Graded decisions in the human brain

The study examines whether human decisions are executed via definitive, all-or-none mechanisms or through graded neural representations. Prior work shows that many brain regions track accumulated decision evidence, often modeled by drift-diffusion processes. Traditional bounded-accumulation accounts posit that choices are made when a decision variable reaches a fixed bound, leading to switch-like commitment. Alternative frameworks—such as collapsing bounds influenced by time constraints and multidimensional attractor network models—predict decision-related activity that remains graded at the time of choice. Distinguishing these models behaviorally is challenging due to overlapping predictions. Moreover, previous neural recordings distinguishing mechanisms have been limited: invasive single-unit recordings in non-human primates focus on specific regions, while non-invasive human studies capture low-frequency, spatially coarse signals. To resolve this, the authors recorded intracranial local field potentials from multiple human cortical regions during perceptual decisions to characterize neural dynamics and representations underpinning decision formation.

The paper situates its investigation within competing models of decision-making. Bounded drift-diffusion models propose that decisions terminate when neural activity reaches fixed thresholds, explaining choice and reaction-time distributions in stationary settings. However, constraints like urgency and time pressure motivate collapsing boundaries and urgency-gating variants, which relax fixed termination criteria. Multidimensional attractor network models suggest decisions emerge in high-dimensional neural state spaces across distributed circuits without requiring explicit bounds. Prior human EEG work supports graded signals but is low in spatial resolution, and primate single-neuron recordings provide precise but region-specific insights. Thus, a whole-brain, invasive human approach is needed to adjudicate whether neural decision representations at choice are graded or bound-limited.

Participants: Eight patients (5 male, 3 female; 6 right-handed; age 15–57, mean ± s.d. 39 ± 15) undergoing invasive monitoring for epilepsy contributed 13 recording sessions (7 congruent, 6 reversed; 5 subjects performed both). All had normal cognition/hearing/vision. IRB-approved with informed consent.

Task: Subjects maintained central fixation while listening to binaural trains of 0.2 ms clicks drawn from a homogeneous Poisson process, with total clicks CL + CR = 50 within up to 2 s. Two task contexts mapped stimulus evidence to effectors: in the congruent context (left-hemisphere implants), more left-ear clicks prompted a saccade (SC) to the left target; more right-ear clicks prompted a right-hand button press (BP). The reversed context flipped this mapping. Subjects could respond any time within the 2 s stimulus window; the stimulus ceased upon response. Trials with fixation breaks (>150 ms), premature responses, both movements, or no response within 2 s were excluded. Valid trials averaged 55.2 ± 18.1% of all trials. Incorrect trials were excluded from analyses. SC and BP trials were analyzed separately to avoid confounding response type and difficulty.

Behavioral/physiological monitoring: Eye position (Tobii T60, 60 Hz; onset corrected by 33 ms), right-hand joystick button presses, and forearm EMG (surface electrodes on flexor digitorum superficialis and first dorsal interosseous; 20–170 Hz band-pass, 60/120 Hz notch; Hilbert envelope) were recorded. Movement onset was defined by button press (BP) or gaze velocity criterion (2% of max; SC).

Neural data acquisition and preprocessing: Intracranial LFPs were recorded at 1200 Hz using subdural grids (7 subjects; 4 mm contacts, 6/10 mm spacing) or stereo-EEG (1 subject; 0.8 mm contacts). Electrodes with epileptiform/artifactual activity were excluded (228/799), leaving 571 electrodes; 79 identified as auditory-responsive (via a passive listening localizer) were further excluded, yielding 492 electrodes. Signals were high-pass filtered at 0.5 Hz, referenced by common average, and notch filtered at 60/120 Hz. Broadband gamma (γ) activity was extracted by 70–170 Hz band-pass and Hilbert-envelope computation, then z-scored per electrode.

Decision variable (DV): A trial-specific DV was defined as the log-likelihood ratio of the observed click sequence favoring SC vs BP: DV(t) = Σ_i log[P(e_i|SC)/P(e_i|BP)], where click-category probabilities were estimated from choice-conditional click counts. DV at choice was defined 100 ms before movement onset. DV slope was computed from 200 ms post-stimulus to choice by linear fit. Trials for graded analyses were binned into within-session terciles of DV magnitude (LO/ME/HI).

Effector-modulated electrode identification: For each electrode and effector, mean γ during a baseline window (50–300 ms after stimulus onset) was compared with an effector-related window (−200 to +50 ms around movement onset) using Spearman’s R and randomization tests with FDR correction. Electrodes significant for SC and BP were labeled SC&BP-modulated; those significant for only one effector were SC- or BP-modulated.

Analyses:

• Model-free graded test: For SC&BP-modulated, SC-modulated, and BP-modulated groups, average γ across electrodes was computed per trial near choice (mean over −100 to 0 ms relative to movement). Spearman correlation between γ and the corresponding DV at choice quantified grading; significance via randomization tests. Time-resolved correlations assessed pre/post-choice dynamics with one-sample t-tests across sessions.
• Model-based bounded null test: Constructed a bounded DV regressor that linearly ramped from 0 to ±1 from 100 ms post-stimulus to 100 ms pre-movement and clamped at ±1 in the final 100 ms (0 elsewhere). Downsampled γ to 100 Hz and regressed γ onto the modeled DV over −200 ms pre-stimulus to movement onset. The predicted γ (from regression) averaged near choice (−100 to 0 ms) was correlated with the raw DV at choice. A significant correlation rejected the fixed-bound null. A stringent circular time-shift randomization preserved autocorrelation to control overfitting.

