A Comparative Study on Data Balancing Methods for Alzheimer's Disease Classification

Abstract

Alzheimer's disease is a prevalent neurological disorder affecting millions of people worldwide, often associated with the aging process, leading to the death of nerve cells in the brain and loss of connections. Recently, promising results have been demonstrated in diagnosing Alzheimer's disease using deep learning models, and various approaches for early diagnosis have been proposed. However, the imbalance in health datasets, particularly those containing rare cases, can lead to performance losses and misleading results during model training. This study focuses on these imbalance issues, evaluating the effectiveness of different balancing methods using the Alzheimer's MRI dataset. In this context, the performance of SMOTE, ADASYN, and Weight Balancing methods is compared using a custom model. Experimental results indicate that, compared to the original imbalanced dataset, Weight balancing outperforms in terms of accuracy, precision, recall, and F1 score. While SMOTE and ADASYN show improvement in various metrics, they are considered inferior to the Weight Balancing method. This study contributes to selecting data-balancing methods to enhance the accuracy of deep learning models in Alzheimer's disease classification and emphasizes the importance of addressing class imbalances in health datasets.