To show that reference layer artefact subtraction (RLAS) has the potential to increase the utility of simultaneous EEG-fMRI methods for investigating brain function.Artefacts present in EEG data acquired concurrently with fMRI can make the recovery of neuronal signals in low-frequency bands (delta [0-4Hz] and theta [4-8Hz]) particularly problematic¹. It is therefore important to improve artefact-removal techniques to extend the frequency range of EEG signals that are recoverable in EEG-fMRI studies. The RLAS method, which uses a reference layer carrying electrodes and leads that pick up similar artefacts to those appearing in scalp recordings (Figure 1), was previously shown to effectively attenuate all EEG artefacts; including the movement (MA) and pulse artefacts (PA) which dominate the low-frequency bands². However, safety considerations prevented the application of RF during proof-of-concept studies. Subsequent tests showed that the RLAS set-up induces no adverse RF-heating effects. Here we explore the benefits of RLAS in a full simultaneous EEG-fMRI experiment, focusing upon detection of low-frequency, driven neural responses.

Custom made Ag/AgCl ring-electrode pairs were set up with one electrode attached to the subject’s scalp and the second connected to a conducting, agar reference layer (Fig. 1). The scalp and reference layer electrodes were clipped together, but separated by a 60µm-thick polyethylene insulating layer which is shaped to fit each subject’s head. Eight electrode pairs were positioned approximately at locations Fpz, T7/8, Tp7/8, O1/2 and Oz in the 10-20 system. A reference electrode pair was placed near Cz. A ground electrode was attached to the scalp near Pz and shorted to the reference layer using conductive gel. Electrode pairs were connected to star-quad cables and fed out around the edge of the reference layer (Fig. 1A).

Recordings were made on 3 healthy-subjects inside a 3T MR scanner. A BrainAmp MRplus system (5kHz sampling-rate, 0.016-250Hz passband) recorded EEG data during an fMRI acquisition (EPI: TR/TE=2000/40ms, 20 slices, 270 volumes, 3mm resolution, 80x80 matrix). Data were collected during nine 30s periods of stimulation with a 2Hz flashing (one complete reversal), full-field, radial checkerboard interleaved with 30s rest periods (fixation cross). VCG recordings were acquired to aid PA correction³.

fMRI data were analysed using standard pre-processing methods. A first level general linear model using a boxcar stimulus model was used to identify areas of BOLD activation.

EEG data were high-pass filtered (cut-off frequency, 0.02Hz) to remove baseline drift. The reference layer voltages were re-referenced to the reference electrode connected to the reference layer before subtraction from the corresponding scalp voltages. Average artefact subtraction (AAS) was used to further attenuate the gradient artefact (GA)⁴ and PA^3,5. AAS_GA and AAS­_PA sliding-window average templates consisted of 61 and 21 intervals, respectively. AAS methods were also applied just to scalp voltages, akin to standard EEG, for comparison.

The root-mean-square (RMS) voltages over time, before and after correction, were calculated for each lead. Artefact attenuation achieved using different methods was then calculated using:

$$-20\log_{10}{\frac{\mathrm{RMS_{corrected}}}{\mathrm{RMS_{uncorrected}}}}$$

Corrected datasets were band-pass filtered (1-40Hz) and re-referenced to an average of all channels. Using the known timings of the checkerboard reversals, 1080 single-trial visually-evoked potentials were obtained and visualised in a stack plot. Magnitude spectra were averaged over stimulation blocks and the relative amplitude of the data recovered using different correction methods compared.

Figure 2 shows undistorted BOLD activation due to the visual stimulus. This data suggests that the presence of the additional electrodes and cabling has not had a significant effect upon the MRI data quality (also confirmed in B₁- and B₀-mapping experiments).

Figure 3 shows EEG artefact attenuations averaged over all non-saturated channels. These reflect the improvement in quality achieved by RLAS-AAS_GA-AAS_PA compared with AAS_GA-AAS_PA. Saturation occurred on some channels due to the large wire-loop sizes. After minimising the wire-loop sizes, similar benefits were realised as shown in Fig. 4. RLAS provides attenuation of MA due to gross movements (Fig. 4, time=3-5s) as well as small artefacts probably generated by cable vibrations or small involuntary head movements⁶ undetected in fMRI realignment (Fig. 4, inset). Consequently RLAS-AAS_GA-AAS_PA affords a more accurate identification of amplitude and latency of single-trial evoked potentials, notably the P100, than AAS alone (Fig. 5A&B). Moreover, in the frequency domain (Fig. 5E) RLAS-AAS_GA-AAS_PA is effective at recovering low-frequency neuronal activity typically obscured by MA and PA when only applying AAS.

This work shows the potential of RLAS to expand the utility of EEG-fMRI by reducing the likelihood of spurious correlations between EEG and fMRI data¹, particularly in the low-frequency domain. Work is now required to increase the number of electrodes and the robustness of the system.