How well can we predict perinatal mortality and its associated risk factors using machine learning models-a case study on malawi

The danger of death is greatest during the newborn phase of life. Determining which babies are most vulnerable before their birth they are born can have a significant impact in turning the fetal tides for the newcomers. Even though the perinatal mortality rate has improved since last two decades, Africa is still the region with highest mortality rates. This study tries to differentiate pregnant women who are at risk of giving birth to vulnerable babies in a poor southeastern country called Malawi. To find the early predictors of newborn mortality, sophisticated computational techniques such as predictive modeling can be handy. This study's goal is to compile, evaluate critically, and analyze causes of perinatal mortality in Malawi using machine learning models. Output variable in this study is binary in nature with either normal or adverse outcome. The data was unbalanced with more normal deliveries than adverse ones. After cleaning and preprocessing the data, unbalanced class issue was resolved through Synthetic Minority Oversampling Technique (SMOTE) and then the models were trained. The algorithmsproved to be handy in predicting perinatal mortality of the African country. Two ensemble models, Random Forest and Gradient Boost, have shown higher accuracy and precision scores while predicting adverse outcome of the pregnant women. Model based feature selection and Shapely Additive Explanations (SHAP) techniques identified most prominent risk factors affecting the mortality. The study included 56 features varying from previous medical record to socioeconomic condition of the women. It is noticeable that study is already conducted on women who had iron deficiency and were being observed after treating them with iron infusion through oral and ferric carboxymaltose (FCM) methods. Given the history, medical reasons dominated the outcome of pregnancies. Respiratory rate, weight of mothers, and blood pressure levels drew clear lines in separating normal births from the contrary. Secondly, economic factors are also highly deciding. Those women who spent less in hospital, whose families did not possess any land , without access to safe drinking water and belonging to daily wage earners were the most vulnerable in giving birth to the unhealthy babies. Social factors like illiteracy, early pregnancies, affiliation with certain tribes and religion also affected the birth outcome in Malawi. Those women who were given oral iron had relatively low birth casualties. Therefore, the study has shown that machine learning models are not just effective in rightly identifying mortality but also in identifying factors playing havoc for the poorly nourished women in Malawi.