Identifying Influential Factors Affecting Pregnancy Events Using Machine Learning Approaches

Abstract

Despite of health awareness campaigns and improvements in education system, neonatal mortality is still a critical issue around the world. Out of 140 million children born annually, 4 million die in the first month of their life. This has also become a severe issue in Pakistan with neonatal rate of 44 per 1000 live births. Pakistan is a country where cousin marriage rate is above 60% and is located in such a region where gender preference is common. Studies have suggested that cousin marriage also impacts the pregnancy events. Hence it is a dire need to find out the causes of high neonatal rates and impact of cousin marriage on pregnancy events in Pakistan. Different researches have been carried out on these issues before. Some researchers have investigated the relationship between cousin marriage and adverse pregnancy outcomes while some explored the determinants of child mortality in Pakistan. While these researches focused on specific pregnancy factors such as birth interval and still births, they ignored other important factors like cousin marriage and preterm birth. Some studies have used data with missing factors, such as birth interval, cousin marriage or gestation period; while other studies have mostly applied bivariate or multivariate regression analysis. These techniques have limitations in terms of dealing with categorical data or data with multiple levels of factors. To resolve short comings of the existing studies, we are proposing a framework that will apply association rules, bayesian network and hidden markov model to find associations among different factors in the Pakistan Institute of Medical Sciences (PIMS) hospital dataset. The objectives of this research are (i) to study the effects of different factors that cause neonatal mortality, (ii) cousin marriage impact on neonatal mortality. Finally (iii) to analyze the impact of cousin marriage on gender determination. Data was preprocessed using imputations and models were applied. In order to identify the factors of neonatal mortality, bayesian network and association rules were applied. Bayesian network (BN) produced an accuracy of 94%. Association rules were applied using ‘rattle library and around 9000 rules were generated but only few hold valuable information, such as chances of caesarean delivery are high for short birth intervals and short birth intervals trend has been observed in first pregnancy. To see cousin marriage impact on gender determination, data was distributed on the basis of cousin marriage and non cousin marriage and was converted into sequential data. Hidden markov model was then applied on each dataset and a comparison was performed to see the impact of cousin marriage on gender identification. Its results indicate that mode of delivery, preterm birth, gestation period and birth interval are the major factors influencing the neonatal mortality. Cousin marriage of couple’s parents , place of residence and mother’s education are secondary influencing factors. The results also suggest that short birth interval is observed in first pregnancy. Moreover this research also claims that cousin marriage does not play any significant role in the determination of gender. The results will help in improving decision making and making better policies for mother care in order to reduce neonatal mortality rates.