Graded decisions in the human brain

The study asks whether human decisions are made in an all-or-none manner when neural activity reaches a fixed decision bound, or whether the underlying neural representations remain graded by accumulated evidence up to the time of choice. Prior work has shown that many brain regions track accumulated decision evidence, formalized by drift-diffusion models. However, how the accumulation process concludes remains debated: traditional models posit fixed bounds leading to switch-like choices, while alternative accounts (e.g., collapsing bounds due to time pressure and multidimensional attractor networks) allow decision-related activity to remain graded at choice. Human non-invasive studies typically access low-frequency, distributed signals, and animal single-neuron recordings have focused on specific regions, leaving a gap in understanding the full human decision process. This study directly records human intracranial local field potentials across multiple cortical regions during a perceptual decision task to test whether decision-related neural activity reaches a bound or remains graded.

The authors review evidence that many cortical areas track decision evidence (references 1–13) and the dominance of drift-diffusion models (14–16). Traditional bounded accumulation models suggest decisions occur when neural signals hit a fixed threshold (17–25). Critiques propose collapsing bounds under time constraints (28, 40) and attractor network models that do not require fixed bounds, instead evolving activity toward stable states shaped by urgency, accuracy and task demands (41–43). Human studies using EEG report time-dependent urgency and accumulated evidence signals but lack spatial resolution (6–13), while animal recordings (e.g., LIP, premotor cortex) show evidence accumulation dynamics (1–5, 54) yet are region-limited. Alternative models (25, 28, 31–33) and theories of embodied cognition suggest decisions may be formed within sensorimotor circuits (60–66). This background motivates intracranial recordings in humans to distinguish graded versus bounded neural decision representations.

Participants: Eight human subjects (5 male, 3 female; 6 right-handed, 2 left-handed; ages 15–57, mean ± s.d. 39 ± 15) with intracranial electrodes implanted for epilepsy monitoring (7 subdural grids, 1 stereotactic depth) participated (IRB-approved; informed consent). Electrodes: Platinum-iridium subdural contacts (4 mm diameter, 2.3 mm exposed; 6 or 10 mm spacing) and stereotactic contacts (0.8 mm diameter, 3.5 mm spacing). Signals were referenced to contacts distant from epileptic foci and areas of interest. Initial 799 electrodes were inspected; 228 excluded for epileptic/artifact, leaving 571; further 79 auditory-related electrodes were excluded based on a passive listening task, yielding 492 electrodes for decision analyses. Electrode localization used CT-MRI co-registration, FreeSurfer cortical models, Talairach transformations, and Talairach Demon labeling for Brodmann areas. Task and stimuli: Subjects fixated a central cross (2 visual degrees), then viewed left/right icons (15° eccentricity) while listening to binaural Poisson-distributed click trains (0.2 ms clicks; CL + CR = 50 over up to 2 s; minimum 5 ms inter-click). Two task contexts: congruent and reversed. In congruent sessions (electrodes in left hemisphere), more clicks to left ear required a left saccade (SC), more to right ear required a right-hand button press (BP). The mapping was flipped in reversed sessions. Subjects could respond any time within 2 s; stimulus ceased upon response. Eye movements were detected with a Tobii T60 eye tracker (60 Hz), onset corrected by 33 ms latency; hand onset defined by first button press. Invalid trials (fixation breaks, early button presses, dual responses, no response by 2 s) were excluded; feedback and inter-trial intervals provided. Sessions and behavior: 13 sessions total across 8 subjects (7 congruent, 6 reversed; 5 subjects performed both). Mean analyzed trials per session: SC 212 (range 84–490), BP 245 (90–434). Valid trials constituted 55.2 ± 18.1% (mean ± s.d., n = 13). Incorrect trials were excluded from main analyses. EMG from forearm muscles (flexor digitorum superficialis and first dorsal interosseous) was recorded with surface electrodes; eye gaze tracked continuously. Signal acquisition and preprocessing: Neural and EMG signals were amplified and sampled at 1200 Hz (g.USBamp/g.Hlamp). Eye tracking at 60 Hz. Signals high-pass filtered at 0.5 Hz, common average reference applied, notch filters at 60/120 Hz. Broadband gamma (γ) extracted by band-pass filtering 70–170 Hz, Hilbert transform envelope, and z-scoring per electrode. EMG processed with bipolar montage, notch at 60/120 Hz, band-pass 20–170 Hz, Hilbert envelope. Decision variable (DV): Defined as the log-likelihood ratio over discrete clicks per signal detection theory: DV(t) = log LR(t) = sum over clicks up to time t of log[P(e_i|SC)/P(e_i|BP)], with probabilities computed from click frequencies across trials for each choice. DV at choice defined as DV value 100 ms before movement onset; trials grouped into terciles of evidence (LO/ME/HI) with equal counts per choice per session. DV slope computed from a linear fit between 200 ms post-stimulus onset and choice time. Effector-modulated electrodes: For each electrode, SC and BP trials were used to compare mean γ during baseline (50–300 ms after stimulus onset) vs effector-related period (−200 to +50 ms around movement onset). Spearman’s R with condition labels (baseline = −1, effector = 1) assessed modulation; significance via randomization tests (1,000 iterations) with FDR correction. Electrodes showing significant change (increase or decrease) were defined as effector-modulated. Electrodes significant for both SC and BP were SC&BP-modulated; significant only for SC or BP were SC-modulated or BP-modulated. Analyses: Model-free: γ was averaged across SC&BP-modulated, SC-modulated, or BP-modulated electrodes. For each trial, mean γ in the 100 ms pre-movement window was correlated (Spearman’s R) with the corresponding DV at choice, significance via randomization tests. Time-resolved correlations between γ and DV at choice were computed, aligned to stimulus and movement onset. Model-based (bounded null): A modeled DV that linearly ramps to fixed bounds (SC: +1; BP: −1) from 100 ms post-stimulus to 100 ms pre-movement, with bound maintained in the final 100 ms and zero otherwise, was used to regress downsampled γ (to 100 Hz) via linear regression over −200 ms pre-stimulus to movement onset. Predicted (regressed) γ was then correlated with raw DV at choice. Overfitting control used circular time-shifts of γ across electrodes to produce null distributions. Brain region analyses identified electrodes with γ significantly graded by DV at choice (p < 0.05, one-tailed randomization tests) and summarized contributions across Brodmann areas.

