Abstract
Value-based decision–making involves multiple cortical and subcortical brain areas, but the distributed nature of neurophysiological activity underlying economic choices in the human brain remains largely unexplored. Specifically, the nature of the neurophysiological representation of reward-guided choices, as well as whether they are represented in a subset of reward-related regions or in a more distributed fashion, is unknown. Here, we hypothesize that reward choices, as well as choice-related computations (win probability, risk), are primarily represented in high-frequency neural activity reflecting local cortical processing and that they are highly distributed throughout the human brain, engaging multiple brain regions. To test these hypotheses, we used intracranial recordings from multiple areas (including orbitofrontal, lateral prefrontal, parietal, cingulate cortices as well as subcortical regions such as the hippocampus and amygdala) from neurosurgical patients of both sexes playing a decision-making game. We show that high-frequency activity (HFA; ɣ and HFA) represents both individual choice-related computations (e.g., risk, win probability) and choice information with different prevalence and regional representation. Choice-related computations are locally and unevenly present in multiple brain regions, whereas choice information is widely distributed and more prevalent and appears later across all regions examined. These results suggest brain-wide reward processing, with local HFA reflecting the coalescence of choice-related information into a final choice, and shed light on the distributed nature of neural activity underlying economic choices in the human brain.
Significance Statement
Economic decision-making depends on multiple brain areas. However, how neural activity in the human brain supports choices is not well understood, due to the difficulty of measuring human neural activity. Here, we leveraged the rare opportunity to record electrophysiological activity from several human brain regions implicated in decision-making from neurosurgical patients to study the neurophysiological basis of economic decisions. We show that neural activity supporting human economic choices under uncertainty is highly distributed across brain areas. However, different relevant calculations, such as the probability of a win or the risk of an uncertain choice, are differentially reflected in across brain regions. This study demonstrates the highly distributed, but regionally specific, nature of choices and reward computations in the human brain.
Introduction
Human decision-making involves the coordinated activity of multiple brain areas to adaptively generate choices. Abundant evidence implicates a variety of prefrontal and subcortical regions in this process, often conceptualized as a sequential transformation from abstract valuation to concrete motor plans (Rangel et al., 2008; Rushworth et al., 2009; Padoa-Schioppa, 2011). However, increasing evidence from animal models supports a distributed view of neural activity underlying reward-guided behavior, with activation across brain regions emerging simultaneously rather than sequentially during choice processes (Wallis and Miller, 2003; Kennerley et al., 2009; Steinmetz et al., 2019; Ottenheimer et al., 2022) and neural activity reflecting specific choice-related computations such as risk (O’Neill and Schultz, 2013) and reward probability (van Duuren et al., 2009; Wallis and Kennerley, 2011). In the human brain, little is known about how distributed neural activity enables adaptive choice. A wealth of functional magnetic resonance imaging (fMRI) studies demonstrating an involvement of individual brain regions in decision-making behavior (O’Doherty et al., 2001; Delgado et al., 2005; Daw et al., 2006; Kable and Glimcher, 2007; Tobler et al., 2007; Preuschoff et al., 2008) suggest discrete involvement of one or a few regions, which is in direct contrast to the growing evidence about distributedness from animal models (Steinmetz et al., 2019; Shin et al., 2021; Ottenheimer et al., 2022). Moreover, our knowledge of the neurophysiological representation of these choice-related signals in the human brain remains sparse, partially due to the difficulty of directly observing human neural activity with sufficient anatomical precision and temporal resolution (Marco-Pallares et al., 2008; Saez et al., 2018; Lopez-Persem et al., 2020). For example, the relative contribution of high-frequency activity (HFA), related to local activation and neuronal spiking, versus low-frequency oscillatory neural activity engaged in a variety of cognitive processes such as memory (Burke et al., 2013; Herweg et al., 2020; Qasim et al., 2023), spatial navigation (Watrous et al., 2013; Park et al., 2014; Qasim et al., 2021), and attention (Jensen et al., 2007; Klimesch, 2012) in reward-based decision–making is unclear. Finally, this lack of knowledge about the nature of neural representation extends to choice-related computations, both related to abstract (i.e., in goods space, e.g., win probability, risk) and to action-oriented (i.e., in action space, e.g., spatial location, motor commands) aspects of choice. Given prior intracranial evidence that HFA is implicated in choice-related processes (Marco-Pallares et al., 2008; Saez et al., 2018; Lopez-Persem et al., 2020) and the relationship between HFA and neuronal spiking known to reflect reward computations in animal models (Wallis and Miller, 2003; O’Neill and Schultz 2010; Wallis and Kennerley, 2011), we hypothesize that HFA plays a central role in neural representation of both choices under uncertainty and underlying choice-related computations.
