EEG analysis using deep learning methods with integrated empirical transformation

Brain cells interact with an average of 30,000 neurons. During these interactions, electrical oscillations occur in the brain, which can be measured using an Electroencephalography (EEG) device. In this study, EEG data collected from 18 healthy individuals through 29 different channels were utilized. To examine spatial attention shifts, the analysis focused on the PO7 and PO8 channels, which are indicated in the literature as being influenced by attention shifts. EEG signals were recorded using 29 standard Ag/AgCl electrodes, referenced to the right mastoid, and sampled at a frequency of 250 Hz. For feature extraction, Empirical Wavelet Transform (EWT) and Ensemble Empirical Mode Decomposition (EEMD) methods were employed. The extracted features were analyzed using deep learning and machine learning algorithms. Both channel-based and Intrinsic Mode Function (IMF)-based comparisons were conducted. In channel-based comparisons, the analysis emphasized PO7 and PO8 channels, with the PO7 channel providing more accurate results for attention shifts. In IMF-based comparisons, values of 6, 8, 10, and 12 were tested. When EEMD was applied, and one-dimensional time signals were transformed into two-dimensional images, the highest accuracy was achieved with IMF=6. In algorithm-based comparisons, the analysis revealed that, among machine learning methods, k-NN and decision tree algorithms achieved the highest test accuracy with 100.0%, while among deep learning methods, the GNN algorithm attained a maximum test accuracy of 99.84%.