Safety and data quality of EEG recorded simultaneously with multi-band fMRI
by Maximillian K. Egan1,2, Ryan Larsen2, Jonathan Wirsich2,3, Brad P. Sutton2,4, Sepideh Sadaghiani1,2
1Psychology Dept., Univ. of Illinois At Urbana-Champaign, Urbana, IL (USA)
2Beckman Institute for Advanced Science and Technology, Univ. of Illinois At Urbana-Champaign, Urbana, IL (USA)
3EEG and Epilepsy Unit, Univ. Hospitals and Faculty of Medicine of Geneva, Geneva (Switzerland)
4Bioengineering Dept., Univ. of Illinois At Urbana-Champaign, Urbana, IL (USA)
Abstract
Simultaneously recorded EEG-fMRI is highly informative yet technically challenging. Until recently, there has been little information about EEG data quality and safety when used with newer multi-band (MB) fMRI sequences. Here, we measured the relative heating of a MB protocol compared with a standard single-band (SB) protocol considered to be safe. We also evaluated EEG quality recorded concurrently with the MB protocol on humans during a resting state scan, as well as with the SB protocol and in an fMRI-free condition to allow for direct comparison. The heating induced by the MB sequence was lower than that of the SB sequence, and the resting state EEG comprised peaks in the power spectrum in the alpha frequency range consistent with offline eyes-closed EEG. Taken together, these results indicate the safety of radiofrequency-related heating, as well as acceptable data quality using traditional artifact removal techniques, for simultaneous EEG-fMRI with MB sequences of the entire brain with high temporal resolution (TR=440ms).
Introduction
Due to the non-invasive nature of recording coupled with ease of access to equipment, the use of simultaneous EEG-fMRI is becoming more commonplace. However, there are many technical challenges in acquiring high quality EEG-fMRI data, and little work has assessed how these challenges are further affected by the use of multi-band (MB) fMRI sequences. Previous investigations into quality of EEG recorded concurrently with MB fMRI have demonstrated suitable data quality for several experimental conditions, but the free parameters used in MB sequences change with the needs of individual experiments.
More important than data quality when considering a new MB EEG-fMRI sequence is ensuring subject safety. For EEG-fMRI the key safety concern is the deposition of radiofrequency (RF) power that causes heating in the EEG leads and electrodes, with only a few studies demonstrating safe RF heating using MB EEG-fMRI [1–4]. Electrode heating has previously been characterized as a function of the Specific Absorption Rate (SAR) [1,5], though using SAR has several problems and limitations [6–8]. Given these problems it has been suggested to characterize safety limits using B1+RMS, which is a fixed characteristic of the sequence and protocol and depends on the time-averaged RF amplitude transmitted by the scanner [7]. Brain Products has recently begun to use B1+RMS to specify safety limits [9], but B1+RMS is a relatively new standard, its use to assess the safety of simultaneous fMRI-EEG has been limited to the most recent literature on the safety of MB sequences [2].
The heating associated with a given experiment can be characterized in absolute or relative terms, with most previous EEG-fMRI safety studies reporting the absolute heating of MB sequences during scanning [1–3]. The few reports that have included both MB and single-band (SB) sequences have not related these differences to differences in B1+RMS [1,3]. Here we measure the relative heating of a MB sequence compared to a SB sequence that has been established as safe under a wide variety of conditions [10]. After demonstrating the safety of the protocol using a relative temperature measurement, we evaluate the efficacy of using traditional artifact rejection methods to determine acceptable EEG data quality.
Methods & Results
We completed two different experiments in this study; 1) a phantom recording for assessing electrode heating and safety using the MB sequence in simultaneous EEG-fMRI, and after establishing safety, 2) eyes-closed resting state recordings in human subjects utilizing simultaneous EEG-fMRI with the same MB sequence to assess EEG data quality.
Human subjects
Nine subjects (four female) underwent simultaneous EEG-fMRI recordings. All subjects gave written informed consent according to procedures approved by the Institutional Review Board of the University of Illinois at Urbana‐Champaign.
fMRI
Data were acquired on a 3 T Prisma scanner with a 64-channel head coil (Siemens, Erlangen, Germany). Our MB protocol was chosen to minimize TR, while still maintaining the resolution and coverage typically used for non-MB sequences (TR= 440ms, 28 slices, MB factor = 4). We compared the MB sequence against a sequence chosen to comply with the maximum intensity limits recommended by the manufacturer of the EEG equipment (Brain Products; [10]), which we name SB (TR: 2000ms, 25 slices). For the MB protocol, B1+RMS was 0.8 µT; whereas B1+RMS of the SB protocol was 1.0 µT. Note that B1+RMS is a property of the sequence and does not depend on the properties of the object being scanned. For full imaging parameters of each sequence see Egan et. al 2021 [11].