Behavior and DV: DV polarity diverged over time according to choice (SC vs BP). One second after stimulus onset, DV reached 3.9 ± 2.4 for SC and −3.4 ± 1.8 for BP (mean ± s.d., n = 13). Psychometric curves relating DV at choice to choice probability explained 92.4 ± 3.2% of the variance (mean ± s.d., n = 13). Reaction time decreased with |DV slope| with slopes −7.4 ± 6.8 (SC) and −10.5 ± 8.5 (BP) (mean ± s.d., n = 13). Eye gaze and hand EMG cleanly dissociated effectors: Cohen’s d around the time of choice was 10.5 ± 3.3 for gaze and 3.0 ± 0.7 for EMG (mean ± s.d., n = 13). Gamma dynamics and grading: Averaged γ activity ramped gradually during deliberation and correlated with DV time course 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, γ at the time of choice was strongly graded by DV, contrary to a fixed bound prediction. In SC&BP-modulated regions, session-average Spearman’s R between γ and DV at choice was 0.12 (SC; t(12) = 3.2, p = 7.3 × 10⁻³) and −0.12 (BP; t(12) = −4.1, p = 1.5 × 10⁻³). In SC/BP-modulated regions, R = 0.12 (SC; t(12) = 3.2, p = 7.5 × 10⁻³) and R = −0.13 (BP; t(12) = −3.5, p = 4.7 × 10⁻³). The correlation was significant up to 300 ms pre-movement for BP and 240 ms for SC in SC&BP-modulated electrodes, and up to 300 ms (BP) and 40 ms (SC) in effector-specific sets. Correlations dropped sharply immediately after choice, indicating cessation of accumulation and arguing against post-decision processing. Bounded null regression: Using a modeled DV that reaches fixed bounds at choice, regressed γ still showed strong grading by raw DV at choice (SC&BP-modulated: R = 0.14 for SC, t(12) = 4.1, p = 1.5 × 10⁻³; R = −0.12 for BP, t(12) = −4.7, p = 4.8 × 10⁻⁴). In SC/BP-modulated regions: R = 0.14 (SC; t(12) = 3.6, p = 3.9 × 10⁻³) and R = −0.14 (BP; t(12) = −4.8, p = 4.3 × 10⁻⁴). Circular-shift randomization controls showed no graded effect, ruling out overfitting. Regional contributions: Brodmann area 40 (parietal cortex) contributed most prominently to grading during both SC and BP choices. For SC choices, BA40 and BA8 (including frontal eye fields) were most prominent; for BP choices, BA40 and BA6 (premotor/supplementary motor areas) were most prominent. Single-trial correlations in graded regions were robust (examples: SC&BP-graded regions: R = 0.17, p = 4.3 × 10⁻³ for SC; R = −0.15, p = 7.8 × 10⁻¹² for BP. SC/BP-graded regions: R = 0.26, p = 3.3 × 10⁻²⁸ for SC; R = −0.19, p = 2.0 × 10⁻²¹ for BP. BA40 SC&BP-graded: R = 0.17, p = 2.9 × 10⁻⁴ for SC; R = −0.16, p = 1.3 × 10⁻³ for BP). Behavioral probability and efferent effects: Choice probability was significantly modulated by evidence level (LO/ME/HI) for both SC and BP (one-way ANOVA: F(2,36) = 27.9, p = 4.8 × 10⁻⁸ for SC; F(2,36) = 43.1, p = 2.8 × 10⁻¹⁰ for BP). Peripheral effector signals were modestly but significantly graded by DV at choice (hand EMG: t(12) = 4.0, p = 0.0017; saccade amplitude: t(12) = 3.1, p = 0.010).

