Neonatal Disease Prediction using Machine Learning Techniques

Abstract

The newborn baby is defined as an infant in the first 28 days of life after birth. In the world, neonatal mortality is a big problem that accounts for the lion’s share of under-five mortality. Neonatal diseases are one of the deadliest diseases for children under five in the world. Countries and health organizations make a great effort to eradicate this global burden of disease. However, some obstacles have a hindering effect on the fight against the disease. This lower number of neonatal health professionals, especially in sub-Saharan countries such as Ethiopia, is the main obstacle to increasing the accessibility of diagnosis for all neonatal patients. Most neonatal health professionals decide the type of illness only on clinical diagnosis (oral questions and answers). This method often led to misdiagnosis because the health professional may not have a complete picture of all the variables that have a contributing effect on neonatal disease as well as, and some health professionals may curiously not diagnose all newborns due to workload. In addition, neonatal diseases often have similar symptoms, which also led to misdiagnosis and high attention. These problems can be minimized with the implementation of a predictive system that can take all the necessary data from the patient and classify the most likely neonatal disease. The availability of relevant historical data and emerging technologies such as machine learning and data science can address this type of problem. Therefore, this study proposed a predictive model of specific neonatal diseases based on data from neonatal patients collected at the Asella Compression Hospital. The newly developed model stacking suite was compared to a machine learning algorithm such as XGBoost (XGB), Random Forest (RF), and support vector machine (SVM) with and without recursive feature removal with cross-validation (RFECV). under cross-stratified k fold validation. The results of the performance evaluation showed that the newly developed model stacking ensemble with the RFECV method performs better than other models, which obtained an accuracy of 97.04. The proposed neonatal predictive model classifies a common neonatal disease based on a multi-class prediction approach to help domain experts