Investigating the Impact of Signal-to-Noise Ratio on EEGResting-State source reconstruction

Introduction: Identifying sources of electroencephalography (EEG) activity is a complex problem that requires models of thehead and tissues [1,2]. The effect of Signal-To-Noise Ratio (SNR) on source localization accuracy is oftenevaluated considering the task-evoked cortical activity [3]. However, elucidating spontaneous activation of thebrain, i.e., in the absence of a stimulus or task, is not immediate as the signal is of low amplitude and theunderlying neural sources are challenging to examine [4]. In the EEG resting-state signal, the effect of SNR iscritical to be determined as prior information. Moreover, many studies have used spherical heads to investigatethe localization errors of dipoles [5]. Here, we present a simulation study to investigate the effect of differentSNR values on the performance of source estimation (SNR LOC) using the Minimum Norm Estimation (MNE)[6] and a realistic head model. Methods: We simulated synthetic resting-state EEG signals with different known SNRs [7]. The signal was 1 min longand sampled at 256 Hz. It was generated from synthetic source time courses, using two non-linear dipolarcoupled sources located in the primary motor cortex and fifty uncorrelated noise sources randomly distributedover the whole cortex. The two non-linearly coupled sources, with quadratic nonlinearity, presented a timedelay of 15 ms [8]. Using a BEM volume conductor model based on the New York Head model [9] andimposing the EEG electrode locations, the leadfield matrix for the simulated sources was computed accordingto [10]. The source space consisted of a cortical layer of 10016 distributed points registered to a commontemplate. The SNR of the simulated EEG signal was set equal to [1, 5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100].The simulated EEG data were analyzed using Independent Component Analysis (ICA) to remove artifacts andretain ICs of brain origin. The inverse solution was obtained using MNE on brain ICs, with differentregularization parameters (defined as λ proportional to 1/SNR2) that were used to balance the accuracy andsmoothness of the solution. The variation of the SNR LOC values [0.1, 1:10, 50, 100] influences the numericalsolution of the inverse problem in terms of the spread and position of the source reconstruction. Theperformance was evaluated using three different metrics: the localization error, the measure of the sourceextension, and the source fragmentation. The localization error was defined as the distance between theinverse solution peak and the true location of the generating source. The source extension was measured asthe Euclidean distance between the point of the source with the highest intensity and all the points of a Regionof Interest, i.e., where the inverse problem solution is higher than 80% of the solution range. To evaluate thesource fragmentation, we apply the K-means clustering with the Calinski-Harabasz criterion, to the outliers ofthe distances from the peak of the highest intensity. Results: Fig.1A shows the localization error: this decreases as the SNR LOC increases, with a 1/SNR trend. Fig.2Bshows the distribution of median distances between the peak of the inverse solution and the true location of thegenerated source. As the SNR LOC increases, the sources become narrower. The repeated-measure ANOVA,with a four-level within-subject factor, indicates statistically significant differences (p<0.05) between thedistributions, and a post-hoc test was carried out with Bonferroni correction. Regarding the sourcefragmentation (Figure 2), the number of clusters for SNR LOC equal to 100 was significantly higher than forSNR LOC equal to 0.1, 3, and 10, which present no statistically significant differences. Conclusions: Evaluating the effect of SNR LOC strongly influences the spatial resolution of the source-level analysis: anSNR loc value of 10 appears to be a good trade-off between the three metrics, as it provides a focused sourcereconstruction and ensures a low localization error.