Comparative Analysis of Machine Learning Algorithms for Predicting Under-Five Mortality: Evidence from Tanzania Demographic and Health Survey

Abstract

Under-five mortality remains a global health challenge with the rates of 43 deaths per every 1000 live births in Tanzania and 37 deaths per every 1000 live births globally. Although child mortality has significantly declined in the last twenty years, the current rates are far from reaching the anticipated Sustainable Development Goal of atmost 25 deaths per 1000 live births in 2030. This study intended to find the best performing classifier of under-five mortality status by comparing ten supervised machine learning algorithms. These machine learning algorithms are Decision Trees, Random Forest, Support Vector Machines, SMOTE-Based Boosted Random Forest, XGBoost, LightGBM, CatBoost, Logistic Regression, K-Nearest Neighbors and Stacked Ensemble Methods. The class imbalance of the dataset detected in the pre-processing stage was addressed using weighted categorical cross-entropy and SMOTE with a 5-folds cross validation and data splitting ratio of 80% for training set and 20% for testing set. With 20 experiments for each of the nine algorithms, the average results were reported to ensure that the findings were not by chance. Further, the stacking ensemble model was developed integrating six of the best performing algorithms using an inclusion criterion of AUC > 0.97. The findings revealed that ensemble algorithm consistently outperformed the other nine algorithms by achieving 100%, 100%, 99.97% and 99.24% for AUC, Accuracy, F1-Score and MCC respectively. This implies that stacking ensemble can uncover more insights than the individual algorithms in predicting under-five mortality status. This study recommends designing policies on under-five mortality that integrate insights from the stacking ensemble algorithm which shows the highest predictive performance.