Alpha and broadband high-frequency activity track task dynamics and predict performance in controlled decision-making

Abstract

Intracranial recordings in human subjects provide a unique, fine-grained temporal and spatial resolution inaccessible to conventional non-invasive methods. A prominent signal in these recordings is broadband high-frequency activity (approx. 70–150 Hz), generally considered to reflect neuronal excitation. Here we explored the use of this broadband signal to track, on a single-trial basis, the temporal and spatial distribution of task-engaged areas involved in decision-making. We additionally focused on the alpha rhythm (8–14 Hz), thought to regulate the (dis)engagement of neuronal populations based on task demands. Using these signals, we characterized activity across cortex using intracranial recordings in patients with intractable epilepsy performing the Multi-Source Interference Task, a Stroop-like decision-making paradigm. We analyzed recordings both from grid electrodes placed over cortical areas including frontotemporal and parietal cortex, and depth electrodes in prefrontal regions, including cingulate cortex. We found a widespread negative relationship between alpha power and broadband activity, substantiating the gating role of alpha in regions beyond sensory/motor cortex. Combined, these signals reflect the spatio-temporal pattern of task-engagement, with alpha decrease signifying task-involved regions and broadband increase temporally locking to specific task aspects, distributed over cortical sites. We report sites that only respond to stimulus presentation or to the decision report, and interestingly, sites that reflect the time-on-task. The latter predict the subject’s reaction times on a trial-by-trial basis. A smaller subset of sites showed modulation with task condition. Taken together, alpha and broadband signals allow tracking of neuronal population dynamics across cortex on a fine temporal and spatial scale.

Introduction

Intracranial recordings in human subjects are gaining increasing relevance for cognitive neuroscience (Jacobs & Kahana, 2010; Lachaux et al., 2012; Parvizi & Kastner, 2018) and provide a unique, fine-grained temporal and spatial resolution inaccessible to conventional non-invasive methods. While recording locations are typically limited and exclusively based on clinical decisions, by combining data from a relatively large cohort of patients, substantial cortical coverage can be obtained. This approach provides intracranial electrophysiology on an unprecedented scale of anatomical resolution in humans, opening up a unique line of research that is otherwise only available in nonhuman animal studies, while allowing for the study of cognitive paradigms otherwise only available in non-invasive human studies.

One aspect that is readily observed in the human intracranial EEG is broadband high-frequency activity (Crone et al., 2006; Jerbi et al., 2009; Lachaux et al., 2012; Uhlhaas et al., 2011), typically reported in a range between 50 up to 200 Hz. This broadband activity is considered to be a relatively direct measure of neuronal ensemble firing patterns, i.e., multiunit activity (Manning et al., 2009; Miller et al., 2014; Ray, Crone, et al., 2008; Ray & Maunsell, 2011), though it likely also indexes additional neuronal processes such as synaptic currents and dendritic spikes (Lachaux et al., 2012; Leszczyński et al., 2020; Suzuki & Larkum, 2017). Critical for the present study, broadband activity provides a reliable measure of local neuronal excitation. Note that “local” here refers to a spatial resolution in the order of several mm, though the question how “local” the local field potential actually is remains actively debated, with estimates varying from several hundred microns to over a cm (Buzsáki et al., 2012; Kajikawa & Schroeder, 2011).

Here we explore the relationship between broadband activity and the alpha rhythm (8–14 Hz), which is thought to play a key role in regulating engagement and disengagement of neuronal populations based on task demands (S. Haegens, V. Nácher, et al., 2011; Jensen & Mazaheri, 2010; Klimesch et al., 2007). Indeed, one current proposal states that the alpha rhythm reflects a mechanism of active inhibition, i.e., alpha functions to suppress neural processing, as evidenced by a negative correlation between alpha power and spike firing rates (S. Haegens, V. Nácher, et al., 2011; van Kerkoerle et al., 2014). This account is further supported by findings showing that both low alpha power in task-relevant regions and high alpha in task-irrelevant regions are associated with better behavioral performance (e.g., contra- vs. ipsilateral alpha modulation in spatial attention tasks; S. Haegens, B. F. Händel, et al., 2011; Haegens et al., 2012; Thut et al., 2006). If alpha indeed reflects functional disengagement via active inhibition, and broadband activity reflects engagement of a neuronal population, these signals should be negatively correlated. While several studies have commented on such a negative relationship (see e.g. Jerbi et al., 2009), few have systematically investigated this, with existing quantitative reports mostly limited to sensorimotor regions (e.g., de Pesters et al., 2016; Jiang et al., 2020; Potes et al., 2014).

