Comparing the Predictive Accuracy of Machine Learning Algorithms for Neonatal Mortality Risk Classification

Abstract

Neonatal mortality is a major concern in developing countries and is a crude indicator of the prevalence of poverty in a society. Neonatal risk analysis is an important decision-making tool for gauging progress and penetration of medical services in a specific region, especially in developing countries. This research compares the accuracy of the predictions of various machine learning algorithms and the impact of different sampling techniques, when attempting to classify neonatal mortality. We have conducted different rounds of experiments with six sampling techniques and seven machine learning algorithms on a neonatal death dataset. Further, groups of three base classifiers chosen among the seven machine learning algorithms were made into six different ensemble combinations and trained to improve the classification accuracy. The proposed diagnostic support system with an ensemble of random forest classifier, Gaussian NB, and AdaBoost with pipeline sampling achieved an accuracy and ROC AUC scores above 98% and 96%, respectively. This decision support tool will facilitate medical practitioner to identify infants at risk of dying before 28 days of life.