Comparative analytics of classifiers on resampled datasets for pregnancy outcome prediction

Abstract

The main challenges of predictive analytics revolve around the handling of datasets, especially the disproportionate distribution of instances among classes in addition to classifier-suitability issues. This unequal spread causes imbalance learning and severely obstructs prediction accuracy. In this paper, the performances of six classifiers and the effect of data balancing (DB) and formation approaches for predicting pregnancy outcome (PO) were investigated. Synthetic minority oversampling technique (SMOTE), resampling with and without replacement, were adopted for data imbalance treatment. Six classifiers including random forest (RF) were evaluated on each resampled dataset with four test modes using Waikato Environment for Knowledge Analysis and R programming libraries. The results of analysis of variance performed separately using F-measure and root mean squared error showed that mean performance of classifiers across the datasets varied significantly (F=117.9; p=0.00) at 95% confidence interval, while turkey multi-comparison test revealed RF(mean=0.78) and SMOTE (mean=0.73) as having significantly different means. The RF model on SMOTE produced each PO class accuracy ≥0.89, area under the curve ≥ 0.96 and coverage of 97.8% and was adjudged the best classifier-DB method pair. However, there was no significant difference (F=0.07, 0.01; p=1.000) in the mean performances of classifiers across test data modes respectively. It reveals that train/test data modes insignificantly affect classification accuracy, although there are noticeable variations in computational cost. The methodology significantly enhance the predictive accuracy of minority classes and confirms the importance of data-imbalance treatment, and the suitability of RF for PO classification.