First Steps to Using Carbon Wire Loops to Correct for Artifacts in simultaneous EEG-fMRI
First Steps to Using Carbon Wire Loops to Correct for Artifacts in simultaneous EEG-fMRI
by Lars Hausfeld, PhD Department of Cognitive Neuroscience, Maastricht University, Maastricht Brain Imaging Centre (M-BIC), The Netherlands
The simultaneous acquisition of EEG and fMRI signals promises linked datasets with high temporal and spatial resolution, respectively. To make optimal use of the data from this advanced acquisition scheme, the attenuation of artifacts in EEG signals is necessary. One type of artifact that arises from blood pulsation, the so-called ballisto-cardiographic artifact, is difficult to correct for due to its variability. Here, a prototype BrainCap MR with carbon wire loops (CWLs) was used to directly measure this artifact while acquiring EEG and MRI data from one participant. Example data from two experimental paradigms show that signals from carbon wire loops can be used to efficiently regress these artifacts out and indicate the impact of such artifact attenuation for both event-related potential and stimulus reconstruction analyses.
Electro-encephalography (EEG) was fundamental in advancing our understanding about brain processes and states in clinical and non-clinical populations and remains a tool of choice today given its precise temporal resolution, relatively low cost and portability. It is nowadays complemented with functional magnetic resonance imaging (fMRI), a technique that provides an increasingly precise spatial resolution that EEG source modelling cannot offer. During the last decade, studies have performed simultaneous measurements of EEG and MRI with the goal of combining the respective strengths of these techniques to better determine when and where neural processes occur at rest and in task contexts as well as in response to sensory stimuli. The strong magnetic field in the MRI scanner, however, induces noise in EEG measurements and proper correction methods are required before analyzing the EEG data. A first correction step for EEG signals is the removal of the MRI artifact related to gradient switching. While this artifact is of high amplitude, effective methods based on template subtraction have been introduced that are able to correct for this artifact (Allen et al., 2000; Niazy et al., 2005).
A different artifact arises from electrode motion due to blood pulsation, the ballisto-cardiographic (BCG) artifact. While this artifact’s amplitude is small compared to the MRI scanning artifact – but still 3-4 times larger than EEG – it is difficult to correct for due to its spatio-temporal non-stationarity. Most preprocessing approaches make use of the simultaneously acquired electro-cardiogram (ECG) and advanced processing (e.g., OBS-ICA; Debener et al., 2007). These techniques first detect occurrences of heart beats and subsequently correct for the BCG artifact by estimating the dynamic BCG artifact observed in EEG signals. In contrast, a new approach not relying on the ECG uses carbon wire loops that are attached to the EEG cap to measure the BCG (Abbott et al., 2015; van der Meer et al., 2016). Once measured, the BCG – and other motion-related artifacts apparent in additional measurements (e.g., from the He-pump or head movements) – can be removed by efficient regression-based approaches.
Here, I will share my experiences with first simultaneous EEG-fMRI measurements and analyses using a prototype BrainCap MR that included CWLs. In the best case, this provides a useful guidance for the (beginning of) cleaning EEG data from the BCG artifact. The correction methods used here were chosen to due to their availability as open-source EEGLAB toolboxes and straight-forward application.
The data were acquired in a 3T Siemens Prisma Fit MRI scanner at the Maastricht Brain Imaging Center (Maastricht, The Netherlands) from one participant. Throughout simultaneous EEG measurement, functional MRI data was acquired using a continuous EPI protocol (time-of-repetition [TR]: 1s, voxel size: 2x2x2.5mm3) that provided coverage of the temporal and large portions of the frontal lobe. The He-pump was not switched off during measurement. The participant was presented with three runs of an optimized MMN paradigm (Näätänen et al., 2004) and two repetitions of two 5-min audiobook excerpts. During each run of the oddball paradigm, the participant was presented with 495 standard tones and 240 deviants (opt1 sequence in Näätänen et al., 2004) and asked to focus on a fixation cross. The two audiobook excerpts were read by a female and male speaker and pauses between sentences or words were kept shorter than 300ms for analysis (Hausfeld et al., 2018).
