Electrophysiological population dynamics reveal context dependencies during decision making in human frontal cortex

A large animal literature has established that neuronal firing rates in multiple brain areas encode subjective value and that these signals adapt to context, including spatial and temporal history of rewards. While human fMRI robustly identifies a valuation network, it has provided limited and sometimes inconsistent evidence in orbitofrontal and parietal cortices and has offered limited insight into context dependency during decision-making due to the BOLD signal’s spatial/temporal constraints. It remains unclear in humans whether OFC activity exhibits temporal context-dependency during valuation as seen in macaques, where recent reward history shapes current value signals. This study asks whether human intracranial electrophysiological signals in reward-related areas exhibit temporal context dependency at the time of valuation, whether such signals can be localized at the single-contact level within individuals, and how these signals are spatially distributed across OFC subregions and other structures.

Electrophysiology in macaques and rodents has mapped reward-related value representations across OFC, ACC, striatum, and parietal cortex, showing context-dependent (adaptive) coding: spatial effects in parietal cortex and temporal effects in OFC/ACC. Behavioral work in humans and animals highlights that context shapes choices, and computational frameworks including range adaptation and divisive normalization formalize such effects. Human fMRI meta-analyses implicate vmPFC and ventral striatum in subjective value, but OFC value signals are inconsistently detected, possibly due to susceptibility artifacts and limited temporal resolution. Prior human intracranial studies demonstrated OFC high-gamma value correlates and broader valuation properties in intracranial signals, but direct human electrophysiological evidence for temporal context dependency at valuation time has been lacking. Models of efficient coding of subjective value predict context-normalized representations consistent with animal electrophysiology and behavioral context effects observed in humans.

Participants: Twenty patients (9 males; 16–51 years; mean 29.2) with drug-resistant epilepsy undergoing stereo-EEG (sEEG) monitoring participated; implantation locations were clinically determined. For behavioral comparison, 35 healthy controls (18 males; 20–27 years; mean 22) performed the same task. Informed consent and IRB approvals were obtained. Task: Subjects performed a Becker-DeGroot-Marschak (BDM) auction to elicit willingness-to-pay (WTP) for 100 snack items, each presented twice (8 blocks × 25 trials). Trial structure: fixation (1 s), stimulus (snack image) for evaluation until a readiness click, a 1 s fixation, then a price matrix (0–200 TWD in steps of 10) for response. Inter-trial interval was randomized (1, 1.5, 2 s). One randomly selected trial was realized at the end per standard BDM rules with an endowment of 200 TWD. Behavioral exclusions included RT outliers (>3 SD). Electrophysiology: Depth electrodes (AdTech) with 6–10 contacts (2.29–2.41 mm contact length; 4–8 mm spacing) recorded at 2048 Hz with 878 Hz low-pass filter, referenced to scalp PFz or an intracranial white matter contact. Preprocessing included band-pass (0.5–250 Hz), 60 Hz notch, rereferencing (bipolar for sEEG), PCA-ICA-based ocular artifact removal, epoching from −1.5 to +2 s (baseline −1.5 to 0 s), interictal spike rejection, and valid-trial selection. Time-frequency analysis: Wavelet transform estimated spectral power 4–200 Hz; trials re-epoched from −1 to +1.5 s (10 ms resolution) with baseline subtraction. Band-specific power time series were extracted for high-gamma (80–150 Hz), gamma (30–80 Hz), beta (13–30 Hz), alpha (8–12 Hz), and theta (4–7 Hz). Electrode localization: CT-MRI coregistration and normalization to MNI space (SPM12); OFC contacts identified using Harvard-Oxford atlas probabilities (>1%) and excluding insula-border contacts with higher insula probability. Statistical analyses: For each contact and time point, General Linear Models (GLMs) regressed band power on current-trial subjective value (SV) and previous-trial SV (GLM-1). Robustness GLMs: GLM-2 (stepwise regressions to address collinearity), GLM-3 (exclude zero bids), GLM-4 (add RT regressor), GLM-5 (add SV two trials back), GLM-6 (add previous-trial power to address autocorrelation), and GLM-1 stratified by ITI. Significance was assessed using permutation tests with threshold-free cluster enhancement (TFCE) and familywise error correction across time points (group-level across contacts and individual-contact levels). Group-level time-frequency maps used TFCE across time-frequency points. State-space analyses: Population activity across 166 OFC contacts was denoised and reduced using PCA; trials were split by subject-specific medians of current and previous SV (2×2 conditions). Trajectories were plotted on the first two PCs. A regression subspace analysis used PCA-denoised data and orthogonalized regression vectors (current and previous SV) to define value-context axes and project condition-wise trajectories. Brain regions analyzed: OFC overall and subregions (medial/central/lateral; anterior/posterior), amygdala, hippocampus, striatum, insula, ACC/MCC, PCC, and IPS.

