Elsevier

NeuroImage

Volume 135, 15 July 2016, Pages 45-63
NeuroImage

Full Length Article
Ballistocardiogram artifact correction taking into account physiological signal preservation in simultaneous EEG-fMRI

https://doi.org/10.1016/j.neuroimage.2016.03.034Get rights and content

Abstract

The ballistocardiogram (BCG) artifact is currently one of the most challenging in the EEG acquired concurrently with fMRI, with correction invariably yielding residual artifacts and/or deterioration of the physiological signals of interest. In this paper, we propose a family of methods whereby the EEG is decomposed using Independent Component Analysis (ICA) and a novel approach for the selection of BCG-related independent components (ICs) is used (PROJection onto Independent Components, PROJIC). Three ICA-based strategies for BCG artifact correction are then explored: 1) BCG-related ICs are removed from the back-reconstruction of the EEG (PROJIC); and 2–3) BCG-related ICs are corrected for the artifact occurrences using an Optimal Basis Set (OBS) or Average Artifact Subtraction (AAS) framework, before back-projecting all ICs onto EEG space (PROJIC-OBS and PROJIC-AAS, respectively). A novel evaluation pipeline is also proposed to assess the methods performance, which takes into account not only artifact but also physiological signal removal, allowing for a flexible weighting of the importance given to physiological signal preservation. This evaluation is used for the group-level parameter optimization of each algorithm on simultaneous EEG-fMRI data acquired using two different setups at 3 T and 7 T. Comparison with state-of-the-art BCG correction methods showed that PROJIC-OBS and PROJIC-AAS outperformed the others when priority was given to artifact removal or physiological signal preservation, respectively, while both PROJIC-AAS and AAS were in general the best choices for intermediate trade-offs. The impact of the BCG correction on the quality of event-related potentials (ERPs) of interest was assessed in terms of the relative reduction of the standard error (SE) across trials: 26/66%, 32/62% and 18/61% were achieved by, respectively, PROJIC, PROJIC-OBS and PROJIC-AAS, for data collected at 3 T/7 T. Although more significant improvements were achieved at 7 T, the results were qualitatively comparable for both setups, which indicate the wide applicability of the proposed methodologies and recommendations.

Introduction

The complementarity between the high temporal resolution of electroencephalography (EEG) and the millimeter spatial resolution of functional magnetic resonance imaging (fMRI) has strongly motivated the integration of these two neuroimaging techniques (for reviews, please refer to Jorge, J., et al., 2013, Murta, T., et al., 2015). Simultaneous EEG-fMRI acquisitions, however, induce two major artifacts on EEG data. Firstly, the rapidly-changing, high-amplitude MR-related artifacts are induced by the switching of magnetic fields during the fMRI acquisition. The gradient-related artifact occurrences are practically time-invariant and the use of standard average artifact subtraction (AAS) methods usually yields acceptable artifact reduction (Allen et al., 2000).

The correction for the ballistocardiogram (BCG) artifact represents a bigger challenge, mainly due to its non-stationary nature. As discussed in (Mullinger, K.J., et al., 2013, Yan, W.X., et al., 2010), several mechanisms contribute to the BCG artifact, the most plausible being: 1) the rotation of the head inside a strong, static magnetic field due to the cardiac pulse (Bonmassar et al., 2002), 2) the Hall effect of the pulsatile blood flow, which is an electrically conductive fluid, inducing changes of the voltage measured on the surface of the scalp (Tenforde et al., 1983) and 3) voltages generated by movement associated with pulse-driven expansion of the scalp (Debener et al., 2008). Studies aiming at a better characterization of the contribution of each of those mechanisms to the BCG artifact have found that most of the artifact variance is explained by flow-induced Hall voltage and pulse-driven head rotation (Mullinger et al., 2013). Additionally, a significant increase of the amplitude of the BCG artifact with the static magnetic field strength, B0, has also been shown (Debener, S., et al., 2008, Neuner, I., et al., 2013), with severely hampered visual inspection of typical EEG patterns at 7 T. The accurate removal of the BCG artifact while preserving the physiological EEG signal is of utmost importance in several applications, particularly when the quantification of EEG features allows the prediction of blood-oxygen-level dependent (BOLD-)fMRI signal changes (e.g., Rosa et al., 2010), including epilepsy applications (Leite et al., 2013).

