Detection and classification of sleep apnea syndrome from EEG signals using deep learning methods

This study is focused on the detection and classification of sleep apnea syndrome from Electroencephalography (EEG) signals using machine learning methods. Sampling, filtering, windowing, normalization and frequency subbanding were processed on EEG signals from Polysomnography (PSG) recordings for using in all experiments. Feature extraction was performed by applying basic statistical parameters to the subbands of the obtained EEG signals. These features were applied as input to a multilayer perceptron neural network (MLPNN) model for sleep apnea syndrome classification. To evaluate the effect of normalization and frequency subband segmentation, the raw EEG signals and normalized signals were reclassified by using the same MLPNN model. The normalization results in less significant increase in classification success, so the normalized signals were segmented into frequency subbands, the feature extraction proccess was repeated and classified by using the same MLPNN model. Normalization and frequency subband segmentation significantly improved the classification success. In another experiment the entropy of spectrograms were applied as input to the MLPNN model. In each class, an increase in success was observed, but the success rate of the severe apnea class was the highest. In another experiment, a convolutional neural network (CNN) model, one of the deep learning models frequently used in the literature, was used in additional to MLPNN. In two different experimental studies, images and spectrograms of EEG signals were applied as input to this model and sleep apnea syndrome classification was performed. In this classification model which does not require preprocessing and feature extraction, the success rate was higher as a result of the experiment with spectrograms. As the last experiment of the study, sleep apnea syndrome classification was performed by using YOLOv5 and YOLOv8 models. For these classifications, spectrograms of windowed EEG signals were used. The performance of the YOLOv8 model was higher than the YOLOv5 model because of its high proccessing speed, less number of layers and parameters and high classification success.