive more aggressive care or be transferred to a better-equipped hospital. "[The monitors] are already measuring all of these signals," says Suchi Saria, a computer scientist at Stanford University who led the work. So this would be a way of "utilizing existing resources to make better use of the data that's already collected," she says.
"This is a huge advance in the field," says Rosemary Higgins, a neonatologist at the National Institute of Child Health and Human Development in Rockville, Maryland. "Predicting the outcome of preterm babies is a major challenge for physicians." Still, she would like to see how the model fares in the tiniest preemies—those weighing less than 1 kg. "That's the group really at highest risk for major developmental problems," she says.
Namasivayam Ambalavanan, a neonatologist at the University of Alabama, Birmingham, says physicians often use their judgment to identify preemies that will fare poorly. He would like to see a study that pits PhysiScore against clinical judgment.
Penn says the same techniques used to create PhysiScore might also work to identify high-risk surgical patients or adults most likely to suffer complications following a heart attack. "One of the most interesting things will be to see if we can apply this modeling to other settings," she says.This article identified Suchi Saria of Stanford University as a co-author on the study. She led the research. The text has been changed to reflect this.