Data acquisition
EEG Equipment and Setup for Phantom and Human Experiments
EEG data was recorded with a 64-channel EEG cap (BrainCap MR) in a 10-20 montage from Brain Products that included 61 scalp electrodes and 3 drop-down electrodes (HEOG, IO, ECG) using the manufacturer’s BrainVision Recorder software (BrainVision Recorder; [12]). The cap was connected to two MR-compatible BrainAmp MR plus 32-channel EEG amplifiers within the scanner bore. All impedances were kept <5 kOhms. For scans using EEG-fMRI, the amplifiers and battery pack were strapped down and weighted with sandbags on a stabilizing sled from Brain Vision LLC to reduce vibration artifacts. The EEG recording hardware was directly connected to the SyncBox. We used a custom-made RTBox [13] connected to the scanner to place scanner pulse (TR) markers directly in the EEG file. Scalp electrodes (10 kOhm built-in resistors) were recorded at 0.5 µV resolution, while drop-down electrodes (20 kOhm built-in resistors) were recorded at 10 µV resolution. All electrodes had a a low cutoff filter with a 10s time constant, high cutoff filter of 250 Hz, and a sampling rate of 5000 Hz. This combination of microvolt resolutions, sampling rate, and cutoffs gave us the highest possible recording resolution while avoiding amplifier overloading. Preventing overloading of the amplifiers is critical for EEG-fMRI to ensure that the peaks of the artifact can be detected for gradient artifact (GA) / ballistocardiogram (BCG) artifact rejection. Additionally, during all human subject runs the scanner’s helium pump was turned off to eliminate vibration artifacts at 42 Hz from the pump.
Setup of Heating Experiments in Phantom
During the electrode heating tests a watermelon ‘phantom’ was fit with the 64-channel cap. The watermelon provides a conductive surface and permits fine-grained control over impedances at electrode contacts [3,14]. We abraded the watermelon with sandpaper prior to placing the cap, which allowed us to maintain impedances <5 kOhm and protect the amplifiers from high voltages induced by the scanner. Given the small size of the watermelon, the ECG electrode was routed underneath the watermelon once and placed in between the HEOG and IOG electrodes to ensure that no loop was created within the magnet (Fig 1). The monitored electrodes were exposed to the ambient air, without cushioning or other barriers to heat transfer. Temperature changes during scanning were measured with a Luxtron 812 two channel fluoroptic thermometer (LumaSense Technologies, Ballerup, Denmark). Fluoroptic probes were placed in the conductive paste between the electrodes and the watermelon surface.
Figure 1. Watermelon ‘phantom’ outfitted with the 64-channel Brain Products EEG cap (BrainCap MR)
Acquisition of Heating Data in Phantom
We measured the relative heating of the MB protocol vs. the SB protocol in three separate phantom experiments. For each experiment the temperature probes were placed beneath the ECG electrode and one other electrode (TP7, TP8, and FP2, respectively). For each experiment, both sequences (MB/SB) were run three times in an alternating fashion. The purpose of alternating the scans was to minimize bias due to long-term drift of the temperature. Each of the runs consisted of approximately 13.5 minutes of scanning and were spaced by approximately 5 minutes of rest between scans.
During the experiments, the fluoroptic thermometer performed two measurements per second; data were recorded by a computer connected to the Luxtron unit via a serial cable.
Processing of Phantom Heating Data
Temperature data were processed using MATLAB® 2020a. Temperature measurements were smoothed using the “smoothdata” function using a Gaussian-weighted moving average filter with a window length of 200 seconds. The first pair of MB /SB runs was discarded from experiment 2 because scanning commenced before the phantom had equilibrated with room temperature.
Acquisition of EEG Data in Humans
We obtained EEG data with simultaneous fMRI from nine human subjects during eyes-closed resting state. Six of the subjects underwent two EEG-fMRI runs of 10 minutes on two separate days using the MB sequence (20 minutes total). To directly compare the quality of EEG acquired during the MB sequence, the SB sequence, and in the absence of active scanning sequences, we collected data from three additional subjects. The three subjects underwent 5 minutes of data acquisition inside the MRI scanner for each of the MB, fMRI-free, and SB conditions, in that order.
