We here compared results achieved by applying popular methods for reducing artifacts in magnetoencephalography (MEG) and electroencephalography (EEG) recordings of the auditory evoked Mismatch Negativity (MMN) responses in healthy adult subjects. We compared the Signal Space Separation (SSS) and temporal SSS (tSSS) methods for reducing noise from external and nearby sources. Our results showed that tSSS reduces the interference level more reliably than plain SSS, particularly for MEG gradiometers, also for healthy subjects not wearing strongly interfering magnetic material. Therefore, tSSS is recommended over SSS. Furthermore, we found that better artifact correction is achieved by applying Independent Component Analysis (ICA) in comparison to Signal Space Projection (SSP). Although SSP reduces the baseline noise level more than ICA, SSP also significantly reduces the signal—slightly more than it reduces the artifacts interfering with the signal. However, ICA also adds noise, or correction errors, to the waveform when the signal-to-noise ratio (SNR) in the original data is relatively low—in particular to EEG and to MEG magnetometer data. In conclusion, ICA is recommended over SSP, but one should be careful when applying ICA to reduce artifacts on neurophysiological data with relatively low SNR.
Recordings of evoked-responses (also known as event-related potentials, ERPs, or event-related fields, ERFs) with electroencephalography (EEG) or magnetoencephalography (MEG) are widely used methods in cognitive and clinical neuroscience. One of the major challenges in research and clinical applications of evoked-responses is the prevalent strongly interfering electromagnetic signals from external objects and devices in the surrounding MEG or EEG measurement environment as well as nearby mechanical and biological electromagnetic sources originating from the head and other parts of the body of the subject. Since the interfering environmental noise from, for example, laboratory mechanics and electronic devices may be several orders of magnitude stronger than the brain signals of interest (for a review, see, e.g., [
In the clinical routine, data from patients having a limited control of muscular activity (such as stroke or dementia patients or preterm infants) or with ferromagnetic implants (such as cochlear implantees) typically contain a considerable amount of artifacts. The time constraints of experiments and tests on clinical populations exclude the possibility of a large number of trials that would allow discarding the artefactual ones. A viable alternative to simply rejecting parts of the recorded data is that of correcting the data. Both in clinical patient recordings and in experimental settings with healthy subjects, strong electromagnetic noise from electronic devices, static electricity, and in particular with regard to EEG also the 50/60-Hz power-line noise may interfere significantly with the measurements [
Apart from external artifact sources, it is important to reduce the influence of the internal artifacts originating from the head and the rest of the body of the subject. Typically, MEG and EEG recordings are contaminated by relatively strong artifacts caused by the eyes [
Different methods have been developed to reduce the influence of externally and internally originating artifacts. The externally originating interference can be minimized by applying physical shielding techniques in the laboratory [
The tSSS method is additionally able to filter out interferences from artifactual sources in the space between the brain and the MEG sensor array, by reducing signals in the common subspace through comparisons of the time series in the internal and external spaces. For instance, it has been shown that tSSS makes it possible to locate brain sources on the cortex with beamformer methods in clinical patients, although these patients wore strongly magnetically interfering dental braces; thereby, tSSS seems to allow extending the clinical population compatible with MEG [
The internally originating artifacts can be reduced by applying band-pass filtering [
To investigate the performances and risks of using different popular artifact correction methods, we compared here results achieved by applying SSS, tSSS, ICA, and SSP. We chose to study the performance of the correction methods on the Mismatch Negativity (MMN) response, which is a well-known evoked-response [
A sample of ten volunteers from a larger database named “Tunteet” was chosen (for a description of the experimental protocol, see Kliuchko et al., submitted). The participants were six females and four males. Three participants were nonmusicians, three were amateur musicians, and four were professional musicians. All participants were right-handed, and their average age was 24.8 years (range 18–35 years). Written informed consent was obtained from each participant, and the study was approved by the local ethics committee.
The participants listened to a melody pattern of 2100 ms, repeated with variations during ~25 minutes. The melody patterns were created from digital piano tones (McGill University Master Samples) and followed the rules of Western tonal music. All melodies started with a triad (300 ms) followed by four single tones (two of 125 ms and two of 300 ms) and an ending tone (575 ms) all separated by 50 ms silent gap. Between all melodies there was a silent gap of 125 ms. Deviations of six types were inserted into the melody patterns to evoke MMN responses to a deviant tone as compared with corresponding unaltered standard. The deviants are explained in Table
Deviant tones.