Here we further examined the proposed inhibitory role of the alpha rhythm and address the generality of this proposed mechanism across cortex, while tying these dynamics to behaviorally relevant processing. In order to do so, we set out to describe the spatial and temporal dynamics of the anticorrelation between alpha and broadband activity. We asked whether these patterns can be generally observed across cortical regions, i.e., does alpha provide a gating mechanism beyond sensory and motor cortex? Furthermore, we explored their combined relation to behavior: the broadband signal allows us to track, on a single-trial basis, the temporal and spatial distribution of task-engaged areas. Finally, to test the utility of the broadband signal on a trial-by-trial basis, we asked whether the broadband signal encodes task-relevant information by studying its correlation with task condition and subsequent performance measures.

We report on a large database of intracranial recordings—both surface/subdural electrocorticography (ECoG) and stereo-EEG (sEEG)—in 19 patients with medically refractory epilepsy, including a total of approximately 1600 recording sites across the brain. Subjects performed the Multi-Source Interference Task (MSIT; Bush & Shin, 2006; Sheth et al., 2012), a Stroop-like paradigm with different levels of distracter and congruence interference, allowing us to address various components of controlled decision-making, including effects of stimulus conflict, response selection, and feedback processing.

Method

Subjects

Subjects were patients (N=19, 9 female) with pharmacoresistant focal epilepsy in whom a non-invasive investigation failed to adequately localize seizure onset. These patients were therefore implanted with intracranial electrodes for the clinical purpose of identifying seizure foci. See Table 1 for overview of electrode coverage. All patients gave written informed consent before participating in the experimental task (before surgical implantation), and procedures were approved by the local Institutional Review Board.

Table 1. Patients.

Overview of electrode type and number per patient.

Patient	Coverage	Electrodes #
X20	sEEG & grid	33
X25	sEEG & grid	72
X27	sEEG	65
X29	sEEG & grid	114
X30	sEEG	35
X31	sEEG	79
X32	sEEG	84
X33	sEEG	110
X34	sEEG	111
X35	sEEG	117
X36	sEEG	114
X37	sEEG & grid	89
X42	sEEG & grid	99
Z33	sEEG & grid	106
Z34	sEEG & grid	79
Z36	sEEG	78
Z38	sEEG	70
Z39	sEEG	76
Z40	sEEG	62

Task

Subjects performed the MSIT task, a Stroop-like (Stroop, 1935) cognitive interference task (Bush & Shin, 2006; Sheth et al., 2012; Smith et al., 2015) in which they viewed a stimulus consisting of three numbers, and had to indicate by 3-forced-choice button press the unique target number that differed from the other two distracter numbers (left button for “1”, middle button for “2”, and right button for “3”; Figure 1A). Importantly, the correct button corresponded to the target identity (i.e., numbers 1–3), regardless of where the target appeared in the sequence (i.e., its spatial position). After the response, feedback was presented, which was valenced on half the trials (green for correct, red for incorrect responses) and unvalenced (blue) for the other half, switching every 10 trials.

Figure 1. Experimental paradigm and behavioral results.

(A) MSIT task overview: subjects were presented with a central fixation point for 500 ms, after which the stimulus was presented until the response (or for max 5 sec). Subjects had to indicate by button press (left button for “1”, middle button for “2”, and right button for “3”) which one of three numbers was the odd-one-out, after which feedback (valenced or unvalenced) was presented for 500 msec. ITI was 2 sec.

