Handling Imbalanced and Overlapped Medical Datasets: A Comparative Study
Mohammad Sarosh Basit, Adeeba Khan, Omar Farooq, Yusuf Uzzaman Khan, Mohammad Shameem
Abstract
An imbalanced dataset with class overlapping is a challenging issue in medical research. Imbalanced data points lead to overfitting for the majority class while overlapped classes cause misclassification for both classes. Hence, this combination makes it challenging for classic machine learning algorithms to define a decision boundary between minority and majority classes. In our study, different algorithms with different techniques have been compared for example oversampling, undersampling, combined over and under sampling, and the ensemble methods to deal with class imbalance along with class overlapping. Two well-known highly imbalanced and overlapped medical datasets are used to compare the performance of different approaches and performance is evaluated by sensitivity and specificity. On the sleep apnea dataset, oversampling combined with ensemble classifier AdaBoost with the specificity and sensitivity of 0.72 and 0.46 which proved better than other techniques and classifiers. On the diabetes dataset, SMOTE-TOMEK oversampling combined with the Random Forrest classifier with the specificity and sensitivity of 0.91 and 0.77 proved to be better than all the combinations that have been tried for the classification with minimal number of features.