Type of deviant | Description | Occurrence | Tone number | Melodies with deviants |
---|---|---|---|---|
Mistuning | 3% pitch frequency increase | Single | 1st, 2nd, 4th | 14% |
Timbre | Flute sound | Single | 1st, 3rd, 4th, ending | 8% |
Timing delay | 100 ms silent gap before the tone | Single | 1st, 2nd, 3rd, ending | 8% |
Melody modulation | Tone replacement | Consistent | 3rd, 4th | 12% |
Rhythm modulation | Duration switch between two successive tones | Consistent | 2nd, 3rd | 7% |
Transposition | Semitone pitch change of a pattern | Consistent | Initial triad | 16% |
In total, the tested sample contained 120 brain responses, which comprised the responses to the six standard and six deviant conditions from each of the ten participants.
The simultaneous MEG and EEG data were collected at the BioMag Laboratory of the Helsinki University Central Hospital. The measurements were performed in an electrically and magnetically shielded room (ETS-Lindgren Euroshield, Eura, Finland) with Vectorview
During the measurement, subjects were comfortably seated and watched a silenced movie with subtitles. The stimuli were presented with Presentation software (Neurobehavioral Systems, Ltd.). The sound was delivered through a pair of pneumatic headphones at individually adjusted loudness.
Elekta Neuromag MaxFilter 2.2 Signal Space Separation (SSS) and temporal Signal Space Separation (tSSS) [
Correction for internal artifacts with Signal Space Projection (SSP) was performed with the MNE Python version 0.11.0 released with the MNE software version 2.7.4-3485 [
Independent Component Analysis- (ICA-) based correction for internal artifacts was achieved by applying the logistic infomax algorithm implemented in the
Event-related EEG and MEG responses were extracted as single-trial epochs with a time window of 0 to 400 ms after the standard or deviant tone onset. The trials were baseline-corrected by applying a baseline of −100 to 0 ms before the tone onset. Since the planar gradiometer sensors measure along two orthogonal directions, the data from each pair of longitudinal and latitudinal sensors were combined by applying the Pythagorean distance formula, as implemented in the FieldTrip toolbox for MATLAB (Donders Institute for Brain, Cognition and Behaviour/Max Planck Institute, Nijmegen, Netherlands) [
For the sake of clarity, we performed the subsequent analyses on one channel of each sensor type, those in which the highest MMN amplitude was measured within a typical MMN latency range of 75–200 ms. In this case, we analyzed the event-related waveforms from EEG channel 012 (frontal site), magnetometer channel MEG 1341 (right temporal site), and the combined planar gradiometer channels MEG 2222 and 2223 (right temporal site) (see Figure
Topographical plots of MMN responses (without correction for internal artifacts). Showing topographical plots of MMN responses to the mistuning deviants, which elicited the strongest and most consistent MMN responses in the tested sample in the EEG (a), MEG magnetometers (b), and MEG gradiometers (c). The plots are based on grand averages from uncorrected data (only preprocessed with tSSS) by applying a time window from 125 to 155 ms after the tone onset.
We measured the MMN amplitude in response to each type of deviant tone by taking the average value across the time window from 125 to 155 ms after the tone onset. To compare the noise levels after utilizing each artifact correction method, we first used a baseline standard deviation measure. Since a flat baseline is desirable, we applied a baseline standard deviation (STD) measure to show the flatness of the baseline in a single trial (where lower baseline STD means a more flat baseline) [
The Mismatch Negativity (MMN) evoked-response is analyzed by comparing the average response to the deviant stimulus with the average response to the standard stimulus [
Statistical comparisons were made with SPSS version 20 (IBM, Armonk, New York, USA). Since the resulting values were not normally distributed, we applied the Wilcoxon signed-rank test to compare the values achieved after utilizing the different artifact correction methods.
A statistically significant and slightly better reduction of the signal standard deviation (STD) is achieved by applying tSSS in comparison to SSS for for both the MEG magnetometer and gradiometer data (see Figure
Percentagewise standard deviation (STD) reduction in the signal latency range achieved by applying the SSS or tSSS methods in comparison to the level (at 0% STD) in the raw data recording without spatial filtering. The differences are shown for the amplitude peak channel of the magnetometers (a) and gradiometers (b) in Tukey box plots. Bold horizontal lines indicate medians, brown boxes indicate the first 50 percent of the case range, and vertical lines indicate cases within the 1.5 interquartile range from the edge of the 50% of the cases. Outliers more than 1.5 (circles) or 3.0 (stars) interquartile ranges from the edge of the 50% of the cases are denoted with circles or stars, and case numbers are provided for each outlier. The median values are shown in femtoteslas (fT) and femtoteslas per centimeter (fT/cm).
