Application of deep learning architectures on sleep staging problems

Sleep relaxes and rests the body by slowing down the metabolism, making us physically stronger and fitter when we wake up. However, in a sleep disorder that may occur in humans, this process is reversed, and various disorders occur in the body. Therefore, the determination of sleep stages is of vital importance for the diagnosis and treatment of such sleep disorders. However, manual scoring of sleep stages is tedious, time-consuming, and requires considerable expertise. It also suffers from inter-observer variability. Deep learning techniques can automate this process, overcome these problems, and produce more consistent results. For this purpose, three different studies were conducted. Firstly, Time-Frequency components obtained by Short Time Fourier Transform, Discrete Wavelet Transform, Discrete Cosine Transform, Hilbert-Huang Transform, Discrete Gabor Transform, Fast Walsh-Hadamard Transform, Choi-Williams Distribution, and Wigner-Willie Distribution were classified with a supervised deep convolutional neural network to perform sleep staging. The results obtained in this study showed that the transformation methods used for the most accurate representation of the input data are much better than the traditional methods based on manual feature extraction, where time, frequency, or nonlinear features are obtained. The second study aims to solve the classification problem in imbalanced datasets with the help of Siamese Neural Networks. Seven distance measurement methods were chosen to calculate the similarity score during the network design: Euclidean, Manhattan, Jaccard, Cosine, Canberra, Bray-Curtis, and Kullback-Leibler Divergence. Thus, a new competitive method for automatic sleep staging based on deep learning and SNNs has been derived. Finally, a new hybrid neural network architecture is proposed using focal loss and discrete cosine transformation methods to solve the training data imbalance problem. The model was trained on four different databases using k-fold cross-validation (subject-wise), and the highest score was 87.11% accuracy, 81.81% Kappa score, and 79.83% MF1 when using two channels (EEG-EOG). The results of our approach are promising when compared to existing methods.