Determination of schizophrenia patients from EEG signals with deep learning methods

The diagnosis of schizophrenia is made through patient interviews by a psychiatrist who is an expert in the field. The mentioned detection process is a time-consuming, costly and error-prone process. In schizophrenia, which is a serious and chronic disease, patients tend to exhibit different behaviors, believe in unreal events and change their personalities by losing their connection with real life. In this life-long disease, the disease can be brought under control with the right treatment. In this way, patients can continue their lives as healthy individuals and be successful in their social relations and business life. The treatment process requires great care and sensitivity as it can trigger the recurrence of the disease in the slightest neglect. Because of all these reasons, diagnosis of the disease is important. In this thesis, it is aimed to determine whether people have schizophrenia from EEG signals in order to accelerate the diagnosis and diagnosis process of schizophrenia. For this purpose, two applications were carried out. The data set used in the applications consists of EEG signals obtained over 16 channels in accordance with the international 10-20 system received from the NNCI platform. In the first application; Spectrogram images of the signals in the relevant data set were obtained. Hybrid models were created with deep learning architectures and machine learning algorithms (AlexNet-SVM, AlexNet-KNN, GoogleNet-SVM, GoogleNet-KNN, VGG16-SVM and VGG16-KNN) and applications were carried out using spectrogram images. By observing the performance of the classification results of each channel, it was observed that the highest accuracy rate was obtained with the VGG16-SVM model. In addition, classifications made over all EEG channels without dividing into channels were performed on the same hybrid models. The AlexNet-KNN hybrid model gave the best result with 99.39% for the application where all channels are combined. The results of the 6 hybrid models tested are emphasized and the advantages of channel-based classification are emphasized. In a second application within the scope of this thesis, the raw EEG data obtained from the same data set were handled and classified together with the LSTM network. In the classification processes performed with the LSTM network, classifications were carried out both over 16 EEG channels and on the data in which all channels were together.