Deep learning based schizophrenia status determination from eeg signals using dimension augmentation methods

Today, many studies have been conducted to detect, understand or interpret brain activities by examining electroencephalogram (EEG) data, ve various results have been presented. Understanding brain activity is an important step in diagnosing many diseases such as schizophrenia ve epilepsy, ve resolving different conditions or diseases such as sleep stages. Therefore, it is of critical importance to develop new models in order to improve the results of such studies or to take them one step further. In this thesis, firstly, it is aimed to reveal methods with higher performance compared to the analyzes obtained from EEG data classified with classical machine learning algorithms, which were previously made in the literature. In the second stage, it was aimed to make classification with higher performance without extracting features by examining the EEG data with deep learning methods ve to depict the difference between schizophrenia (SZ) patients ve healthy individuals visually. In this thesis, it was aimed to examine brain activities by using data analysis methods using EEG signals. For this purpose, EEG records of especially SZ patients were classified using both classical machine learning algorithms ve deep learning algorithms. In this study, the accuracy of the methods applied using two different datasets of SZ patients was confirmed. In the first dataset (Data set A), there are 39 healthy children ve a 16-channel EEG record with 45 SZ diseases. The second dataset (Data set B) contains 19-channel EEG recordings of 28 adults (14 control ve 14 SZ patients). In the literature, especially classical machine learning algorithms have been widely used in the classification of SZ patients. However, at the first stage of this thesis, better classification performance results were obtained by using different feature extraction ve feature selection methods. In addition, the performance results obtained with different classification algorithms were compared. VII In the second stage of the thesis, a study was conducted with deep learning algorithms, in which a method aiming to automatically diagnose SZ patients using EEG records was conducted. Unlike many literature studies, the proposed method converts raw EEG to 2D using Short Time Wavelet Transform (STFT) ve Continuous Wavelet Transform (CWT) to have a useful representation of frequency-time properties rather than manually extracting features in EEG recordings. This study is the first to use 2D time frequency properties for automatic diagnosis of SZ patients in the relevant literature. It has been trained with the VGG-16 algorithm, which uses a state-of-the-art Convolution Neural Network (CNN) architecture to extract the most useful features from all the features available in the 2D space ve to classify samples with high accuracy. The results were also tested with different CNN algorithms (such as CIFAR, VGG-19, ResNet, DenseNet, MobileNet, Xception…), but the best performance value was obtained with VGG-16 with the least iteration ve time. The experimental results show that the presented method is successful in classifying SZ patients ve healthy controls in two datasets of different age groups, spectrogram images with 95% ve 97%, ve scalogram images with 98% ve 99.5% classification accuracy. With this performance, the proposed method outperforms most literature methods. The experiments of the study also reveal a relationship between the frequency components of an EEG recording ve SZ disease. In addition, with the AktivizationMaximization, Grad-CAM ve Saliency Map methods presented in our thesis, interpretable images were obtained while separating an SZ patient from a healthy control. With these images, the difference between SZ patients and healthy controls, which support the expert opinion, can be easily understood. Keywords: Schizophrenia, EEG, Clasification, Spectrogram, Deep Learning, Scalogram