The BCG artifact occurrences are known to be approximately time-locked with the cardiac cycle; therefore, an AAS algorithm can be employed whereby an artifact template is extracted by averaging across multiple cardiac cycles, and a time-domain subtraction is then performed for artifact correction (Allen et al., 1998). In the most common approach, a moving average template is computed from successive artifact occurrences, assuming that the BCG artifact occurrences change slowly over time. The BCG variability can be more accurately taken into account by computing the temporal Principal Component Analysis (tPCA) over all the time-locked occurrences of the artifact in order to build an optimal basis set (OBS), comprising a given number of principal components (PCs) that explain the BCG artifact variance to some extent (Niazy et al., 2005). This basis set is then fitted to, and subtracted from, each artifact occurrence. More direct approaches to correct for the BCG artifact have been proposed, which are based on the use of motion sensors (Bonmassar, G., et al., 2002, Chowdhury, M.E.H., et al., 2014, Jorge, J., et al., 2015a, Masterton, R.A.J., et al., 2007). However, currently these techniques require hardware-related changes of the experimental apparatus and may therefore not be applicable in general.

A third category of BCG artifact correction methods is based on blind source separation. Independent Component Analysis (ICA) is commonly used for this purpose. By removing the contribution of independent components (ICs) reflecting artifact-related processes in the back-reconstruction of the EEG signal, an artifact-corrected signal can be obtained (Bénar, C.-G., et al., 2003, Mantini, D., et al., 2007, Srivastava, G., et al., 2005). Most of the ICA algorithms, however, assume spatial stationarity of the sources, ignoring the spatio-temporal variability of the BCG artifact occurrences (Debener, S., et al., 2008, Vanderperren, K., et al., 2007). Many approaches nevertheless rely on ICA-based methods, since alternative methods present their own limitations as well (Vanderperren et al., 2010). When using ICA-based EEG de-noising methods, several questions arise regarding: 1) the implementation of the ICA decomposition of EEG data to be used; 2) the number of ICs to be estimated; and 3) the identification of the ICs to be classified as BCG-related. All three issues have been addressed in (Vanderperren et al., 2010); however, the objective and accurate classification of ICs remains a major concern and several criteria can be found in the literature for that purpose, which will be presented and discussed later in this paper.

Due to the relative advantages and disadvantages of OBS and ICA-based methods, the combination of both has already been proposed (Debener, S., et al., 2005, Debener, S., et al., 2007). The OBS method is used to remove most of the BCG artifact contribution, followed by an ICA decomposition of the OBS-corrected EEG data to further remove residual artifacts. Despite its effectiveness in removing the artifact, the application of OBS directly on the EEG signal may also induce undesirable physiological signal reductions, which would be exacerbated if an additional step of ICA-based correction is applied. Recently, it has also been proposed to use a modified version of OBS in the IC space instead (Liu et al., 2012). For that purpose, the EEG data is first decomposed into a set of ICs and the mutual information (MI) (Bell and Sejnowski, 1995) between them and the electrocardiography (ECG) data is computed. The number of ICs with the highest MI is then chosen by means of a modified method (Peng et al., 2005) that finds the optimal number of ICs yielding minimal error in leave-one-out cross-validation. The BCG-related ICs are removed from the back-reconstruction of the EEG signal and all the remaining ICs are singular value decomposition (SVD-)corrected before back-projecting them onto the EEG space, as it is hypothesized that the BCG artifact contributes to all ICs to varying degrees due to its non-stationary nature (Liu et al., 2012).

Regardless of the BCG artifact correction method used, the quality of the correction should be assessed (Freyer, F., et al., 2009, Grouiller, F., et al., 2007). Typically, event-related simultaneous EEG-fMRI studies are conducted and the performance of a given method can be computed based on features extracted from the event-related potentials (ERPs) of interest, such as: the inter-trial variability (Vanderperren et al., 2010), the signal-to-noise (SNR) ratio (Debener et al., 2007) or the difference between the ERPs extracted from the inside-MR EEG datasets and those that are obtained from the BCG artifact-free outside-MR EEG data (Mantini et al., 2007). When the frequency content of a task-modulated EEG signal is known, the power of the EEG within that frequency band can also be computed before and after BCG artifact correction (Xia et al., 2014). In resting-state fMRI (rs-fMRI) studies, the quality of the correction can be assessed by comparing the BCG artifact occurrences before and after correction, based on the root mean square (RMS) of the BCG waveform or their peak-to-peak (PTP) values (Chowdhury et al., 2014). Additionally, the total spectral power within windows around the cardiac fundamental frequency and its first harmonics (Liu et al., 2012) can be computed, and a ratio expressing the loss in normalized spectral power after the correction is used to quantify the amount of BCG artifact that was removed, measured by the Improvement in terms of Normalized Power Spectrum (INPS) (Tong et al., 2001).