Regional analyses: Electrodes with γ significantly positively graded by DV at choice (per effector) were localized to Brodmann areas (BA) via CT-MRI coregistration, cortical surface modeling (FreeSurfer), AC-PC alignment, and Talairach labeling (Talairach Demons). Proportions per BA were quantified. Single-trial scatter plots illustrated within-region grading.

Statistics: Across-session means reported with s.e.m./s.d. as indicated; tests included Spearman correlations, t-tests, one-way ANOVA, FDR corrections, and randomization procedures (1,000 permutations).

• Behavior validated the DV: DV diverged by choice over time; psychometric fits of choice probability vs DV at choice explained 92.4 ± 3.2% of behavioral variance across sessions (n=13). Reaction time was anti-correlated with DV slope (SC slope −7.4 ± 6.8; BP slope −10.5 ± 8.5; mean ± s.d., n=13). Eye gaze and EMG confirmed single-effector responses with large separations near choice (Cohen’s d: gaze 10.5 ± 3.3; EMG 3.0 ± 0.7; n=13).
• γ dynamics ramp during deliberation and are graded at choice: Averaged γ tracked DV time courses within decision periods (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, near choice (−100 to 0 ms), γ was significantly graded by DV: • SC&BP-modulated regions: session-average R=0.12 (SC; t(12)=3.2, p=7.3×10⁻³) and R=−0.12 (BP; t(12)=−4.1, p=1.5×10⁻³). • Effector-specific regions: R=0.12 (SC; t(12)=3.2, p=7.5×10⁻³), R=−0.13 (BP; t(12)=−3.5, p=4.7×10⁻³). Grading was detectable up to ~240–300 ms before movement onset and dropped sharply after choice, indicating cessation of accumulation and arguing against post-decision processing.
• Bounded-null regression rejected: After regressing γ on a fixed-bound modeled DV, the predicted γ remained significantly graded by the raw DV at choice: • SC&BP-modulated: R=0.14 (SC; t(12)=4.1, p=1.5×10⁻³) and R=−0.12 (BP; t(12)=−4.7, p=4.8×10⁻⁵). • Effector-specific: R=0.14 (SC; t(12)=3.6, p=3.9×10⁻³), R=−0.14 (BP; t(12)=−4.8, p=4.3×10⁻⁴). A circular-shift randomization showed no graded effect, ruling out overfitting.
• Regional contributions: Parietal cortex (BA40) showed the strongest graded effects, with additional contributions from BA8 (including frontal eye fields) for SC choices and BA6 (premotor/SMA) for BP choices. Single-trial grading was robust (examples: SC&BP-graded regions R=0.17 and −0.15; SC/BP-graded regions R=0.26 and −0.19; BA40 SC&BP-graded R=0.17 and −0.16; all Spearman p-values highly significant).
• Behavior and periphery reflect grading: Choice probabilities varied systematically with DV terciles (LO/ME/HI) for both effectors (SC: F(2,36)=27.9, p=4.8×10⁻⁷; BP: F(2,36)=43.1, p=2.8×10⁻¹⁰). Efferent measures were modestly but significantly graded by DV at choice (BP EMG: t(12)=4.0, p=0.0017; SC saccade amplitude: t(12)=3.1, p=0.010).
• Overall: Across contexts, effectors, and analyses, neural activity did not converge to a fixed bound at choice but remained graded by accumulated evidence, supporting analog decision representations.

Findings indicate that human cortical γ activity encodes developing perceptual decisions in a graded manner up to the moment of commitment, contradicting canonical fixed-bound termination. This graded representation is consistent with models incorporating urgency/collapsing bounds or attractor dynamics and aligns with flexible decision policies required in dynamic environments. The task’s context manipulation (congruent vs reversed) underscored flexibility in stimulus–response mapping while preserving evidence-based decisions.

Spatially, graded evidence signals spanned parietal, frontal, and premotor/motor cortices, with prominent effects in parietal BA40 and effector-relevant premotor/FEF regions, supporting distributed, embodied decision formation in circuits that also plan and execute actions. Multiple controls argue against sensory or motor confounds: exclusion of auditory-responsive electrodes, separation by effector choice, presence of grading in regions modulated by either/both effectors, early pre-movement grading (~300 ms), gradual γ ramping, and strongest effects in higher-order parietal areas.

The coupling between graded γ and graded choice probability suggests that neural grading may reflect certainty/confidence, consistent with previous reports linking accumulated evidence signals to confidence. Compared with prior human EEG and animal single-unit studies, intracranial γ offers a higher-fidelity, regionally resolved view demonstrating that human decisions can be implemented without fixed neural bounds at commitment.

Intracranial recordings across multiple human cortical regions provide direct evidence that decision-related neural activity remains graded by accumulated evidence up to choice, supporting an analog, flexible decision framework rather than all-or-none bound crossing. This graded representation may underlie adaptive behavior in dynamic contexts and reflects distributed, embodied computations within parietal and premotor/oculomotor circuits. Future work should disambiguate which decision-related variables (DV, urgency, confidence) are encoded, extend recordings to additional (including deep) regions and more naturalistic decision settings, and further link γ/LFP dynamics to microcircuit mechanisms.

• Variable specificity: The γ grading at choice may reflect DV, confidence, urgency, speed–accuracy tradeoffs, or correlated variables; the study cannot isolate the exact factor(s).
• Ecological validity: Laboratory sensory–motor decisions may not fully capture naturalistic decision-making; recordings in more real-world contexts are needed.
• Spatial coverage: Only a subset of recorded regions showed DV grading; unrecorded cortical and deep structures may contribute, and broader implantation would refine circuit-level conclusions.