New series: Explaining papers [#1]
Lately, we published a new paper on estimating functional connectivity from electrophysiological data (Westner, Kujala, Gross, & Schoffelen, 2024). Let me give you a short description of what we did!
Estimating functional connectivity from electrophysiological (i.e., EEG or MEG) data means that one wants to see if any brain areas show activity that is somehow synchronised (although this synchronisation can be shifted in time: think of two people clapping together or clapping alternatingly - both is considered a synchronisation in the neuroscience world). EEG or MEG measures activity from outside of the head, so from rather far away: that means that active brain regions are measured by many different EEG electrodes or MEG sensors - and that each electrode or sensor measures activity from more than just one brain region. For the estimation of connectivity, this has consequences: our measures are tainted by false connectivity: e.g., because the activity at two electrodes that appears to be synchronous actually comes from the same underlying source!
Part of this problem can be solved by first estimating where the activity at each electrode or sensor comes from within the brain. We call that source reconstruction. However, also this step cannot get rid of all those spurious interactions that are reflecting true connectivity.
Our goal in the paper was to find a way to deal with this spurious connectivity without over-correcting, i.e. by preserving the true connectivity.
To achieve this, we combined some tools and tricks from different fields: first, we use a special source reconstruction method that estimates the interaction between sources instead of the single sources themselves (called a two-dipole beamformer). Second, we formulated a way to estimate an adequate baseline from this model - to find out how much noise our total estimate inherently contains. And lastly, we borrowed a trick from machine learning: We subjected the source reconstruction and baseline estimation to Ensemble Learning logic. The reasoning here is that estimating many, many noisy instances and averaging across them is better than estimating one less-noisy-but-not-perfect instance. To achieve that, we subsampled the MEG sensors and iterated: we computed our source reconstruction, baseline estimate, and subsequent corrected connectivity estimate many, many times from varying sets of MEG sensors and then averaged across the results.
We show in simulation and with one real data set, that our method fares well in comparison to other established methods. Our approach indeed works in making the connectivity estimate more robust towards noise and spurious estimates!
Paper: Westner, Kujala, Gross, & Schoffelen (2024). Towards a more robust non-invasive assessment of functional connectivity. Imaging Neuroscience. https://doi.org/10.1162/imag_a_00119