Classification of motor and cognitive imagery based magnetoencephalography signals using time-frequency analysis

In this thesis, a two-stage method was developed for the binary classification of motor imagery (MI) and cognitive imagery (CI) tasks using open-access magnetoencephalography (MEG) data. The main objective of the study was to extract integrated time–frequency and spatial features from MEG signals and to achieve high classification accuracy. The dataset consisted of 306-channel MEG recordings collected from 17 participants in two sessions, including four different tasks: imagining the movement of hands (H) or feet (F), performing arithmetic subtraction (S), and word generation (W). In the first stage, the MEG signals were divided into two main frequency bands using a bandpass filter. For each sub-band, time–frequency features were extracted using the Continuous Wavelet Transform (CWT), while spatial features were obtained through the Common Spatial Pattern (CSP) method. The extracted features were selected using a t-test-based feature selection approach, and the resulting features were classified using an Artificial Neural Network (ANN). In this stage, high classification accuracies were obtained, particularly for complex task pairs (H–W: 71%, H–S: 66%). These results demonstrated that the combined use of CWT and CSP is effective in distinguishing between motor and cognitive processes. In the second stage, the MEG signals were decomposed into subcomponents using five different signal decomposition methods: Empirical Wavelet Transform (EWT), Maximal Overlap Discrete Wavelet Transform (MODWT), Empirical Mode Decomposition (EMD), Variational Mode Decomposition (VMD), and Multivariate Variational Mode Decomposition (MVMD). Spatial features obtained from the decomposed signals through CSP were subjected to t-test-based feature selection and tested using Support Vector Machines (SVM), k-Nearest Neighbors (KNN), ANN, and Naive Bayes (NB) classifiers. In particular, for the MVMD method, subject- and task-specific feature selection was also applied to determine the optimal feature set for each participant and task. The results revealed that the MVMD and MODWT methods, when combined with the ANN classifier, achieved the highest accuracy rates. The MVMD method exhibited superior performance for the H–W (79.2%), F–S (75.8%), and F–W (73.8%) task pairs. Moreover, the proposed MVMD- and CSP-based, subject- and task-specific feature selection approach achieved more balanced and higher accuracy values compared to studies utilizing the same dataset. Notably, the achieved accuracy rates for the H–W (81.24%) and F–W (80.06%) tasks were among the highest reported in the literature.