Adverse Pregnancy Outcome Prediction Using Deep Learning

Abstract

Pregnancy outcome, also known as birth outcome, refers to the long-term effects of fertilization on the newborn child, extending from the beginning of fetal viability (about 28 weeks gestation) to the first weeks following birth. Adverse pregnancy outcome is an encompassing term that comprises health concerns that affect either the expectant mother, the baby, or both, and throughout the whole pregnancy, childbirth, and the postpartum period. Nearly, 94% of these complications are treatable and preventable. Ethiopia ranks among the nations that has the highest number of maternal fatalities. pregnant women from developing countries are thirty-six times more prone to experience complications associated with pregnancy than those of developed countries. Early prediction of these pregnancy complications has a strong correlation with women's survival. This study was carried out in accordance with experimental and design science approaches. The objective of this study, was to develop a deep-learning based model to predict adverse pregnancy outcomes. The model construction process involved conducting five experiments, each of which utilized a dataset comprising a total of 10,220 instances and featured 15 variables, one of which is a target variable. We have developed a dataset for model development gathered from EDHS. The data collected from EDHS were prepared and preprocessed by using various preprocessing techniques to get quality data that is suitable for the proposed deep-learning approaches to predict adverse pregnancy outcomes. To construct the predictive model, we trained distinct models, by using deep learning algorithms such as LSTM, FNN, BiLSTM, FNN_LSTM, and FNN_BiLSTM for comparison of their performance of prediction of adverse pregnancy outcomes. We have examined the effect of algorithm hybridizations by using FNN with LSTM and BiLSTM. The performance of the model is assessed under stratified 10-fold cross-validation. Based on the experiment performed the FNN_BiLSTM outperformed other models for the prediction of adverse pregnancy outcomes. It achieved 98.94% of accuracy with 99% of F1-Score. The proposed predictive model classifies adverse pregnancy outcomes based on a multi-class prediction approach. Finally, we recommend future researchers to focus on enhancing the predictive model by utilizing a substantial dataset for more extensive investigation