Machine Learning Reveals That Prior Pregnancy Experiences, Race/Ethnicity, and Emotion Dysregulation Predict Newborn Health Outcomes

• In Event: The impact of stress and coping on development: Measurement and mechanisms

Abstract

Background: The United States has roughly six infant deaths per 1,000 live births, and many of these deaths are caused by birth defects and low birthweight (CDC, 2019; Chen et al., 2016). Poor maternal mental and physical heath predispose newborns to risk for aberrant outcomes, including mortality and neurodevelopmental delay (Glover, 2011). Transdiagnostic mental health markers, such as emotion dysregulation, may be especially useful for predicting newborn outcomes (Ostlund et al., 2019). Furthermore, there are numerous obstetric health disparities across racial/ethnic identities and socioeconomic status, suggesting that infant mortality may be socially influenced as well (Lorch & Enlow, 2016).
Aims: We used a robust machine learning approach to predict newborn head circumference and birthweight, potential indices of adverse outcomes (e.g., mortality, chronic illness, psychopathology), with several prenatal maternal health markers, including emotion dysregulation, obstetric health (e.g., BMI, number of preterm births), and potential social determinants of health (e.g., socioeconomic status, race/ethnicity).
Method: We recruited 594 third-trimester pregnant women and their newborns. Data were primarily gathered via verified hospital records. Outcomes were newborn head circumference and standardized birthweight (Oken et al., 2003). Pregnant women also completed the Difficulties in Emotion Regulation Scale (DERS; Gratz & Roemer, 2004); the six subscale and total scores were used as predictors. Other predictors included maternal third trimester BMI and age; binary indicators of psychiatric medication use, smoking, alcohol use, gestational diabetes, preeclampsia, and newborn sex; number of preterm births, abortions, and living children; maternal race/ethnicity (as three binary variables: White and Non-Latina; Latina; Other [e.g., Black/African American, Asian, American Indian]); and coded occupational prestige, a socioeconomic status index (Hout et al., 2016).
Analyses: We ran a series of linear regularized regression models, first for predictor selection and second for model performance. If fit was poor, we used support vector machine models to allow for nonlinear predictions. To determine fit, we examined R2 values and correlations between predicted outcome values on k-fold cross-validated training and withheld testing datasets (James et al., 2013).
Results: See Table 1 for final models. We also created individual conditional expectancy plots for the support vector machine (Figure 1). We found that mothers’ pregnancy experiences and health (i.e., number of children, preterm births, BMI, preeclampsia) predicted both newborn outcomes. Several of the predicted effects on birthweight were nonlinear or unique from person to person. Consistent with prior research (Lorch & Enlow, 2016), mothers’ race/ethnicity also predicted both outcomes. Newborns with Latina mothers had smaller head sizes than other participants. Newborns with mothers who were Black/African American, Asian, American Indian, and other races/ethnicities were smaller in birthweight. However, occupational prestige was the overall best birthweight predictor. Difficulties being clear about emotions also predicted greater birthweight.
Discussion: To our knowledge, this is the first study to use a robust machine learning approach to examine prenatal maternal health and emotion dysregulation as predictors of neonatal outcomes. Results suggest that obstetricians should consider not only physical well-being but also socially-mediated health indicators and transdiagnostic mental health markers to predict newborn outcomes and potential risk.

Authors

• Robert Vlisides-Henry, University of Utah
  Presenting Author

• Sarah Terrell, University of Utah
  Non-Presenting Author

• Marilynn Lape Santana, University of Utah
  Non-Presenting Author

• Lee Raby, University of Utah
  Non-Presenting Author

• Uma P. Dorn, University of Utah; New York University
  Non-Presenting Author

• Mengyu (Miranda) Gao, Faculty of Psychology, Beijing Normal University, China
  Non-Presenting Author

• Parisa R. Kaliush, University of Utah
  Non-Presenting Author

• Marcela Smid, University of Utah Health
  Non-Presenting Author

• Liz Conradt, Duke University
  Non-Presenting Author

• Sheila E Crowell, University of Utah
  Non-Presenting Author