Predicting high risk pregnancies in Pakistan- a demographic assessment using predictive machine learning

Abstract

Pakistan is unable to meet its maternal and child health targets. Predictive machine learn- ing has the potential to predict high risk pregnancies based on data from women who have

had a miscarriage or stillbirth. This would help advise better healthcare plans at primary and tertiary level and help achieve Sustainable Development Goal targets in the country. The aim of this study was to evaluate several machine learning models to measure their ability to detect high risk pregnancies. The Pakistan Demographic Health Survey (2018) has been used which includes data from 15,068 women across Pakistan. Fourteen machine

learning classifiers have been employed to predict high risk pregnancies, with the follow- ing evaluation metrics reported: precision, recall, false positive rate (FPR), accuracy, and F1-score. We find that five models have the highest overall performance: (i) Deep Neural Network, (ii) SELU Network, (iii) Multilayer Perceptron, (iv) Gradient Boosting, and (v) AdaBoost, exhibiting near good precision (73.0-76.0%), effective recall (83.0-86.0%), ro- bust accuracy (89.0-90.0%), and decent F1-Scores (79.0-80.0%). This study recommends the integration of low-cost online models to predict high risk pregnancies as a critical tool to help achieve maternal health targets in the country.