The Infant Mortality Rate (IMR) is defined as the number of infants for every thousand infants that do not survive until their first birthday. IMR is an important metric not only because it provides information about infant births in an area, but it also measures the general societal health status. In the United States of America, the IMR is higher than many other developed countries, despite the high level of prosperity. It is important to note here that the U.S.A. exhibits strong and persistent inequalities in the IMR across different racial and ethnic groups (Kochanek et al. in Natl Vital Stat Rep 65(4):1–122, 2006). In this paper, we study predictive models in the problem of infant mortality. We implement traditional machine learning models and state-of-the-art neural network models with various combinations of features extracted from birth certificates. Those combinations include features that can be summed as socio-economic and ethical features related to the mother and the father of the infant and medical measurements during the pregnancy and the delivery. We approach the classification problem of infant mortality, whether an infant will survive until her first birthday or not, both as binary and multi-class based on the time of death. We focus on understanding and exploring the importance of features extracted from the birth certificates. For example, we test the performance of models trained on the general population to models trained in subsets of the population, e.g., for individual races. We show in our experimental evaluation comparisons between different predictive models (including those used by epidemiology researchers), various combinations of features, different distributions in the training set and features’ importance.