Direct intracranial LFP recordings in humans show that broadband gamma activity, a surrogate of local multi-unit discharges, encodes developing perceptual decisions with gradual ramps and remains graded by accumulated evidence at the moment of choice. This finding challenges fixed-bound implementations of the drift-diffusion model and supports alternative graded frameworks (e.g., collapsing bounds, attractor dynamics) that flexibly integrate urgency, constraints, and mapping demands. The graded nature bridges neural activity and behavior: γ grading parallels graded choice probabilities and modest efferent signal grading. Decision evidence was represented across parietal, frontal, premotor, and motor areas, with parietal BA40 showing the strongest effects, consistent with distributed, embodied decision formation in sensorimotor circuits rather than a central, effector-independent decision module. Control analyses rule out sensory confounds (reversed mapping, exclusion of auditory-responsive electrodes) and motor confounds (effects observed when sorting by fixed effector, presence in SC&BP-modulated regions, ramping dynamics, pre-movement timing, and strongest effects in parietal cortex). The immediate drop in γ–DV correlation after choice indicates the accumulation process ends at choice without continued post-decision processing. Together, the results provide convergent neural evidence that human decisions can be represented in an analog, graded manner.

The study demonstrates that intracranial broadband gamma activity in humans ramps with deliberation and is graded by accumulated decision evidence up to the time of choice, contradicting a definitive bound at choice. The strongest graded signals arise in parietal cortex and effector-specific premotor/frontal regions, supporting distributed, embodied decision-making. These analog neural representations may underlie flexible human choice behavior. Future work should clarify which decision-related variables (e.g., DV, confidence, urgency) are specifically encoded by γ activity, extend recordings to more and deeper brain regions, and test naturalistic decision-making contexts to assess generalization beyond laboratory settings and further explore potential bound-related mechanisms elsewhere in the brain.

Three main limitations: (1) The study cannot disentangle which precise decision-related factors the DV-graded gamma activity encodes (e.g., the DV itself, confidence, urgency, reward expectation, speed-accuracy trade-offs), so graded effects should be interpreted broadly as decision-related. (2) While lab-based sensorimotor decisions are argued to reflect real-life deliberation, direct neural recordings in naturalistic decision scenarios are needed to demonstrate generalization. (3) Only a subset of recorded regions exhibited DV-related grading at choice; other brain areas (including deep structures) may implement evidence accumulation toward a bound. Wider implantation across regions would help assess these possibilities.