Abstract
Alpha activity (8–14 Hz) is the dominant rhythm in the awake brain and is thought to play an important role in setting the internal state of the brain. Previous work has associated states of decreased alpha power with enhanced neural excitability. However, evidence is mixed on whether and how such excitability enhancement modulates sensory signals of interest versus noise differently, and what, if any, are the consequences for subsequent perception. Here, human subjects (male and female) performed a visual detection task in which we manipulated their decision criteria in a blockwise manner. Although our manipulation led to substantial criterion shifts, these shifts were not reflected in prestimulus alpha band changes. Rather, lower prestimulus alpha power in occipital-parietal areas improved perceptual sensitivity and enhanced information content decodable from neural activity patterns. Additionally, oscillatory alpha phase immediately before stimulus presentation modulated accuracy. Together, our results suggest that alpha band dynamics modulate sensory signals of interest more strongly than noise.
SIGNIFICANCE STATEMENT The internal state of our brain fluctuates, giving rise to variability in perception and action. Neural oscillations, most prominently in the alpha band, have been suggested to play a role in setting this internal state. Here, we show that ongoing alpha band activity in occipital-parietal regions predicts the quality of visual information decodable in neural activity patterns and subsequently the human observer's sensitivity in a visual detection task. Our results provide comprehensive evidence that visual representation is modulated by ongoing alpha band activity and advance our understanding on how, when faced with unchanging external stimuli, internal neural fluctuations influence perception and behavior.
Introduction
Identical stimuli may lead to variable percepts. When presented with the same threshold-level stimulus repeatedly, an observer sometimes detects it and sometimes misses it. The internal state of the brain, largely determined by moment-to-moment neural activity fluctuations, contributes to this perceptual variability (for review, see Harris and Thiele, 2011; Samaha et al., 2020).
Alpha activity (8–14 Hz) is the dominant rhythm in the awake brain. One prevailing hypothesis is that the alpha rhythm sets the excitability level of the internal state via functional inhibition (Klimesch et al., 2007; Jensen and Mazaheri, 2010). Because of a tight link with internal excitability, alpha band oscillatory activity (and temporal fluctuations) is a plausible key source of perceptual variability. Indeed, extensive work reported that prestimulus alpha band power modulates perception of simple sensory stimulations (Thut et al., 2006; Hanslmayr et al., 2007; van Dijk et al., 2008; Haegens et al., 2011; van Ede et al., 2011; Barne et al., 2020).
Only recently researchers started to investigate how internal excitability changes indexed by alpha oscillations manifest in subjects' behavioral outcomes (Iemi et al., 2017; Samaha et al., 2017), using signal detection theory (Macmillan & Creelman, 2004). In this theoretical framework, two metrics are used to quantify the perceptual process: sensitivity (d′), a metric of how well the observer distinguishes sensory signals of interest from noise, and criterion (c), a metric of the observer's tendency for one perceptual decision over the others. Depending on potentially differential impacts on sensory signals of interest versus noise, internal excitability changes indexed by alpha activity may modulate different aspects of the perceptual process.
One hypothesis is that decreased alpha activity amplifies the sensory signal of interest and noise to the same extent (e.g., the gain is additive to the response amplitude) without making them more distinct from each other and therefore only results in a criterion shift in the observer's behavior (Fig. 1). This view is mainly supported by empirical work correlating ongoing, as opposed to experimentally induced, alpha activity changes with perceptual reports. Several studies showed that when prestimulus alpha activity was low, subjects reported higher subjective visibility (Benwell et al., 2017) and confidence (Samaha et al., 2017) and had a higher tendency to see a target (i.e., more liberal criterion) but displayed no change in perceptual sensitivity (Limbach and Corballis, 2016; Iemi et al., 2017). If a low state of alpha activity indeed amplifies signal and noise similarly and thereby shifts the observer's criterion, a key question is whether (top-down) changes in criterion are implemented via (and/or reflected by) shifts in the alpha state.
Hypothesis schematic. A, Internal response distribution in different states, with state 1 the initial or baseline state (internal response distribution of signal modeled as a normal distribution with mean
Another hypothesis is that decreased alpha activity amplifies the signal of interest to a larger extent relative to the noise (e.g., the gain scales multiplicatively as a function of response amplitude), making them more distinct from each other and hence resulting in improved perceptual sensitivity (Fig. 1). This view is largely inferred based on studies using classical attentional-cuing paradigms, in which prestimulus alpha band activity patterns (i.e., alpha lateralization, activity difference between task-relevant and task-irrelevant regions) and perceptual sensitivity covary following attentional cuing (Thut et al., 2006; Haegens et al., 2011; van Ede et al., 2011). Although informative, it is unclear whether the observed correlation between alpha lateralization and performance is mainly driven by an alpha amplitude decrease in task-relevant regions or and increase in task-irrelevant regions. Furthermore, if alpha oscillations indeed modulate sensory signals of interest more strongly relative to the noise, an immediate prediction is that decreased alpha activity leads to enhanced information coding of the task-relevant stimuli.
To differentiate between these hypotheses, we tested human subjects in a magnetoencephalography (MEG) experiment, during which we experimentally manipulated their decision criterion using a novel visual detection paradigm. We examined whether and how changes in prestimulus alpha dynamics (1) correlate with criterion shifts, (2) influence sensitivity while the criterion is controlled, and (3) modulate information coding, indexed by the representational content decodable in the neural signal.