(B) Four different conditions were used: Type 0, with spatial congruence and easy distracters; Type 1A, with spatial incongruence; Type1B, with difficult distracters; and Type 2, with spatial incongruence and difficult distracters.

(C) Normalized reaction times per condition, showing average over 19 subjects. Significant corrected post-hoc tests indicated with asterisk (p<0.05). Error bars indicate standard deviation versus the mean, clouds represent distribution of datapoints.

By varying the position of the target and the identity of the distracters, the task contains two types of cognitive conflict or interference: Simon interference is introduced by the presence of spatial incongruence between the position of the target number in the stimulus sequence and the position of the correct button on the button pad. For example, the correct choice for both stimulus sequences “200” and “020” is the second button (i.e., the correct target number is “2” in both examples); the stimulus “200” contains spatial interference (incongruence between the position of the target number in the stimulus and position of the correct button choice), whereas the stimulus “020” does not, as here the position and identity of the target number are congruent. Eriksen flanker interference is introduced by the presence of distracters that are possible targets. For example, the correct choice for both stimuli “133” and “100” is the first button; but whereas “133” contains distracter interference, as “3” is a possible button choice, stimulus “100” does not, since “0” is not a possible button choice. Combining these two forms of interference creates four distinct trial types (Figure 1B), ranging from zero (type 0; easy) to one (types 1A and 1B; intermediate) to two levels (type 2; difficult) of cognitive interference. These trial types were presented to the subject in random order. Note that while all subjects did the easy (type 0) and difficult (type 2) conditions, not all subjects participated in the intermediate type 1A and 1B trials (3 patients did not perform conditions 1A&B, 4 others did not do 1B).

Data acquisition

Data were collected using the clinical Xltek system and sampled at 500 Hz, high-pass filtered at 0.5 Hz and low-pass filtered at 250 Hz. Clinical data were synchronized with patient behavior via a task control computer that sent analog synchronization signals to the Xltek system. The task was implemented using the monkeylogic toolbox for Matlab (Asaad & Eskandar, 2008). Electrophysiological data, including time-locked synchronization signals were exported offline from the clinical database for subsequent analyses. In total we recorded data from 1899 electrodes across 19 patients (9 female). All patients participated in at least one full experimental session; some underwent multiple sessions (on subsequent days). The average number of trials per session was 285 trials (range: 134–550 trials).

Data analysis

For data analysis, we used custom-built Matlab code and the FieldTrip toolbox (Oostenveld et al., 2011), as well as SPM8. Data from bad channels (16% of total) were rejected with visual inspection, leaving 1593 channels that were reliably in the brain and not contaminated by interictal epileptiform discharges or recording artifacts. All remaining channels were re-referenced to the common average (per patient, per recording session). Line noise (at 60 and 120 Hz) was removed using a band stop filter, and data were normalized by subtracting the mean signal per channel per trial. Remaining bad trials were removed after visual inspection of the preprocessed data (19% of total).

Spectral analysis

Time-frequency representations (TFRs) of power were computed using a sliding time window fast Fourier transform (FFT) approach. For lower frequencies (4–30 Hz) we used an adaptive window of four cycles length (Δt = 4/f), while for higher frequencies (70–150 Hz) we used a fixed 200-ms window length. Each window was multiplied with a Hanning taper before estimating power using FFT.

For the subsequent analysis we specifically targeted two frequency bands: the alpha band and broadband high-frequency activity (70–150 Hz), often referred to as “high gamma”. Note that this latter activity likely goes up to higher frequencies (>200 Hz); the cap at 150 Hz here was simply chosen because of the low-pass filters applied during data acquisition. Furthermore, as we believe this frequency band to reflect multi-unit activity, rather than an oscillatory process, we refer to this activity from here on as broadband, rather than gamma activity.

For the alpha band, we determined each patient’s alpha peak frequency across channels and trials (local maximum of power spectrum in the 7–15 Hz range), and used a +/− 3 Hz window centered at the individual peak frequency for further analysis (Haegens et al., 2014). Spectra were normalized (per electrode) with average power across all frequencies before this procedure. In order to address the relationship between alpha and broadband activity, we binned the data from low to high alpha power—either binning across channels to assess the spatial dynamics, or binning trials per channel to assess temporal dynamics. Binning consisted of sorting all channels (or trials) based on alpha power (averaged within the subject-specific band and over the time window of interest) and splitting them in five equal-sized, non-overlapping bins, and computing the average normalized broadband activity per bin (i.e., averaged over all data points per bin).

