Classification of Sleep Arousals Based on Multiphysiological Signals Using Deep Learning

Sleep is crucial to the wellbeing of humans. Lack of quality sleep leads to many health risks such as diabetes, obesity, and heart conditions. Statistics showed that around 20% of the population suffers from more than 60 types of sleep disorders (SD) that causes sleep arousals. In 2018, PhysioNet introduced a challenge that utilizes data collected during polysomnography (PSG) studies from 1,985 subjects that includes 13 physiological signals to solve a binary classification problem of the existence of certain types of arousals (non-apnea). This thesis employed EEGNet, a compact convolutional neural network (CNN) that relies on Depthwise and Separable Convolutions layers for classification of sleep arousals. In this study, using a network with 594 trainable parameters only. The selected input signals are EEG (three channels), EMG (two channels), EOG, and AIRFLOW. Signals were preprocessed, down-sampled, and segmented to overcome the large classes imbalance ratio between the target arousal and the no-arousal and no-target-arousal classes. The model was trained on 80% of the segments generated from the data of 100 subjects. The achieved area under the precision-recall curve (AUPRC) was 0.677 for the intra-subject test (20% of the data of the 100 subjects), and 0.183 on the inter-subject test on the data of another 12 hidden test subjects. This result falls within the range of the official scores of the challenge; indicating a promising application in using this lightweight model for automated sleep arousals classification.