Therefore, we set out to explore (1) the encoding of choices and choice-related computations across prefrontal, parietal, and limbic regions and (2) the nature of this encoding in neurophysiological activity. We hypothesize that the neural representation of choices and choice-related computations is distributed across multiple brain regions. Furthermore, we hypothesize the primary role of HFA, rather than low-frequency oscillations, in representing choices. Because testing these hypotheses requires an anatomically accurate, temporally precise depiction of distributed neurophysiological activity, we leveraged human intracranial electroencephalography (iEEG) recordings from neurosurgical patients, which overcome some of the limitations of noninvasive human neural methodologies such as limited signal-to-noise ratio, temporal resolution (fMRI), or anatomical accuracy (EEG; Cohen et al., 2007; Buzsáki et al., 2012). Subjects played an economic risky decision-making game during iEEG recording from multiple brain areas, including orbitofrontal, lateral prefrontal, and premotor and motor cortices, insula (Ins), and amygdala (Amy). We related local field potentials (LFPs) to economic choices and choice-related computations, such as risk and win probability during the deliberation period.
Our results show that during deliberation, neural activity across frequencies is modulated in different subcircuits (limbic, prefrontal, and frontoparietal). However, choice information (i.e., safe bet or risky choice) and individual choice-related computations (e.g., risk, win probability) were represented primarily in higher-frequency activity (mostly ɣ and HFA). Choice encoding was more widespread and appeared later than choice-related computations, which were also more regionally specific suggesting the coalescence of region-specific choice–related information into a widely represented final decision. Overall, these results support the notion that decision-making processes engage widespread power modulations, with highly distributed ɣ/HFA playing a prominent role in encoding choice-related computations (i.e., risk, win probability) and choices, and demonstrate the distributed nature of neural activity supporting reward-guided decisions in the human brain.
Materials and Methods
Subjects
Data were collected from 34 patients (19 female) with intractable epilepsy who were implanted with chronic subdural grid or strip electrodes (electrocorticography, ECoG) or stereotactic EEG (sEEG) electrodes as part of a procedure to localize the epileptogenic focus. Electrode placement was based solely on the clinical needs of each subject. Data were recorded postoperatively in epilepsy monitoring units at five hospitals: the University of California (UC), San Francisco Hospital (n = 3), the Stanford School of Medicine (n = 3), UC Irvine Medical Center (n = 23), and UC Davis Medical Center (n = 5). As part of the clinical observation procedure, subjects were off antiepileptic medication during these experiments. All participants gave written informed consent to participate in the study in accordance with the University of California, Davis, or University of California, Berkeley Institutional Review Board. Subjects understood that they could decline participation at any time, and verbal assent was reaffirmed prior to each experimental task.
Electrophysiological data acquisition
ECoG and sEEG activities were recorded, deidentified, and stored at the same time as behavioral datasets. We collected data using Tucker-Davis Technologies, Nihon Kohden, or Natus System. Data processing was identical across all sites: channels were amplified 10,000 times, analog filtered (0.01–1,000 Hz) with >1 kHz digitization rate, re-referenced to a common average off-line, and high-pass filtered at 1.0 Hz with a symmetrical (phase true) finite impulse response filter (∼35 dB/octave roll-off). Behavioral data were simultaneously collected using a PC laptop-running Python (v.2.7) and PsychoPy (v.1.85.2) and synchronized with a timed visual stimulus (trial start) recorded by a photodiode through an analog input to the electrophysiological system. We excluded four patients because of incomplete or missing data, either behavioral (n = 3) or electrophysiological (n = 1), leaving n = 30 patients for further analysis.