Channel selection

First, we selected task-engaged channels based on the alpha response. Single trial spectra were normalized per channel with the average spectral power over that trial, and then averaged over the alpha range using each individual alpha peak frequency +/− 3 Hz. Statistical testing on the TFRs was done using a non-parametric cluster-based permutation approach (Maris & Oostenveld, 2007). By clustering neighboring time samples that show the same effect, this test deals with the multiple-comparisons problem taking into account the dependency of the data. We tested whether there was a significant decrease of alpha activity during task (t=[0.5 2] sec) vs. baseline (t=[−1 −0.5] sec). By randomizing the data across the two conditions and recalculating the test statistic 2000 times, we obtained a reference distribution to evaluate the statistic of the actual data. All channels with a significant alpha decrease (cluster-corrected p<0.05 per subject) during task vs. baseline entered the subsequent analysis steps.

Channel classification

In order to classify channels based on response type (see e.g., Haller et al., 2018) and to test for correlation with reaction times and task condition, we used a generalized linear model (GLM) approach in which we applied a multivariate, multiple linear regression analysis.

For channel classification, per channel, the data were represented as trials-by-time series. Single trial spectra were normalized per channel with the average spectral power across trials, and averaged over the broadband high-frequency range. We then modeled the data with predictors that reflected (1) stimulus processing, (2) response preparation/execution and (3) feedback processing. Each predictor was modeled as a 500-ms long boxcar event aligned with stimulus onset, or centered at response time or feedback presentation, respectively. The output of our GLM was a set of beta weights (averaged over repetitions and time points), for each channel. Channels were classified based on which predictors had positive beta weights passing a threshold determined by simple t-tests per predictor (set at p<0.05). If both the stimulus and response predictors passed the threshold, and activation was sustained between these two time windows, the channel was labeled “task-duration” type. The feedback channel type was allowed to be present in addition to the other three types.

GLM

After selecting and classifying the task-engaged sites, we used a GLM approach with regressors RT and task condition to test whether the broadband response was predictive of behavior and/or type of interference (Type 0, 1A, 1B, or 2). By definition, the response and task electrodes (see above) correlated with RT; however, testing these two factors at the same time allowed us to disentangle RT-driven effects (which differed across conditions) from genuine condition differences (see also Results section). Here the output of our GLM was a set of beta-weight time series per channel. This allowed us to evaluate per channel, whether the broadband response correlated with RT and/or condition, and when these effects occurred. Significance was evaluated by computing t-tests on the regression coefficients, corrected for multiple comparisons using FDR correction (Benjamini & Hochberg, 1995; Groppe et al., 2011) over tests per subject, and Bonferroni corrected for number of subjects. In order to threshold the beta-weight time courses for visualization purposes, we randomly permuted the regressors with respect to the trials and recomputed the GLM. We collected the beta weights of each permutation (N=100) to construct a time course to threshold the beta weight time course of the experimental data.

Results

We studied alpha band activity and broadband high-frequency activity in a distributed cortical network using intracranial recordings in 19 patients with intractable epilepsy performing the MSIT task. MSIT is a Stroop-like task with Simon and Eriksen flanker distraction elements, in which subjects view a stimulus consisting of three numbers and have to indicate the unique target number that differs from the other two numbers (Bush & Shin, 2006; Sheth et al., 2012). By varying target position and distracter identity, the four trial types introduce different levels of cognitive interference.