Many studies continue to be dedicated to the development of more efficient methods for BCG artifact correction, but the associated, unwanted removal of true physiological signal in the background has not been systematically assessed. Here, we propose a novel method for the selection of BCG-related ICs (PROJection onto Independent Components, PROJIC), after which three ICA-based approaches for the removal of the BCG artifact are explored. First, the contribution of the BCG-related ICs is removed in the back-reconstruction step of the EEG signal (PROJIC). Alternatively, we propose the use of OBS (PROJIC-OBS) or AAS (PROJIC-AAS) for correcting the BCG-related ICs before back-projecting them, instead of simply removing their contribution. In parallel, a novel evaluation pipeline that assesses both artifact and background signal removal is presented, and used to compare the novel approaches with previous approaches, on data collected from a group of epilepsy patients (imaged at 3 T) and a group of healthy volunteers (imaged at 7 T). Data quality improvements on ERPs of interest collected with both setups (inter-ictal epileptiform discharges, IEDs, and visual-evoked potentials, VEPs, collected at, respectively, 3 T and 7 T) were also assessed by means of an inter-trial variability measure.

Section snippets

Materials and methods

The main steps of the processing pipeline proposed in this work for the optimal BCG artifact correction and subsequent evaluation, in terms of not only artifact removal and physiological signal preservation, but also data quality improvements on ERPs of interest collected with both datasets, are depicted in Fig. 1.

Results

The results obtained using the different algorithms tested on the data collected using Setups #1 and #2 are presented and discussed jointly in this section. The IC selection criterion underlying PROJIC was found to successfully select BCG-related ICs in both setups, as evidenced by the clear presence of artifact occurrences in the corresponding time-courses upon direct visual inspection. Moreover, the subsequent application of the three proposed approaches was also found to substantially reduce

Discussion

In this paper, a novel method for the selection of BCG-related ICs was proposed, following which three ICA-based approaches for BCG artifact removal were explored on EEG data recorded simultaneously with fMRI. These novel methods were compared with other ICA-based approaches in the literature, as well as with the commonly used AAS and OBS methods. We found that group-level optimization of the algorithms' parameters significantly improved performance of all methods, and that, when optimized,

Conclusion

We have proposed a novel method for the selection of BCG-related ICs, with three associated approaches for the correction of the BCG artifact in simultaneous EEG-fMRI, and have shown that they outperform both ICA-based and non-ICA-based state-of-the-art methods for most physiological background preservation weights. The PROJIC-OBS method returns the strongest attenuation of the BCG artifact, while the PROJIC-AAS method manages to preserve most of the physiological signal in the background with

Acknowledgments

We acknowledge the Portuguese Science Foundation (FCT) for financial support through Project PTDC/SAUENB/112294/2009, Project PTDC/EEIELC/3246/2012, Grant UID/EEA/50009/2013 and the Doctoral Grant PD/BD/105777/2014. We also acknowledge Centre d'Imagerie BioMédicale (CIBM) of the UNIL, UNIGE, HUG, CHUV, EPFL and the Leenaards and Jeantet Foundations for providing the data acquired at 7 T.

References (55)

  • S. Debener et al.

    Improved quality of auditory event-related potentials recorded simultaneously with 3-T fMRI: removal of the ballistocardiogram artefact

    NeuroImage

    (2007)
  • W. Deburchgraeve et al.

    Automated neonatal seizure detection mimicking a human observer reading EEG

    Clin. Neurophysiol.

    (2008)
  • A. Delorme et al.

    EEGLAB: An open source toolbox for analysis of single-trial EEG dynamics including independent component analysis

    J. Neurosci. Methods

    (2004)
  • F. Freyer et al.

    Ultrahigh-frequency EEG during fMRI: pushing the limits of imaging-artefact correction

    NeuroImage

    (2009)
  • F. Grouiller et al.

    A comparative study of different artefact removal algorithms for EEG signals acquired during functional MRI

    NeuroImage

    (2007)
  • J. Jorge et al.

    Towards high-quality simultaneous EEG-fMRI at 7 T: detection and reduction of EEG artefacts due to head motion

    NeuroImage

    (2015)
  • J. Jorge et al.

    Simultaneous EEG-fMRI at Ultra-high field: artefact prevention and safety assessment

    NeuroImage

    (2015)
  • K. Kobayashi et al.

    Separation of spikes from background by independent component analysis with dipole modeling and comparison to intracranial recording

    Clin. Neurophysiol.

    (2001)
  • P. LeVan et al.

    Modulation by EEG features of BOLD responses to interictal epileptiform discharges

    NeuroImage

    (2010)
  • Z. Liu et al.

    Statistical feature extraction for artefact removal from concurrent fMRI-EEG recordings

    NeuroImage

    (2012)
  • D. Mantini et al.

