What if artificial intelligence could help us predict the risk for newborns of developing serious diseases? According to the results recently published in the magazine Science Translational Medicine, there would be promising prospects. A group of researchers at Stanford University has developed a system that is based on the data contained in the mother-infant medical record and collected in the months preceding and in the moments immediately following birth. Here’s how it works.
Neonatal mortality
Neonatal mortality is still a worrying issue today: according to WHO (World Health Organization), in 2020 almost half of infant deaths occurred within the first 28 days of life. The most frequent causes are related to congenital defects, infections or other intrapartum complications and, very often, to premature birth. A tool that helps us predict which newborns (especially those born prematurely) are more likely to develop serious diseases could perhaps save more lives.
There is no more time: update newborn screenings as soon as possible
by Simone Valesini
I study
With this aim, the scientists of the Stanford University have “feed” to a neural network system the contents of the medical records of over 22,000 pairs of mothers and newborns, born between 2014 and 2018. In this way it was possible train the system by teaching it to recognize the most probable associations between the various parameters contained in the records (lifestyle, pathologies, drug intake, etc.) and the onset of 24 serious neonatal pathologies, including necrotizing enterocolitis, cerebral palsy , intraventricular hemorrhage and bronchopulmonary dysplasia. The authors then validated the prediction ability of the system on another 10,250 pairs of mothers and newborns, this time born between 2019 and 2020. The neural network was found to be able to predict with high accuracy 10 of the 24 pathologies on which it was trained, and 7 others with a slightly higher degree of error.
The research team also created a interactive website and made it available to independent researchers interested in exploiting the dataset collected during this study for further investigation. “The machine learning methodology employed here – the article reads – has allowed us to build predictive models for neonatal outcomes and will potentially serve as an important resource for clinicians and researchers to independently examine”. However, the authors stress that further validation studies will be needed before their predictive model can be used in the clinical setting.
Breast cancer, artificial intelligence increases the sensitivity of ultrasound
by Mara Magistroni
Limits and future prospects
One of the limitations underlined by the researchers themselves is the fact that the prediction system seems more reliable for some subgroups of infants than for others. That is, the model is able to provide very accurate predictions for newborns possessing a certain phenotype and does not have the same predictive capacity for other subgroups. This is an algorithm bias that could perhaps result from the lack of data for certain subgroups compared to others. Recognizing it – the researchers write – is a first step towards perfecting the model. Thanks to this study, the authors conclude, “we have acquired a greater knowledge of the role of the fetal environment and its contribution to the risk of neonatal diseases. Future prospective studies are now needed to evaluate the clinical impact of the model.”