Processing of Human EEG Data
EEG data was preprocessed using the BrainVision Analyzer software (Version 2.2) [15]. GA subtraction was performed first followed by BCG artifact rejection. The GA subtraction used marker detection from the trigger pulse markers obtained directly from the scanner with a continuous artifact [16]. A baseline correction over the whole artifact was used with a sliding average calculation of 21 marker intervals. Bad intervals were corrected with the average of all channels. The data was not downsampled (to minimize preprocessing). A lowpass finite impulse response (FIR) filter was applied at 100 Hz. The data was then segmented to include only artifact-free resting-state data, and BCG artifact rejection was performed using semi-automatic mode.
After correcting all marked heartbeats manually, artifact removal of the heartbeat template was performed using sequential 21 pulse templates as the template average.
EEG data were analyzed in MATLAB 2018b® using EEGLAB (Version 2019.1) [17]. Prior to spectral analysis, the data was passed through a second lowpass FIR filter at 70 Hz. The spectrograms for scalp channels of each subject were computed at a frequency resolution of 0.2 Hz using the Multitaper approach. We chose not to do any further processing of the data (i.e., Independent Component Analysis decomposition) as we wanted to show the quality of the data with the minimum amount of cleaning resulting from the BrainVision Analyzer artifact rejections.
Results
Relative heating of the MB vs. SB sequences
In all experiments the temperature of the electrodes increased during scanning periods, as shown in Fig 2. Superposition of the temperature changes during each of the scans reveal consistently greater heating of the SB protocol for all electrodes, as shown in Fig 3. The relative heating of the two sequences was estimated by dividing the rate of heating during each MB sequence with that of the SB sequence that was recorded immediately following. Average values of these ratios from the four electrodes were in the range of 0.52 to 0.82, and the combined heating ratio from all electrodes is 0.73 ± 0.38, as shown in Table 1. The relative heating of the two sequences therefore is in approximate agreement with the RF power deposition ratio of 0.64 derived from the scanner. This implies that the differences in heating between the two sequences are captured by the total RF power deposition.
Statistical analysis (cf. main publication) revealed that heating for the MB sequence was significantly higher than for the SB sequence. Because the empirically measured heating was below that of the safe SB sequence, our MB sequence can be considered to be safe.
Figure 2. Temperature measurements in the watermelon “phantom”: The total course of consecutive measurements during the different conditions are color coded for the MB sequence (red), the SB sequence (blue), and rest periods of no scanning (green). The ECG electrode, which has the longest lead and highest potential of heating, was included in all three experiments A through C. Experiments A, B, and C additionally measured temperature at TP7, TP8, and FP2, respectively.
Table 1. Comparison of average heating rate of the MB and SB protocols at the ECG, TP7, TP8 and FP2 electrodes in the watermelon phantom.
Figure 3. Superposition of the individual temperature measurements in the watermelon “phantom”. Traces from the individual conditions that were recorded consecutively are superimposed for better comparison (MB sequence = red, SB sequence = blue).
EEG Data quality validation
To check EEG signal quality, we assessed the spectral dominance and topography of the alpha frequency band upon GA and BCG artifact cleaning. Alpha-band oscillations are uniquely positioned for this purpose because their exceptionally high power rises above the 1/f aperiodic component of the EEG spectrum during eyes-closed resting state. Fig 4 shows the log-power spectrogram of the EEG during eyes-closed resting state in a single subject (Subject 2). We demonstrate that each successive step substantially improves data quality, first showing the raw data without GA or BCG artifact subtraction (Fig 4A), then with only the GA artifact cleaned (Fig 4B), and finally the fully cleaned data with both GA and BCG artifacts removed (Fig 4C).The cleaned spectrogram showed a clear power peak in the alpha range (~10Hz) while displaying only a minimal residual power increase related to the GA at the RF excitation repetition frequency (Fig 4D).
