Precise spatial tuning of visually driven alpha oscillations in human visual cortex

Alpha-band (~8–13 Hz) oscillations are historically associated with reduced cortical activity, increasing during eyes-closed or uniform-field viewing and decreasing with arousal and attention. Noninvasive EEG/MEG work shows spatially specific alpha decreases contralateral to attended locations, implying a negative relation between alpha power and cortical engagement. In contrast, invasive studies in macaques and some human intracortical work have reported alpha power increases to stimuli within receptive fields, suggesting a positive association. The authors hypothesize that stimulus-driven measurements confound two concurrent processes: a decrease in true alpha oscillatory power and an increase in broadband power that extends into low frequencies, making raw power at the alpha frequency ambiguous. The study asks: (1) Do visual stimuli increase or decrease the alpha oscillation locally? (2) How spatially specific is the modulation of alpha, quantified via population receptive fields (pRFs)? By separating oscillatory alpha from broadband components and fitting spatial pRF models, they test whether alpha suppression is precisely tuned to stimulus location.

Classical EEG studies (Berger, Adrian & Matthews) linked alpha increases to cortical idling, with reductions during arousal/attention. Numerous EEG/MEG studies show retinotopically specific alpha decreases during spatial attention (e.g., Worden 2000; Kelly 2006; Sauseng 2005; Foster 2017; Popov 2019). Invasive studies have differed: macaque ECoG/LFP reports showed alpha increases for stimuli inside RFs (Takaura 2016; Klink 2021) and human ECoG sometimes reported alpha increases near electrode RFs (Harvey 2013; Luo 2024). Broadband (aperiodic) power increases accompany neural activation and extend across frequencies (Crone 1998; Miller 2009; Winawer 2013; Hermes 2017), potentially masking narrowband alpha decreases if not decomposed. Prior work also separated broadband from narrowband gamma/SSVEP (Winawer 2013; Hermes 2015, 2017, 2019) and provided tools to parameterize periodic vs aperiodic spectra (Donoghue 2020). These literatures motivate a modeling approach to disentangle alpha oscillations from broadband changes and to map their spatial tuning via pRFs.