Behavioral task
We probed risk–reward trade-offs using a simple gambling task described previously (Saez et al., 2018). Briefly, subjects chose between a safe bet ($10 USD, fixed) or a gamble for potential higher winnings (between $15 and $30); all subsequent amounts are in USD. Gamble win probability varied per trial based on an integer between 0 and 10 shown at game presentation randomly generated using a uniform distribution. Subsequently (t = 550 ms postchoice), a second number (also 0–10) is revealed. The gamble results in a win if the second number is greater than the first. Therefore, a “2” had a win probability of 80%, and an “8” had a win probability of 20%; ties were not allowed. Location of a safe bet and gamble options (left/right) were randomized across trials. Subjects played 10 practice trials, repeated as many times as necessary, to ensure they had full knowledge of the (fair) structure of the task prior to game play (200 trials). Trials started with a fixation cross (t = 0), followed by a game presentation screen (t = 750 ms). Subjects had up to 8 s to respond (mean reaction time, 1.4 s). Gamble outcome presentation appeared 550 ms after button press (choice) on each trial regardless of choice. A new round started 1 s after outcome reveal. The experimental task lasted 12–15 min. This gambling task minimized other cognitive demands (working memory, learning, etc.) on our participants while allowing us to test for decision-making under risk.
Behavioral performance was assessed by examining the proportion of trials in which the subject chose to gamble as a function of win probability; the proportion of gamble trials was calculated for each win probability value (0–100% in 10% increments) and fit with a logistic curve (Fig. 1B; Extended Data Fig. 1-1). Behavioral fits were similar using expected utility, suggesting that the patients were largely insensitive to gamble magnitude and behavior was driven primarily by sensitivity to reward probability. Results from a subset of these subjects playing the same task were published previously (Saez et al., 2018). As a control for behavioral data quality, we excluded subjects in which a logistic function did not appropriately fit the relationship between the percentage of gambles and win probability (p < 0.05, logistic fit; Extended Data Fig. 1-2); 10 subjects did not show a significant fit and were removed from further analysis, leaving 20 subjects with behavioral data of sufficient quality. The high exclusion rate largely reflects the demands of data collection with subjects in a clinical environment as well as the importance of stringent behavioral criteria.
Experimental design and analytical strategy. A, Subjects played a gambling game in which they chose between a safe $10 reward and a risky gamble for a higher reward. Gamble win probability varied parametrically on a trial-by-trial basis (0–100%). B, Average subject behavior (yellow line, shading indicates SEM), shown as the proportion of risky choices (gambles) for each win probability, averaged in 10% increments. Subjects gambled more often as the win probability increased. Red line represents a sigmoidal fit to the average behavior. See Extended Data Figures 1-1 and 1-2 for individual patient behavior; see Extended Data Figure 1-3 and Extended Data Table 1-1 for reaction time results. C, Anatomical iEEG coverage. We collected data from subjects that underwent either ECoG (n = 9) or sEEG (n = 11) during epilepsy surgery (n = 20 subjects and 1,042 electrodes total). We focused our analyses on electrodes located in gray matter in regions involved in reward-related behavior: OFC, LPFC, PrG, CC, PoG, PC, Amy, HC, and Ins. See Extended Data Figure 1-4 and Extended Data Table 1-2 for anatomical coverage and electrode counts for individual patients. D, Analytical strategy: (1) LFPs from each gray matter electrode were cleaned and preprocessed, and (2) subsequently decomposed into power across frequency bands [δ (1−4 Hz), θ (4–8 Hz), α (8–12 Hz), β (12–30 Hz), ɣ (30–70 Hz), and HFA (70–200 Hz)]. After (3) trialing around behavioral events of interest (subject choice), we (4) determined which electrodes presented task modulation (significant power change vs baseline) or (5) choice activity (significantly different power between gamble and safe bet trials), in any frequency band, and (6) reward computations in HFA (risk, win probability).
Figure 1-1
Individual patient gambling behavior, win probability. Light grey datapoints show individual trial choices (light grey; 1 = gamble, 0 = safe bet). Black points show averaged gamble proportion for all trials in 10% win probability bins. All patients gamble more often when win probability is higher. Red line indicates sigmoidal logistic fit (all p < 0.01), indicating that patients understood the task and played to maximize their winnings. Patients were excluded if their behavioral data could not be fit with a sigmoidal curve. Download Figure 1-1, TIF file.
Figure 1-2
Individual patient gambling behavior, excluded patients. Average subject behavior for patients that did not meet behavioral quality control criteria (n = 10). X-axis represents gamble win probability, y-axis the proportion of gambles, and the lines represent the sigmoidal fit to the average behavior for each excluded patient. Unlike patients in our final dataset (Fig. 1-1), these patients’ behavior did not show sensitivity to gamble win probability. Download Figure 1-2, TIF file.
Figure 1-3
Distribution of reaction times across all patients and trials. Histogram shows distribution of reaction times over all trials as above completed for all patients (3287 trials, n = 20 patients). Download Figure 1-3, TIF file.