Figure 4. Log-power spectrogram of the EEG during eyes-closed resting state in a single human subject during concurrent fMRI recordings with the MB sequence. The spectrogram is shown for the (A) raw EEG data without artifact removal, (B) the EEG data with GA artifact rejected but prior to BCG correction, and (C) the cleaned EEG data after both GA and BCG artifact correction. Each trace corresponds to one of 60 scalp channels (1 channel excluded due to excessive noise). (D) shows the scalp topography at 10Hz for the cleaned EEG data. The prominent power peak at ~10Hz emerging more clearly after artifact correction (in C) and the posterior topography (in D) are consistent with the spectral dominance of the alpha rhythm in eyes-closed resting state. The dotted line represents the frequency of the RF repetition artifact at 15.9 Hz. This subject corresponds to Subject 2 in Fig 5.
Fig 5a shows the log-power spectrum for the 9 subjects averaged across posterior electrodes typically capturing the highest power in the alpha range. At each subject’s individual alpha peak frequency, a posterior topography (derived from all channels) was detectable in all individual subjects (Fig 5B), while other power peaks in lower frequencies were greatly attenuated. The large spikes shown in Subjects 1 and 5 near 47 Hz were due to the scanner bore fan being on during the scan for subject comfort.
Fig 6 shows the direct comparison across the log-power spectrum of the MB, SB, and fMRI-free conditions for the final three subjects after removing GA and BCG artifacts. For all subjects the EEG recorded during the MB sequence showed a similarly clean power spectrum compared to the SB sequence. EEG in both fMRI conditions were comparable to the fMRI-free condition, except for an overall drop in power of matching magnitude for both sequences broadly across frequencies, a likely side effect of GA and BCG artifact subtraction.
We conclude that EEG is of sufficient quality for cognitive neuroscience research using the MB sequence after application of artifact rejection methods originally developed for SB sequences, specifically template subtraction [16,18].
Figure 5. EEG log-power spectrogram and topographies during eyes-closed resting state for all human subjects during concurrent fMRI recordings with the MB sequence. (A) Power spectral density averaged across 8 posterior channels of each subject (O1, O2, Oz, PO3, PO4, PO7, PO8, POz). The stars denote the individual subjects’ maximum values within the alpha band (8-12 Hz) i.e. individual alpha peak. The dotted line represents the scanner artifact at 15.9 Hz. (B) Corresponding scalp topographies using all channels for the 9 subjects at the individual alpha peak frequency.
Figure 6. Comparison of EEG log-power spectrogram across different concurrent fMRI conditions during eyes-closed resting state for three human subjects. Power spectral density averaged across all channels directly comparing the MB sequence, SB sequence, and fMRI-free conditions for Subjects 7-9. The dotted black line denotes the RF excitation repetition frequency for the MB and SB sequences (15.9 Hz and 16 Hz respectively, indistinguishable on this figure).
Discussion
Simultaneously recorded EEG-fMRI is a powerful tool that can provide information beyond what unimodal approaches are able to [19–21]. Although EEG-fMRI using traditional SB sequences is well established, safety and data quality of EEG-fMRI imaging using modern MB fMRI sequences is less understood. Here we demonstrate that a particular MB sequence with high temporal resolution (TR = 440 ms) produces less RF heating at the EEG electrode sites than a traditional EPI sequence while maintaining acceptable EEG data quality. Our results also demonstrate the potential of a relative measurement to compensate for drift, the value of alternating between sequences multiple times while performing such measurements, and as a whole support the usefulness of B1+RMS as a benchmark for assessing protocol safety. Additionally, after GA and BCG artifact rejection a clear posterior topography at a readily identifiable individual alpha peak frequency was observed in all human subjects, indicating that traditional MR artifact rejection techniques [16,18] are sufficient for use in extended MB EEG-fMRI recordings. This research adds additional support to the growing body of literature on the safety and efficacy of multiband simultaneous EEG-fMRI imaging.
References
1. Foged MT, Lindberg U, Vakamudi K, Larsson HBW, Pinborg LH, Kjær TW, et al.
Safety and EEG data quality of concurrent high-density EEG and high-speed fMRI at 3 Tesla.
PLoS One. 2017;12. doi:10.1371/journal.pone.01784092. Uji M, Wilson R, Francis ST, Mullinger KJ, Mayhew SD.
Exploring the advantages of multiband fMRI with simultaneous EEG to investigate coupling between gamma frequency neural activity and the BOLD response in humans.
Human Brain Mapping. 2018;39: 1673–1687. doi:10.1002/hbm.239433. Chen JCC, Forsyth A, Dubowitz DJ, Muthukumaraswamy SD.
