Noise removal in electroencephalogram (EEG) using deep learning algorithms

The development of brain abnormalities is tracked using the EEG. It is also used to evaluate individuals who are dealing with brain-related problems, such as patients with schizophrenia. Noises of many kinds contaminate EEG recordings, which makes it difficult to properly analyze the brain signals. It is critical to eliminate unwanted noise from the signals since patient EEG records are crucial for assessing any type of brain disorder. To obtain reliable data on brain activity and prevent errors in its interpretation, the EEG noise removal technique aims to decrease physiological interference. Maintaining a clear and functional signal requires noise cancellation and reduction. Because some signals, like the EEG, are non-stationary, the statistical feature of noise is difficult to understand. In the past year, a variety of methods have been put out to reject various types of noise in EEG signals. However, optimized results have not yet been obtained. In this thesis, a new method will be developed to improve noise removal results using deep learning algorithms. Hybrid deep learning algorithms were created. Noise removal was performed on EEG signals using methods such as Autoencoders(AEs), Convolutional Neural Networks (CNN), and Long Short-Term Memory Networks (LSTM). Two hybrid models, CNN-AEs and LSTM-AEs, were created. EEG Signal-Noise Ratio(SNR) values obtained by trying different numbers of layers were compared with the Principal component analysis(PCA) method. At the end of this study, better training results were obtained with appropriate hyperparameters in hybrid deep learning models.