Comparison of SMOTE and GAN Data Augmentation Methods in Sleep Stage Classification

Abstract

Sleep is a vital process that occupies a significant portion of human life. Poor sleep quality can lead to the emergence of neurological and cardiovascular diseases. Therefore, sleep staging plays a crucial role in the diagnosis of sleep-related disorders. Traditionally, sleep staging is performed manually by experts in clinical settings by evaluating biomedical signals recorded via polysomnography (PSG) in accordance with the American Academy of Sleep Medicine (AASM) standards. However, since this process is both labor-intensive and prone to human error, its automation is of great importance. Excluding the wake stage, sleep is divided into two main phases: Rapid Eye Movement (REM) and Non-Rapid Eye Movement (NREM). The distribution of these stages throughout the sleep period provides essential information regarding sleep quality. However, these stages are not distributed evenly, which leads to the Class Imbalance Problem (CIP) in automatic classification and can negatively affect model performance. In this study, the MIT-BIH Polysomnographic Database—a challenging dataset due to significant class imbalance and signal variability—was utilized to automatically determine sleep stages. The effects of data augmentation approaches, specifically the classical SMOTE method and the deep learning-based GAN, were investigated using a proposed end-to-end CNN - BiLSTM architecture. Performance evaluations were conducted for both 3-class and 5-class staging scenarios. The experimental results are as follows: The accuracy values achieved for 3-class and 5-class staging were 87.58% and 75.85% without data balancing, 85.91% and 71.82% with SMOTE, and 86.60% and 75.31% with GAN-based augmentation, respectively. The results indicate that data augmentation methods do not always yield performance improvements in sleep staging tasks, particularly due to the complex temporal structure of EEG signals. Nevertheless, the 87.58% accuracy achieved without any data balancing demonstrates that the proposed CNN–BiLSTM architecture possesses a high capacity for learning sleep staging patterns.