Explainable feature learning for predicting Neonatal Intensive Care Unit (NICU) admissions

Abstract

Neonatal Intensive Care Units (NICU) service costs are rapidly growing due to the higher resource utilization intensity. This in turn increases the healthcare costs for NICU patients besides the inaccessibility and unpreparedness of both NICU service providers and patient caretakers hence an increase in neonatal mortality and morbidity. There a lot of contributors to NICU admissions but the exiting methods consider very limited features to precisely predict NICU admissions. In this paper, we present a robust Explainable Artificial Intelligence approach that allows machines to interpretably learn from a pool of possible contributing features in order to predict an NICU admission. Our machine learning approach interpretably illustrates the thought process of admission prediction to the physician and patient. This provides transparent and trustable insights for the precise, proactive, personalized and participatory NICU medical diagnostics and treatment plans for the patient. We statistically and visually present Random Forest and Logistic Regression prediction explanations using SHAP, LIME and ELI5 techniques. This predictive technological approach can preventively increase success of maternal and neonatal monitoring and treatment plans. It can also enhance proactive management of NICU facilities (resources) by the responsible facility administrators most especially in resource constrained settings.