Sleep scoring in physiological signals by deep learning-based algorithms

Sleep is a complex and dynamic process that is essential for human health. Sleep disorders are a group of medical conditions characterized by disturbances in the quality, quantity, or timing of sleep. The diagnosis of sleep disorders is crucial for the effective management and treatment of these disorders. Polysomnography is a non-invasive test used for the diagnosis of sleep disorders in which various physiological parameters such as brain waves, eye movements, muscle tone, heart rate, and breathing are recorded during sleep. Sleep stage and arousal scoring are two important components of polysomnography that help sleep specialists assess a patient's sleep patterns. However, sleep scoring is a tedious and time-consuming process that requires trained personnel to carefully review and score at least six hours of recordings according to predetermined rules. In this thesis, efficient and effective deep learning-based methods for sleep stage classification and arousal detection from polysomnography signals are presented in two studies. The first study aims to evaluate and validate a novel approach proposed for sleep stage classification. The approach is based on the use of local pattern transformations and convolutional neural networks. The signals were transformed into new signals containing local patterns using one of the local pattern transformation methods. The transformed signals are divided into 30-second epochs. Classification is performed using a model that can accept multiple consecutive epochs. The model learns features from multiple epochs and takes into account the context between epochs during classification. Performance analysis of the model was performed with the Sleep-EDF Expanded dataset and leave-one-out cross-validation. For a comprehensive analysis, a total of 60 experiments were conducted using four local transformation methods, five signal combinations, and three different values for the number of epochs. Overall, the best performance results were obtained with the combination of two-channel EEG and single-channel EOG, one-dimensional local binary pattern method, and five epochs of input, with an accuracy of 0.848, an F1 score of 0.782, and a Cohen kappa of 0.790. In the second study, we propose a novel multi-task learning approach based on fully convolutional neural networks for sleep stage and arousal detection. The model accepts single-channel EEG signals recorded overnight and generates sleep stage and arousal labels. It comprises convolution, recurrent, attention, and segmentation modules to extract local features, establish long-term dependencies, focus on important regions of the inputs, and generate prediction masks. We treat sleep stage and arousal detection tasks as segmentation problems, providing a unified solution for both tasks. The performance evaluation on the Sleep Heart Health Study (SHHS) and Multi-Ethnic Study of Atherosclerosis (MESA) datasets demonstrated strong results. For sleep stage scoring, the model achieved high accuracy, F1 score, and kappa of 0.875, 0.803, and 0.826 for SHHS and 0.829, 0.761, and 0.755 for MESA, respectively. In terms of arousal scoring, the model obtained impressive AUPRC and AUROC of 0.695 and 0.973 for SHHS, with similar scores for MESA. These results demonstrated the model's effectiveness in accurately detecting sleep stages and arousals, surpassing previous literature benchmarks and significantly improving arousal scoring performance.