Graded decisions in the human brain

The study addresses whether human decisions terminate via an all-or-none bound-crossing process or are represented in a graded manner. Classical drift-diffusion models propose that choices are committed when neural activity reaches fixed bounds, predicting definite, switch-like termination. Alternative frameworks, including collapsing bounds and multidimensional attractor network models, allow decision-related activity to remain graded at choice time. Prior work in non-human primates and human EEG suggests evidence accumulation and urgency effects, but non-invasive human recordings lack spatial specificity and animal single-unit studies are region-limited. The authors aim to resolve how the human brain represents the culmination of deliberation by recording intracranial local field potentials across multiple cortical regions during a perceptual decision task, testing whether neural activity reaches a bound or remains graded at the moment of choice.

The paper situates the work within a large literature on decision-making that demonstrates evidence accumulation across brain regions and formalizes behavior with drift-diffusion and sequential sampling models. Traditional bounded models (fixed bounds) have been challenged by evidence for time constraints and urgency, modeled as collapsing bounds, and by attractor network accounts that do not require bounds and predict graded activity at choice. Human EEG studies have observed low-frequency, spatially broad signatures linked to accumulation, urgency, and prior information, while non-human primate single-unit studies provided detailed but region-specific insights. However, a decisive neural test in humans with high spatial and temporal resolution, spanning multiple regions, has been lacking. This gap motivates the present intracranial LFP study focusing on broadband gamma activity as a proxy for local neuronal population spiking.

Participants and sessions: Eight human epilepsy patients (5 males, 3 females; 6 right-handed; age 15–57, mean ± s.d. 39 ± 15) with clinical intracranial electrodes performed a perceptual decision task across 13 sessions (7 congruent, 6 reversed; 5 subjects performed both contexts). Valid trials comprised 55.2 ± 18.1% of all trials; incorrect trials were excluded. Task design: Subjects fixated centrally and listened to binaural Poisson click trains (0.2 ms clicks; stereo sequences summing to 50 clicks over up to 2 s; minimum 5 ms inter-click). They decided whether more clicks occurred in the ear contralateral or ipsilateral to the recording hemisphere and indicated choice by either a saccade (SC) or a two-button press (BP) on a joystick. In the congruent context, contralateral-ear evidence mapped to BP and ipsilateral-ear evidence to SC; mapping was reversed in the reversed context. Subjects responded at will during the stimulus; the stimulus stopped upon response. Eye position (Tobii T60, 60 Hz), hand EMG (surface electrodes over forearm), and behavior were recorded and synchronized. Neural recordings and preprocessing: Intracranial LFPs (subdural grids in 7 subjects; stereo-EEG in 1) were acquired at 1200 Hz. Electrodes near epileptic foci or with artifacts were excluded (228/799), leaving 571 electrodes; auditory-responsive electrodes identified via a separate passive listening task were further removed (79), yielding 492 electrodes for analysis. Signals were high-pass filtered at 0.5 Hz, common average referenced, and notch filtered at 60/120 Hz. Broadband gamma (γ) activity was extracted by band-pass filtering 70–170 Hz, computing the Hilbert envelope, and z-scoring per electrode. Effector-modulated electrodes: For each electrode and effector (SC, BP), γ activity during a baseline window (50–300 ms post-stimulus) was compared to an effector-related window (−200 to +50 ms around movement onset) using Spearman correlation with condition labels and randomization testing (FDR-corrected). Electrodes with significant modulation (increase or decrease) were labeled SC-modulated, BP-modulated, or SC&BP-modulated. Approximately 72 ± 24% (SC) and 73 ± 18% (BP) of effector-modulated electrodes showed increased γ at response. Decision variable (DV): A signal-detection-theoretic DV was computed as the cumulative log-likelihood ratio of left vs right clicks given the eventual choice, updated with each click. DV at time of choice was defined as the value 100 ms before movement onset. DV slope was computed by linear fit from 200 ms post-stimulus to choice. Analyses:

• Behavior: DV time courses, psychometric choice functions relating DV at choice to choice probability, and chronometric analyses relating RT to DV slope.
• Model-free neural analysis: For SC and BP trials separately, averaged γ across selected electrode groups (SC&BP-modulated; SC- or BP-modulated) was correlated (Spearman’s R) with the trial-wise DV at choice using a 100 ms pre-movement window. Significance was assessed via randomization tests and across-session t-tests. Time-resolved correlations examined when γ–DV coupling emerged relative to movement onset.
• Model-based test of bounded hypothesis: A null modeled DV that linearly ramps to a fixed bound (value ±1) at choice (with 100 ms pre-movement plateau) was constructed. γ signals (downsampled to 100 Hz) from effector-modulated electrodes were linearly regressed onto the modeled DV in a window from −200 ms pre-stimulus to movement onset to obtain predicted γ (regressed γ). If a fixed-bound hypothesis were correct, residual grading by the raw DV at choice should be absent. The mean regressed γ in the 100 ms pre-movement window was correlated with the raw DV at choice. A stringent control randomized the temporal alignment by circularly shifting γ to generate null distributions. Regional analyses: Electrodes graded by DV at choice were localized to Brodmann areas (BA) via CT–MRI coregistration, cortical surface reconstruction (FreeSurfer), and Talairach-based labeling (Talairach Demon). Proportions and contributions of SC&BP-graded and effector-specific graded electrodes were quantified across BA40 (parietal), BA8 (FEF), BA6 (premotor/SMA), and other areas. Single-trial scatter analyses illustrated graded effects. Peripheral measures: EMG amplitude (hand) and saccade amplitude were tested for DV grading at choice. Statistics: Spearman’s correlations per session with randomization tests; across-session two-tailed t-tests; one-sample t-tests for time courses; one-way ANOVAs for behavioral grading; FDR correction for multiple electrodes; Kolmogorov–Smirnov tests for null distribution normality.