Reaction times increase with interference load

First, we analyzed the behavioral performance on the MSIT task. The average performance over all conditions in 19 subjects was 95.7% ± 7.8% correct, with a significant difference across conditions according to a repeated measures ANOVA (N=12, F(3,33)=6.177, p=0.002, p=0.013 after Greenhouse-Geisser correction for violation of sphericity assumption). However, Bonferroni-corrected post-hoc tests showed no significant pairwise differences between the conditions. Note that only 12 patients had data for all conditions. A paired t-test contrasting condition type 0 and type 2 (for which all 19 patients had data) showed a significant difference, with higher accuracy for type 0 (mean=98.7%, sd=2.1%) than type 2 (mean=93.0%, sd=10.9%; t(18)=2.632, p=0.017).

Average RT, computed on correct trials only, was 1198.3 ms ± 450.6 ms, which was significantly different across conditions (N=12, F(3,33)=43.322, p<0.001, also after Greenhouse-Geisser correction; Figure 1C; note that statistics were computed over subject-normalized data). Post-hoc tests showed significant pairwise differences between all conditions (Bonferroni-corrected p<0.05) except type 1A versus 1B, which showed a trend (corrected p<0.069). A t-test directly comparing type 0 (mean=1014.8 ms, sd=384.3 ms) and type 2 (mean=1400.0 ms, sd=567.6 ms), including data from all subjects, further confirmed this pattern of faster responses for the no-conflict condition (t(18)=−11.877, p<0.001).

In sum, subjects were faster on the no-conflict condition than on the other conditions, with slowest reaction times for the condition with both sources of interference. Additionally, there was a small effect of higher accuracy for the no-conflict condition compared to the double interference condition.

Task-related alpha decrease accompanied by broadband increase

We next tested our hypothesis regarding the gating function proposed for alpha: specifically, a negative relationship between alpha and broadband activity. To do so, we first computed TFRs of power for the task, across conditions, for each electrode location. A total of 1593 electrodes (Figure 2A) were included in the initial analysis. Generally, we found that decreased alpha activity was accompanied by an increase of broadband activity, while sites with task-related alpha increase saw little to no broadband activity (Figure 3AB).

Figure 2. Electrode locations.

(A) Standard brain showing locations of recording electrodes that were included in the analysis (N=1593, across 19 patients).

(B) Standard brain showing locations of sites that showed task-engagement, as determined by a significant decrease of alpha activity vs. baseline (N=621, across 19 patients). Note that the few electrode contacts that appear to be above the cortical surface are likely artefactually displaced due to small patient-patient registration errors.

Figure 3. Alpha power vs. broadband activity.

(A) TFR of lower frequency range (5–30 Hz) showing alpha decrease (absolute change vs. baseline), in two example parietal electrodes in two different patients (upper panels) versus alpha increase in example electrode in another patient (lower panel).

(B) TFR of higher frequency range (70–150 Hz) showing high-frequency broadband increase (relative change vs. baseline; upper panels) versus no change (lower panel), for the same locations as in A.

(C) Negative relationship between alpha and broadband activity over electrodes; data for all patients combined, computed over all recording sites, showing non-linear relationship (N=1593 sites, see Figure 2A). Blue (portions of) bars reflect data from electrodes within the task-engaged channel selection (i.e., significant alpha decrease during task vs. baseline). Bins comprised 100%, 82%, 14%, 0%, and 0% of data from the selected subset, respectively.

(D) Similar as C, computed over all sites with significant task-related alpha decrease (N=621 sites, see Figure 2B).

(E) Negative relationship between alpha and broadband activity over trials; data for all patients combined, selected subset of channels only.

We then binned all electrodes from low to high alpha power (equally sized bins, based on average alpha power over trials, in each electrode) and computed the average normalized broadband activity per bin (i.e., averaged over all electrodes in each bin). We found a significant negative relationship between alpha and broadband activity (F(4,1585)=60.78, p<0.001; Figure 3C), such that electrodes with lower alpha during the task showed the highest broadband increase. This dynamic was found across the brain. In addition to sites where alpha power decreased during the task, this analysis also included electrodes with task-related alpha power increase, and these showed virtually no broadband modulation, or even a decrease vs. baseline. As a further check we repeated this analysis including only channels with a distinct alpha peak (local maxima within the alpha range) to prevent mixing in of other low-frequency band effects: this confirmed a negative alpha-broadband correlation (F(4,905)=35.5, p<0.001).