Figure 1-4
Anatomical coverage for individual patients. Anatomical coverage for individual patients (p01-p20). Patients p01 through p09 underwent electrocorticography (ECoG) implantation, whereas patients p10-p20 underwent stereotactic EEG (sEEG) implantation. Areas of interest with 2 + electrodes were included for all patients. See Table 1-2 for details on electrode count by area and patient. Download Figure 1-4, TIF file.
Table 1-1
Reaction times. Reaction times means and standard deviation (seconds) for each patient. All completed trials with the exception of one outlier trial (p17, rt = 193.5 secs) are included. Overall mean reaction time = 1.00 seconds, stdev = 0.017. Download Table 1-1, DOCX file.
Table 1-2
Count of electrodes per patient and region. Table shows the number of electrodes in each anatomical regions of interest for each patient in our final sample (n = 20), as well as the total count of electrodes per ROI across all patients and the number of patients with coverage for each ROI. Download Table 1-2, DOCX file.
Anatomical analyses
Electrode localization was based strictly on clinical criteria for each patient: 9 of 20 had ECoG grids, predominantly in orbitofrontal, lateral prefrontal, and parietal regions, whereas 11 of 20 had sEEG coverage, predominantly of deep temporal lobe regions (Amy, hippocampus; Extended Data Fig. 1-4). For each patient, we collected a preoperative anatomical MRI (T1) image and a postimplantation CT scan. The CT scan allows identification of individual electrodes but offers poor anatomical resolution, making it difficult to determine their anatomical location. Therefore, the CT scan was realigned to the preoperative MRI scan following a previously described procedure (Stolk et al., 2018). Briefly, both the MRI and CT images were aligned to a common coordinate system and fused with each other using a rigid body transformation. Following CT–MR coregistration, we compensated for brain shift, an inward sinking and shrinking of the brain tissue caused by the implantation surgery. A hull of the patient brain was generated using the FreeSurfer analysis suite, and each grid and strip was realigned independently onto the hull. This step was necessary to avoid localization errors of several millimeters common in ECoG subjects. Subsequently, each brain and the corresponding electrode locations were normalized to a template using a volume-based normalization technique and snapped to the cortical surface (Stolk et al., 2018). Finally, the electrode coordinates are cross-referenced with labeled anatomical atlases (Brainnetome Atlas) to obtain the gross anatomical location of the electrodes, verified by visual confirmation of electrode location based on surgical notes. We selected all gray matter electrodes across a broad set of regions known to be involved in reward-related behavior for analysis (Fig. 1C): lateral prefrontal cortex (LPFC; 391 electrodes in n = 19 subjects), orbitofrontal cortex (OFC; 193; n = 18), cingulate cortex (CC; 84; n = 13), hippocampus (HC; 65; n = 13), Amy (32; n = 11), Ins (46; n = 8), precentral gyrus (PrG; 108; n = 11), postcentral gyrus (PoG; 88; n = 9), and parietal cortex (PC; 78; n = 8; Fig. 1; Extended Data Table 1-2).
Electrophysiological analyses
Quality control and preprocessing
Epileptogenic channels and channels with excessive noise (low signal-to-noise ratio, 60 Hz line interference, electromagnetic equipment noise, amplifier saturation, poor contact with cortical surface) were identified and removed. Out of 1,194 electrodes localized to regions of interest, 1,085 were artifact-free and included in subsequent analyses. Additionally, all channels were visually inspected to exclude epochs of aberrant or noisy activity (typically <1% of data points). Data analysis was carried out using custom scripts written in MATLAB and FieldTrip toolbox (Oostenveld et al., 2010). Data for each channel were downsampled to 1 kHz. Each channel was low-pass filtered (200 Hz), high-pass filtered (1 Hz), and notch filtered (60 Hz and harmonics) to remove line noise. Electrode channels were re-referenced to a common average reference of all electrodes in each strip/grid. Even though bipolar derivations or white matter referencing are often used for sEEG electrodes, we opted to use a single (CAR) re-referencing strategy for both ECoG and sEEG electrodes for analytical consistency. Trials were epoched to the time of decision using a [−4,3] s window around events of interest (options presentation and subject choice), and the leading and trailing 1 s of data were discarded to remove edge effects. Time–frequency representations [discrete prolate spheroidal sequences (DPSS) taper method] were plotted for each region and subject (averaged across electrodes and trials) and visually inspected for artifacts.
