MEG-based Machine Learning Semantic Classification of Observed Words

Machine learning methods are starting to be widely used in the analysis of neuroimaging data. Apart from playing a crucial part in the development of Brain-Computer Interface technologies, machine learning can be also used in academic context to link cognitive phenomena to their neurophysiological sources. In this study we attempted to use a SVM model to classify fragments of MEG recording according to the semantic categories of the words that were presented to the subject at the moment. The preprocessed data was clustered in spatial and temporal domains and the clusters were subject to the permutational F-tests. A three-dimensional epochs array was cropped to the time intervals of significant clusters from the selected channels and had its dimensionality reduced with Principal Component Analysis (PCA) or Uniform Manifold Approximation and Projection (UMAP). The resulting vector was used to fit the model to solve the binary classification problem.