Predicting Neonatal Mortality in Kenya Based on the 2022 Kenya Demographic and Health Survey Using Machine Learning Models
John Kung’u *
Department of Mathematics and Actuarial Science, Murang’a University of Technology. P.O. Box 75-10200, Murang’a, Kenya.
Ayub Anapapa
Department of Mathematics and Actuarial Science, Murang’a University of Technology. P.O. Box 75-10200, Murang’a, Kenya.
*Author to whom correspondence should be addressed.
Abstract
Neonatal mortality poses a major hurdle in global health efforts. As per the 2022 Kenya Demographic and Health Survey (KDHS), the neonatal mortality rate was 21 deaths per 1000 live births. Although the rates have been declining over the years, the rate remains significantly higher than the target set by Sustainable Development Goal 3.2 of 12 deaths per 1,000 live births by 2030 for preventable deaths endorsed by the United Nations meeting in 2015. To achieve this target, we need a predictive model that could provide actionable insights and timely interventions to predict and avert neonatal deaths. Machine learning has great potential to improve health outcomes for newborns, especially in areas with limited resources. The objective of this study was to develop a predictive model based on ensemble machine learning algorithms to predict neonatal mortality in Kenya using the 2022 KDHS dataset and identify the critical features contributing to the prediction of neonatal mortality. The Synthetic Minority Over-Sampling Technique was used to address the imbalance in the dataset. The Categorical gradient boosting model outperformed others in terms of accuracy (98.00%), precision (97.46%), recall (98.00%), F1 score (97.57%), and the Receiver Operating Characteristic (ROC) area under the curve (92.24%). The major factors identified included the number of children below five in the household, the total births recorded over the past five years, region, religion, sex of the newborn, source of drinking water, occupation, sex of the household head, use of mosquito nets, place of residence, ever-terminated pregnancy and duration of pregnancy. The research findings support the integration of machine learning into healthcare approaches, such Kenya Health Information System, to facilitate the accurate prediction of neonatal deaths and enhance survival for our newborns. The government should also promote family planning in the country to lower the number of children below five years in households and reduce the number of births recorded per woman in the last five years. The model's generalizability might be constrained by the exclusion of crucial clinical variables from missing data despite high performance.
Keywords: Kenya Demographic and Health Survey, machine learning algorithms, neonatal mortality