Time–frequency representation of neural activity (bandpass estimates)
To examine the role of individual oscillatory bands, we decomposed the neural activity into canonical, discrete activity bands: [δ (1–4 Hz); θ (4–8 Hz); α (8–12 Hz); β (12–30 Hz); ɣ (30–70 Hz); HFA (70–200 Hz)] for each gray matter electrode for each subject using the Filter–Hilbert method. Power in the six bands was calculated by applying a Butterworth bandpass filter (order 3 for δ and order 4 for all other power bands) and Hilbert transform and multiplying the resultant complex signal by its complex conjugate (Cohen, 2014). As before, 1 s is removed from the beginning and end of the data to reduce edge effects.
Task-active electrodes
To examine patterns of electrodes that showed significant task-related power modulation, we compared average power estimates during deliberation period (−1 to 0 s prechoice) with baseline power estimates (−0.5 to 0 s prestimulus onset) for each electrode and each frequency band, independently (paired t test across trials, α = 0.01). Data were analyzed and plotted using custom scripts in MATLAB and R. To investigate patterns across subjects, regions, and powerbands, we summarized and plotted the proportion of task-active electrodes for each subject and then calculated the mean proportion of task-active electrodes across subjects for each power band and region. Mean subject proportions and standard errors provide a depiction of population activity across subjects that would be obscured in the aggregate. Finally, in order to probe the apparent homogeneity of the results, we plotted proportions of task-active electrodes that increased or decreased power per power band and regions which revealed patterns that were quantified with clustering methods described below.
Region clustering
We applied an unsupervised clustering algorithm (k-nearest neighbor) to evaluate the degree of similarity in power modulations across regions. We started by parameterizing activity in each brain region by estimating the proportion of electrodes that showed increases or decreases in power modulation during the deliberation period, separately for each subject, region, and frequency band (see above, Task-active electrodes). Therefore, each subject–region combination (104 total data points) is initially represented by a single point in 12D space (increase or decrease in power × six frequency bands). To reduce noise or redundancy, we performed dimensionality reduction using Uniform Manifold Approximation and Projection (UMAP; McInnes et al., 2020). UMAP converts the data into a k-neighbor graph and then identifies a projection into a lower-dimensional space by minimizing the cross-entropy between the two representations while maintaining the fundamental characteristics of the original graph. Once the data were projected into a lower dimension, we applied the k-nearest neighbor algorithm to sort the data into a specified number of clusters. In order to ensure the clustering algorithm was meaningfully separating regions into super region clusters, we bootstrapped our data by repeatedly (1,000 times) randomly shuffling the labels on each region and rerunning the clustering algorithm to test the likelihood that regions would be randomly separated into super region clusters. We used that distribution to evaluate the significance of each cluster. This method effectively separated the regions into three hypothesized functional groups (bootstrapped p < 0.05; Extended Data Table 4-1).
Choice-active electrodes
We defined choice-active electrodes as those that showed a significant period of activation during deliberation (−1 to 0 s prechoice) between trials where a gamble bet was chosen compared with a safe bet. This was determined by a permutation test. Trials were separated by subsequent choice (gamble or safe bet), and then a two-sample t test was applied to every time point (1 ms resolution) in the 1 s deliberation epoch. MATLAB’s bwconnect function was used to find contiguous suprathreshold clusters of points (α < 0.05). An electrode was categorized as choice-active by comparing the sum of the t statistic for the largest cluster to a trial-shuffled null distribution at α = 0.001 (two-tailed). To determine the null distribution, we shuffled the trial labels 10,000 times, and the sum of the t statistic of the largest supratheshold cluster was recorded on each iteration. This was repeated for each electrode, for each power band, and for each subject.
Electrode-wise regression analysis
In order to identify which behavioral and choice-related computations were reflected in HFA (60–200 Hz) in each of the nine regions of interest, we performed a stepwise regression analysis on an electrode-by-electrode basis. Data from 14 subjects were included; 6/20 subjects had low trial number counts and were excluded from this analysis. As described for other analysis, power-time vectors for each trial were smoothed using a rolling window average (200 ms window width, 50 ms step increments). The resulting trials × power (over time) matrix was used to carry out stepwise regression as described below.
Regressors of interest
We chose a set of five regressors related to decision/choice and action, which showed low to moderate colinearity: win probability, gamble risk, risky side index, gamble index, and button press index (Extended Data Table 7-1). Briefly, win probability represents the probability of the gamble option resulting in a win (winprob = (10 − number shown) / 10); gamble risk represents the level of uncertainty associated with a gamble (risk = [winprob * (1 − winprob)]; 0 if winprob is 0 or 1, and maximal if winprob = 0.5); the risky side is a Boolean regressor indicating which side corresponded to the gamble choice (right, 1); the choice side is a Boolean regressor indicating which side was chosen (right, 1); and choice represents the subject choice as above (1, gamble; 0, safe bet).
