Time-series prediction of adverse birth outcomes in the U.S. using multilayer perceptron neural networks

Adverse birth outcomes (ABOs), including preterm birth, low birth weight, low Apgar scores, and neonatal mortality, remain major public health challenges in the United States and disproportionately affect racial, socioeconomic, and demographic subgroups. A time-series multilayer perceptron (MLP) neural network was developed to predict ABOs and identify populations with elevated risk using U.S. Vital Statistics Natality Birth Data from 2009 to 2023, covering 57.9 million births. Unlike standard methods such as ARIMA, the MLP model captures nonlinear relationships among predictors and learns seasonal patterns directly from the data. Key predictors included maternal age, body mass index, adequacy of prenatal care, education, race/ethnicity, and short interpregnancy intervals. Monthly aggregated ABO rates were modeled using data from January 2009 to December 2022 (168 monthly observations), while data from 2023 (12 monthly observations) were reserved as an independent holdout test set. Predictive performance on the holdout set was assessed using root mean square error (RMSE = 0.181), and forecasts were projected through 2030. Between 2009 and 2023, ABO prevalence increased from 14.9% to 15.5%, with Black and American Indian/Alaska Native mothers consistently exhibiting the highest rates. Pregnancy-related medical conditions, advanced maternal age, underweight status, low educational attainment, and short birth intervals were among the variables most strongly associated with model predictions. Projections suggest continued increases in ABO risk among underweight mothers and American Indian/Alaska Native populations. Performance was lower for smaller or structurally disadvantaged subgroups, highlighting challenges in predicting outcomes for high-risk populations. These findings demonstrate that time-series MLP models can support forecasting of adverse birth outcomes and help identify groups experiencing elevated risk. Predictive models may support public health surveillance, maternal health planning, and resource allocation, although forecasts should be interpreted cautiously and are not intended for individual-level clinical decision-making.