<p>Sleep quality directly affects human health. Poor sleep quality is associated with various cardiovascular and neurological disorders, which underscores the importance of accurate sleep staging for the diagnosis of sleep-related diseases. Sleep staging is traditionally performed by experts using polysomnography (PSG) recordings according to the standards of the American Academy of Sleep Medicine (AASM). Sleep is categorized into the wake, rapid eye movement (REM), and non-rapid eye movement (N1, N2, and N3) stages. This study investigates the effects of data balancing and batch normalization placement in deep learning models for automatic sleep staging using Electroencephalogram (EEG) signals from the PhysioNet Sleep-EDF Expanded dataset. To address class imbalance, the NearMiss undersampling algorithm was applied, and its effect on classification performance was evaluated. Three dataset versions were generated: an original standardized dataset, a 5-class balanced dataset, and a 3-class balanced dataset. Three convolutional neural network architectures with different batch normalization configurations were developed and compared. The experimental results obtained from an epochwise dataset that contained recordings from 30 participants revealed that the architecture without batch normalization performed the best and reached 95.99% accuracy for 3-class classification and 93.99% accuracy for 5-class classification. To assess generalization ability, a subjectwise analysis was also performed using all the participants in the SleepEDF dataset. In this setting, the architecture with batch normalization in three layers achieved the best results, with accuracies of 94.16% and 90.31% for 3-class and 5-class classification, respectively. Finally, fivefold cross validation was performed to assess the robustness of the proposed approach.</p>