Around 10% of all live births, about 15 million babies per year, are preterm; that is, they happen before 37 weeks of gestation are completed. Preterm births are a leading cause of newborn mortality. Moreover, many preterm babies suffer from long-term morbidity, including permanent neurological damage. Because treatments can delay preterm births and improve their outcomes, identifying pregnant mothers at high risk of preterm birth is compelling, as the World Health Organization (WHO) recognizes. 

Although several methods can predict preterm births, they have limitations. Broad historical risk factors, such as previous preterm births or multiple gestations, can identify mothers at higher risk of preterm birth. Still, these risk factors alone cannot accurately predict which individual mothers will deliver preterm. 

Common technique 

In clinical practice, preterm birth is usually predicted by measuring cervical length or the concentration of cervicovaginal fibronectin alpha. In mothers with symptoms of preterm labour, these minimally invasive tests can predict births that will occur within one week. Moreover, the combination of these tests has been reported to produce more accurate results than each method. It could be used to predict preterm births in symptomatic mothers within two weeks of testing.  

These measurements are helpful because they inform physicians and guide treatments to reduce the risk of preterm labour and improve outcomes. However, these measurements are not cost-effective screening tools for the general population of pregnant mothers because they have low predictive values among mothers at low risk for preterm labor, such as nulliparous women with singleton pregnancies. 

The advent of AI 

Researchers from the Department of McKelvey School of Engineering, Washington University, are developing a deep learning model for predicting preterm births from electrogastrogram recordings. Their recently published report shows that their model is not sensitive to varying implementations of specific features or how uterine contractions are segmented. 

The researchers developed their work using EHG measurements and supplementary clinical information from two public databases. Importantly, the models were developed with care to avoid data leakage. Their models could predict births in pregnant mothers around their 31st week of gestation.  

The model's predictive accuracy was close to that achieved by using cervical length and fibronectin alpha measurements to predict preterm labor in mothers with symptoms of preterm labor and within one week of delivery.  

In the report, the researchers stated that "by investigating the measurement components that contribute to the predictions of our model, we showed that it is possible to predict preterm births using short recording times, thus facilitating clinical adoption and at-home implementation of EHG measurements".  

According to the researchers, this finding is aligned with the observations of Jager et al., who proposed that preterm births can be predicted from short contractile or noncontractile intervals of EHG measurements with similar accuracy as when using 30-minute-long recordings. 

The results 

"Our work and results encourage using EHG measurements and deep learning for predicting preterm births in real-world scenarios. Their successful employment could help reduce newborn morbidity and mortality, especially in populations with limited access to healthcare, who suffer more from preterm birth", stated the published report. 

The methods used by the researchers predicted preterm births more accurately than existing technologies. They also showed that preterm births can be predicted using short EHG recordings. Their work and results are useful for developing applications to predict preterm births early during pregnancy and ultimately improve their outcomes. 

Report - "Predicting preterm births from electrohysterogram recordings via deep learning."