Development of a Sliding Window-Based Oversampling Framework for Convolutional Neural Network Sleep Stage Classification Using Polysomnographic Spectrogram Fusion

This study proposed an oversampling framework based on a sliding window approach for CNN-based sleep stage classification using PSG spectrogram fusion. Continuous Wavelet Transform (CWT) was employed as the feature transformation method to convert raw EEG and EOG signals into time-frequency spectrograms. To address the issue of class imbalance commonly found in sleep datasets, the sliding window method was applied as a data augmentation technique, resulting in a more balanced distribution of samples across sleep stages. Multi-channel, multi-epoch spectrograms were used as input to three CNN-based classifiers: AlexNet with SVM, AlexNet, and ResNet-18. The results demonstrated that the sliding window method significantly enhanced classification performance across all stages, particularly in the N1 stage, achieving an average accuracy of 98.53%. The highest overall average accuracy of 95.81% was obtained using ResNet-18.