Functional and effective connectivity methods are key to understanding how brain regions interplay to perform complex cognitive processes. Yet, most current connectivity methods summarise activity within brain regions to unidimensional measures, resulting in a loss of information. Only recently, new functional connectivity methods have been introduced that exploit multidimensional information, i.e. pattern-to-pattern relationships across regions. To date, these methods have mostly been applied to functional Magnetic Resonance Imaging (fMRI) data, and no method allows the estimation of vertex-to-vertex transformations with the temporal specificity of electro-/magnetoencephalography (EEG/MEG) data. The current thesis introduces novel multidimensional pattern-based connectivity methods for event-related EEG/MEG applications. I introduced time-lagged multidimensional pattern connectivity (TL-MDPC), nonlinear TL-MDPC (nTL-MDPC), and multivariate TL-MDPC (mvTL-MDPC), as novel bivariate and multivariate functional connectivity metrics for EEG/MEG research. They detect linear as well as nonlinear dependencies between patterns, through estimation vertex-to-vertex transformations, for pairs of brain regions and latencies. I evaluated my methods on simulated data as well as on an existing EEG/MEG dataset. Particularly, I asked: 1) whether multidimensional connectivity methods capture more information than unidimensional ones, 2) whether a nonlinear multidimensional connectivity method captures more and different information than its linear counterpart, and 3) whether moving from a bivariate multidimensional connectivity method towards its multivariate version yields more realistic results. In numerical simulations I could show that: 1) none of my novel methods are prone to producing false positives for independent random patterns, 2) TL-MDPC captures both unidimensional and multidimensional connectivity, and performs better than its unidimensional version in the case of multidimensional effects, 3) nTL-MDPC captures both linear and nonlinear dependencies, and performs slightly better than its linear version when nonlinear dependencies exist, 4) mvTL-MDPC produces more realistic results than its bivariate version, TL-MDPC.

I used my new methods to gain novel insights into the brain semantic network. In order to do so, I compared two semantic visual word processing tasks, varying the depth of semantic processing of words by contrasting a semantic decision (SD) and a lexical decision (LD) task. I used both conventional unidimensional approaches, including evoked responses and coherence, as well as my novel multidimensional connectivity methods, and found 1) my multidimensional methods provide a more complete picture of the brain dynamics compared to coherence and another unidimensional connectivity method, by producing richer connectivity among ROIs, 2) four semantic regions, lATL, rATL, PTC, and IFG, showed rich connectivity with each other, 3) lATL and rATL showed strong connectivity throughout using all unidimensional and multidimensional connectivity methods, confirming the essential role of a bilateral ATL hub for semantic representation, and finally 4) I did not find a key semantic role for AG.

Importantly, to deal with the limited spatial resolution issue of EEG/MEG, I proposed a novel spatial subsampling method to select the most informative vertices within each ROI’s patterns, based on a k-means clustering algorithm. I also provided a quantitative assessment of homogeneous and non-homogeneous leakage among my ROIs.

In summary, this thesis proposes novel connectivity methods for exploiting the multidimensional information of EEG/MEG patterns and demonstrates their usefulness for the investigation of the brain semantic network. My key findings support the principles of the Controlled Semantic Cognition (CSC) framework, and highlight the usefulness and advantages of multidimensional connectivity methods for cognitive neuroscience.