Materials and Methods
Subjects
Thirty-four healthy volunteers completed the full experiment. Two of them were excluded from analysis because of excessive head movement during the MEG recording, resulting in the planned sample size of 32 subjects in the reported analysis (mean age = 25.8, SD = 6.0; 23 females). The study was approved by the ethics committee (Commissie Mensgebonden Onderzoek, Arnhem-Nijmegen). All subjects gave informed consent before the experiment and received monetary compensation for their participation.
Procedure
Subjects reported to the lab on 2 d in 1 week for a training session on day 1 and an MEG recording session on day 2.
Training session on day 1
The training session on day 1 served to familiarize subjects with the task and to prepare them for the MEG recording session on day 2. We introduced the task with 10 slow trials in which we set the interstimulus interval (ISI) between the target and the backward mask to 100 ms and provided feedback at the end of each trial. Before starting the up–down staircase, subjects practiced 20 normal-speed trials with feedback, to become familiar with the task. This introduction phase was repeated if necessary. Two runs of three down the staircase and one up were then used to titrate the contrast of the backward mask, one run for each target orientation [clockwise (CW) and counterclockwise (CCW) gratings, order counterbalanced across subjects]. Subjects were explicitly informed about the orientation and grating presence rate (50%) at the start of each staircase run. No feedback was provided during the staircase procedure. After the staircase procedure, subjects completed four experimental blocks. Notably, to maximize the priming effect for the MEG session, gratings were presented 80% (or 20%) of the time in both the priming and main task phase, and feedback was provided at the end of each block (i.e., accuracy and mean reaction time). We aimed to convince subjects that using the provided target-presence rate in their decisions was beneficial for task performance.
MEG testing session on day 2
On day 2, similar to the training session, subjects first completed two runs of three down the staircase and one up inside the MEG to determine the contrast level of the backward mask that resulted in ∼80% detection accuracy when an equal number of target-present and absent trials was presented. The staircase runs were done after a brief recap of the task in which subjects completed 10 slow trials and 20 normal speed trials with trial-by-trial feedback (i.e., same as the introduction phase of the training session on day 1). They then performed eight main experimental blocks and three grating localizer blocks, during which neurophysiological data were recorded.
Experimental paradigm
Main experimental block
Two stimuli, a target and a mask, were presented sequentially in each trial, and the subject's task was to report as accurately as possible whether a grating was present or not in the trial (Fig. 2A, trial schematic). To introduce criterion shift in the subject's perceptual decisions, we designed a paradigm inspired by Crapse et al. (2018). Each main experimental block started with an instruction screen, informing subjects of the orientation (CW or CCW) and grating presence rate (20% or 80%) in the upcoming block. Every block consisted of 40 priming trials and 80 main task trials. For the first 40 priming trials, gratings were presented either 20% or 80% of the time, consistent with the block instruction. The backward masks were of 5% Michelson contrast, which had little masking effect on the target grating. We purposely presented the targets in a clearly discriminable way during the priming phase to encourage subjects to shift their decision criteria accordingly. Before starting the 80 main task trials, a reminder was presented on the screen for 3 s, highlighting target orientation of the current block and presence rate again. From the subject's point of view, other than a contrast increase in the backward mask (titrated for each subject), everything remained the same between priming and main task trials. However, unbeknown to the subjects, gratings were presented 50% of the time regardless of the instructed target-presence rate (and contrary to the priming part of the block). Critically, the increased perceptual ambiguity because of strong backward masks allowed us to keep the grating presence rate the same between conservative and liberal priming blocks while introducing a bias in the subject's perceptual decisions.
Grating localizer block
Interleaved with the main task blocks, subjects performed three blocks of a grating localizer task (Fig. 2A). Each block consisted of 120 trials, 18 of which contained a brief fixation blink (i.e., the central fixation dot turned from black to white for 50 ms). The subjects' task was to detect these occasional fixation blinks and to press a button as soon as possible when they saw them. Clockwise and counterclockwise gratings identical to those presented during the main task were presented in a pseudorandom order for 250 ms, followed by a uniformly jittered intertrial interval (ITI) of 1–1.5 s. We used a longer presentation duration (relative to that used in the main task trials) to increase the signal-to-noise ratio in the stimulus-driven response. These trials enabled us to identify brain regions most responsive to the bottom-up stimulation by grating stimuli. Trials containing fixation blinks, which served exclusively to maintain the subject's attention on the fixation point, were excluded from later analyses.
Stimuli
A bull's-eye [outer black ring = 0.5 × 0.5° degree of visual angle (dva), innermost black dot = 0.25 × 0.25° dva] was presented at the center of the screen throughout each block as the fixation point. Subjects were instructed to always maintain fixation and not blink during the presentation of the stimuli. A 2 s fixation window was presented before the start of the first priming and the first main task trial in each block. Each trial started with a brief presentation of the target (16.67 ms), within which either an oriented grating (Michelson contrast, 40%; spatial frequency, 1 cycle per degree; orientation, 45° CW or CCW relative to vertical, randomized spatial phase) or a bandpass-filtered noise patch (Michelson contrast, 40%; spatial frequency, 1 cycle per degree; randomly generated in each trial) was shown in an annulus (inner radius = 1.5°, outer radius = 7.5°; contrast of the stimuli decreased linearly to 0 over the outer and inner 0.5° radius of the annulus) around the central fixation (Fig. 2A). After a brief (16.67 ms) ISI window, a backward mask consisting of a bandpass-filtered noise patch (Michelson contrast, titrated for each subject; spatial frequency, 1 cycle per degree; randomly generated in each trial) was presented for 100 ms (Fig. 2A). After another ISI window of 800 ms, the letters P (for grating present) and A (for grating absent) were presented on each side of the screen, respectively, indicating the response mapping of the current trial, which changed pseudorandomly and unpredictably across trials to avoid motor preparation confounds. Subjects had a maximum of 2 s to provide their response (i.e., grating present vs absent). The ITI started as soon as subjects committed a response. The ITI duration was jittered across trials between 1.8 and 2.4 s for the main task trials and between 1.5 and 2.0 s for the priming trials. Note that we titrated the contrast of the backward mask via a three-down and one-up staircase procedure at the beginning of each session so that overall accuracy in a detection task with 50% target presentation rate was comparable across subjects and across sessions.