EEG data was acquired with the prototype 32-channel BrainCap MR including an ECG electrode and additional channels for five carbon wire loops Data were recorded with a BrainAmp MR plus and BrainAmp ExG MR using BrainVision Recorder software (Brain Products GmbH, Gilching, Germany) with the recommended recording parameters for EEG-fMRI. The data was collected referenced to FCz.
EEG data were preprocessed in MATLAB (The Mathworks, Natwick, MA, version 18.104.22.1684444, [R2018b]) using EEGlab (Delorme and Makeig, 2004, version 14.1.2) and additional toolboxes. First, the MRI gradient artifact was removed from EEG data using the FMRIB toolbox (version 1.21, http://fsl.fmrib.ox.ac.uk/eeglab/fmribplugin/; see Niazy et al., 2005). Here, I used the pop_fmrib_fastr.m function with the following parameters: a low-pass filter of 70Hz, 5-fold interpolation/upsampling (resulting in 25kHz signals), an averaging window size of 10 occurrences, specification of the MRI trigger (‘R128’) and flags for volume triggers and not performing adaptive noise cancellation (default parameters otherwise).
Copy to Clipboard
This resulted in a robust attenuation of the gradient artifact as can be seen from Figure 1A that shows both corrected EEG as well as CWL time-courses. Channels P4, TP10 and POz (not presented here) showed abnormal EEG signals and were interpolated after the CWL-based correction.
Figure 1: Example EEG and CWL signal traces before and after CWL-based correction. A) The upper panel shows 10-s example EEG traces from the different electrodes. In the lower panel, the CWL traces are presented. Note the clear heart-beat related artifacts at about 60bpm in CWLs. The EEG recordings show the large impact that the BCG artifact has on electrodes. B) Same as in A) but after CWL-based correction. Note here the very strong attenuation for the BCG artifact observed in A) as well as the eye blinks at Fp1/Fp2.
Unfortunately, the ECG signal time-course showed aberrant signals during this MRI acquisition (without MRI acquisition, typical ECG patterns were obtained). Distortion of the ECG signal inside the scanner can occur and occasionally the signal can be unusable. However, the preprocessing of EEG and removal of the BCG based on CWL measurements does not rely on ECG recordings and thus could be performed. In contrast, other approaches (Debener et al., 2007; Niazy et al., 2005) to attenuate BCG artifacts require temporal information on the heartbeat occurrences (QRS complex) extracted by a robust detection algorithm and manual or semi-automated setting/correction of these triggers (e.g., BrainVision Analyzer 2, FMRIB toolbox). If present, CWL signals could be used as alternative signals for detecting heartbeat events when necessary, especially when facing distorted ECG recordings.
After downsampling of the EEG and CWL data to 500Hz, the BCG and other potential motion artifacts were corrected by using the cwleegfmri toolbox (http://sccn.ucsd.edu/eeglab/plugins/CWRegrTool.zip; van der Meer et al., 2016). More specifically, this approach creates windowed regressors from CWL time-courses that are used to regress out – channel-by-channel – the signal traces that are related to the CWL signals. As these do not measure any EEG but motion-induced currents amongst which the BCG artifact but also other potential motion artifacts, this step corrects for these non-brain signals (it has been suggested that residuals of gradient artifacts might also be attenuated as these will be present in the CWL signals as well, Abbott et al., 2015). The function pop_cwregrssion.m was called with mainly default parameters specifying the sampling rate (500Hz), window duration (4s), delay (0.021s), tapering factor (1), tapering function (‘hann’), channel indices of carbon wire loops, channel indices of EEG, and the regression approach (‘taperedhann’):
Copy to Clipboard
The result of this correction can be seen in Figure 1B that shows the CWL-corrected data and showcases the attenuation of the BCG artifact. Please note the residual BCG – though small – that remained after correction (e.g., in channels Oz or TP9) and that further optimization of regression parameters might lead to increased attenuation of the BCG artifact. Furthermore, it becomes clear that eye-related artifacts remain in the corrected dataset (here at 5s and 8s most prominent in channels Fp1 and Fp2). These can now be effectively corrected for by a method of choice. From this point onwards, researchers should be able to follow their desired preprocessing and analysis scheme. For these two datasets, I chose to use Independent Component Analysis (ICA) to decompose and further process the CWL-corrected EEG data and remove independent components (ICs) that represented blinks and eye-movements.
