| Abstract [eng] |
Sleep staging is a time-consuming and resource-intensive process. Automating this task could significantly enhance clinicians' productivity, allowing them to focus on diagnosing and treating patients. Classifying complex, non-stationary signals is inherently challenging, and convolutional neural networks (CNNs) are widely used for this purpose. Recent research has shifted focus toward classifying the scalogram of EEG signals, rather than the raw signal itself. Wavelet transformation is particularly relevant for constructing these scalograms, as it enables the representation of the signal in both time and frequency domains. However, there has been limited research on how the choice of mother wavelet affects classification performance. Additionally, contrastive learning frameworks and data augmentation techniques hold potential for improving predictions of sleep stages, but no studies have applied these methods to scalogram data for sleep stage classification. This study aims to examine how the choice of wavelet influences sleep stage classification. Furthermore, the SimCLR contrastive learning framework was implemented, simple data augmentation techniques were tested, and results were compared with a baseline model. Notably, this study uses data from a single electroencephalogram (EEG) channel, aligning with the goal of developing portable, cost--effective polysomnographic devices for home diagnostics. Additionally, the study compares classification results across four distinct age groups, highlighting the impact of age on classification performance. Based on a comprehensive review of relevant research, the following wavelets were selected: Morse, complex Morlet, complex Gaussian, and Mexican Hat. Two CNN architectures, SqueezeNet and ResNet-18, were evaluated to determine the optimal baseline model. The combination of ResNet-18 and the complex Morlet wavelet was chosen for further improvement using the SimCLR pretraining framework and data augmentation. To enhance interpretability, the GradCAM method was applied to visualize the scalogram regions influencing the model's decisions. The study found that sleep stage classification metrics varied with patient age, the youngest patients achieving the highest classification metrics and the oldest patients the lowest. In specific cases, using Morse and/or complex Morlet wavelet showed improved metrics compared to other wavelets. Additionally, ResNet-18 architecture has been shown to be more stable. Contrastive learning proved beneficial, especially for classifying REM sleep stage across most age groups, except for the oldest patients. This discrepancy may be attributed to the inappropriate selection of data augmentation techniques for this age group. Additionally, SimCLR with augmentations enhanced the classification of sleep stage I, but only for one specific age group. These findings underscore the importance of considering patient age when improving sleep stage classification methods, particularly in developing more accurate and personalized diagnostic tools. |
