Data Resampling and Feature Selection in Diabetes Prediction

ABSTRACT Diabetes prediction aims to identify people at high risk of developing diabetes in the early stage. However, in practice, the collected diabetes prediction datasets are usually class imbalanced, which means that the size of the diabetic class is much smaller than that of the nondiabetic class. To address this, data resampling, including under, over and hybrid sampling methods, can be used to rebalance class‐imbalanced datasets. However, related studies have considered only the first two types of data resampling methods. Therefore, the first research objective of this paper is to examine the performances of these three types of methods in diabetes prediction. On the other hand, conducting feature selection by selecting some representative features from a given training set has shown potential for improving prediction performance. Therefore, the effect of performing feature selection on the resampled data is further investigated. The experimental results obtained by comparing three undersampling, two oversampling and four hybrid sampling methods over six related datasets show that the top three methods are random oversampling (ROS), random undersampling (RUS) and the combination of ROS and RUS. Moreover, performing feature selection by classification and regression tree (CART) first and data resampling second is a better choice than the other combination orders, among which the hybrid sampling method using Tomek links and SMOTE performs the best. However, applying the feature selection and data resampling steps does not necessarily mean that the prediction model provides better performance than models based on data resampling alone.