MEG language mapping using a novel automatic ECD algorithm in comparison with MNE, dSPM, and DICS beamformer

Introduction The single equivalent current dipole (sECD) is the standard clinical procedure for presurgical language mapping in epilepsy using magnetoencephalography (MEG). However, the sECD approach has not been widely used in clinical assessments, mainly because it requires subjective judgements in selecting several critical parameters. To address this limitation, we developed an automatic sECD algorithm (AsECDa) for language mapping. Methods The localization accuracy of the AsECDa was evaluated using synthetic MEG data. Subsequently, the reliability and efficiency of AsECDa were compared to three other common source localization methods using MEG data recorded during two sessions of a receptive language task in 21 epilepsy patients. These methods include minimum norm estimation (MNE), dynamic statistical parametric mapping (dSPM), and dynamic imaging of coherent sources (DICS) beamformer. Results For the synthetic single dipole MEG data with a typical signal-to-noise ratio, the average localization error of AsECDa was less than 2 mm for simulated superficial and deep dipoles. For the patient data, AsECDa showed better test-retest reliability (TRR) of the language laterality index (LI) than MNE, dSPM, and DICS beamformer. Specifically, the LI calculated with AsECDa revealed excellent TRR between the two MEG sessions across all patients (Cor = 0.80), while the LI for MNE, dSPM, DICS-event-related desynchronization (ERD) in the alpha band, and DICS-ERD in the low beta band ranged lower (Cor = 0.71, 0.64, 0.54, and 0.48, respectively). Furthermore, AsECDa identified 38% of patients with atypical language lateralization (i.e., right lateralization or bilateral), compared to 73%, 68%, 55%, and 50% identified by DICS-ERD in the low beta band, DICS-ERD in the alpha band, MNE, and dSPM, respectively. Compared to other methods, AsECDa’s results were more consistent with previous studies that reported atypical language lateralization in 20-30% of epilepsy patients. Discussion Our study suggests that AsECDa is a promising approach for presurgical language mapping, and its fully automated nature makes it easy to implement and reliable for clinical evaluations.

. Stimuli design for the word recognition task used in this study.

Language regions-of-interest
We used language specific regions-of-interest (ROIs) in the Brainnetome atlas (Fan et al., 2016) for AsECDa and DICS beamformer (Table S1-a). For MNE and dSPM, the Destrieux standard atlas (Destrieux et al., 2010) was used for the language specific ROIs (Table S1-b). Table S1. Language ROIs from (a) the Brainnetome volumetric atlas, which were used for automatic single equivalent current dipole algorithm (AsECDa) and dynamic imaging of coherent sources (DICS) beamformer, and (b) the Destrieux standard atlas, which were used for minimum norm estimation (MNE) and dynamic statistical parametric mapping (dSPM).

Effect of bandpass filters on MEG evoked magnetic fields (EMFs)
We used a 0.1-170 Hz bandpass filter for the DICS beamformer and 0.1-20 Hz bandpass filter for the other three methods (i.e., ECD, MNE, and dSPM). As shown in Figure S2, using a 0.1-170 Hz bandpass filter will not significantly change the time course of the average EMFs in the MEG sensors. Applying a 0.1-20 Hz bandpass filter instead of a 0.1-170 Hz bandpass filter on the MEG signals will reduce high frequency noise in the time course of the average EMFs. It is important to note that a 0.1-20 Hz bandpass filter cannot be used for the DICS beamformer if we wish to measure the power of brain signals in high beta (20-30 Hz), low gamma (30-50 Hz), and high gamma (50-110 Hz) bands.

Impact of tSSS and Regularization on DICS Beamformer Results
The tSSS pre-processing reduces the rank of the cross-spectral density (CSD) of the MEG sensors, which is defined as Q(f) in Equation (9) of the manuscript, from 206 to approximately 70 (see Figure S3). To invert Q(f) in DICS (as described in Equation (8)), we handled the rank deficiency of CSD after tSSS by using a regularization parameter in the Fieldtrip toolbox, as "cfg.dics.lambda = '10%'". With this regularization parameter, we calculated the inversion of Q(f) Avearge EMF in "MEG0212" [pT/cm] using the following formula in MATLAB: "invQf = pinv( Qf + lambda * eye(size(Qf)) )", where lambda is calculated as 10% of the average of the eigenvalues of Q(f).
We performed experiments to evaluate the effect of different values of lambda on the DICS beamformer results, and found that using various values did not significantly improve the results.
We also applied the DICS beamformer on MEG data without using the tSSS filter, and the results did not improve either.
To further investigate these findings, we reanalyzed MEG data of Pt# 18, who had a Wada test with a left lateralized language, using the DICS beamformer with different values of lambda and with/without tSSS filter. The results presented in Table S2 demonstrate that different values of lambda did not significantly change the laterality index (LI) in MEG data with tSSS filter. Moreover, using MEG data without applying the tSSS filter reduced the consistency of the LI across different values of lambda, and generated an LI that showed a stronger tendency towards right lateralization, despite the fact that the subject was left lateralized based on the Wada test. We have included these results in the Supplementary Material. Figure S3. Singular value decomposition of the cross-spectral density (CSD) of the MEG sensors in the alpha band for a representative subject (Pt#18) in two pre-processing approaches: (a) with applying the tSSS filter and (b) without applying the tSSS filter. The blue line represents the CSD before applying regularization, while the red line represents the CSD after applying regularization at 10% of the average of the eigenvalues of the CSD.