Data acquisition
Stimuli were displayed on an LCD screen during the training session and on a semitranslucent screen (1920 × 1080 pixel resolution, 120 Hz refresh rate) back projected by a PROpixx projector (VPixx Technologies) during MEG recordings. The experiment was programmed with Psychtoolbox (Brainard, 1997) in MATLAB (MathWorks) and ran in a Linux environment. Brain activity was recorded using a 275 channel axial gradiometer magnetoencephalography (MEG) system (VSM/CTF Systems) at 1200 Hz in a magnetically shielded room. Six permanently faulty channels were disabled during the recordings, leaving 269 recorded MEG channels. Three fiducial coils were placed at the subject's nasion and both ear canals to provide online monitoring of the subject's head position (Stolk et al., 2013) and to serve as anatomic landmarks for off-line coregistration with structural MRI scans. Eye position and pupil size were recorded using an infrared eye tracker (EyeLink, SR Research) during the MEG recordings. On completion of the MEG session, the subject's head shape and the location of the three fiducial coils were digitized using a Polhemus 3D tracking device. T1-weighted MRI scans were acquired using a 3T MRI system (Siemens). Earplugs with a drop of vitamin E were placed at the subject's ear canals during MRI acquisition to facilitate coregistration with MEG data.
Behavioral data analysis
We focused our analysis on main task trials (i.e., trials with high-contrast backward masks) recorded during the MEG session. In addition to accuracy and reaction times (RT), our main behavioral variables of interest were sensitivity (d′) and criterion (c), which were estimated following signal detection theory (SDT), assuming the internal signal and noise distributions shared the same variance (Macmillan & Creelman, 2004) as follows:
To compare d′ and c between priming conditions at the group level, we first computed for each condition the pooled d′ and c estimates as well as the corresponding variances, using the averaged hit rate
To capitalize on trial-by-trial fluctuations and to capture the relationship between prestimulus state and behavioral outcomes, we adapted and extended the GLM formulation of signal detection theory (DeCarlo, 1998).
For each subject, we defined the subject's responses (target presence/absence) determined by a combination of factors (stimulus presence, the oscillatory state of interest, and the interaction between them) via a probit link function as follows:
Replacing z(Hit) and z(FA) in Equation 1 with Equation 6 and 7 results in the following:
At the group level, we extended the above model and constructed a generalized linear mixed model (GLMM) to better account for between-subject variation. Specifically, we fitted the model with subjects as the only random grouping factor and included for each fixed effect the corresponding random slope coefficient and intercept. This was done using the fitglme function in MATLAB with the quasinewton optimizer. We reported results of the estimated coefficients
MEG data analysis
MEG preprocessing
MEG data were preprocessed off-line and analyzed using the FieldTrip toolbox (Oostenveld et al., 2011) and custom-built MATLAB scripts. Trials of the main task blocks and grating localizer blocks were segmented and processed separately. All data were down sampled to 400 Hz, after applying a notch filter to remove line noise and harmonics (at 50, 100, and 150 Hz). Trials with excessive noise were rejected via visual inspection before independent component analysis (ICA). ICA components were visually inspected, and those representing eye and heart artifacts were then projected out of the data (Jung et al., 2000). Finally, outlier trials with extreme variance were removed, leaving on average 297 of 320 (SD = 14.6) priming trials, 595 of 640 (SD = 33.5) main task trials, and 331 of 360 (SD = 18.0) localizer trials per subject.
MRI processing
MRI data were coregistered to the CTF coordinate system using the fiducial coils and the digitized scalp surface. Volume conduction models were constructed based on single-shell models of individual subjects' anatomic MRIs (Nolte, 2003). Dipole positions were defined using a cortical surface-based mesh with 15,784 vertices created using FreeSurfer version 6.0 (RRID:SCR_001847) and HCP workbench version 1.3.2 (Connectome Workbench; RRID:SCR_008750). The vertices were grouped into 374 parcels based on a refined version of the Conte69 atlas (Van Essen et al., 2012), allowing us to reduce the dimensionality of the data, similar to Schoffelen et al. (2017). For each dipole position, lead fields were computed with a reduced rank, which accommodates the fact that MEG is blind to radial sources.
Event-related fields
Before calculating the event-related fields (ERFs), singe-trial data were baseline corrected using a time window [−0.5–0 s] for the priming and main task trials, and [−0.2–0 s] for the grating localizer trials. To avoid the confounding influence of noise (in the planar transformation) because of unequal trial numbers across conditions, trial numbers were equated via subsampling. Specifically, we subsampled an equal number of trials from each condition before averaging over trials so that the number of trials per condition matched that in the condition of the fewest trials. Planar gradients of the MEG field distribution were then calculated, which makes spatial interpretation of the sensor-level data easier and facilitates comparison of ERF topographies across subjects. We repeated the above-mentioned procedure 10 times per condition, to ensure every trial was used at least once. We then averaged across all corresponding planar-combined averages to obtain ERFs per condition.
