Large-scale dynamics of perceptual decision information across human cortex

The study investigates how the brain transforms fluctuating sensory inputs into perceptual decisions by accumulating evidence over time into a decision variable that can guide action. Prior work indicates that sensory cortex encodes instantaneous evidence while downstream association and motor regions reflect accumulated evidence and action plans. However, the large-scale cortical distribution and dynamics of these computations, and their relation to behaviorally inferred evidence accumulation, remain unclear. Moreover, classical feedforward accounts predict that decision-related signals in sensory cortex solely reflect stimulus-driven feedforward influences, whereas hierarchical circuit models with feedback predict that sensory cortical choice-predictive activity contains both stimulus-dependent (feedforward) and endogenous (feedback) components. The authors aim to integrate behavioral psychophysics with region-specific MEG decoding to map, across human cortex, the time courses of (i) sensory evidence encoding, (ii) accumulated evidence, and (iii) choice-related build-up, and to disentangle stimulus-driven versus endogenous decision signals, particularly within the visual hierarchy.

Evidence accumulation models successfully explain choices, reaction times, and confidence across species and tasks, mapping the decision variable onto motor planning. Sensory neurons encode momentary evidence, while motor and parietal regions show build-up signals consistent with integration. Few studies have simultaneously probed dynamics across multiple cortical areas and linked them to behaviorally inferred accumulation time courses. Feedback connections from (pre-)motor/association to sensory cortex are abundant and hierarchical decision models posit continuous feedback of the evolving decision variable, predicting mixed feedforward/feedback choice signals in sensory cortex. Monkey single-unit findings in visual cortex support decision-related feedback, but sources and dynamics across human cortex remain to be established.

Participants: Fifteen adults (8 female), normal or corrected-to-normal vision, provided informed consent (ethics: Medical Association Hamburg). Compensation: 10 €/hour. Four MEG sessions per participant (plus one training session), with structural MRI acquired separately when needed. Task: Two-alternative forced choice with confidence. Each trial presented a 400 ms reference grating at fixed contrast (0.5), followed after 1–1.5 s by a test stimulus comprising 10 consecutive 100 ms contrast samples whose values fluctuated around a trial-specific mean. Participants judged whether the mean test contrast was stronger or weaker than the reference and simultaneously reported high/low confidence via four-button responses (left/right hands mapped to choice; fingers to confidence; counterbalanced). Test offset cued response; auditory feedback followed after 0–1.5 s. The test mean was titrated to ~75% accuracy with a QUEST staircase; per-trial sample SD randomly drawn from {0.05, 0.1, 0.15}. Stimuli: Full-field expanding or contracting circular gratings (radius >10°; generative radius 12.5°, truncated at 11.3°), implemented via radial sinusoidal modulation blended to achieve desired contrast; inner annulus (1.5°) set to gray. Motion speed 4/3°/s; no within-trial direction changes. Presentation at 1920×1080, 60 Hz; luminance linearized. Data acquisition: MEG (CTF 275 axial gradiometers) at 1200 Hz in a shielded room; concurrent eye tracking (EyeLink 1000, 1000 Hz), ECG, EOG, continuous head position monitoring. Preprocessing included artifact detection (blinks, muscle, jumps, environmental), notch filtering for line noise, epoching, and downsampling to 600 Hz. Time-frequency and source analysis: Multi-taper TFRs: low frequencies 1–9 Hz (0.25 s windows, 8 Hz smoothing) and 10–150 Hz (0.1 s windows, 20 Hz smoothing). Source reconstruction via LCMV beamforming using individual 3-layer head models (FieldTrip/MNE), cortical surfaces (FreeSurfer), and an atlas-based ROI approach. Source space: 4096 vertices/hemisphere; power projected and averaged within ROIs; baseline-normalized using -250 to 0 ms pre-test interval. Regions of interest: A focused visuo-motor pathway set including retinotopic visual clusters (V1, V2–V4, V3A/B, LO1/2, MT/MST, VO1/2, PHC, IPS0/1, IPS2/3) and action-related regions (aIPS, IPS/PostCeS, M1 hand), and a cortex-wide set of 180 parcels (Glasser et al.). Behavioral analyses: Logistic regression related sample-wise contrasts to choice and to confidence; psychophysical reverse correlation computed ROC AUC per sample position within stimulus categories to yield psychophysical kernels (AUC deviation from 0.5 indicates influence of sample contrast on choice). Neural measures:

