Predicting the risk of spontaneous premature births using clinical data and machine learning

Abstract

Background:: Spontaneous preterm birth (sPTB) is a worldwide public health issue that affects millions of infants per year and causes long-lasting effects. Prediction of sPTB is critical for clinical management and patient referral to centers capable of treating preterm infants. The outstanding capabilities of machine learning to detect patterns and make predictions have motivated recent studies to improve the prediction of sPTB. However, there is still a high level of uncertainty, especially among patients without a prior sPTB. Objective:: The objectives of this study were to formulate more effective and accurate predictions of sPTB and to identify relevant sPTB risk factors. Methods:: From the data set collected by the NICHD Maternal Fetal Medicine Units (MFMU) Network between 1992 and 1994, we selected a cohort of women with no previous preterm birth (n = 2390). We then developed a pre-processing protocol that prioritized original information by limiting interpolation criteria to individual feature averages, reducing complexity and cross-feature correlation to increase generalization. Next we created a machine learning model based on support vector machines and a radial basis function kernel to: (i) predict the occurrence of spontaneous preterm birth, and (ii) accommodate clinical considerations for treatment. Results:: In clinically conservative, moderate, and aggressively models, prediction of sPTB reached average true positive rate of 0.36, 0.62, and 0.82, respectively, and corresponding average false positive rate of 0.03, 0.38, and 0.56, respectively in an independent testing partition. The area under the receiver operating characteristics curve (AUC) was 0.75. The models benefited most from features collected in the earlier stages of the study protocol and pregnancy, generally ¡26 weeks gestational age, and involved a combination of socioeconomic and biological contributors identified as risk factors in previous studies. Conclusions:: Our findings provide a framework for improving clinical decision support system models to assist providers in predicting sPTB. Updating such models with more recent data that include multiple levels of individual and contextual factors could prove useful to predict sPTB.

Keywords

• Spontaneous preterm birth
• Machine learning
• Nulliparous women