Large-scale collaboration in ENIGMA-EEG: A perspective on the meta-analytic approach to link neurological and psychiatric liability genes to electrophysiological brain activity.

The organization of the work required in our investigations of EEG genetics. Much of the work is performed by the collaborating sites (columns in black), including EEG recording, preprocessing, phenotype extraction, and performing the genetic association. The role of ENIGMA‐EEG is to regularly hold teleconference calls to create the protocols for EEG analysis, QC, and genetics analyses (blue). Lead groups of ENIGMA‐EEG members are formed to perform centralized quality control (QC) of the EEG features and to meta‐analyze of the summary statistics provided by the sites. The summary statistics are then distributed to the individuals who will perform genetic follow‐up analyses. Finally, a manuscript is prepared. Note that most of the genetics work is not included in this workflow, thus excluding a huge amount of work on taking biological samples (blood, saliva), DNA extraction and storage, sending for genotyping, data management, imputation, quality control. EEG, electroencephalography; ENIGMA, Enhancing NeuroImaging Genetics through Meta‐Analysis; GWAMA, genome‐wide meta‐analysis; QC, Quality control; Sumstats, Genetic summary statistics from genome‐wide association

Effect of reference on EEG coherence and power. We calculated power and coherence in the alpha band (8–12.5 Hz) for the 128 channels available in this sample of 39 subjects (data from (Smit et al., 2013)). Data were initially analyzed with average reference. (a) Changing to mastoid reference biases alpha power upward (left inset bar graph). The correlation between mastoid and average reference is very high (>0.90). Therefore, a GWAS of EEG alpha power will be marginally impacted despite the large bias. (b) Changing to mastoid reference also biases channel average coherence upward (inset bar graph). The correlation across subjects is low (r < .30, right topoplot). This will substantially affect genetic association and indicates that reference needs to be harmonized across studies. (c) Local bipolar derivations show similar low correlation with the average reference setup (r < .28). (d) A selected channel pair (C3, C4) showed variable connectivity between the reference setups. Markedly, mastoid reference showed negative correlation with the average reference and local bipolar derivations

Spherical interpolation for quality control of a dataset of 765 subject in a 17 channel montage with A1/A2 reference using the data from (Smit et al., 2005), eyes‐close resting condition, and cleaned by visual inspection, filtering 1–30 Hz, and ICA decomposition with visual rejection (Pion‐Tonachini et al., 2019). Theta power (4–8 Hz, left), beta power (13–21 Hz, middle), and theta–beta ratio (right) were calculated for channel Cz. Next, the same power values are calculated for a spherical interpolation of channel Cz using 16 remaining channels (implemented in EEGLAB (Delorme & Makeig, 2004)). Even at this low‐density montage, oscillatory power is generally quite well imputed (r ≥ .97), and outliers easily detected by statistical methods (false discovery rate). For theta power, ten observations were considered suspect at FDR q = 0.01. For beta power, three observations were considered suspect. These values may be replaced with the imputed values. For theta–beta ratio, five values were considered suspect. Retracing the subjects' signals revealed that three of these were affected by some residual artifacts in channel Cz, and their values replaced by the interpolated values. It shows that highly automated algorithms of multichannel EEG data can produce high‐quality data and flag errors in visual cleaning