Gradient and network structure of lagged correlations in band-limited cortical dynamics

Normal brain function results from directed causal interactions between brain regions. However, the large-scale spatial and temporal structure of these interactions remains unclear. Lagged correlations between the activity of different brain regions may reflect their directed causal interactions. Thus, we combined magnetoencephalography (MEG) and machine learning approaches to characterize and disentangle the structure of lagged correlations in the human brain.

We computed lagged correlations between all pairs of cortical regions and frequencies of neural activity. We employed curvature-regularized autoencoders to disentangle the resulting space. A second step of the autoencoder approximated the manifold of frequency-specific interactions, enabling the interpretation and comparison of identified dynamical components.

We found consistent temporally irreversible lagged interactions within and across frequency bands. While some lag-frequency patterns were spatially ubiquitous, other patterns involve specific cortical connections and networks. These clusters of connections emerged naturally from the embedding, were consistent between subjects, and could be interpreted in terms of known functional networks.

Most components showed similar dynamics in the resting state and during task performance, which suggests that the underlying lagged interactions reflect a fundamental causal structure of brain dynamics that persists across mental states. Components were also selectively modulated by task demands, suggesting they play specific roles in different cognitive functions.

Our results show that lagged correlations of neural activity are structured in components that are identifiable using machine learning, involve specific cortical networks and frequencies, and may reflect brain-wide causal interactions.