Prediction of under-5 children mortality in Rwanda using machine learning techniques

Abstract

Controlling under-5 children mortality is crucial as it impacts the future of country’s economic growth worldwide. Under support of the World Health Organization (WHO), regular Demographic Health Surveys (DHS) are conducted in several countries to gather socio-demographic data, including mortality rates. In Rwanda, the Demographic and Health Survey is conducted every five years. The mortality rate for under-5 children remains a concern in Rwanda, for instance, the Rwandan Demographic and Health Survey (RDHS-2019/2020) indicatedthattheinfantmortalityratestoodat33deathsper1,000livebirths,under-5childrenmortalityat45 deaths per 1,000 live births, and child mortality at 13 deaths per 1,000 live births. A deep analysis of this reporthowever,portraysthatthedatasetsbeingextensive,theyfocusedonnewbornmortalityratherthanunder-5 children mortality which is a high concern for the present study. Again, the survey used traditional analysis techniques which really manifested incompetent to predict mortality rates accurately. The present study havingusedmachinelearningmethods,itensuresmoreconventionalandaccuratemethodsinforecastingunder-5 childrenmortalityinRwanda. Inthisstudy,thirteenindependentvariablesrelatedtochildrenmortality(Highest Educational level, Births in Last Five Years, Exposure of the mother, Currently Breastfeeding, Number of Living children, Wealth Index combined, Total children Ever Born, Desire for More children, Sex of child, BirthOrdernumber,NumberofAntenatalVisitsduringpregnancy,HadDiarrhearecently,SourceofDrinking Water) and the dependent variable “child is Alive”, were considered. Leveraging Machine Learning, analysis of this large dataset, containing both numerical and categorical data, was conducted using Python 3.12 and various packages such as Pandas, Matplotlib, and NumPy. Through training and testing the under-5 children dataset,itwasdiscoveredthatamongsevenmachinelearningalgorithmsusedsimultaneously,RandomForest wasthebestpredictor,outperforminglogisticregressionwithanaverageaccuracyof98.2%. ExceptforNa¨ıve Bayes, all classifiers used scored greater than 95%, indicating their suitability for predicting under-5 children mortality. Features such as the number of living children in the home, source of drinking water, and number of antenatal visits were identified as important predictors of under-5 children mortality. Health care providers shouldpayattentiontothesefeaturestoforecastthelivesofchildrenunder-5yearsold. Additionally,SHapley Additive exPlanation (SHAP) values revealed that breastfeeding, number of living children in a family, and birth order were significant factors in predicting under-5 children mortality in Rwanda.