To gain quantitative insight into the artifact handling using the CWL signals, Figure 2 presents a coarse metric that indicates the strong attenuation of the BCG artifact and other motion artifacts as picked up by the CWLs. While the time course of EEG channels showed a high correlation between CWL and EEG signals of |r| = 0.15 for most channels, the highest correlation after CWL-based correction among all channels was |r| = 0.021 and |r| = 0.041 for the MMN and audiobook dataset, respectively. For ICs that were computed separately on data with and without CWL-based correction one can observe a similar effect with the maximally correlated component of the CWL-corrected dataset at |r| = 0.026 and |r| = 0.047 while only 5 and 4 components remained below this value when performing ICA on the dataset without CWL signals regressed out. While this was only observed with two datasets of one participant, it suggests an advantage of the CWL-based correction over using ICA only: a few independent components were highly correlated (|r| > 0.2) with the CWL signals and would be chosen as “BCG-components” to be removed. However, most other components also showed a high similarity with CWL signals (i.e., motion-related artifacts) that would remain in the corrected data.
A comparison with the OBS or OBS-ICA approach (Debener et al., 2007; Niazy et al., 2005) was unfortunately not possible due to the poor quality of the ECG signal (see above). However, in terms of computational effort, the CWL approach is faster and was in my experience less sensitive to changes of parameters.
Figure 2: Similarity of CWL and EEG signals before and after CWL-based correction. A) As an indication of artifacts detected by CWLs in the acquired EEG data, the 5 CWL signals were correlated with signals from EEG channels. For simplicity only the highest absolute correlation is presented here. Blue lines show the similarity of EEG and CWL signals for uncorrected data, red lines show the similarity for data after CWL regression. Solid (MMN) and dotted lines (audiobook) indicate the two sets of experimental data.
When analysing the data acquired during the oddball paradigm, ERPs computed from the dataset without CWL-based artifact correction look noisy – in particular for deviants relying on less trials – with some typical ERP components being identifiable (Figure 1A). In contrast, the data after regressing out CWL signals showed cleaner ERPs with typical components clearly discernible (Figure 1B). Please keep in mind that this is data from a single participant; in usual studies grand average ERPs are computed from 20+ participants.
Figure 3: ERPs from the MMN paradigm before and after CWL-based correction. A) Solid lines show the recorded ERPs for a fronto-central electrode cluster (Fz, FCz, Cz, FC1, and FC2) for the standard (black) and deviant (orange), the dotted grey line presented the difference deviant – standard before CWL-based correction. B) same as in A) but for after the CWL signal were regressed out. In addition, topographic plots are presented at selected latencies. These show typical auditory evoked patterns with fronto-central vs lateral polarity inversion.
In a final step, the data acquired during audiobook presentation was analyzed with a stimulus reconstruction approach that estimates so-called multivariate temporal response functions (mTRFs). In short, using the continuously acquired EEG data, part of the data was selected to train an mTRF model using the MATLAB-based mTRF toolbox (Crosse et al., 2016). After training, this model was applied to unseen test data to reconstruct the speech envelope presented while this unseen data was acquired. The predicted and the actual envelope are then compared using the correlation coefficient. For this single-participant dataset, the stimulus could be reconstructed well with rmTRF = 0.116 for the CWL-corrected data. This is in the higher range of reconstruction performances in comparison to a previous EEG-fMRI study (Puschmann et al., 2017). Perhaps surprising, the reconstruction for the dataset without CWL-based correction showed a higher reconstruction of rmTRF = 0.136. I speculate that this is due to the high sensitivity of the mTRF models, which are susceptible to exploiting artifacts (here: BCG); this exemplifies that thorough cleaning of EEG data will provide researchers with more confidence that their results reflect neural processing, which is even more important in the EEG-hostile environment of the MRI scanner.