Stepwise regression strategy
As an initial step, and to reduce processing time, we ran independent linear regressions for each electrode for each regressor of interest and each time point. Regressors that were significant at this stage (p < 0.05) were passed to a sum of F statistic permutation tests (1,000 iterations; see below for details), and the resulting p value was multiple comparison corrected over the number of electrodes in our sample (n = 978). The regressors that survived permutation and multiple comparisons were then passed to a stepwise regression model. The best regressor order was selected by starting with the regressor with the highest association (R2) and then adding each possible subsequent regressor. The model that explained the most variance was chosen based on model comparison of the reduced model (reg1) to the complex model (reg1 + reg2) using an ANOVA test. If the ANOVA was significant, meaning the complex model shows a better fit than the reduced model, the steps above were repeated adding the subsequent regressor in the list for a newer complex model with one more component. If not, model fitting is concluded. Once the best regressor order was determined, the stepwise regression continued with permutation tests to select the optimal number of regressors: incrementally adding regressors in the optimal order until no more variance was captured (ANOVA p > 0.05).
Sum of F stat statistic
To determine whether activity in a given frequency band encodes a specific regressor, we employed a sum-of-F-stat strategy. Briefly, a linear regression was calculated at each time point, resulting in a regression goodness-of-fit and significance value at each time point. Next, we examined the whole time course to determine the maximum number of consecutive regressions that resulted in a significant association (p < 0.05). We then calculated a summary statistic by summing the absolute value of the F stat for each of those consecutive significant data points. This calculation has several benefits: it is agnostic to the exact timing of the activation, since it does not impose any priors or temporal constraints on when it happens; it is sensitive to sustained activations, since those will result in several adjacent time points that cross the significant threshold; and it is sensitive to the strength of the activation, since greater correlations will result in higher F stat values that add weight to the summary statistic. The resulting sum-of-F-stat therefore reflects the strength of the neural–behavioral association.
To obtain a significance estimation for the summary F stat, we generate a null sum-of-F-stat distribution by recalculating it after shuffling the behavioral labels 1,000 times and carrying out the same calculation for each permutation. The resulting permutation p value is the proportion of shuffled F stats with a value higher than the observed sum-of-F-stat. If the activation is robust, this will result in a small p value since the majority of shuffled sum-of-F-stats will be well below the observed value.
General linear model
For our simple general linear model (GLM) analysis, we included all choice-related computations (choice index, risky side index, choice side, win probability and risk) in a single GLM, without carrying out a stepwise selection strategy. As in the stepwise regression strategy, we used a permutation strategy to determine whether the association between neural data and individual regressors was significant.
Timing analyses
To analyze the timing of encoding across electrodes, we extracted the encoding peak, corresponding to the time point in our time-resolved regression results in which encoding was maximal (maximum absolute T stat; see above, Regression of interest) for each electrode and regressor. We then compared the resulting distributions across regions or regressors using Kruskal–Wallis tests. When a significant result was identified, we followed up with post hoc testing using Dunn's test, which we selected because it handles tied ranks better than pairwise Wilcoxon rank sum tests.
Results
We collected iEEG recordings from 34 subjects with medication-refractory epilepsy while playing a postoperative gambling task. After data and behavioral quality control, the final dataset comprised n = 20 patients (see Materials and Methods). Patients made trial-by-trial choices between a safe prize and a risky gamble with varying win probabilities (Fig. 1A). Patients gambled more often in trials with higher win probability, as expected in reward-maximizing behavior (Niv, 2009; Doll et al., 2012; Fig. 1B; Extended Data Figs. 1-1, 1-2). Patients underwent either ECoG, providing subdural coverage predominantly in frontoparietal regions (9/20 subjects; Fig. 1C), or sEEG, predominantly in deep temporal lobe regions (Amy, hippocampus, Ins; 11/20 subjects; Fig. 1C; Extended Data Fig. 1-4). We analyzed electrophysiological recordings from electrodes located in the gray matter of regions involved in reward-related behavior: OFC, LPFC, CC, PrG, PoG, PC, Amy, HC, and Ins (n = 1,085 electrodes total; Fig. 1C; Extended Data Fig. 1-4; Extended Data Table 1-2).