Spectral Analysis
Time-frequency representations (TFRs) of power were calculated for each trial by applying a fast Fourier transform to short sliding time windows. We applied Hanning tapers of four cycles in length in time steps of 50 ms to single-trial data before computing the spectra (4–30 Hz). For sensor-level analyses, spectral decomposition was applied to synthetic planar gradient data and combined into single spectra per sensor. For visualization purposes, condition-average TFRs were expressed relative to a frequency-specific baseline window, which started four cycles and ended two cycles before stimulus onset (e.g., for 10 Hz activity, a 400–200 ms prestimulus time window was used as baseline window) to prevent leakage of poststimulus activity into the baseline window. Note that for all statistical analyses, noncorrected TFR data were used. To compute TFRs of an anatomic parcel (source level), spectral decomposition was performed on the response time series of the parcel (see below, Source reconstruction). To account for intraindividual and interindividual variability in alpha peak frequency (Haegens et al., 2014), we defined the individual alpha peak frequency for each brain area (parcel) and each subject separately. This was done with the priming trials by computing the Fourier spectrum of the prestimulus 1 s time series of each parcel and finding the local maximum within the 7–14 Hz band of the resulting Fourier spectrum. We computed the oscillatory alpha power within each parcel by averaging the above-mentioned power estimate within a
Phasic modulation of accuracy
We assessed whether behavioral accuracy, both of the subjects and of the neural decoder (see below, Multivariate pattern analysis) changed as a function of oscillatory alpha phase. We grouped the trials based on their corresponding phase estimates of a specific time window and parceled into 16 equidistant phase bins [bin centers = 0, 22.5, …, 337.5°; bin width = 45°) and computed the average accuracy for each bin. We combined all trials (conservative and liberal) in this analysis because it is unlikely that subjects endogenously aligned their instantaneous alpha phase with respect to the stimulus onset as onset was randomly jittered and hence unpredictable to the subjects. We used accuracy as an index of interest for the current analysis because accuracy, compared with d′, is a more robust measure of performance when trial number is limited (in this case, <80 trials per phase bin). We then fitted a cosine function of one cycle and unknown amplitude and phase to the resulting 16 accuracy scores. The amount of phasic modulation of accuracy was defined as the fitted amplitude of the cosine function. Intuitively, an estimated amplitude of 3% suggested that the model predicted an accuracy difference of 6% between trials of the optimal phase and those of the suboptimal phase. This procedure was applied to every time window and parcel of interest. To obtain a null distribution of the fitted amplitudes for statistical inference, we circularly shifted the phase time course of every trial with a random value between −360 and 360° before phase binning the data, which allows randomization without breaking the temporal (autocorrelation) structure within a trial.
Source reconstruction
Source-level analyses (see below, Results) are based on a sample of 29 subjects, whose T1-weighted anatomic images were available. The linearly constrained maximum variance beamformer approach (Van Veen et al., 1997) was used to obtain the source reconstruction of the event-related response to the localizer gratings. The data covariance matrix was computed over a window of [−1.0, 1.0 s] for priming and main trials and [−0.5, 0.5 s] for localizer trials, time locked to stimulus onset, and was subsequently used to construct common spatial filters. To estimate the amplitude of the single-trial stimulus evoked response, we projected the trial data computed over [0, 0.2 s] through the spatial filters, and then normalized the resulting amplitude by that of a baseline window of [−0.2, 0 s] time locked to stimulus onset. Neural response time series of an anatomic parcel were computed by taking the first principal component of the time series over all dipole positions within the parcel.
We used the partial canonical coherence beamformer approach (Schoffelen et al., 2008) to localize the sources of oscillatory alpha band activity in response to the main task trials. To estimate the spatial distribution of alpha band power, we first extracted 500-ms data segments [time window = (−0.5 0), time locked to stimulus onset], then computed cross-spectral density (CSD) matrices using Slepian tapers (Mitra and Pesaran, 1999) centered at a frequency of 10 (±4) Hz. With the CSD matrices and the lead fields, a common spatial filter (for all trials across both conditions) was constructed for each dipole position for each subject. By projecting sensor-level CSD through the common spatial filter, the spatial distribution of power was then estimated for each trial. To reduce data dimensionality, we took the mean estimated power of all dipoles within each parcel as the single-trial power estimate of the parcel.
Multivariate pattern analysis
To probe the representational content of neural activity, we performed multivariate pattern analysis using the MVPA-light toolbox (Treder, 2020). Before running the classification analysis, we first applied a low-pass filter at 30 Hz and down-sampled the data to 100 Hz. Multivariate logistic regression (L2 regularized) was then applied to predict the presence of the grating (present vs absent), based on the sensor-level spatial distribution of neural activity at each time point. To avoid overfitting, we trained the classifier with priming trials and tested on the main task trials. This procedure was done within-subject in a time-resolved manner, resulting in a temporal generalization matrix (King and Dehaene, 2014).
