Abstract
Objective. To assess the potential of using machine learning
and administrative birth data for predicting premature births.
Design. The performance of ordinary least square (OLS) and deep
neural network (DNN) classifiers for predicting low birth weight (LBW)
and preterm birth (PTB) was compared using randomly selected two million
birth records from the US CDC between 2016 and 2018. One million records
from 2016 and 2017 were used to train the classifiers, while another
million records from 2018 were utilized to test them. For hyperparameter
tuning, a grid search with varying numbers of hidden layers, class
weights on positive cases, and thresholds, was undertaken.
Setting and Population: All births in the US Methods:
ordinary least squares regression, deep neural networks Main
Outcome Measures. LBW (<2,500g) and PTB(<37 weeks)
Results. The classifiers generally showed high accuracy and specificity,
however, the DNN classifiers showed much improvement in increasing
sensitivity. Based on the results, the highest sensitivity with
comparable specificity was 0.71 for LBW and 0.65 for PTB.
Conclusion. These findings highlight that a ML approach could
benefit PCHV programs by helping identify mothers with a high risk of
premature birth. In particular, the DNN classifiers with administrative
data can provide accessible solutions for public agencies and nonprofit
organizations providing PCHV services that are not likely to possess
massive clinical data or highly accurate genetic testing equipment.