• Choice-specific lateralization: Power contra- vs. ipsilateral to the effector for stronger choices, contrasted within stimulus categories to isolate choice effects (cluster-based permutation testing).
• Stimulus-specific activity: Hemisphere-averaged power contrasted between stimulus categories within choices. Decoding approaches: Linear SVM-based classifiers using frequency-resolved power features (1–145 Hz), trained separately per stimulus category, to decode choice over time for ROIs (10-fold CV; ROC-AUC performance; Bayesian estimation of group means). A cortex-wide version extended to 180 parcels. A fine-grained approach used vertex-level phase and power with dimensionality reduction for selected prefrontal ROIs and M1. Encoding/decoding of evidence: Ridge regression decoded individual sample contrast and the running mean (accumulated contrast) from hemisphere-averaged (visual ROIs) or lateralized (action-related ROIs) power over time; performance measured by correlation between decoded and true values. A complementary linear encoding regression related frequency-specific power at 190 ms post-sample onset to sample contrast. Neural-activity kernels: ROC-AUC between band-limited power (gamma 40–75 Hz; low frequency 0–20 Hz; sampled at 190 ms post-sample onset) and choice within stimulus categories to quantify choice-predictive signals in visual ROIs. Residual kernels computed after regressing out contrast-driven power fluctuations to isolate endogenous components. Cross-area coupling: Cross-correlation between M1-hand choice decoding time course and visual low-frequency kernels recomputed at matched temporal resolution across lags -215 to +215 ms; compared against a conservative within-M1 zero-lag reference to bound signal leakage. Statistics: Parametric tests (t-tests), cluster-based permutation tests (TFCE), and Bayesian estimation of decoding time courses. Multiple comparisons controlled where applicable.

• Behavior indicates evidence accumulation across all 10 samples: choices predicted by mean test contrast (mean prediction accuracy 76%, range 61–81%); psychophysical kernels showed significant AUC>0.5 at all positions with declining influence over time (slope -0.009, t(14) = -5.3, p = 0.0001).
• Visual cortex encodes instantaneous sample contrast: V1 contrast decoding from spectral patterns peaked ~190 ms after sample onset and decreased from the first to the ninth sample; strongest encoding in gamma/high-frequency bands; regression weights significant in gamma/high-frequency only.
• Accumulated contrast is represented downstream: Decoding of running mean (accumulated contrast) was strongest in IPS/PostCeS and (pre-)motor cortex (PMd, M1-hand), mirroring choice-predictive build-up. IPS/PostCeS choice signal remained significant after regressing out M1-hand power (AUC = 0.53 at 1.1 s; t(14) = 2.01, p = 0.032), indicating partial independence.
• Choice decoding in action-related regions: M1-hand and IPS/PostCeS showed robust, frequency-specific build-up of choice-predictive activity, faster when confidence was high.
• Dissociation of stimulus-driven vs endogenous choice signals in visual cortex: • Gamma-band (40–75 Hz) V1 kernels were >0 early (samples 1–2; no baseline effect t(14) = -1.24, p = 0.24) and decayed over time (slope -0.017, t(14) = -3.1, p = 0.007), correlating with psychophysical kernel time course (r = 0.45). After removing contrast-driven fluctuations, residual gamma kernels were not different from zero, indicating a stimulus-dependent (feedforward) source. • Low-frequency (0–20 Hz, peaking near alpha ~10 Hz) V1 kernels predicted choice for later samples, showed no baseline effect (t(14) = -1.32, p = 0.21), were negatively correlated with psychophysical kernels (r = -0.23), and were unchanged after removing contrast-driven fluctuations, indicating an endogenous (feedback) source. Similar patterns extended across the visual hierarchy. • Spectral profiling showed choice-predictive activity confined to low-frequency and gamma bands; kernel magnitude increased over time in alpha but decreased in gamma. • Cross-frequency interaction: In V2–V4, a negative interaction emerged in the second half of the trial (t(14) = -2.4, p = 0.0305), suggesting low-frequency activity suppressed the impact of gamma responses on choice late in decision formation.
• Coupling between downstream choice signals and visual low-frequency activity: Cross-correlations revealed that M1-hand choice-predictive activity led visual low-frequency kernels by ~150 ms (e.g., V1 peak lag 0.15 s; V2–V4 0.15 s; other visual maps 0.12–0.20 s). Positive-lag correlations were significantly greater than zero and exceeded conservative leakage references, supporting genuine feedback-related coupling. Overall, early and intermediate visual areas primarily encoded instantaneous sensory evidence (gamma/high-frequency), while parietal and (pre-)motor regions encoded accumulated evidence and choice. An endogenous, low-frequency choice-predictive component emerged in visual cortex that lagged behind and tracked action-related build-up, consistent with feedback.