Participants and data availability: Nine epilepsy patients (NYU Grossman School of Medicine: 7; UMC Utrecht: 2) implanted with clinical subdural electrodes participated with informed consent and IRB/ethics approvals. Data and MATLAB code are openly available (BIDS dataset ds004194; ECoG_utils and ECoG_alphaPRF repositories). Electrodes and recording: Arrays included standard subdural grids (8×8, 2.3 mm exposed diameter, 10 mm spacing), linear strips (4–12 contacts), depth electrodes (1.1 mm diameter, 5–10 mm spacing), and high-density (HD) grids (16×8, 1 mm exposed diameter, 3 mm spacing; Patients 8 and 9). Amplifiers: NicoletOne or Neuroworks Quantum at NYU (recorded 2048 Hz then downsampled to 512 Hz; bandpass ~0.01–682.67 Hz or 0.16–250 Hz), MicroMed at UMCU (2048 Hz, 0.15–500 Hz). Stimulus triggers were recorded via audio/serial links; responses via keypads. Stimuli and task (pRF mapping): Grayscale band-pass noise patterns were presented within a 2°-wide bar aperture sweeping the visual field in eight directions. Stimulus field: circular, 16.6° diameter (radius 8.3°). Each bar step lasted 850 ms (500 ms stimulus, then 350 ms mean-gray blank). Each run: 224 steps plus 3 s blanks at start/end, totaling 196.4 s. Participants fixated and pressed a button upon fixation-cross color changes (1–5 s intervals). Each participant had 2–6 identical runs. Preprocessing: Bad/noisy or epileptiform electrodes excluded after visual inspection. Data were re-referenced to common average per electrode group (grid/HD/strip/depth). UMCU data were downsampled to 512 Hz; a stimulus-onset delay of ~72 ms (95% CI 63–85 ms) was corrected via ERP cross-correlation against NYU data. For one NYU patient with trigger issues, step timing was reconstructed from the first valid trigger and known schedule (850 ms steps). Voltage was epoched from −0.2 to 0.8 s per trial and baseline-corrected using the −0.2 to 0 s pre-stimulus window. Artifact rejection and electrode selection: Epochs with excessive residual power after projecting out the evoked potential (ERP) were excluded using an inverse Gaussian model per electrode; ~0.8% epochs were rejected. Twelve high-variance electrodes were removed; some runs with fixation breaks were excluded. Electrodes for pRF analyses included all HD grid contacts and grid/strip contacts with ≥5% probability of belonging to any visual area (Wang et al. probabilistic atlas). Final dataset: 366 electrodes from 9 participants (334 in defined visual areas). Electrode localization: Post-implant MRI (NYU) or CT (UMCU) was co-registered to pre-op T1 MRI; electrodes were projected to the pial surface. Visual areas were assigned probabilistically using the Wang et al. (2015) retinotopic atlas (25 maps merged into 12 groups, e.g., V1–V3, hV4, LO, TO, V3AB, IPS, SPL). Probabilistic assignments were retained and propagated to uncertainty estimates. Spectral analysis and ERP regression: To avoid spectral contamination from ERPs, stimulus and blank ERPs were estimated and regressed from each epoch. Power spectral density (PSD) was computed during the 0–500 ms stimulus window (1 Hz bins) using Welch’s method (200 ms Hann window, 50% overlap). Baseline PSDs were geometric means across blank epochs. Model-based decomposition of alpha and broadband: For each stimulus position, the log power ratio log10(Ps/Pb) in 3–26 Hz was modeled as the sum of a broadband elevation term and a Gaussian bump capturing alpha suppression centered at the per-electrode alpha peak (constrained 8–13 Hz), with parameters: linear broadband shift and Gaussian amplitude/width (Equation 1). For two participants with prominent beta (~24 Hz) peaks, an additional constrained Gaussian term (15–30 Hz) was included (3–32 Hz fit; Equation 2) as a nuisance to improve alpha estimation. Broadband elevation for pRF fitting was independently estimated as the stimulus/blank geometric mean power ratio across 70–180 Hz excluding line-noise harmonics (Equation 3). Time-series construction and decimation: For each electrode, 224 trial-wise summary values were extracted: alpha suppression metric (from model) and high-frequency broadband elevation. To improve SNR, these time series were low-pass filtered/decimated (Chebyshev Type I IIR, order 3) to 75 time points. The same decimation was applied to the binary aperture sequences. Population receptive field (pRF) modeling: Two pRFs were fit per electrode—one to alpha, one to broadband—using a circular Difference-of-Gaussians (DoG) formulation with an effectively infinite surround (implemented as a Gaussian center plus constant offset). Five parameters were estimated: center position (x,y), center size σ, center gain g1, and surround offset g2. Predictions were the dot product of aperture and pRF (Equation 4). Parameters were fit to maximize R² with twofold cross-validation: parameters estimated on one half (temporal halves with all orientations), predicting the other half; cross-validated R² was computed on concatenated halves (Equation 6). pRF centers were constrained to ±16.6°. Electrode grouping and goodness thresholds: Electrodes were grouped into visually responsive vs non-responsive based on broadband pRF fits at high frequency (70–180 Hz). Null distributions for alpha and broadband goodness were obtained by shuffling pRF parameters across electrodes (5000 iterations), setting thresholds at the 95th percentile of null R² (31% broadband; 22% alpha). Electrodes exceeding thresholds and with pRF centers within 8.3° were included in alpha–broadband comparisons. Probabilistic area assignments were respected via resampling across 5000 bootstraps. Inter-electrode coherence (HD grids): For Patients 8 and 9 (3 mm spacing grids), magnitude-squared coherence (MSC) was computed after ERP regression using sliding 500 ms windows (75% overlap), across −200 to 800 ms, in 1 Hz bins. Alpha coherence was taken at the peak alpha frequency; broadband coherence was averaged across 70–180 Hz. Coherence was averaged over runs and analyzed as a function of inter-electrode distance (binned in 3 mm steps), with bootstrapped exponential decay fits and baseline convergence estimates. Analyses were repeated across trial subsets (stimulus overlapping vs not overlapping the pRF) with similar outcomes.

• Separating oscillatory alpha from broadband is essential: Visual stimulation increased broadband power while decreasing the alpha oscillation; without decomposition, alpha-band power at 8–13 Hz can appear unchanged or increased due to broadband contamination.
• pRF prediction accuracy: For visually selective electrodes in V1–V3, alpha pRFs explained a substantial fraction of variance (mean cross-validated variance explained ~37%), while non-visual electrodes were near 0%. In dorsolateral maps, alpha pRFs explained ~24% in visually responsive electrodes and ~0% in non-responsive. Example electrode fits achieved cross-validated R² of 90.2% (broadband) and 69.5% (alpha).
• Alpha vs broadband pRFs—location vs size: Alpha and broadband pRF centers were highly co-localized in polar angle (V1–V3 r=0.98; dorsolateral r=0.95) and correlated in eccentricity (r=0.68 in V1–V3; r=0.09 dorsolateral), but alpha pRFs were 2–4× larger in size (size correlations: r=0.33 V1–V3; r=0.24 dorsolateral). The broadband pRF was largely contained within the alpha pRF: 92.3% containment (V1–V3) and 98.0% (dorsolateral); after shuffling alpha/broadband pairings, containment dropped to 30.0% and 25.7%, respectively.
• Size–eccentricity scaling: Both alpha and broadband pRF sizes increased with eccentricity, but with steeper slopes for alpha (V1–V3: 0.53 alpha vs 0.15 broadband; dorsolateral: 1.17 alpha vs 0.40 broadband).
• Temporal-frequency control: Low-frequency broadband (3–26 Hz) pRFs resembled high-frequency broadband (70–180 Hz) pRFs in size and positive gain and remained much smaller than alpha pRFs (V1–V3 sizes approx.: high-freq broadband 1.0°, low-freq broadband 1.1°, alpha 2.3°), indicating that larger alpha pRFs are not due to frequency differences or baseline correction artifacts.
• Importance of model-based baseline correction for alpha: Compared to band-limited alpha power without correction, the model-based approach improved median cross-validated R² about 4× (43% vs 10%), yielded consistently negative gains (29/31 electrodes negative vs mixed signs without correction), and recovered the expected positive size–eccentricity relationship (which was absent without correction).
• Spatial coherence: Alpha coherence was higher and spread farther than broadband coherence. Neighboring HD electrodes (3 mm) showed alpha coherence ~0.6, decaying toward a baseline ~0.35 by ~10 mm; broadband coherence was lower at all distances. Coherence peaked at the per-electrode alpha frequency and was similar whether or not the stimulus overlapped the pRF, suggesting larger-scale spatial synchrony for alpha than for broadband.
• Behavioral relevance: The alpha pRF profile (negative center with larger spatial spread into the periphery) mirrors classic asymmetric spatial cueing effects in behavior (Downing & Pinker, 1985; Shulman et al., 1986), supporting a mechanistic link between stimulus-driven alpha suppression and exogenous attention.