We then selected all electrodes (N=621) that were marked as task-engaged as reflected by a significant alpha band decrease vs. baseline (i.e., approx. 39% of electrodes, tested per patient, cluster-corrected p<0.05; Figure 2B). In these sites we observed a graded modulation of broadband activity with alpha level (one-way ANOVA, F(4,615)=17.45, p<0.001; Figure 3D). Note that these selected electrodes entered the subsequent analysis steps.

Next, we asked how this dynamic played out over time, now binning the single trials (per electrode, including only the task-engaged subset) based on average alpha power per trial, and again computing the average broadband activity per bin (Figure 3E). We found that also on a trial-by-trial basis, alpha power levels negatively correlated with broadband activity (F(4,3056)=5.56, p<0.001).

In sum, we found task-related alpha decrease in approximately 39% of all electrodes across widespread brain regions, which was accompanied by an increase in broadband activity. This dynamic was observed on a trial-by-trial basis and across brain regions, providing further evidence for the generality of the alpha inhibition hypothesis.

Broadband activity tracks task components

For the 621 electrodes showing significant task engagement, we then asked in which part of the task they were involved. We used a GLM approach to classify the electrodes in the following predefined categories: (1) stimulus, (2) response, (3) feedback, or (4) task duration selective electrode (Figure 4 and Table 2). Classification was determined based on the time course of broadband activation (see Methods). Out of 621 selected channels, 453 (73%) could be successfully classified using this approach. The activity in the remainder of the channels was not clearly time-locked to one of our predefined task aspects. Of the classified electrodes, 30% were stimulus-selective, 18% were selective to the response, 47% to task-duration (i.e., these sites responded to both the stimulus and the response), and 39% to feedback (6% were exclusively feedback-selective, the remaining 33% responded to feedback as well as other task elements). The location of these classified electrodes was widely distributed over recording sites (see Figure 4C for location of each electrode type). We make the following observations regarding location: the stimulus- and feedback-selective sites had a posterior to frontal distribution, predominantly along the ventral path, while response-selective sites tended to have a posterior to central distribution, more along the dorsal path. Task duration-selective sites were found along a combination of both these locations, and especially in the parietal area.

Figure 4. Electrode classification.

(A) Examples of stimulus-, response-, feedback- and task duration-selective electrodes (from left to right), showing one example location in one (different) representative example patient each. Upper panel: TFR of broadband activity (relative change vs. baseline). Middle panel: broadband event-related responses per condition. Lower panel: broadband response per trial, sorted for RT (overlaid in black).

(B) Three more examples for each electrode class, showing broadband response per trial, for different patients and locations. NB: the upper panel for feedback-selective sites in B and the example in A are from the same electrode in the same patient, on successive recording days. The lower panel for feedback-selective sites shows a location that is responsive to all visual stimulation onsets (fixation cross, target and feedback, respectively).

(C) Localization of electrode types, for all patients combined, showing left lateral and superior view of standard brain projection (stimulus, N=135; response, N=82; feedback, N=176; task, N=211).

Table 2. Electrode classification.

Overview of number of electrodes (absolute and percentage of total, N=621) per class: (1) stimulus selective, (2) response preparation/execution, (3) feedback (feedback-only and total feedback, including combination with other types), and (4) task duration, as well as non-classified electrodes (N/C). Last column indicates per type how many electrodes were selective for condition.

Type	Number of electrodes	Condition selective
Stimulus	N = 135 22% of total	N = 27 20% of this type
Response	N = 82 13% of total	N = 21 26% of this type
Feedback	N = 25 (176) 4 (28) % of total	N = 4 (64) 16 (36) % of this type
Task duration	N = 211 34% of total	N = 71 34% of this type
N/C	N = 168 27% of total	N = 37 22% of this type