To conclude, carbon wire loop signals can substantially aid the cleaning of EEG datasets obtained during fMRI data acquisition. As these loops measure the artifact related to blood pulsation and other motion sources directly at the level of EEG channels, a fast regression algorithm is able to attenuate these artifacts substantially, which would require more advanced approaches otherwise. Unfortunately, a direct comparison with these techniques was not possible but previous results suggest added benefits from CWL signals in comparison to other approaches (Abbott et al., 2015; van der Meer et al., 2016). Taken together, the fast and intuitive use and artifact measurements close to EEG channels make CWLs a powerful tool for cleaning EEG data obtained from simultaneous EEG-fMRI measurements.
Abbott DF, Masterton RAJ, Archer JS, Fleming SW, Warren AEL, Jackson GD. 2015. Constructing Carbon Fiber Motion-Detection Loops for Simultaneous EEG-fMRI. Front Neurol 5:260. doi:10.3389/fneur.2014.00260
Allen PJ, Josephs O, Turner R. 2000. A Method for Removing Imaging Artifact from Continuous EEG Recorded during Functional MRI. Neuroimage 12:230–239. doi:10.1006/nimg.2000.0599
Crosse MJ, Di Liberto GM, Bednar A, Lalor EC. 2016. The Multivariate Temporal Response Function (mTRF) Toolbox: A MATLAB Toolbox for Relating Neural Signals to Continuous Stimuli. Front Hum Neurosci 10:604. doi:10.3389/fnhum.2016.00604
Debener S, Strobel A, Sorger B, Peters J, Kranczioch C, Engel AK, Goebel R. 2007. Improved quality of auditory event-related potentials recorded simultaneously with 3-T fMRI: Removal of the ballistocardiogram artefact. Neuroimage 34:587–597. doi:10.1016/j.neuroimage.2006.09.031
Delorme A, Makeig S. 2004. EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. J Neurosci Methods 134:9–21. doi:10.1016/j.jneumeth.2003.10.009
Hausfeld L, Riecke L, Valente G, Formisano E. 2018. Cortical tracking of multiple streams outside the focus of attention in naturalistic auditory scenes. Neuroimage 181:617–626. doi:10.1016/j.neuroimage.2018.07.052
Moosmann M, Schönfelder VH, Specht K, Scheeringa R, Nordby H, Hugdahl K. 2009. Realignment parameter-informed artefact correction for simultaneous EEG-fMRI recordings. doi:10.1016/j.neuroimage.2009.01.024
Näätänen R, Pakarinen S, Rinne T, Takegata R. 2004. The mismatch negativity (MMN): towards the optimal paradigm. Clin Neurophysiol 115:140–144. doi:10.1016/j.clinph.2003.04.001
Niazy RK, Beckmann CF, Iannetti GD, Brady JM, Smith SM. 2005. Removal of FMRI environment artifacts from EEG data using optimal basis sets. Neuroimage 28:720–737. doi:10.1016/j.neuroimage.2005.06.067
Puschmann S, Steinkamp S, Gillich I, Mirkovic B, Debener S, Thiel CM. 2017. The Right Temporoparietal Junction Supports Speech Tracking During Selective Listening: Evidence from Concurrent EEG-fMRI. J Neurosci 37:11505–11516. doi:10.1523/JNEUROSCI.1007-17.2017
van der Meer JN, Pampel A, Van Someren EJW, Ramautar JR, van der Werf YD, Gomez-Herrero G, Lepsien J, Hellrung L, Hinrichs H, Möller HE, Walter M. 2016. Carbon-wire loop based artifact correction outperforms post-processing EEG/fMRI corrections—A validation of a real-time simultaneous EEG/fMRI correction method. Neuroimage 125:880–894. doi:10.1016/j.neuroimage.2015.10.064