Accurate sleep stage classification from EEG signals is essential for sleep disorder diagnosis and neuroscience research. In this study, we conduct a comparative analysis of four deep learning models ResNet50-1D, VGG19-1D, LSTM, and Xception-1D trained on preprocessed time-series EEG data from the OpenNeuro Bitbrain Open Access Sleep (BOAS) dataset. Prior to model training, the EEG signals were filtered using bandpass filtering and cleaned using Independent Component Analysis (ICA) to remove artifacts and high-frequency noise. These models represent both convolutional and recurrent architectures, enabling the evaluation of their capabilities in learning spatial and temporal patterns from minimally processed signals. All models were trained and evaluated under identical experimental conditions to ensure a fair comparison. The results highlight the trade-offs between model performance, complexity, and training efficiency, offering valuable insights for selecting suitable architectures in automated sleep staging using EEG time series data.