Power modulation patterns during deliberation vary across brain regions
We first sought to evaluate whether neural activity was significantly modulated during the deliberation period. Specifically, we characterized power modulations in each of six canonical neural frequency bands [δ (1–4 Hz), θ (4–8 Hz), α (8–12 Hz), β (12–30 Hz), ɣ (30–70 Hz), and HFA (70–200 Hz)] by comparing the deliberation epoch (−1 to 0 s before choice button press; see Materials and Methods, Extended Data Table 1-1, and Extended Data Fig. 1-3 for reaction times) to the baseline (−0.5 to 0 s prior to stimulus presentation) for each electrode. We identified significant modulations in 76.8% of electrodes overall (75.3% ± 0.25 of electrodes on average across subjects; p < 0.01; paired t test; Fig. 2A; Extended Data Table 2-1). These modulations were due to increases or decreases in power that occurred in one or more frequency bands. In addition, power modulations at a single electrode were often evident in multiple frequencies (2.09 ± 1.70 frequency bands per electrode; mean ± SD). The proportion of electrodes with significant power modulations was similar across frequency bands (34.9% ± 0.05; mean ± SD; Fig. 2B), suggesting broad power modulations during deliberation. However, the proportion of electrodes with significant power modulation across regions was more variable (Fig. 2C) ranging from 63.0% in Amy to 81.3% in PC. These data suggest a wide regional distribution and broad range of frequencies for task-related power modulations during the deliberation period.
Proportion of task-active electrodes showing significant power modulation during deliberation. Power was compared between deliberation (−1 s to 0 s prechoice) and baseline (fixation cross) epochs. A, Proportion of task-active electrodes per subject. Horizontal lines show mean ± SEM. See Extended Data Table 2-1 for per patient results. B,C, Mean proportion of task-active electrodes across subjects, grouped by frequency across all regions (B) and anatomical region across all frequencies (C).
Table 2-1
Proportion of task- and choice-active electrodes, per patient. Table shows the number of electrodes in each anatomical region of interest: total number and number/proportion showing significant power modulation in any frequency band during deliberation, per patient (task-active electrodes) and the number/proportion showing a significant association between power and choice (choice-active electrodes). Download Table 2-1, DOCX file.
Because individual electrodes could show either increases or decreases in power, we separately quantified the proportion of electrodes showing power modulation in either direction (Fig. 3). Of those electrodes showing significant power modulation, 50.5% of electrodes showed power increases, 55.6% showed decreases, and 29.3% of electrodes showed a combination of increases and decreases in separate frequency bands (e.g., an increase in δ accompanied by a decrease in β). There was substantial heterogeneity in the proportion of electrodes that showed power increases/decreases across regions (Fig. 3B). For example, electrodes in Amy/HC predominantly showed power increases, whereas power decreases were most common in PC/PrG/PoG (Fig. 3B). Grouping activations by frequency band instead of region shows a complementary depiction of power modulations. For example, power modulation in δ/θ consisted predominantly of power increases, whereas β/ɣ modulations consisted predominantly of power decreases (Fig. 3C). Finally, HFA was unique in that it showed bidirectional modulation in all regions, with individual electrodes within a region showing either increases or decreases (Fig. 3C).
Task-active electrodes showed regional patterns of power increases and decreases during deliberation. A, Mean proportion of task-active electrodes across subjects, grouped by both region and frequency band. The proportion of electrodes that showed significant increases or decreases relative to the baseline is shown as positive and negative bars on the vertical axis, respectively. B,C, Point plots showing the proportion of electrodes showing increases (x-axis) or decreases (y-axis) during deliberation period, separated in individual panels for each ROI (B) and frequency band (C). B, Subject-averaged proportion of electrodes showing increases (y-axis) or decreases (x-axis) separated by regions (panels) and frequency bands (points). C, Subject-averaged proportion of electrodes showing increases (y-axis) or decreases (x-axis) separated by frequency bands (panels) and regions (points).