On the Quality, Statistical Efficiency, and Safety of Simultaneously Recorded Multiband fMRI/EEG.
Brain Topogr. 2020;33: 303–316. doi:10.1007/s10548-020-00761-w4. Kyathanahally SP, Wang Y, Calhoun VD, Deshpande G.
Investigation of True High Frequency Electrical Substrates of fMRI-Based Resting State Networks Using Parallel Independent Component Analysis of Simultaneous EEG/fMRI Data.
Front Neuroinform. 2017;11. doi:10.3389/fninf.2017.000745. Nöth U, Laufs H, Stoermer R, Deichmann R.
Simultaneous electroencephalography-functional MRI at 3 T: An analysis of safety risks imposed by performing anatomical reference scans with the EEG equipment in place.
Journal of Magnetic Resonance Imaging. 2012;35: 561–571. doi:10.1002/jmri.228436. Baker KB, Tkach JA, Nyenhuis JA, Phillips M, Shellock FG, Gonzalez‐Martinez J, et al.
Evaluation of specific absorption rate as a dosimeter of MRI-related implant heating.
Journal of Magnetic Resonance Imaging. 2004;20: 315–320. doi:10.1002/jmri.201037. Faulkner W.
New MRI Safety Labels & Devices. Signals.
2016;5. Available: https://www.ismrm.org/smrt/E-Signals/2016FEBRUARY/eSig_5_1_hot_2.htm8. Nitz WR, Brinker G, Diehl D, Frese G.
Specific Absorption Rate as a Poor Indicator of Magnetic Resonance-Related Implant Heating.
Investigative Radiology. 2005;40: 773–776. doi:10.1097/01.rli.0000185898.59140.919. Warbrick T.
Exciting news for our EEG-fMRI customers: Update of MRI sequence conditions.
Brain Products GmbH. Apr 2020 Products&Applications.10. Brain Products. BrainAmp MR: Operating Instructions for use in an MR environment. 014 ed. Gilching, Germany; 2012.
11. Egan MK, Larsen R, Wirsich J, Sutton BP, Sadaghiani S.
Safety and data quality of EEG recorded simultaneously with multi-band fMRI.
PLOS ONE. 2021;16: e0238485. doi:10.1371/journal.pone.023848512. Brain Products. BrainVision Recorder 1.20: User Manual. 007 ed. Gilching, Germany; 2012.
13. Li X, Liang Z, Kleiner M, Lu Z-L.
RTbox: A device for highly accurate response time measurements.
Behavior Research Methods. 2010;42: 212–225. doi:10.3758/BRM.42.1.21214. Balasubramanian M, Wells WM, Ives JR, Britz P, Mulkern RV, Orbach DB.
RF Heating of Gold Cup and Conductive Plastic Electrodes during Simultaneous EEG and MRI.
Neurodiagn J. 2017;57: 69–83. doi:10.1080/21646821.2017.125672215. Brain Products. BrainVision Analyzer V. 2.0.2: User Manual. 004 ed. Gilching, Germany; 2012.
16. Allen PJ, Josephs O, Turner R.
A Method for Removing Imaging Artifact from Continuous EEG Recorded during Functional MRI.
NeuroImage. 2000;12: 230–239. doi:10.1006/nimg.2000.059917. Delorme A, Makeig S.
EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis.
Journal of Neuroscience Methods. 2004;134: 9–21. doi:10.1016/j.jneumeth.2003.10.00918. Allen PJ, Polizzi G, Krakow K, Fish DR, Lemieux L.
Identification of EEG Events in the MR Scanner: The Problem of Pulse Artifact and a Method for Its Subtraction.
NeuroImage. 1998;8: 229–239. doi:10.1006/nimg.1998.036119. Ritter P, Villringer A.
Simultaneous EEG–fMRI.
Neuroscience & Biobehavioral Reviews. 2006;30: 823–838. doi:10.1016/j.neubiorev.2006.06.00820. Calhoun VD, Sui J.
Multimodal fusion of brain imaging data: A key to finding the missing link(s) in complex mental illness.
Biol Psychiatry Cogn Neurosci Neuroimaging. 2016;1: 230–244. doi:10.1016/j.bpsc.2015.12.00521. Sadaghiani S, Wirsich J.
Intrinsic connectome organization across temporal scales: New insights from cross-modal approaches.
Netw Neurosci. 2020;4: 1–29. doi:10.1162/netn_a_00114