LAUSR

Sleep state detection and classification play a critical role in understanding sleep patterns and diagnosing related disorders. Traditional methods like polysomnography are accurate but labor-intensive, time-consuming, and reliant on manual intervention, which limits their scalability for large-scale or home-based applications. This study proposes an improved 1D U-Net model for efficient and accurate sleep state detection and classification. The proposed model processes 1D physiological data such as EEG signals, employing an encoder-decoder structure for efficient feature extraction. To extract local and global features, residual blocks are added in the encoder part. This also helps to improve the deep feature extraction and enhances the feature representation. These features are passed to the decoder part of the U-Net for further processing. The decoder contains squeeze and excitation blocks, which consist of powerful attention mechanisms to enhance channel-wise features. Evaluations on the Sleep-EDF-20, Sleep-EDF-78, and SHHS datasets demonstrate the model’s effectiveness, achieving an accuracy of up to 94%, a mean F1 score of 87.1%, and Cohen’s kappa values of 0.84, with reduced computational overhead compared to baseline methods. The results highlight the practical significance of the model for scalable and efficient sleep state classification, suitable for both clinical and home-based applications. The implementation code is available at https://www.kaggle.com/code/irtazasheik/se-res-unet.snip