Tracking dynamic brain networks using high temporal resolution MEG measures of functional connectivity

Fluctuations in functional interactions between brain regions typically occur at the millisecond time scale. Conventional connectivity metrics are not adequately time-resolved to detect such fast fluctuations in functional connectivity. At the same time, attempts to use conventional metrics in a time-resolved manner usually come with the selection of sliding windows of fixed arbitrary length. In the current work, we evaluated the use of high temporal resolution metrics of functional connectivity in conjunction with non-negative tensor factorisation to detect fast fluctuations in connectivity and temporally evolving subnetworks. To this end, we used the phase difference derivative, wavelet coherence, and we also introduced a new metric, the instantaneous amplitude correlation. In order to deal with the inherently noisy nature of magnetoencephalography data and large datasets, we make use of recurrence plots and we used pair-wise orthogonalisation to avoid spurious estimates of functional connectivity due to signal leakage. Firstly, metrics were evaluated in the context of dynamically coupled neural mass models in the presence and absence of delays and also compared to conventional static metrics with fixed sliding windows. Simulations showed that these high temporal resolution metrics outperformed conventional static connectivity metrics. Secondly, the sensitivity of the metrics to fluctuations in connectivity was analysed in post-movement beta rebound magnetoencephalography data, which showed time locked sensorimotor subnetworks that modulated with the post-movement beta rebound. Finally, sensitivity of the metrics was evaluated in resting-state magnetoencephalography, showing similar spatial patterns across metrics, thereby indicating the robustness of the current analysis. The current methods can be applied in cognitive experiments that involve fast modulations in connectivity in relation to cognition. In addition, these methods could also be used as input to temporal graph analysis to further characterise the rapid fluctuation in brain network topology.