Predicting preterm births using US national birth data: a deep learning approach

Shuying Zhu
E-mail: zshuying@connect.hku.hk
University of Hong Kong

Team: Shuying Zhu, Andrew Dellinger, Karen Chan, Wendy Lam, and Herbert Pang

Purpose
Preterm birth is the leading cause of neonatal morbidity and mortality globally. Deep learning has been shown to perform well in big biomedical data.National preterm birth prediction using deeplearning remains to be explored.We aim to develop a deep learning model for preterm birth predictionbased on US national birth data.

Methods Artificial neural networks with modern deep learning techniques were developed using national birth data (n = 17,378,139) of women who began theirfirst prenatal visit no later than week 24of pregnancy in the United States between 2014 and 2018 to predict preterm births. For the prediction,26 and 34 routinely collected variables that are known before the 24 weeks of pregnancy were used fornulliparouswomen and multiparous women, respectively. Training, validation, and test data sets wereused to build models and assess performance. Predictive performance was assessed by AUC. Theproposed models were compared with logistic regression and other machine learningmethods.Additional analysis was performed to identify important variables.

Results
The proposed models achieved an AUC of 0.682 (95% CI: 0.680, 0.684) for nulliparous women, and AUC of 0.766 (95% CI: 0.764, 0.767) for multiparouswomen. The AUCs of the proposedmodel were higher than those of the logistic regression, random forests and XGBoost methods.

Conclusions
The utilization of deep learning techniques on a big national birth data set demonstrates an improved ability to predict preterm births comparedto traditional statistical and machine learningmethods. Similar models with further refinements could be incorporated into routine health informationsystems to allow better risk assessments and management of pregnant women.