From grand average waveforms of the event-related responses (across all participants and conditions), it can be seen that the SSP-based artifact correction reduces the signal amplitude, whereas the signal amplitude is similar before and after the artifact correction based on ICA (see Figure
Grand average event-related waveforms. Showing the channels with the signal peak amplitude among the EEG channels, magnetometers, and gradiometers, without any artifact correction (black) and after artifact correction based on Independent Component Analysis (ICA) (red) and Signal Space Projection (SSP) (blue).
The SSP method results in lower baseline standard deviation (STD) and signal STD in comparison to the ICA method (see Figure
Percentagewise baseline standard deviation (STD) and signal STD reduction achieved by applying ICA or SSP methods in comparison to the level (at 0% STD) in the raw data recording with tSSS only. Median values are shown in microvolts (
Signal-to-noise ratios (SNR) achieved by applying artifact correction based on Independent Component Analysis (ICA) or Signal Space Projection (SSP). Tukey box plot showing the SNR levels achieved by applying the ICA or SSP method in the signal amplitude peak channel of the EEG, magnetometers (Mag.), and gradiometers (Grad.) (the SNR shows the relationship between the signal and noise level in single trials). Outliers more than 1.5 (circles) or 3.0 (stars) interquartile ranges from the edge of the 50% of the cases are denoted with circles or stars, and case numbers are provided for each outlier.
The SNR achieved by applying the ICA-based artifact correction is similar to that achieved by applying the SSP method with respect to the EEG and magnetometer channels. However, for the gradiometers applying the ICA method results in statistically significantly and slightly better SNR than that achieved by applying the SSP method (see Figure
We compared the noise suppression results achieved with SSS and tSSS on healthy subjects not wearing magnetized material. MaxFilter with tSSS resulted in better suppression of artifacts from external and nearby noise sources in comparison to SSS. In particular, the application of tSSS instead of SSS was important with respect to the MEG gradiometers, since SSS correction in 6% of the cases resulted in an increase of the noise level in the MEG gradiometer data, and thus the reliability of the gradiometer data decreased in comparison to that before SSS. We also compared the performance of ICA and SSP in reducing internal electrophysiological artifacts, originating from eye movements and heart beats of the participants. The ICA-based artifact correction performed better than the SSP method. The SSP method reduced part of the signal of interest along with the artifacts, and the SNR was slightly higher after applying the ICA method than after applying the SSP method. However, after ICA-based artifact correction the baseline noise level increased, in particular in the EEG and magnetometer channels, which have relatively low SNR in the original data. These findings support both the importance of reducing the bias on measures of evoked-responses with EEG and MEG caused by artifacts and the importance of minimizing the bias introduced by errors in artifact correction methods.
With regard to the suppression of external noise, we here observed that the averaged MMN evoked-response in 6% of the cases with the MEG gradiometers even became more unreliable after applying SSS correction than before. Possibly, the influence of nearby artifacts on the evoked-response can become stronger after the external artifacts have been reduced with the SSS. When tSSS is applied the influence of such nearby artifacts on the evoked-responses would be reduced. In general, it seems relevant to further investigate whether correction methods for reducing one type of artifact, such as SSS, in some cases might enhance the influence of other artifacts on the averaged event-related waveform.
The comparisons of the ICA and SSP methods for suppression of internal artifacts revealed particular biases appearing after the corrections. For the SSP method, there is a risk that the artifacts and signals of interest are not described by orthogonal components [
In summary, our test results suggest that tSSS is recommendable for reducing the influence of artifacts originating from external and nearby sources instead of SSS only. We find that the noise level decreases more with tSSS than with SSS in this sample of EEG and MEG data from healthy participants despite the fact that they were not wearing strongly magnetized materials. For the reduction of internal physiological artifacts, we showed that the highest signal-to-noise ratio (SNR) is achieved with ICA-based artifact correction on the tested sample. However, both ICA- and SSP-based artifact corrections are subject to certain limitations. In particular, one must be aware of the risk when processing data with relatively low SNR, such as EEG and MEG magnetometer data, that artifact correction based on ICA may decrease the interference from artifacts while simultaneously increasing the noise level, due to increasing errors in estimating the mixing matrix in the context of data with lower SNR levels.
The authors declare that they have no competing interests.