Benchmarking metrics for inferring functional connectivity from multi-channel EEG and MEG: A simulation study

Abstract

I present a systematic evaluation of different types of metrics, for inferring magnitude, amplitude, or phase synchronization from the electroencephalogram (EEG) and the magnetoencephalogram (MEG). I used a biophysical model, generating EEG/MEG-like signals, together with a system of two coupled self-sustained chaotic oscillators, containing clear transitions from phase to amplitude synchronization solely modulated by coupling strength. Specifically, I compared metrics according to five benchmarks for assessing different types of reliability factors, including immunity to spatial leakage, test-retest reliability, and sensitivity to noise, coupling strength, and synchronization transition. My results delineate the heterogeneous reliability of widely used connectivity metrics, including two magnitude synchronization metrics [coherence (Coh) and imaginary part of coherence (ImCoh)], two amplitude synchronization metrics [amplitude envelope correlation (AEC) and corrected amplitude envelope correlation (AEC_c)], and three phase synchronization metrics [phase coherence (PCoh), phase lag index (PLI), and weighted PLI (wPLI)]. First, the Coh, AEC, and PCoh were prone to create spurious connections caused by spatial leakage. Therefore, they are not recommended to be applied to real EEG/MEG data. The ImCoh, AEC_c, PLI, and wPLI were less affected by spatial leakage. The PLI and wPLI showed the highest immunity to spatial leakage. Second, the PLI and wPLI showed higher test-retest reliability and higher sensitivity to coupling strength and synchronization transition than the ImCoh and AEC_c. Third, the AEC_c was less noisy than the ImCoh, PLI, and wPLI. In sum, my work shows that the choice of connectivity metric should be determined after a comprehensive consideration of the aforementioned five reliability factors.