Sleep Stage Classification Based on Uniform Manifold Approximation and Recursive Feature Elimination Using EEG Signals

Sleep stage classification is a critical component in diagnosing and managing sleep disorders. This research presents a comprehensive workflow for sleep stage classification utilizing advanced EEG signal processing and deep learning techniques. The process begins with data collection from a standardized Physionet database, and extraction of EEG signals from the database, Then the extracted EEG signal is passed through a Butterworth bandpass filter to eliminate noise components, and a Z-score normalization is applied on clean EEG signals for data standardization. By employing the Short-Time Fourier Transform (STFT), on standardized data Spectrograms are generated. Spectrograms play a crucial role in sleep stage classification by providing a detailed time-frequency representation of EEG signals. This representation is essential for capturing the dynamic changes in brain activity that occur during different sleep stages. Since EEG recordings for sleep stage classification are collected over an entire night, the resulting database is substantial in size. As the model needs to train on multiple samples, the volume of information becomes even larger. Therefore, a dimensionality reduction process is recommended to manage the data effectively and improve computational efficiency. In this work, Uniform Manifold Approximation and Projection (UMAP) is employed for dimensionality reduction, preserving the data's intrinsic structure. Further Recursive Feature Elimination with Cross Validation (RFECV) is used to select the most informative features, enhancing the model’ s efficiency. As a final step, DenseNet-201 is selected as the model and the model's hyperparameter tuning is done via grid search with cross-validation, which optimizes the model's parameters. This systematic approach ensures robust, reliable, and precise classification of sleep stages, contributing to improved diagnosis and management of sleep-related disorders. The proposed method for sleep stage classification demonstrates superior performance across various metrics compared to existing models such as HS-LSTM, C-SVM, CNNs, and GN-BiLSTM. It achieves an accuracy of 91.25%, a specificity of 90.19%, and an F1 score of 90.11 %, indicating enhanced reliability and precision. Additionally, the proposed model shows lower false positive and false negative rates, contributing to more accurate and robust sleep stage classification.