A Deep Learning Model for Predicting Under-Five Mortality in Zimbabwe

Abstract

The death of children before they reach five years old (under-five mortality or U5M) is a global scourge that has attracted the attention of many governments, including the World Health Organisation and the United Nations. Children under-five in Sub-Saharan Africa are disproportionately susceptible to death, with a fifteen-fold likelihood of death compared to their counterparts in developed countries. Regardless of the numerous efforts by the Zimbabwean Government to improve child health, such as free access to care, provision of nutritional supplements, immunisation programmes and prevention of mother-to-child transmission, the country still has high under-five mortality rates (U5MRs). Zimbabwe's failure to reduce U5MRs to acceptable levels suggests that the current methods must be complemented. Identifying contextual risk factors and children at risk of death could help paediatricians to make timely and targeted interventions and policymakers to review existing and craft new policies to save children's lives. Therefore, this study applied deep learning to Zimbabwe's 2019 Multiple Indicator Cluster Survey data to predict under-five mortality and identify its associated risk factors. The study used a deep neural network with four hidden layers, k-fold cross-validation and the stochastic gradient descent (SGD) optimiser. All layers used the Rectified Linear Unit activation function except the output layer, which used the sigmoid activation for binary classification. The model produced a 90.04% accuracy, 92.39% precision, 87.30% recall and 95.04% area under the curve. Though the model predicts under-five mortality, it does not prescribe the appropriate interventions to save lives, a gap that future studies could fill.