Machine Learning Approaches for Handling Imbalances in Health Data Classification

Abstract

Imbalanced data classification is an important area of statistical learning that has gained research attention over the years. Despite extensive research, imbalanced data classification remains one of the most challenging problems in data science and machine learning, especially for large data sets. Health data imbalance poses challenges for machine learning models to classify data points accurately and can lead to biased and inaccurate predictions with severe consequences in medical settings. The majority of real world datasets are skewed. Reports from previous researchers suggest that sampling approaches are effective for data imbalances. To tackle this problem, in this study, we discuss sampling approaches, including oversampling and undersampling methods, such as Random Oversampling, SMOTE, ADASYN, Random Undersampling, Tomek links, NearMiss and so on, and conduct experiments on four different skewed health datasets to achieve promising performances. The four imbalanced secondary health data sets used are on Diabetics, Anaemia, Lung Cancer, and Obesity classification respectively. Results show that Repeated Edited Nearest Neighbours (RENN) undersampling technique with logistic regression is more effective in handling data skewness. RENN technique should be adopted in cases of imbalance in health research.