Implementing Predictive Model for Low Birth Weight in Afghanistan

Birth weight is a significant determinant of the likelihood of survival of an infant. Low Birth Weight (LBW) has become an increasingly common problem, especially in developing and underdeveloped countries. Therefore, an ability to predict the LBW is beneficial and a good preventive measure. It is also a good indicator of future health risks of that infant. This study is concerned about the implementation of predictive LBW models for rural and urban areas of Afghanistan based on the data obtained from the Afghanistan Demographic and Health Survey 2015. The main objective of the study is to identify the most suitable machine learning techniques (i.e. classifiers) among the five popular ones. These are K-Nearest Neighbor (K-NN), Naïve Bayes, Neural Network, Random Forest, and Support Vector Machine (SVM). Prior to implementing the predictive models, data preprocessing is carried out. This is done by means of data cleansing and feature selection. The well-known Correlation based Feature Selection algorithm (CFS) is employed to select the top ten attributes. The classifier in this study comprises two categories, Normal and LBW. The preparation of the dataset is carefully done to ensure well-balanced samples in each category. The study reveals that Random Forest is the best classifier with an accuracy of 84.70% and 85.2% and with the Area Under the Curve (AUC) of 91.0% and 92.1% for both rural and urban areas, respectively. The study has a direct benefit to the health and prevention policy making in Afghanistan.