The findings map the cortical dynamics of perceptual decision formation by linking behaviorally inferred evidence weighting to MEG-derived, region-specific signals. Early visual cortex encodes rapid, sample-wise contrast fluctuations predominantly in gamma/high-frequency bands, while action-related parietal and motor regions represent the running mean (accumulated evidence) and show choice-predictive build-up consistent with motor preparatory signals. Within visual cortex, choice-predictive activity splits into two components: a stimulus-dependent feedforward component in the gamma band that mirrors psychophysical weighting and a stimulus-independent endogenous component in low frequencies (alpha-centered) that rises later and closely tracks, with ~150 ms delay, the build-up of choice-predictive activity in M1. This temporal and spectral dissociation supports hierarchical models in which feedback conveys the evolving decision variable from downstream action-related regions back to sensory cortex. The results also suggest that adaptation in sensory cortex contributes to the observed behavioral primacy (decreasing evidence sensitivity over time), complementing bounded accumulation accounts. More broadly, the work aligns with theories that feedforward and feedback information flow utilize distinct frequency channels (gamma vs. alpha/low frequency), here linked explicitly to momentary evidence encoding and decision-variable feedback during perceptual choice.

This study provides a cortex-wide, temporally resolved account of perceptual decision information in humans by integrating psychophysical reverse correlation with atlas-based MEG decoding. Key contributions include: (i) demonstration that early visual cortex predominantly encodes instantaneous evidence in gamma/high-frequency bands; (ii) identification of accumulated evidence and choice build-up in parietal and (pre-)motor regions; (iii) dissociation of stimulus-dependent (gamma) and endogenous (low-frequency/alpha) choice-predictive components in visual cortex; and (iv) evidence that endogenous visual low-frequency activity lags and tracks action-related choice signals, consistent with feedback. Future work should causally test cross-frequency and feedback mechanisms, probe laminar and subcortical contributions, generalize beyond high-contrast gratings to diverse stimuli, and refine models that incorporate both sensory adaptation and decision bounds within hierarchical feedback circuits.

• The endogenous component in V1 gamma-band residuals was not detected; linear regression assumptions (linearity of contrast-to-power and additive superposition of stimulus/endogenous components) and low MEG gamma SNR may have limited sensitivity or mismatched the true neural code.
• Choice build-up largely reflected motor preparatory formats; accumulation may occur partially outside measured cortical ROIs (e.g., striatum), potentially underdetected with MEG.
• Cross-correlation lags support feedback but do not establish direct (monosynaptic) pathways; intra-visual cortical dynamics may also contribute.
• Source reconstruction and ROI leakage are inherent concerns; conservative controls were applied but cannot eliminate all confounds.
• Sample size (N=15) and no independent replication; task/stimulus specifics (full-field gratings inducing strong gamma) may limit generalizability to other sensory evidence forms.