Using Machine Learning to Develop a Predictive Model of Infant Hypoglycemia Based on Maternal and Infant Variables in an Electronic Health Record

All newborns experience low blood glucose levels when they first initiate carbohydrate metabolism. Some levels remain low, with potential seizures and severe brain injury. Predicting newborns at higher risk is clinically useful because newborns can have their blood sugar raised with breastfeeding, donor milk, formula, or oral dextrose gels. Additionally, informing parents of this higher risk can enhance shared decision-making in the first 48 hours after birth. To address this, we propose three predictive models using binary logistic regression for newborns receiving treatment with oral dextrose gels for hypoglycemia. The first is a parsimonious model, where a high-risk newborn's first blood glucose value is highly predictive of requiring an oral dextrose gel treatment. The second model can be used earlier in the clinical workflow. It is based on the most predictive variables that are also electronically available for all newborns and do not change much in the electronic health record. The third model explores the most predictive variables based on a conceptual model of factors associated with health disparities. These three models are informed from insights gleaned by an exploratory analysis of alternative outcome measures, variables, and threshold cutoffs using a standard heuristic of greedily finding the highest average difference for records on both sides of partitions. We discuss how the dynamics of when data are available during a hospital stay in the postnatal care unit for all patients impact the selection of useful variables for electronically-based decision support. We plan to modify handouts for postnatal care nurses that detail treatment guidance and support shared decision-making. We plan to embed stratified guidance, recommended scripts for high and low-risk cohorts, orientation materials for float and junior nurses, and patient-facing educational materials.