Cluster-based permutation tests
Statistical significance was evaluated using cluster-based permutation tests (Maris and Oostenveld, 2007) at the sensor level. The time interval of interest was [−0.8, 0 s], that is, the 800 ms window before stimulus onset. For ERFs, data of different preselected sensors were first combined, resulting in a single time series per subject. These time series data were then compared univariately at each time point. Neighboring time points for which a two-tailed paired t test (or a repeated-measures ANOVA) resulted in a nominal p value smaller than 0.05 (uncorrected) were clustered. A similar procedure was applied to the TFRs, with clustering across spectral, temporal, and spatial dimensions. The sum of the t values (or F values) within a cluster was then computed as cluster-level statistics. The cluster with the maximum sum was subsequently used as test statistic. By randomly permuting the condition labels (or by shuffling the predictor variable of interest across trials) and recalculating the test statistic 10,000 times (or 100 times for computationally expensive operations such as the GLMM and phasic modulation analysis), we obtained a reference distribution of maximum cluster t values (or F values) to evaluate the statistic of the actual data (alpha = 0.05). For reference, we also report the smallest p value obtained for cluster-based permutation tests that failed to reject the null hypothesis.
Data availability
All data and code for stimulus presentation and analysis are available from the Donders Repository at https://doi.org/10.34973/w1k5-sm41.
Results
Priming leads to criterion shifts
Subjects finished eight blocks of the main experimental task and three blocks of a localizer task during the MEG session. For the localizer task, they had to detect and report as fast as possible occasional fixation changes, which occurred in 15% of all trials. Subjects correctly reported fixation changes on 99.5% (between-subjects SD = 0.5%) of all trials with mean median RT of 406 ms (between-subjects SD = 49 ms), confirming close engagement in the task.
For the main experimental blocks, subjects' task was to report the presence (or absence) of an oriented grating as accurately as possible. They consistently performed at ceiling during the priming phase when the masking effect was minimal (mean accuracy = 95.2%, with a between-subjects SD of 5.1%; mean median RT = 421 ms, with a between-subjects SD of 73 ms). For the main task trials, in which the mask was presented at a contrast level titrated for each subject, overall accuracy dropped to ∼75% as aimed for. Note the mean accuracy was lower than that expected from a three-down, one-up staircase procedure (i.e., ∼80%), likely because the instructed target-presence rate was different from the actual target-presence rate. Two-way (priming type × grating orientation) repeated-measures ANOVA showed no significant differences across conditions in RT or accuracy (F values < 1; Fig. 2B). Results of the SDT analysis (Fig. 2C) showed that subjects were significantly more liberal in reporting target present in the liberal priming blocks than in the conservative priming blocks (p < 10−6), validating our experimental manipulation. Critically, their sensitivity (d′) was comparable between the liberal and conservative blocks (p = 0.082). Together, our behavioral results show that the current paradigm successfully introduces criterion shifts in subjects' detection responses.
Experimental paradigm and behavioral results. A, Schematic of the task trials (top) and localizer trials (bottom). Example grating and noise stimuli are shown for illustration purposes. B, C, Accuracy and reaction times (B) and sensitivity and criterion (C) for the main task trials during the MEG recording and the corresponding hit and false alarm rates (mean and between-subjects SDs). Shaded areas denote within-subject standard errors in B and the variance of the group-level pooled d′ and c in C. Dots represent individual subjects (N = 32). *p < 0.05). n.s., not significant.
Priming does not lead to modulation of prestimulus oscillatory power or stimulus-evoked activity
According to the additive hypothesis, prestimulus alpha power modulates the decision criterion. Therefore, one may expect (top-down) changes in criterion induced by different priming conditions to be reflected in alpha power differences. We first compared prestimulus TFRs between the conservative versus liberal trials, without making any assumptions about the frequency range of potential effects. We observed no significant differences in prestimulus TFRs (p = 0.366, cluster-based permutation test with frequencies from 1 to 30 Hz, time points from 800 ms prestimulus to stimulus onset, including all sensors) between the two priming conditions. We then repeated our statistical tests by constraining our frequencies of interest to around the alpha range (6–14 Hz) or by constraining the channels of interest to 20 subject-specific channels with strongest evoked response to the localizer gratings in a time window of [0.1, 0.2 s], relative to baseline. We found no significant prestimulus TFR differences between the two priming conditions in either analysis (Fig. 3C).
ERFs and TFR of selected sensors and classification results. A, ERFs for the priming trials (showing data for the 20 selected sensors most responsive to the localizer gratings). Black and gray lines (bottom) indicate statistically significant time points identified by cluster-based permutation tests. B, Same as A for main task trials. C, TFR of the main task trials (left) and the TFR of the difference between conservative and liberal trials (right). D, Temporal generalization matrix of classification accuracy. E, Time courses showing the diagonal of the temporal generalization matrices. Right, Subjects' response to grating present and absent trials as a function of the amount of evidence in the classifier trained and tested at time point 0.12 s (marked by a triangle above the time series). Dots represent group averages of each bin.
In addition to prestimulus activity, we also examined whether priming modulates the sensory response amplitude to the stimuli. We constrained these analyses to the above-mentioned 20 subject-specific channels, which primarily reflect the stimulus-evoked responses. We did not observe any significant differences in the average ERF amplitude (averaged across all selected sensors within each subject) between different priming conditions for the main task trials (Fig. 3B). Interestingly, when the same contrasts were repeated for the priming trials, we found that for target-present trials, conservative priming led to larger ERF amplitude 400–650 ms after stimulus onset (p < 10−4; Fig. 3A, solid lines) compared with liberal priming; and this pattern was reversed for target-absent trials (p < 10−4; Fig. 3A, dotted lines). These observations may reflect a general P3 surprise effect (Mars et al., 2008) to the less frequently presented stimuli, given that target-present and target-absent trials were rare events (i.e., 20% of trials) in the conservative and liberal priming blocks, respectively. These ERF results also suggest that subjects incorporated the different probabilistic information in sensory processing during the priming phase.
