A novel approach for imbalanced instance handling toward better preterm birth classification

Abstract

Preterm birth (PTB) is a major cause of child and mother mortality, a PTB classification model can assist in assessing the health condition ahead of time and help avoid complications during childbirth. Mother’s significant feature (MSF) dataset created for this study has features derived from mother’s physical, lifestyle, social and stress attributes. MSF dataset consists of 119 features of 1,000 mothers with 172 preterm and 828 full-term deliveries, resulting in issues of dataset imbalance namely class inseparability and classification bias. To overcome the imbalance issue, a novel algorithm named majority penalizing minority upsampling (MPMU) is proposed. MPMU forms clusters looking into the degree of dataset imbalance, it analyses the composition of each cluster individually and computes the varied penalty for majority class instances. It further balances dataset composition by oversampling minority class instances. MPMU processed dataset is further used to train the proposed 6L-ANN network which finds the probability of occurrence of PTB. The proposed model has shown efficient results on MSF sub-datasets with precision values ranging from 0.90 to 0.97, area under the curve (AUC) between 0.86 to 0.99, and prediction accuracy ranging from 93.04% to 99.47%. Experiment results show that a mother’s lifestyle and stress features have a strong influence on the childbirth outcome.