To correct for susceptibility distortions in fMRI data, researchers collected additional scans using the same T2*-weighted sequence but with inverted phase-encoding direction (inverted readout/phase-encoding (RO/PE) polarity) while participants were resting at multiple points throughout the experiment.
The study evaluated predictions of neurobiological theories of consciousness using 256 participants performing the same behavioral task across three neuroimaging modalities: functional magnetic resonance imaging (fMRI; n = 120), magnetoencephalography (MEG; n = 102), and intracranial electroencephalography (iEEG; n = 34).
The fMRI decoding methodology involved fitting a GLM per event using the NiBetaSeries (0.6.0) package, including 24 nuisance regressors, and performing decoding via whole-brain searchlight with a 4-mm radius and ROI-based approaches constrained by functional and anatomical ROIs.
The adversarial study of integrated information theory (IIT) and global neuronal workspace theory (GNWT) involved 256 human participants who viewed suprathreshold stimuli for variable durations while researchers measured neural activity using functional magnetic resonance imaging, magnetoencephalography, and intracranial electroencephalography.
Within-task category and orientation decoding were performed using a leave-one-run-out cross-validation scheme for fMRI data and a k-fold cross-validation scheme for MEG and iEEG data.
The Cogitate consortium distributes M-EEG, fMRI, and iEEG datasets through downloadable data bundles and an XNAT instance (https://cogitate-data.ae.mpg.de), with raw and BIDS formats available at https://www.arc-cogitate.com/data-bundles.
fMRI data quality control was performed using MRIQC (0.16.1) and custom scripts for data rejection.
Functional MRI analysis found no selective increase in interareal connectivity between object-selective nodes and the prefrontal cortex or V1/V2, even when separating task conditions.
The 1.2% increase in fMRI decoding accuracy observed when including frontal areas was present only in the combined features analysis and was not observed in the combined models' analysis.
Twelve participants were excluded from the fMRI dataset: eight due to motion artifacts, two due to insufficient coverage, and two due to incomplete data.
The study's testing framework involved an initial optimization phase on one-third of the MEG (n = 32) and fMRI (n = 35) datasets to evaluate data quality and optimize analysis pipelines, followed by a preregistered replication phase on novel datasets (MEG n = 65 and fMRI n = 73).
For the MEG and fMRI datasets, one-third of the data that passed quality tests (the optimization dataset) was used to optimize analysis methods, while the remaining two-thirds (the replication dataset) were used for the reported study results.
The fMRI decoding strategy utilized a multivariate pattern analysis approach on the pattern of BOLD activity over voxels, similar to the strategy used for iEEG and MEG data.
The fMRI sample in the study included 108 healthy participants with a mean age of 23.28 ± 3.46 years, consisting of 70 females and 105 right-handed individuals.
In the study's complementary results for decoding conscious content, fMRI searchlight decoding accuracies for letters versus false fonts were evaluated using pattern classifiers trained on relevant stimuli and tested on irrelevant stimuli, or vice versa, with significance evaluated through a cluster-based permutation test (p < 0.05; two-sided).
fMRI source DICOM data were converted to BIDS format using BIDScoin (v3.6.3), which includes converting DICOM data to NIfTI using dcm2niix and creating event files using custom Python codes.
The study utilized fMRI with a sample size of 73 participants to analyze cross-task generalization of face–object decoding.
The study combined iEEG, MEG, and fMRI techniques to mitigate the limitations of using single data modalities, creating a cross-compensating approach for testing consciousness theories.
The fMRI experimental protocol utilized a mean inter-trial interval of 3 s (range 2.5–10 s) and a trial length of approximately 5.5 s to avoid non-linearities in the BOLD signal that could affect fMRI decoding.
The fMRI study consisted of 576 total trials, organized into 8 runs with 4 blocks each, containing 17–19 trials per block.
The MEG and fMRI laboratories used the MEG-compatible and fMRI-compatible EyeLink 1000 Plus Eye-tracker system (SR Research) to collect data at 1,000 Hz.
The fMRI version of the experiment was conducted on an MSI laptop at Yale and a Dell Desktop PC at DCCN.
In the fMRI analysis, connectivity was assessed using generalized psychophysiological interaction (gPPI) implemented in SPM119.
fMRI Bayesian analysis showed substantial-to-very-strong support for the null hypothesis of no face orientation decoding in 34–55% of prefrontal voxels (BF01: 3–71.5), with support for the alternative hypothesis in only 1–9% of voxels.
Decoding of face orientation (left, right, or front views) was achieved in posterior but not in prefrontal regions of interest using iEEG (approximately 95% with pseudotrial aggregation) and fMRI searchlight approach (approximately 45%).
The researchers tested the IIT prediction on fMRI data by selecting the 150 most selective voxels within each of the two regions of interest (300 voxels total) for each participant, using face versus object contrast masking.
The researchers performed analysis-specific fMRI data preprocessing using FSL 6.0.2, SPM 12, and custom Python scripts (NiBabel 3.2.2 and SciPy 1.8.0). For univariate analyses, functional data were spatially smoothed with a Gaussian kernel (5 mm full-width at half-maximum), grand mean scaled, and temporal high-pass filtered (128 s), while multivariate analyses were performed without spatial smoothing.
For MEG experiments, eye tracking data were acquired binocularly, while for fMRI experiments, data were acquired monocularly from either the left or the right eye in DCCN and Yale, respectively.
fMRI searchlight decoding of category (faces-objects), collapsed across durations, shows significantly above-chance (50%) decoding in both task-relevant and task-irrelevant conditions, as visualized on inflated cortical surfaces.
In fMRI decoding of faces versus objects, including frontal areas resulted in a 1.2% increase in classification accuracy compared to excluding frontal areas, an effect observed in 56% of participants.
The researchers acknowledge that their fast-event-related fMRI design might be suboptimal for detecting activity changes in the salience network, potentially leading to an underestimation of regions involved in conscious processing.
fMRI data were acquired using a 3 T Prisma scanner with a 32-channel head coil, including high-resolution anatomical T1-weighted MPRAGE images (GRAPPA acceleration factor = 2, TR/TE = 2,300/3.03 ms, 8° flip angle, 1-mm isotropic voxels) and a whole-brain T2*-weighted multiband-4 sequence (TR/TE = 1,500/39.6 ms, 75° flip angle, 2 mm isotropic voxels).
fMRIPrep is a robust preprocessing pipeline designed for functional MRI data.
The paper 'The pulse: transient fMRI signal increases in subcortical arousal systems during transitions in attention' by Li, R. et al. was published in NeuroImage, volume 232, 117873, in 2021.
Statistical significance for ROI-based fMRI decoding was determined using a one-sample permutation test with FDR correction for multiple comparisons, while whole-brain decoding significance was evaluated using a cluster-based permutation test (P < 0.05) complemented by Bayesian analysis.
The researchers performed a series of conjunction analyses on fMRI data to identify three categories of brain areas: those responsive to task goals, those responsive to task relevance, and those putatively involved in the neural correlates of consciousness (NCC).
In the study, fMRI orientation decoding involved 64 front and 32 left and right trials per category.
All fMRI data were preprocessed using fMRIPrep (20.2.3), which is based on Nipype (1.6.1).