The task-duration electrodes reflect, on a trial-by-trial basis, the time between stimulus onset and response. Hence, these electrodes, as well as the response selective electrodes, predict (by definition) the subject’s RT (Figure 5A). In addition to reaction time, a subset of electrodes reflected the trial type (condition). Here, it is critical to separate out RT effects, since RTs increased with task difficulty, and thus to a certain extent covary with condition; i.e., in terms of differences in neuronal activation, “pure” condition-selective effects should be reflected by differences in peak height of the response, whereas differences in latency or duration would be reflective of “RT-mediated” condition effects (e.g., compare the right most panels in Figure 5A and B, respectively). We performed a GLM with both RT and condition as regressors to distinguish between these effects. A subset of sites (26%) significantly reflected task condition (Figure 5C, Table 2). Here we observed different types: sites that increased vs. decreased their activity with task difficulty, but also more complex sites that preferred one condition strongly over the others (see Figure 5B for examples). These different condition-selectivity profiles were fairly evenly distributed across the previously identified electrode classes.

Figure 5. High-frequency broadband activity selective for RT and task condition.

(A) Examples of task-duration and response-selective electrodes, which are (by definition) predictive of RT. Upper panels: broadband event-related responses per condition (time points for which GLM showed significant effects for RT shown in light blue on top). Lower panels: broadband response per trial, sorted for RT (overlaid in black).

(B) Examples of task-duration selective electrodes that are additionally selective for task condition. Same conventions as in A, showing significant effects for condition in dark blue.

(C) Localization of condition-selective electrodes, for all patients combined, showing left lateral and superior view of standard brain projection (N=160 sites).

Discussion

We recorded intracranial EEG in 19 patients with intractable epilepsy who were implanted with subdural and depth electrodes for seizure monitoring. We asked whether alpha-band activity functions as an inhibitory (gating) mechanism across a range of cortical regions. Indeed, across recording sites, we found a negative relationship between alpha power and broadband high-frequency activity. Combined, these activation patterns allowed us to track task-engaged regions, showing a distributed cortical network involved in controlled decision-making in different conflict situations. We found different types of responses, distributed across brain regions, including sites that only responded to the stimulus presentation or to the response, and interestingly, sites that reflected the time-on-task. The latter sites closely tracked the subject’s reaction times on a trial-by-trial basis. Furthermore, we found a small subset of sites that showed modulation with task condition (i.e., conflict type).

Previous studies have shown conflict monitoring effects in the dorsal anterior cingulate cortex and dorsolateral prefrontal cortex (Botvinick et al., 2004; Horga et al., 2011; Mansouri et al., 2017; Sheth et al., 2012). Here we found conflict effects for a distributed subset of electrodes, with certain sites showing varying activation levels to the different conflict conditions. Considering that the different conflict types were also associated with reaction time differences, it is important to disentangle reaction time-related modulations (i.e., latency differences) from pure condition effects (e.g., power modulations independent of latency differences), which, as we show, is not trivial. In that sense, the fine temporal resolution of intracranial EEG can complement prior findings from fMRI studies, in which it is harder to separate these two factors.

Broadband high-frequency activity is a prominent feature of human intracranial EEG recorded at the pial surface of the cortex (Canolty et al., 2006; Crone et al., 2006; Miller et al., 2014; Miller et al., 2007; Ray, Niebur, et al., 2008). It can also be observed in LFPs recorded within the cortex (Saskia Haegens et al., 2011; Ray, Crone, et al., 2008; Ray & Maunsell, 2011) and MEG recordings (e.g., Haegens et al., 2010; Osipova et al., 2006), and while complex, clearly does index neuronal firing (Manning et al., 2009; Ray, Crone, et al., 2008) along with other excitatory processes (Leszczyński et al., 2020). Alpha oscillations have been linked to functional inhibition (S. Haegens, B. F. Händel, et al., 2011; Haegens et al., 2012; Thut et al., 2006), supported by a negative correlation with spike firing (S. Haegens, V. Nácher, et al., 2011; van Kerkoerle et al., 2014). Linking these two observations, previous intracranial EEG studies have shown a negative correlation between alpha and broadband activity for sensory and motor regions (de Pesters et al., 2016; Potes et al., 2014). Here, we expand these previous findings by showing a strong negative relationship between alpha and broadband high-frequency power, observed across cortical recording sites, i.e., not limited to sensory/motor areas. We find that in sites with high alpha power, broadband activity is suppressed altogether, whereas in sites with decreased alpha, there is a graded modulation of broadband activity (with highest activity for sites with lowest alpha). Furthermore, this pattern holds on a trial-by-trial basis, such that trials with higher alpha have lower broadband activity and vice versa. Together, these observations provide further evidence for alpha reflecting a (graded) gating mechanism, which is general across cortex.