Together, we show for the main task trials that neither the prestimulus oscillatory activity nor the stimulus-evoked activity reflects the (top-down) criterion changes introduced by priming.
Prestimulus alpha power modulates d′: region of interest analysis
If excitability modulation by alpha oscillatory power has any behavioral consequence, its magnitude immediately before stimulus onset within the neuronal populations most responsive to the stimuli should be most relevant for subsequent detection responses. Following this rationale, we zoomed in on the anatomic parcel most responsive to localizer gratings (defined for each subject; Fig. 4B, spatial topography of the selected parcels) and asked whether and how alpha power immediately before the stimulus onset in these functionally defined visual ROIs predicts the subject's response. Although we observed prominent oscillatory activity within the alpha band during the prestimulus time window (Fig. 4B, right), we found no significant difference between the two conditions (i.e., conservative vs liberal) in the prestimulus TFR (cluster-based permutation test with frequencies from 1 to 30 Hz, time points from 800 ms prestimulus to stimulus onset, p = 0.257) nor in the power spectrum (cluster-based permutation test with all frequencies, p = 0.078).
ROI-based analysis of relationship between alpha and performance. A, The prevalence (indexed by the number of subjects) of alpha oscillatory activity across brain regions. B, Spatial topography of the visual regions of interest most responsive to grating stimuli and the corresponding alpha peak frequencies and the prestimulus power spectrum. C, Left, GLMM results, estimated coefficients and the corresponding 95% confidence intervals and the distribution of bootstrap estimates for the parameters. Middle, Visualization of the fitted coefficients for the conservative priming trials. Dots represent group-level average estimates and the corresponding within-subject SEs. The line shows the fitted model. Right. Same as the middle but for the liberal trials. D, GLMM results, estimated coefficients and the corresponding 95% confidence intervals and the distribution of bootstrap estimates for the parameters. E, Decoding accuracy (diagonal of the temporal generalization matrix) as a function of prestimulus alpha power for conservative (left) and liberal (right) trials. Gray shaded area marks the time window during which prestimulus alpha power significantly modulated decoding accuracy indicated by the corresponding cluster-based permutation test. Shaded area around the time series denotes within-subject SE. Bar graphs show mean decoding accuracy (within the gray shaded area) as a function of prestimulus alpha power. Error bars indicate within-subject SE.
Next, we asked, Do trial-by-trial changes of prestimulus oscillatory alpha power influence the subject's detection responses? We applied GLMMs (see above, Materials and Methods), which account for both between-subjects and within-subject trial-by-trial response variability to address this question. Our modeling results (Fig. 4C) showed a significant negative effect of prestimulus alpha power on d′, which was exclusively driven by conservative trials [conservative,
Prestimulus alpha power enhances sensory representation
To evaluate the quality of visual information coding, we used MVPA, operationalizing the quality of visual representation as the classification performance of the neural classifier. We first confirmed that classifiers trained on priming trials (classifying target-present vs target-absent trials) are generalizable (i.e., performed significantly above chance) to the main task trials (Fig. 3D). Zooming in to the diagonal of the matrix, we found that classification accuracy did not differ significantly between priming type [p = 0.092, cluster-based permutation test on time window (0, 1.0); Fig. 3E]. Moreover, we found the subjects' responses can be predicted by the decision evidence of the classifier (i.e., distance to the hyperplane) at the time classification accuracy peaked (Fig. 3E). Hence, the neural classifiers trained on priming trials not only generalizes to main task trials but also tracks subjects' responses quantitatively and qualitatively.
We examined whether prestimulus alpha power modulates the quality of visual information coding by first focusing on the response of the neural classifier at the time when classification accuracy peaked (trained and tested at the time point 120 ms after stimulus onset). Similar to the above-mentioned analyses linking prestimulus alpha power with the subject's behavior, we applied the GLMM to the responses of the classifier. We found a negative correlation between prestimulus alpha power and the d′ of the classifer for the conservative trials [Fig. 4D; conservative,
We next asked whether the modulatory effect on classification accuracy by prestimulus alpha power manifested at time points other than when classification accuracy peaked. For each priming condition, we divided trials into low, medium, and high alpha power bins (with ∼100 trials per bin, allowing for robust estimates of classification performance), using the prestimulus alpha power within our visual ROIs. We then compared the classification accuracy time course across the three bins using cluster-based permutation test [on the diagonal of the temporal generalization matrix, time window = (0, 0.5 s)]. This analysis showed that lower prestimulus alpha power in the visual ROIs was associated with higher classification accuracy ∼120 ms after target onset and that this modulation was present only for the conservative trials (p = 0.038; Fig. 4E), not the liberal ones (p = 0.159).
Together, our classification results show that a state of decreased prestimulus alpha power leads to enhanced visual representation, which likely results in better detection sensitivity.