Behavior: • WTP distributions were positively skewed; 23% zero bids. • Current WTP positively depended on previous-trial WTP despite random item order, indicating temporal context dependency. At the subject level, 10/20 patients showed significant previous-SV effects; group mean regression coefficient was >0 (t=4.77, p<0.001). The two-trials-back bid had no significant effect (t=0.41, p=0.341). In healthy controls, 25/35 showed the same effect (t=3.81, p<0.001). • No reliable relationship between WTP and RT at the group level (t=1.60, p=0.063). OFC high-gamma (80–150 Hz): • Across 166 OFC contacts, high-gamma power positively correlated with current SV and negatively with previous SV (permutation TFCE, FWE-corrected). • Approximately 29% of OFC contacts showed significant SV representations; significant contacts clustered in the quadrant indicating positive current SV encoding and negative previous SV encoding. Among significant contacts: 52% encoded only current SV, 31% only previous SV, 17% both. • Results were robust across GLM variants: after accounting for collinearity (GLM-2), excluding zero bids (GLM-3), adding RT (GLM-4), adding two-trials-back SV (GLM-5; no significant effect), adding previous-trial power (GLM-6), and separately by ITI (1, 1.5, 2 s). Gamma (30–80 Hz) showed similar patterns. OFC subregions: • Medial (area 14) and central (areas 11/13) OFC: positive current SV and negative previous SV at the group level; lateral OFC showed robust current SV but not previous SV at the group level. • Along the anterior–posterior axis (y=35 boundary), both anterior and posterior OFC showed positive current SV and negative previous SV at the group level; posterior OFC single-contact patterns resembled overall OFC results. Cross-frequency in OFC: • Time-frequency maps revealed high-gamma/gamma positively encode current SV and negatively encode previous SV post-stimulus. In contrast, low frequencies (beta, alpha, theta) negatively encoded current SV and positively encoded previous SV, including pre-stimulus bias in alpha associated with lower bid variability and stronger behavioral dependency on previous bids. Other regions: • Subcortical: Amygdala encoded current SV at the group level (with many non-significant single contacts), hippocampus encoded both current SV and previous SV at group and single-contact levels; striatum showed no significant group-level effects. • Cortical: Insula encoded both current and previous SV at group and single-contact levels (~27–30% significant contacts), whereas ACC/MCC and PCC encoded previous SV at the group level; IPS showed no significant group-level effects. Population dynamics: • PCA on OFC population activity showed four distinct trajectories (high/low current × high/low previous SV) diverging from a common pre-stimulus origin. PC-1 predominantly captured previous SV (context), and PC-2 captured current SV, with early separations after stimulus onset. • Regression subspace analysis confirmed orthogonal axes for current and previous SV, yielding similar trajectories and showing context as a major variance component.

The study demonstrates that human OFC electrophysiological activity encodes both current subjective value and recent reward context, mirroring animal electrophysiology and linking behaviorally observed context effects in humans to neural mechanisms. High-gamma and gamma activity provide robust correlates of value and context, while low-frequency bands show inverse encoding directions, consistent with theories of inhibitory control and attentional modulation (e.g., alpha power reflecting inhibition of previous-value information and attentional engagement for current valuation). The patchy distribution of significant contacts (~30% per region) and the observation that most contacts encode either value or context but rarely both suggest parallel, separable population-level representations. This aligns with divisive normalization and range adaptation models, although linear analyses cannot distinguish subtractive from divisive computations. The population-level state-space analyses highlight that context may explain more variance than current value during valuation, offering a dynamical systems perspective on human OFC akin to findings in non-human primates. Together, these results reconcile discrepancies between human fMRI and animal electrophysiology by leveraging intracranial signals, and extend prior human intracranial studies from region-level to single-contact mapping across multiple brain structures.

Human OFC and associated regions simultaneously represent current subjective value and temporal context in separable, patchily distributed electrophysiological populations. High-frequency activity encodes current value positively and previous value negatively, whereas low frequencies show reversed patterns, including pre-stimulus biases. Beyond OFC, hippocampus and insula robustly encode both current and previous subjective values. Population state-space dynamics reveal orthogonal value and context axes, with context explaining substantial variance. These findings bridge animal electrophysiology and human neurobiology, supporting models of adaptive coding of value. Future work should: compare elicitation methods (BDM, choice, ratings) within-subject to assess method-specific neural signatures; develop paradigms minimizing motor confounds (e.g., randomized response mappings); expand coverage and apply source modeling to mitigate sparse sEEG sampling; apply multivariate decoding in fMRI to detect intermixed positive/negative encoding; and examine free-choice, multi-option tasks to test attentional modulation and alpha-band mechanisms in valuation.

Key limitations include: (1) Potential motor-related confounds: despite separating evaluation from response, subjects could prepare movements during evaluation given a fixed response matrix; future tasks should randomize response mappings. (2) Sparse and heterogeneous electrode coverage dictated by clinical needs limits generalizability and interpretation of null effects; some OFC subregions may be undersampled. (3) Signal autocorrelation could contribute to previous-value effects; multiple controls (adding previous-trial power, ITI stratification) suggest robustness, but residual confounds cannot be entirely excluded. (4) GLM linearity means divisive normalization may appear subtractive; the true computational form remains undetermined. (5) Limited depth on longer temporal histories: two-trials-back effects were non-significant, potentially due to power constraints; the extent of temporal integration beyond one trial remains open. (6) Task specificity: results obtained with BDM on snack items may not fully generalize to other reward modalities or elicitation methods.