One longstanding question is whether these alpha dynamics are indeed, as we suggest, reflective of an active (top-down) process, rather than the mere switching between two states (i.e., information processing vs. “idling”). If a network can either produce alpha oscillations or broadband activity, but not both at the same time, switching between these states need not necessarily be reflective of a modulatory active mechanism. While to fully solve this larger question spatially localized recordings (combined with causal manipulation) directly targeting these circuits are required, the graded nature of the effects (both spatially and temporally) suggests this goes beyond a passive on/off switch scenario. Importantly, this graded modulation has previously also been observed in relation to anticipatory alpha dynamics (Gould et al., 2011; S. Haegens, B. F. Händel, et al., 2011). Indeed, anticipatory alpha modulation, in the absence of concurrent information processing, would further argue in favor of a top-down modulated mechanism. Another alternative explanation for our observations is that shifts in the slope of the power spectrum (i.e., 1/f or aperiodic signal) manifest as simultaneous decrease in low-frequency and increase in high-frequency power, reflective of shifts in excitation/inhibition balance (Freeman & Zhai, 2009; Gao et al., 2017). While we cannot fully exclude any aperiodic contributions, we observe a narrowband effect, limited to the alpha range, rather than a broadband shift in the low-frequency range, arguing for a role for genuine oscillatory alpha activity. As for the underlying mechanism, one potential way alpha rhythms might exert their inhibitory role is by saturating excitatory synapses of pyramidal cell dendrites, thereby diminishing information relay (Sherman et al., 2016).

It has been suggested that high-frequency broadband activity reflects multi-unit spike activity (Manning et al., 2009; Miller et al., 2014)—especially synchronously firing cells (Ray, Crone, et al., 2008)—as well as synaptic currents (Pettersen et al., 2008; Ray & Maunsell, 2011). Although recent findings in both rodents (Suzuki & Larkum, 2017) and monkeys (Leszczyński et al., 2020) suggest that multiunit activity and broadband activity can be spatially, temporally and pharmacologically dissociated, these findings continue to support the idea that the broadband signal is a reliable proxy for local neuronal excitation in the cortex. Our results provide further evidence for this interpretation, in the form of the negative correlation between broadband activity and alpha. Furthermore, our findings underscore the utility of the broadband measure: (i) the observation that broadband activity can be used to track task-engaged sites on a fine temporal scale (see also e.g., Haller et al., 2018), as well as (ii) provide readout of task-specific information (e.g., conflict type). These observations are all in line with broadband activity being a direct measure of local neuronal activation. Thus, the broadband high-frequency activity signal provides a powerful measure with high spatial and temporal resolution, allowing tracking of trial-by-trial activation patterns otherwise only available in intracranial spike recordings.

The combination of these signals allows the assessment of widespread dynamics, with alpha reflecting a mechanism of resource allocation and providing us a window on engaged vs. disengaged areas and more fine-grained (in terms of spatio-temporal dynamics) broadband activations within those networks directly tied to perceptual and cognitive processing and subsequent behavior. Critically, this approach provides an avenue to tie together work showing that alpha impacts stimulus representation (e.g., Barne et al., 2020; Griffiths et al., 2019) and that alpha organizes neural activity both through power (our current focus) and phasic modulation (Chapeton et al., 2019; S. Haegens, V. Nácher, et al., 2011); i.e., we propose that the effect of alpha dynamics on information processing and subsequent behavior is mediated by a modulation of neuronal excitability. One outstanding question is whether this constitutes an active top-down mechanism or a reflection of underlying dynamics.