Rejection of pulse related artefact (PRA) from continuous electroencephalographic (EEG) time series recorded during functional magnetic resonance imaging (fMRI) using constraint independent component analysis (cICA)
Introduction
In recent years, the combination of functional Magnetic Resonance Imaging (fMRI) and electroencephalography (EEG) has been increasingly used to characterize human brain function, with the objective of harvesting the high spatial resolution of the former and the exquisite temporal precision of the latter. Although also its use to characterize evoked response (Czisch et al., 2002, Liebenthal et al., 2003), the full benefit of combining EEG and fMRI is particularly obvious when one is interested in spontaneous brain states or events characterized by specific EEG oscillations or transients, the time course of which is hardly predictable. Accordingly, EEG–fMRI has proven particularly successful in characterizing the cerebral correlates of normal brain oscillations as waking brain rhythms (Goldman et al., 2000, Laufs et al., 2003, Moosmann et al., 2003, Niazy et al., 2004), or sleep transients (Schabus et al., 2007), as well as in localizing the hemodynamic correlates of epileptic activities (Seeck et al., 2001).
A major problem of EEG–fMRI consists of the contamination of EEG recordings by gradient switching (gradient artefacts) and pulse related artefacts (PRA). The former is related to the echoplanar sequence used in fMRI. The latter is thought to have multiple causes related to the interaction between the static magnetic field and the heart beat, such as pulsatile motion of recording electrodes, cardio-ballistic head movements and Hall effects in blood vessel (Allen et al., 1998, Goldman et al., 2000, Srivastava et al., 2005, Nakamura et al., 2006). Both artefacts are larger than the genuine EEG signal and have to be removed before EEG can be properly analyzed.
Because the gradient artefact is very reproducible across consecutive volume acquisitions, it is satisfactorily removed from raw data by subtracting an averaged artefact waveform (Average Artefact Subtraction (AAS)), followed by adaptive noise cancellation to reduce any residual artefact (Allen et al., 1998). In contrast, PRA is much harder to reject. Indeed, its morphology and its topography substantially vary from beat to beat. Because of this non stationarity, methods based on a simple AAS, although reasonably efficient, result in residual artefacts in the data because they assume that the PRA is a slowly changing signal that can be accurately characterized by a moving average. Various methods were derived from AAS modelled PRA using exponential weighted average (Goldman et al., 2000), median filter (Sijbers et al., 2000) or adaptive filters (Han et al., 2004, Wan et al., 2006) but eventually produced similar results (Grouiller et al., 2007).
To allow more flexibility in the characterization of PRA, other channel-wise model-explicit rejection algorithms were based on temporal Principal Component Analysis (PCA, Niazy et al., 2005) or a moving General Linear Model (mGLM, Vincent et al., 2007). Others used estimates of the motion artefact noise by piezoelectric motion sensors (Bonmassar et al., 2002, Masterton et al., 2007) and an adaptive filtering algorithm based on the Kalman filter (Bonmassar et al., 2002). These led to relatively efficient solutions but required extra equipment to be placed on and around the subject.
Another class of PRA removal techniques is based on blind source separation. Methods using Independent Component Analysis (ICA) decompose the EEG recordings into independent components (ICs) and suppress the PRA-related ones (Benar et al., 2003, Nakamura et al., 2006, Finelli et al., 2003, Han et al., 2004, Mantini et al., 2007). The problems of these methods are non-reproducibility of the results (Briselli et al., 2006, Grouiller et al., 2007) and the difficulty in objectively selecting the PRA-related ICs. To address the issue of ICA variability, it was suggested to apply the ICA algorithm iteratively and to average the results across multiple executions (MICA) but this procedure requires an overwhelming computational power (Briselli et al., 2006, Phillips et al., 2006) and is very time-consuming. To address the issue of IC selection, an automatic method based on the correlation between the IC and the ECG was developed (Srivastava et al., 2005) but this type of detection does not always perform well in detecting PRA-related ICs (Debener et al., 2007). Finally, it has also been argued that ICA could not be used to reject PRA for recordings obtained in a high-field scanner without additional pre-processing because ICA is a fixed spatial filter and the PRA was observed to have a variable within-cycle topography (Debener et al., 2008).
In the framework of sleep studies, the conditions for PRA rejection can be particularly difficult. During the deepest stages non rapid eye movement (NREM) sleep, also called slow-wave-sleep (SWS), EEG recordings are characterized by abundant high-amplitude slow waves. In that case, the difficulty in rejecting PRA from EEG recordings arises from the similarity between slow waves and PRA both in the temporal and frequency domains. Slow waves and PRA typically recur about every second and their amplitudes are in the same order of magnitude (typically several hundred µV peak to peak at 3 T). Similarly, the bulk of the spectral power of both sleep slow waves and PRA lies between 0.5 and 5 Hz. Consequently, the estimation of a model of PRA using classical methods (non-ICA based) is likely to be confounded by genuine neurally-generated slow waves and the ability of these methods to retrieve slow waves during NREM sleep is questionable.
Yet, slow waves of deep NREM sleep represent a fundamental brain activity associated with important functional properties. For instance, they have been used as a measure of sleep depth and their power density (between 0.75 and 4 Hz) is used as a reliable parameter quantifying the dissipation of the homeostatic sleep pressure (Dijk et al., 1997). Slow waves have also been implicated in offline memory processing either because they would allow for a local synaptic downscaling (Huber et al., 2004) or because they might promote the replay of neuronal firing sequence supposedly representing the learned material (Buzsaki, 1996).
In order to allow the study of SWS using EEG–fMRI, we developed a new efficient, robust, and computationally efficient PRA-rejection technique based on constrained ICA (cICA). This novel method is equally efficient in rejecting PRA for EEG data recorded during wakefulness and NREM sleep.