Prestimulus alpha power modulates d′ predominantly in occipital-parietal areas: whole-brain analysis
Although the above ROI analysis enabled us to maximize statistical power to address our research questions, it may have precluded us from observing interesting effects taking place outside our visual ROIs. To explore whether and how prestimulus alpha power fluctuations in different brain regions modulated subjects' detection responses, we applied the GLMM analysis to all anatomic parcels, and asked whether alpha power changes in any of them modulated the subject's d′ or criterion. We applied cluster-based permutation tests to correct for multiple comparisons across time (from 600 to 200 ms prestimulus) and space (374 anatomically defined parcels, whole brain). Under the conservative priming condition, five parcels along the left ventral visual pathway showed significant negative alpha power modulation on d′ (p < 0.01; Fig. 5A), whose effect was most pronounced at time points immediately before stimulus onset. Under the liberal priming condition, a cluster of ∼20 parcels extending bilaterally from the somatosensory cortex to superior parietal lobule showed significant negative alpha power modulation on d′ (p < 0.01; Fig. 5B). In contrast, we observed a positive criterion modulation by alpha power for brain regions overlapping with this cluster, now for the conservative priming condition (p < 0.01; Fig. 5C), although this modulation disappeared for time points immediately preceding the stimulus presentation. There was no significant criterion modulation by alpha power under the liberal priming condition.
Whole-brain GLMM analysis of relationship between alpha and performance. A, Estimated
Prestimulus alpha phase modulates accuracy in a phasic manner
Finally, we asked how detection accuracy fluctuated as a function of oscillatory alpha phase. We estimated the extent of phasic modulation of accuracy by taking the amplitude of the fitted cosine function that best described accuracy changes as a function of oscillatory alpha phase. We first examined the amplitude of the fitted cosine over time during the prestimulus time window in the functionally defined visual ROIs. The estimated amplitudes were significantly above chance ∼200 ms before the stimulus onset, suggesting significant phasic modulation of detection accuracy (Fig. 6A,B). Interestingly, such a phasic modulation of the subject's accuracy seemingly faded out immediately before the stimulus onset. In contrast, the phasic modulation of the responses of the neural classifier peaked at ∼100 ms immediately before stimulus onset, suggesting that oscillatory alpha phase in these visual ROIs may have different impacts on the sensory representation and the subject's detection responses.
Phasic modulation of accuracy. A, Estimated amplitude of the phasic modulation of subjects' accuracy and the accuracy of the neural classifier over time. Shaded areas around the curve denote between-subjects SDs, and dots (bottom) denote statistical significance; the lighter ones correspond to p < 0.05 uncorrected, and the darker ones correspond to those corrected (false discovery rate = 0.05) for multiple comparisons. B, Estimated amplitude of the phasic modulation of subjects' accuracy and accuracy of the neural classifier at the time point 100 ms before stimulus onset in the functionally defined ROIs. Thinner lines denote data from individual subjects, with the gray ones denoting cases where the estimated amplitude failed to exceed that of the null distribution. Thick black lines denote the group average amplitudes. C, Left, Parcels showing significant phasic modulation on the subjects' accuracy and the accuracy of the classifier (when trained and tested at 0.12 s poststimulus onset). Right, Of the highlighted (circled) parcels in the left panel, the corresponding time courses of the estimated amplitude. Same conventions as in A. Gray shaded area denotes the [−0.2, −0.1 s] time window used by the left spatial topography.
Next, we asked whether any anatomic parcels exhibited significant phasic accuracy modulation by oscillatory alpha band activity. We focused on the amplitude of the fitted cosine corresponding to the time window immediately before stimulus onset, whose phase estimates were unaffected by poststimulus responses [−0.2, −0.1 s]. After correcting for multiple comparisons (false discovery rate 5%), two occipital parcels in the dorsal visual stream displayed significant phasic modulation of the subject's detection response, whereas several parcels along the dorsal visual stream extending toward the superior parietal lobe (Fig. 6C) displayed significant phasic modulation of the peak response of the classifier (i.e., classifier trained and tested at 120 ms poststimulus).
These analyses of oscillatory alpha phase suggest that accuracy was modulated in a phasic manner, and that such phasic modulations are most likely driven by occipital-parietal alpha phase along the dorsal visual pathway.
Discussion
To understand how alpha oscillations modulate perception, we recorded MEG in human subjects while they performed a visual detection task. We found that criterion shifts induced by our priming manipulation were not reflected in a modulation of prestimulus alpha activity. Rather, we found that trial-by-trial changes of prestimulus alpha power modulate perceptual sensitivity and that alpha phase in regions along the dorsal visual pathway modulates behavioral accuracy.
By providing the target-presence rate and presenting clearly visible targets at the start of each block (i.e., the priming trials), we successfully changed subjects' perceptual expectations and thereby introduced criterion shifts in their decisions. Although work has demonstrated that the human brain implements perceptual expectations by potentiating neural activity representing the expected stimuli (Ekman et al., 2017; Kok et al., 2017), we did not observe such a potentiation, nor did we observe any neural activity difference between the two priming conditions. Unlike a previous study where criterion was manipulated via changing the stimulus-response reward contingencies (Kloosterman et al., 2019), our current paradigm minimized the effect of reward contingency on neural activity and arguably allowed a cleaner reconstruction of sensory responses. Moreover, we manipulated subjects' perceptual expectations in a blockwise manner; therefore, our null finding cannot be explained by the brain not having enough time to flexibly adjust its internal state. As indicated by the significant and sustained priming effect on behavior, subjects incorporated different perceptual expectations consistently throughout the recording session. However, the endogenous shift of decision criterion based on different perceptual expectations does not seem to require or rely on prestimulus alpha modulations.