• Behavior tracked the DV: DV diverged by choice over time; psychometric fits relating DV at choice to choice probability explained 92.4 ± 3.2% of variance (mean ± s.d., n=13). Reaction time was negatively related to DV slope (slopes: SC −7.4 ± 6.8; BP −10.5 ± 8.5; mean ± s.d., n=13). RTs were right-skewed and choices occurred well within 2 s.
• γ activity ramped with deliberation and was graded by evidence at choice: Averaged γ in effector-modulated regions gradually increased and tracked DV time courses during deliberation (SC: average R = 0.05, t(2319)=9.4, p=8.4×10⁻²¹; BP: average R = −0.06, t(2698)=−13.4, p=8.1×10⁻⁴⁰). Critically, immediately prior to choice, γ was strongly graded by DV (not at a bound): • In SC&BP-modulated regions, session-average Spearman R at choice: SC R=0.12 (t(12)=3.2, p=7.3×10⁻³); BP R=−0.12 (t(12)=−4.1, p=1.5×10⁻³). • In SC- or BP-modulated regions, similar grading: SC R=0.12 (t(12)=3.2, p=7.5×10⁻³); BP R=−0.13 (t(12)=−3.5, p=4.7×10⁻³). • Cross-effector grading was also present (e.g., BP-modulated regions graded at SC choice time and vice versa).
• Temporal profile: Significant γ–DV correlations emerged hundreds of ms before movement onset and dropped immediately after choice, indicating accumulation ceases post-choice. Windows of significance: SC&BP-modulated electrodes—BP: up to 300 ms before movement; SC: up to 240 ms before; SC/BP-modulated electrodes—BP: up to 300 ms; SC: up to 40 ms before movement.
• Null fixed-bound hypothesis rejected: After regressing γ onto a modeled fixed-bound DV, the regressed γ remained graded by the raw DV at choice: • SC&BP-modulated: SC R=0.14 (t(12)=4.1, p=1.5×10⁻³); BP R=−0.12 (t(12)=−4.7, p=4.8×10⁻⁴). • SC/BP-modulated: SC R=0.14 (t(12)=3.6, p=3.9×10⁻³); BP R=−0.14 (t(12)=−4.8, p=4.3×10⁻⁴). Randomization (circular shift) controls showed no graded effect, ruling out overfitting.
• Regional contributions: The parietal cortex (BA40) contributed most prominently to graded γ at choice; for SC choices, BA8 (FEF) and BA40 were most involved; for BP choices, BA6 (premotor/SMA) and BA40 dominated. Single-trial scatter plots showed robust grading (e.g., SC&BP-graded BA40: SC R=0.17, p=2.9×10⁻⁴; BP R=−0.16, p=1.3×10⁻⁴).
• Peripheral efferent measures modestly reflected grading: EMG amplitude (BP: t(12)=4.0, p=0.0017) and saccade amplitude (SC: t(12)=3.1, p=0.010) were graded by DV at choice. Overall, intracranial γ activity indicates that human decision representations are graded and do not converge to a fixed bound at the time of commitment.

The findings directly address the central question of how deliberation concludes in the human brain. Intracranial γ-band activity tracked evolving evidence and, crucially, remained graded at the moment of choice across response modalities (saccade, button press), task contexts (congruent, reversed), and analytic approaches (model-free correlations, model-based regression against a fixed-bound DV). The absence of bound-like saturation at choice, together with extended pre-movement γ–DV correlations that terminate post-choice, supports a graded, analog decision representation rather than all-or-none bound crossing. The broad cortical coverage shows that decision evidence is encoded in a distributed network, with strong contributions from parietal (BA40), premotor/supplementary motor (BA6), and frontal eye fields (BA8), aligning with theories of embodied cognition whereby decision variables can be instantiated within effector-related circuits. This distributed, analog coding likely affords flexibility required by natural, dynamic contexts (e.g., urgency, changing mappings, varying payoffs) that are not well captured by fixed-bound models. The design mitigated sensory and motor confounds by reversing stimulus–response mappings, excluding auditory-responsive electrodes, separating effector types, demonstrating pre-movement grading, and localizing strongest grading to higher-order regions (parietal). Results align with and extend EEG and animal single-unit literature by providing high-fidelity, spatially localized human evidence for graded decision dynamics.

This work provides direct human intracranial evidence that perceptual decisions are represented in a graded, analog manner rather than by reaching a fixed bound at commitment. Broadband γ activity ramps with deliberation and remains graded by accumulated evidence at choice, with prominent contributions from parietal and premotor/oculomotor regions. These results suggest a neural substrate supporting flexible choice behavior in dynamic contexts. Future directions include: (1) delineating which specific decision-related variables (e.g., DV, urgency, confidence, reward expectation) are encoded in γ activity; (2) extending recordings to more and deeper brain regions to capture additional nodes that may implement different termination policies; (3) linking intracranial γ/LFP signals to EEG and microcircuit dynamics; and (4) testing in naturalistic decision scenarios to assess ecological generalizability.

The authors note three main limitations: (1) The γ activity graded by DV at choice could reflect DV, confidence, urgency, or correlated variables; thus effects are interpreted as decision-related without isolating a single construct. (2) Generalization from controlled sensory-motor laboratory tasks to naturalistic decision-making remains to be demonstrated with neural recordings in more ecological settings. (3) Only a subset of recorded regions showed DV grading at choice; other regions, including deep structures, may support bound-like accumulation not sampled here. Additionally, empirical links between EEG, LFPs, and microcircuit dynamics are under-explored, so interpretations should be grounded in localized high-frequency discharges.