Section snippets
PRA rejection algorithm
In this section we will introduce our cICA-based PRA rejection algorithm as well as former ones which we will use for comparison. Since each of these techniques uses a QRS peak detection algorithm in ECG, we decided to use the same algorithm for each of them and describe this one in the next subsection. Note that we use the FMRIB plug-in for EEG, provided by the University of Oxford Centre for Functional MRI of the Brain (http://users.fmrib.ox.ac.uk/~rami/fmribplugin/index.html), for the
Methods and materials
The section is organized in two parts. In the first one, EEG data recorded on 2 sleeping volunteers are processed. They were simply placed in the static field of the MRI scanner, but in the absence of fMRI acquisitions (in order to avoid any gradient artefact in the EEG signal) during NREM sleep. The second part uses EEG data acquired on 9 volunteers in various consciousness states (wakefulness, light and deep non rapid eye movement sleep), during simultaneous fMRI imaging. All volunteers were
EEG recordings without gradient artefact
The PRA was obviously much larger when the head of the subject was placed within the MR scanner than outside the scanner (Fig. 2). The comparison of these recordings thus allowed for the estimation of the sensitivity and specificity of various artefact rejection methods.
We first characterized the power spectrum of PRA and its scalp topography (Fig. 3). We computed the power spectrum of data recorded inside and outside the scanner bore. The power spectrum of data recorded outside the scanner
Discussion
On EEG recordings, deep non rapid eye movement sleep is characterized by large amplitude slow waves which typically lasts about 1 s. In the framework of EEG–fMRI studies on deep NREM sleep, removal of PRA from EEG recordings are problematic because the period and amplitude of sleep slow waves and PRA overlap. In this paper, we introduce constrained ICA as a novel method for PRA rejection from EEG recorded simultaneously with fMRI during slow wave sleep. We first evaluated cICA in datasets
Acknowledgments
This study was supported by the Belgian Fonds National de la Recherche Scientifique (F.R.S.-F.N.R.S.), the Fondation Médicale Reine Elisabeth, the Research Fund of ULg, and PAI/IAP Interuniversity Pole of Attraction. Y.L., E.B., T.D., P.M. and C.P. are supported by F.N.R.S. (Belgium).
References (40)
- et al.
Identification of EEG events in the MR scanner: the problem of pulse artifact and a method for its subtraction
Neuroimage
(1998) - et al.
A method for removing imaging artifact from continuous EEG recorded during functional MRI
Neuroimage
(2000) - et al.
Quality of EEG in simultaneous EEG–fMRI for epilepsy
Clin. Neurophysiol.
(2003) - et al.
Motion and ballistocardiogram artifact removal for interleaved recording of EEG and EPs during MRI
Neuroimage
(2002) - et al.
An independent component analysis-based approach on ballistocardiogram artifact removing
Magn. Reson. Imaging
(2006) - et al.
Altered processing of acoustic stimuli during sleep: reduced auditory activation and visual deactivation detected by a combined fMRI/EEG study
Neuroimage
(2002) - et al.
Improved quality of auditory event-related potentials recorded simultaneously with 3-T fMRI: removal of the ballistocardiogram artefact
Neuroimage
(2007) - et al.
Properties of the ballistocardiogram artefact as revealed by EEG recordings at 1.5, 3 and 7 T static magnetic field strength
Int. J. Psychophysiol.
(2008) - et al.
Acquiring simultaneous EEG and functional MRI
Clin. Neurophysiol.
(2000) - et al.
A comparative study of different artefact removal algorithms for EEG signals acquired during functional MRI
Neuroimage
(2007)
Independent component analysis: algorithms and applications
IEEE Trans. Neural Netw.
Improved ballistocardiac artifact removal from the electroencephalogram recorded in fMRI
Neurosci. Methods
EEG-correlated fMRI of human alpha activity
Neuroimage
Simultaneous ERP and fMRI of the auditory cortex in a passive oddball paradigm
Neuroimage
Complete artifact removal for EEG recorded during continuous fMRI using independent component analysis
Neuroimage
Measurement and reduction of motion and ballistocardiogram artefacts from simultaneous EEG and fMRI recordings
Neuroimage
Correlates of alpha rhythm in functional magnetic resonance imaging and near infrared spectroscopy
Neuroimage
Removal of FMRI environment artifacts from EEG data using optimal basis sets
Neuroimage
Reduction of ECG and gradient related artifacts in simultaneously recorded human EEG/MRI data
Magn. Reson. Imaging
ICA-based procedures for removing ballistocardiogram artifacts from EEG data acquired in the MRI scanner
Neuroimage
Cited by (30)
Artifact removal using deep WVFLN for brain signal diagnosis through IoMT
2022, Measurement: SensorsBallistocardiogram artifact removal in simultaneous EEG-fMRI using generative adversarial network
2022, Journal of Neuroscience MethodsCitation Excerpt :The method of removing BCG artifact based on ECG signal is difficult to deal with all the artifacts because ECG and BCG are not completely linear correlation (Marino et al., 2018), positioning deviation will lead to reduced artifact removal effect. The method of removing BCG artifact by selecting components also has a similar situation, and component selection (Leclercq et al., 2009; Liu et al., 2012) plays a vital role in the artifact removal effect. The brain system is so complex that it is difficult to find a model to accurately describe the brain.
Time-frequency analysis of resting state and evoked EEG data recorded at higher magnetic fields up to 9.4T
2015, Journal of Neuroscience MethodsEEG-fMRI: Dictionary learning for removal of ballistocardiogram artifact from EEG
2015, Biomedical Signal Processing and Control