Detection of alzheimer disease through deep learning analysis of EEG signals' time-frequency representations

Alzheimer Disease (AD) is a neurodegenerative disease and the most common type of dementia. It is a progressive and irreversible disease that characterized by symptoms such as loss of cognitive functionality, difficulty in daily activities and memory deficit. Although there is no definitive treatment for Alzheimer Disease, early diagnosis of the disease is important for effective treatment and management of the disease. Electroencephalography (EEG) is one of the most commonly preferred clinical tools to investigate the effects of Alzheimer Disease on the brain. Trying to detect Alzheimer's disease by training EEG data with Convolutional Neural Network (CNN) architectures, one of the deep learning architectures is one of the research areas that has been intensively studied recently. In this thesis study, early diagnosis of the disease is aimed by classifying the EEG data of Alzheimer's patients and control volunteers using deep learning architectures and time-frequency approaches. For this purpose, EEG signals from 15 patients diagnosed with early-stage Alzheimer's disease through diagnostic tests and 11 healthy control volunteers were analyzed. Spectrogram images of EEG segments were obtained with Short Time Fourier Transform (STFT) and scalogram images with Continuous Wavelet Transform (CWT), Bump, Amor and Morse wavelets, and these images were trained with ResNet18, Modified ResNet18_1, Modified ResNet18_2, GoogleNet, AlexNet and DarkNet Convolutional Neural Network (CNN) architectures and the results obtained were evaluated. When the results of all architectures are evaluated on a channel based and brain region based, it is seen that the highest performance rates are obtained with the CWT Bump time-frequency approach from the P4 channel located in the posterior region of the brain. Classification achievements of 88.11% and 80.64% were achieved in order for ResNet18, which has the highest validation and test accuracy (ACC) results, and for Modified ResNet18_2, which features fewer layers and lower filter sizes derived from the modification of the ResNet18 architecture, and exhibits the fastest and most balanced performance. For the channel-based evaluation of the two architectures, validation accuracy (ACC) values of 96.90% and 93.70% were recorded from the P4 channels, respectively. Although the modified ResNet18_2 architecture has a little lower ACC value compared to the classical ResNet18, when looking at the training progress results, it is seen that it shows a balanced and robust performance and it does not show overfitting. Additionally, the proposed architecture completed the training process much faster than other models, thereby achieving important time savings. These results show that the Modified ResNet18_2 architecture is a good alternative in terms of accuracy and training efficiency in detecting Alzheimer's disease.