Communications of the ACM
Predicting Preterm Birth
"There is an opportunity for machine learning to identify novel risk factors of preterm birth," says You Chen, an assistant professor of biomedical informatics at Vanderbilt University Medical Center in Nashville, TN.
Credit: raisingchildren.net.au
Preterm birth, when an infant is born before 37 weeks of pregnancy, is the leading cause of death and disability in newborn babies. In the U.S., almost 10% of babies are born prematurely, many requiring special medical care resulting in hospital stays longer than babies born full-term. Although therapeutic interventions can help prevent mortality in premature infants, doctors often are unable to predict preterm birth early enough to use them. In addition, screening tests that measure cervical length, or a protein called fibronectin that is thought to keep the amniotic sac glued to the uterus lining, are not always effective. Even though several risk factors have been identified, including diabetes, obesity, or multiple pregnancies, in many cases there is no known cause for preterm birth.
"The biological mechanism behind preterm birth is still a mystery," says You Chen, an assistant professor of biomedical informatics at Vanderbilt University Medical Center in Nashville, TN.
As a result, machine learning is of great interest to better predict preterm birth. Algorithms are able to learn from large amounts of data, such as the past health records of pregnant women, and could help determine which variables contribute the most to premature babies. "There is an opportunity for machine learning to identify novel risk factors of preterm birth," says Chen.
Research by Chen and his Vanderbilt University colleagues aimed to investigate whether deep learning could provide better predictions of extreme preterm birth, when a baby is born prior to 28 weeks of gestational age. These very early births carry the highest risk, accounting for the majority of infant deaths. "We wanted to validate our assumption that patient historical health information contains evidence that can indicate extreme preterm birth," says Chen.
The team used data from the electronic health records of 25,689 women who had delivered a baby at the Vanderbilt University Medical Center over a 12-year period. The dataset contained information such as health conditions, medications taken, surgeries, and lab test results. Extreme preterm births accounted for about 1% of deliveries.
Chen and his colleagues then trained a deep learning recurrent neural network with part of the data, while retaining a subset for testing. They also trained a few traditional machine learning models with the same data, so they could compare the results.
The team found its deep learning model performed better; it was able to use the data to predict which cases would result in preterm birth, and it was accurate 96% of the time. Since deep learning models learn on their own, they are often able to identify novel patterns in data that can help with predictions. In this case, the model identified four known risk factors: twin pregnancy, systemic lupus erythematosus (an autoimmune disease), short cervical length, and hypertensive disorder, as well as a potentially novel fifth: the presence of hydroxychloroquine sulfate, an antimalarial drug used to treat an autoimmune disease associated with preterm births.
Chen emphasizes this work is just a first step. The data his team used came from a single site, so it is unclear whether the model would be able to correctly predict a preterm birth at a different hospital, for example. In addition, the model has not been tested with data from different time periods, such as five or 10 years later, to see if it is still valid. "We need to test the model in different places and in different timeframes," says Chen.
The team plans to follow up by combining data from different health institutions and using it to train their model. Chen thinks if these types of models can be proven to be generalizable, they could be used to support doctors when assessing the risk of a preterm birth. "I think in the next five or 10 years, AI models will be applied in clinical practice for (predicting) preterm birth," says Chen.
Another team also is investigating how machine learning can help improve preterm birth predictions. In a recent study, Irene Díaz, a professor of computer science at the University of Oviedo in Spain, and her colleagues aimed to identify novel factors that may contribute to premature births. They focused on chronodisruption, the alteration of circadian rhythms possibly due to exposure to artificial light after dark, which some studies have found can be a factor in preterm births, and how it might contribute when combined with known risk factors. "Our goal was to produce a system to help doctors prevent preterm births by identifying key variables," says Díaz.
The Oviedo team used a dataset with information on 380 births at the Central University Hospital of Asturias in Spain, or which 157 were preterm. The dataset contained details about the women, such as their age, weight, and whether it was their first pregnancy. Díaz and her colleagues also collected information about the women's sleeping habits through a telephone survey, with some questions related to light exposure before going to bed or during the night to determine whether they suffered from chronodisruption.
The researchers performed statistical analyses to determine how different factors might relate to preterm birth, then trained several machine learning models with the data. The dataset was randomly split into a training and testing subset, with the process repeated 30 times to cross-validate the data and help account for any noise.
The Oviedo team's results suggest that factors related to light exposure at night play a role in preterm birth. Their models found that using electronic devices before sleeping, and thus being exposed to artificial light after dark, was a risk factor when combined with a high body mass index. Similarly, shift work also seemed to play a contributing role in preterm births due to light exposure at night, which can disrupt circadian rhythms.
Díaz and her colleagues plan to train and test their models with a larger dataset. They are working with other hospitals in Spain to obtain data, and hope to include health records from other countries as well. "We are currently making the dataset larger to produce a more general conclusion," says Díaz.
If exposure to artificial light at night continues to be identified as a contributing factor, it suggests doctors should recommend pregnant women be mindful of their nighttime habits to reduce the risk of preterm birth. Díaz thinks these models could be an important tool for doctors if used in conjunction with their expertise. "(Machine learning) methods can help to save lives and improve health," she says.
Sandrine Ceurstemont is a freelance science writer based in London, U.K.
No entries found