    Complete artefact removal for EEG recorded during continuous fMRI using independent component analysis

    NeuroImage

    (2007)
  • R.A.J. Masterton et al.

    Measurement and reduction of motion and ballistocardiogram artefacts from simultaneous EEG and fMRI recordings

    NeuroImage

    (2007)
  • K.J. Mullinger et al.

    Identifying the sources of the pulse artefact in EEG recordings made inside an MR scanner

    NeuroImage

    (2013)
  • D. Nayak et al.

    Characteristics of scalp electrical fields associated with deep medial temporal epileptiform discharges

    Clin. Neurophysiol.

    (2004)
  • I. Neuner et al.

    EEG acquisition in Ultra-high static magnetic fields up to 9.4 T

    NeuroImage

    (2013)
  • R.K. Niazy et al.

    Removal of FMRI environment artefacts from EEG data using optimal basis sets

    NeuroImage

    (2005)
  • N. Novitskiy et al.

    The BOLD correlates of the visual P1 and N1 in single-trial analysis of simultaneous EEG-fMRI recordings during a spatial detection task

    NeuroImage

    (2011)
  • Cited by (34)

    • The relationship between EEG and fMRI connectomes is reproducible across simultaneous EEG-fMRI studies from 1.5T to 7T

      2021, NeuroImage
      Citation Excerpt :

      In the case of simultaneous EEG-fMRI, signal quality is diminished by 1) artifacts on fMRI data induced by the EEG electrodes interacting with the static magnetic field and the MR-pulses (Mullinger et al., 2008b) 2) gradient- and pulse-related EEG artifacts induced by the magnetic gradients on the EEG electrodes and cables (Abreu et al., 2018; Allen et al., 2000). At 7T, EEG-fMRI can be affected by stronger recording artifacts (Abreu et al., 2016; Jorge et al., 2015b; Mullinger et al., 2008a; Neuner et al., 2013). Overcoming those limitations would encourage analysis of the dynamic electrophysiological correlates of BOLD signal at much higher temporal and spatial resolutions (Meyer et al., 2019; Scheeringa and Fries, 2017).

    • Investigating the variability of cardiac pulse artifacts across heartbeats in simultaneous EEG-fMRI recordings: A 7T study

      2019, NeuroImage
      Citation Excerpt :

      Given the spatiotemporal properties of PAs, their correction is essential to the vast majority of EEG-fMRI applications (Mullinger and Bowtell, 2011). Beyond the techniques explored in this work, other relatively popular approaches include independent component analysis (Benar et al., 2003; Mantini et al., 2007; Abreu et al., 2016) and methods based on independent reference signals (Bonmassar et al., 2002; LeVan et al., 2013; Chowdhury et al., 2014), for example. These more recent techniques are often used in combination with AAS, and generally found to produce superior results, but all have their own limitations as well, including suboptimal assumptions regarding the spatiotemporal properties of the PA, and/or the underlying EEG activity (Debener et al., 2008; Mullinger et al., 2013), or the need for additional hardware, for example.

    • An automatic pre-processing pipeline for EEG analysis (APP) based on robust statistics

      2018, Clinical Neurophysiology
      Citation Excerpt :

      However, with the advent of both high-density EEG arrays (64-256 channels) and studies of large populations, yielding increasingly greater amounts of data, supervised methods have become excessively time consuming. To cope with this, and to minimize subjectivity, automatic methods have recently been presented (Abreu et al., 2016a,b; Bigdely-Shamlo et al., 2015; Hatz et al., 2015; Nolan et al., 2010). Fully automated statistical thresholding for EEG artifact rejection (FASTER; Nolan et al., 2010), for instance, enables a fully automated pre-processing of ERP data, based on computing z-scores of different signal metrics, and threshold them in order to detect bad channels, bad epochs and artifactual ICs.

    • EEG synchronization measures predict epilepsy-related BOLD-fMRI fluctuations better than commonly used univariate metrics

      2018, Clinical Neurophysiology
      Citation Excerpt :

      All patients or their legal representatives gave written informed consent and the study was approved by the local ethics committee.” ( Abreu et al., 2016a). Patients were selected for the current study if their EEG-fMRI examination fulfilled the following criteria: (1) no anesthesia was employed, because of its known effects on the BOLD signal, mainly related with the suppression of fluctuations of neuronal origin (Bernal et al., 2012); and (2) clear inter-ictal epileptic activity was recorded on the EEG, in order to maximize the reliability of the hypothesized epileptic networks, which were used as part of the validation of the results.

    View all citing articles on Scopus
    View full text