Next, we asked whether trial-by-trial fluctuations of prestimulus alpha oscillations lead to changes in criterion or sensitivity. By focusing first on the visual regions most responsive to our grating stimuli, we showed that both the subject's perceptual sensitivity and information content decodable from the neural activity patterns are enhanced when prestimulus alpha power in early visual areas is low. Interestingly, the time window during which information content was modulated by prestimulus alpha power (i.e., ∼120–160 ms poststimulus onset) overlaps with that during which local (as opposed to top-down) recurrent processes dominate (Wyatte et al., 2014), suggesting that lower prestimulus alpha power may lead to enhanced local recurrent processing. In line with a large body of spatial attention literature (Haegens et al., 2011; Thut et al., 2006; van Ede et al., 2011) showing that the extent of cue-induced alpha power lateralization correlates with behavioral performance, our current results extend previous work and reveal that ongoing trial-by-trial alpha power fluctuations predict sensitivity changes. Collectively, these findings support the multiplicative hypothesis, namely, that lower prestimulus alpha power amplifies the neural response so that the signal of interest becomes more distinct from noise, thereby improving perceptual sensitivity.
We also asked whether and how alpha oscillations in different brain areas may have different modulatory effects on behavior. We showed that trial-by-trial changes of alpha power within the ventral visual pathway predicted behavioral sensitivity when subjects were primed to adopt a more conservative detection criterion and that fluctuations within the parietal and somatosensory regions predicted behavioral sensitivity when subjects were primed to adopt a more liberal criterion. These different spatial distributions may reflect that when subjects adopt different detection criteria, the processing bottleneck lies at different processing stages. When sensory information travels downstream, task-irrelevant information (i.e., noise) accumulates and scales to a larger extent compared with task-relevant information (i.e., signal), resulting in decreased signal-to-noise ratio of task-relevant information. If gating of sensory noise is ineffective at the start of sensory processing (e.g., in primary visual cortex), the internal response used for perceptual decision-making is unlikely to reach the criterion later on, regardless of the effectiveness of gating of downstream noise. Therefore, when subjects adopt a more conservative criterion, alpha power fluctuations indexing the extent of effective gating in early visual areas become more determinant of perceptual sensitivity. Conceivably, although effective gating of upstream and downstream noise indexed by spontaneous alpha power fluctuations jointly determine perceptual decisions, it is possible that the respective importance of these information gating mechanisms is dependent on subject's decision strategy (i.e., criterion). This proposal is speculative, however, and in need of further empirical evidence.
Our whole-brain exploratory analysis further revealed that under the conservative priming condition, trial-by-trial changes of prestimulus alpha power within the somatosensory regions predicted the subject's detection criterion. Curiously, this relationship was most pronounced for alpha activity 500 ms preceding the stimulus onset and disappeared for time points closer to the stimulus onset. The current observation appears to mirror previous reports showing that subjects adopt a more conservative criterion when alpha power increases (Limbach and Corballis, 2016; Iemi et al., 2017; Samaha et al., 2017), and it further suggests that prestimulus alpha power modulation on d′ and criterion have different time courses. The current findings also highlight a methodological challenge faced by studies using conventional detection paradigms with variable prestimulus windows (Limbach and Corballis, 2016; Iemi et al., 2017), in that the definition of the prestimulus time window in such designs is arbitrary as no stimulus is presented on target-absent trials and consequently there is no stimulus onset to reference. Arguably, this can be reconciled if the stimulus onset is made predictable to the subject (e.g., by using temporal cues and fixed prestimulus windows). However, the immediate question is to what extent do top-down processes contribute to prestimulus alpha fluctuations recorded using paradigms with predictable stimulus onsets? Such paradigms probably capture different neural dynamics than ours (i.e., ongoing fluctuations vs top-down modulation) as previous work found that human subjects endogenously adjust their oscillatory alpha state, both in terms of power (Rohenkohl and Nobre, 2011) and phase (Samaha et al., 2015) before the predictable onset of relevant stimuli to optimize sensory processing.
Finally, having demonstrated the relationship between alpha power and perception, we asked whether prestimulus alpha activity modulates behavioral accuracy in a phasic manner as predicted by the inhibition hypothesis (Klimesch et al., 2007; Jensen and Mazaheri, 2010). We found that the alpha phase along the dorsal visual pathway immediately before stimulus onset modulated the subject's response accuracy and the response accuracy of the classifier. Although one may expect the alpha phase effect to be dependent on and additive to the alpha power effect (Mathewson et al., 2009), our findings are consistent with previous work showing that alpha phase and power effects have different spatial topographies (Busch and VanRullen, 2010). Further research is needed to understand how the interplay between oscillatory alpha power and phase modulates perception.
To summarize, the current study shows with complementary univariate and multivariate results that changes of prestimulus alpha band power modulate the quality of visual information coding at the neural level and perceptual sensitivity at the behavioral level. These findings provide important insights into how ongoing neural activity sets the internal state and subsequently influences perception.
Footnotes
This study was supported by The Netherlands Organization for Scientific Research Vidi Grants NWO 016.Vidi.185.137 awarded to S.H. and 864.14.011 to J.-M.S., Chinese Scholarship Council Grant CSC20170800036 awarded to J.Z., and European Union Horizon 2020 Program European Research Council Starting Grant 678286, ContextVision, awarded to F.P.d.L. We thank Kristina Baumgart and Felix Linsen for assistance with data acquisition.
The authors declare no competing financial interests.
- Correspondence should be addressed to Saskia Haegens at shaegens{at}gmail.com
