Parametric Versus Non-Parametric Models for Predicting Infant Mortality within Communities in Uganda using the 2016 Uganda Demographic and Health Survey Data

Abstract

Machine learning techniques have been infrequently used to identify community-based infant mortality risks. Achieving SDG 3 Targets 3.2 and 3.3 could be expedited by early detection of at-risk infants within communities. This study aimed to devise a community-centric algorithm for predicting infant mortality. We analysed UDHS 2016 data containing birth records for 22,635 children born within the five years preceding the survey, excluding those born within a year of the interview date. Twelve machine learning models were evaluated for their predictive capabilities using the area under the receiver operating characteristic curve (AUC ROC) in Python. Data subsets were divided into training and testing sets in a 2:1 ratio. Among the evaluated models, CatBoost showed superior performance with an AUC ROC of 74.9%. The five most influential variables for the CatBoost model were postnatal care utilisation, paternal age, household size, preceding birth interval, and maternal age. While the algorithm’s best performance was achieved using 28 variables, it still exhibited robust predictive power when limited to the top 8 or 10 variables. Hence, CatBoost stands out as an effective tool for identifying community-based infant mortality risks