Real-world medical datasets are usually highly imbalanced. However, recent developments in science and technology have elicited tremendous hope in the ongoing research results on the problem of imbalanced learning. Dealing with imbalanced data is a prevalent challenge that has attracted growing attention from both industry and academia while performing classification on datasets. In this study, novel data-level sampling methods are employed to balance the class distribution on three health datasets and one simulated dataset using six machine learning algorithms: Logistic Regression (LR), Random Forest (RF), Decision Trees (DT), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and Multilayer Perceptron (MLP) for classification. It was observed that among the undersampling methods, Edited Nearest Neighbor (ENN) gives the most accurate results, while Synthetic Minority Sampling (SMOTE) outperforms all the techniques in terms of providing more accurate classification results and shows the best improvement on the RF classifier. The results of the experimental comparative study of the data-level methods demonstrate that, in general, oversampling methods outperform undersampling methods in terms of overall performance considering Balanced Accuracy (BA) and AUC-ROC. Therefore, SMOTE should be considered the best option when dealing with imbalanced data in the health domain.