Neural-network-based resampling method for detecting diabetes mellitus

Abstract

Diabetes has been the fifth leading cause of death in Taiwan since 1987, and the complications of this disease are a burden to patients, their families, and society. Recent studies have tried to build a classifier that can easily identify diabetes mellitus by employing data mining approaches. However, these studies have encountered a class imbalance problem caused by skewed data, in which almost all of the instances are labeled as one class (healthy) while only a few instances are labeled as the other class (diabetic). When learning from this type of data, machine learning algorithms tend to produce predictive results with a high level of accuracy for the majority class, but poor predictive accuracy for the minority class. This study proposes the neural-networkbased resampling method, which dramatically improves the detection of diabetes. Real diabetes data from a regional hospital in Taiwan and several biological data sets are used to demonstrate the effectiveness of the proposed method.