The study resolves a key discrepancy between invasive and noninvasive findings by showing that visually driven local responses comprise two separable components: a broadband increase and an alpha oscillatory decrease. When conflated, alpha-band power can appear to increase or remain unchanged, masking true alpha suppression. pRF modeling of the decomposed signals shows that alpha suppression is precise in location, aligned with broadband pRF centers, yet integrates over a broader spatial extent (2–4× larger), implying a coarser control of cortical excitability at alpha timescales. The robust negative gains and size–eccentricity scaling reinforce the view of alpha as pulsed inhibition or gating that, when suppressed by stimuli, increases cortical excitability beyond the directly driven zone. Elevated alpha coherence over millimeters of cortex further supports a large-scale synchronizing mechanism. Functionally, the large, negative alpha pRF can account for several hallmarks of exogenous spatial attention: rapid timing (~100 ms), increased sensitivity near the cue with asymmetric spread into the periphery, and center–surround effects (benefit near center, cost in surround). The authors propose that stimulus-driven changes in alpha modulate excitability, producing attention-like behavioral effects, flipping typical causal assumptions that attention drives alpha changes. The thalamocortical generators (e.g., LGN alpha pacemakers modulated by corticothalamic feedback) may underlie the spatial coherence and spread. Methodologically, decomposing periodic and aperiodic components yields more accurate spatial tuning estimates and reconciles prior invasive reports that found mixed alpha increases/decreases or no size–eccentricity scaling. Differences across areas (e.g., low-frequency broadband extending to ~3 Hz in V1–V3 but not dorsolateral) imply regional variability in broadband generators and stimulus effectiveness. Overall, the findings generalize the importance of spectral decomposition beyond vision, wherever alpha suppression and broadband elevations co-occur.

The work demonstrates that visual stimulation elicits two independent signals in human visual cortex: a broadband power elevation and a suppression of the alpha oscillation. Separating these components is critical to accurately estimate spatial tuning with pRFs. Alpha suppression is finely retinotopic yet integrates over a larger spatial extent than broadband responses, aligning with and potentially explaining key properties of exogenous spatial attention via increased cortical excitability. The approach reconciles prior invasive and noninvasive discrepancies and underscores the need to model periodic vs aperiodic components in neural spectra. Future directions include jointly mapping alpha specificity across cortical surface and depth, characterizing inter-areal alpha-mediated feedback pathways, testing generalization to other modalities/areas where alpha suppression occurs, and integrating behavioral manipulations (endogenous/exogenous attention) with decomposition-based ECoG/MEG to further test causal links. Extending models to parameterize aperiodic exponent changes and feature-specific surround dynamics may refine interpretations across cortical hierarchies.

• Clinical sample and coverage: Only nine epilepsy patients with clinically driven implant locations; coverage is incomplete and variable across individuals and areas.
• Electrode localization uncertainty: Areas were assigned probabilistically; some electrodes contribute to multiple area groups, introducing localization uncertainty.
• Site/methodological differences: UMCU vs NYU instrumentation required latency alignment; residual timing inaccuracies could affect ERP-related measures, though ERPs were regressed out.
• Model assumptions: The alpha decomposition assumes a Gaussian bump on log–log spectra and a linear broadband shift; two participants required an added beta component. The pRF model assumes circular Gaussian centers with an effectively infinite surround.
• Frequency range choices: Broadband for pRF was restricted to 70–180 Hz; low-frequency broadband was separately analyzed but not used for main broadband pRFs.
• SNR differences: Alpha signals have lower SNR than broadband high-frequency power, potentially biasing parameter variance; however, size–eccentricity scaling and containment analyses argue against noise-driven size inflation.
• Lack of laminar data: The study maps surface spatial specificity but does not assess cortical depth-specific alpha/beta feedback mechanisms.
• Task constraints: Fixation task minimizes eye movements, but micro-saccades or unmeasured oculomotor factors could still modulate alpha.