We observed a rich pattern of region–frequency-specific task–related power modulations, with similarities among different subsets of regions. For example, PoG/PrG/PC electrodes predominantly showed power decreases across most frequency bands (δ/θ/α/β/ɣ; average decrease, 35.5%); Amy/HC showed the opposite pattern, with widespread power increases across power bands (δ/θ/α/β/ɣ/HFA; average increase, 27%); OFC/LPFC/CC showed concomitant power increases in lower frequencies (δ/θ/α) and power decreases in higher frequencies (β/ɣ/HFA). These observations suggested that the direction of power modulation (significant increases or decreases in power) depended on both frequency and region. We identified three subcircuits whose power modulations were more similar within than between each subcircuit, namely, prefrontal (OFC/LPFC/CC), frontoparietal (PrG/PoG/PC), and limbic (Amy/HC/Ins), and carried out an unsupervised clustering analysis to test this hypothesis. Briefly, power modulations were parameterized by the proportion of electrodes that showed increases or decreases separately for each subject and region in each frequency band resulting in a 12D representation (increase or decrease in power × six frequency bands) for each subject and region (see Materials and Methods). A UMAP approach was used to reduce the dimension of the data, and the k-nearest neighbor (k = 3) clustering algorithm was applied to test our hypothesis that the patterns of power modulations we observed did indeed represent three distinct patterns supported by the three subcircuits. To ensure the clustering algorithm was meaningfully separating each region into functional clusters, we carried out a permutation analysis (see Materials and Methods; comparison to bootstrapped clustering with p < 0.05; Fig. 4; Extended Data Table 4-1). Thus, we validated three functional clusters, corresponding to prefrontal (OFC/LPFC/CC), frontoparietal (PrG/PoG/PC), and limbic (Amy/HC/Ins) circuits originally observed (Fig. 4). These results support the notion that patterns of power modulation are distinct across these proposed subcircuits, with prefrontal, parietal, and limbic circuits showing specific and distinct patterns of power modulations during the task.
Power modulation profiles define functional subcircuits. Clustering of regions into functional groups: limbic (Amy, HC, Ins), prefrontal (OFC, LPFC, CC), and frontoparietal (PrG, PoG, PC). Clustered functional groups significantly above chance (bootstrapped p < 0.05); see Extended Data Table 4-1 for results.
Table 4-1
Clustering task-active electrodes. Alignment results of k-nearest neighbor clustering UMAP projections with the three clusters (columns) with the three proposed groups of regions (rows): 46.8% of the points in cluster one belonged to frontoparietal regions (PrG, PoG, PC), 45.8% of cluster 2 datapoints belonged to limbic regions (Amy, HC, Ins), and 53.6% of cluster 3 datapoints belonged to prefrontal regions (OFC, LPFC, CC). Download Table 4-1, DOCX file.
Distributed high-frequency neural activity is related to choice behavior
Next, we sought to assess whether power modulations encoded choice information. To do this, we examined whether power in each frequency band showed significant differences between safe bet and gamble trials during deliberation, separately for each electrode (choice-active electrodes; see Materials and Methods). Overall, 42.7 ± 8.9% of electrodes were choice-active in one or multiple frequency bands (Fig. 5; Extended Data Table 5-1). There was a higher proportion of choice-active electrodes in higher frequencies (mean, β = 6.05%; ɣ = 13.26%; HFA, 19.72%) than in lower frequencies (δ = 2.76%; θ = 2.14%; α = 3.24; two-tailed t test between low and high frequencies, p < 10−6; Fig. 5B). In contrast, the overall proportion of choice-active electrodes was largely homogeneous across anatomical regions (average, 37 ± 7.7%) but greatest in frontoparietal regions (52.2%), followed by prefrontal regions (40.5%) and limbic regions (33.1%; Fig. 5C).
Proportion of choice-active electrodes showing a significant difference in power between safe bet and gamble trials. Power was compared between safe bet/gamble trials during deliberation (−1 s to 0 s prechoice) using an analytical strategy that allowed differences in timing or intensity; an electrode was considered choice-active if there was a significant difference in power between gamble and safe bet trials in any frequency band (see Materials and Methods). A, Proportion of choice-active electrodes per subject (dots). Horizontal lines show mean ± SEM. See Extended Data Table 2-1 for per patient results. B,C, Mean proportion of choice-active electrodes across subjects, grouped by frequency band (across all regions, B) and anatomical region (across all frequencies, C). See Extended Data Table 5-1 for more information.
Table 5-1
Proportion of choice-active electrodes across frequency bands, per region of interest. Counts of electrodes showing a difference in power across choice conditions (safe bet vs gamble; choice-active electrodes), per region. The table indicates how many electrodes in each region showed choice encoding across one or more frequencies (regardless of in which specific frequencies), or none (0). Only one electrode was choice active in 4 frequency bands; no electrodes were significant for 5 or 6 bands. On average, 42.7% electrodes were choice-active (1 or more bands), with an average of 1.29 active power bands. Overall, there were 0.55 active frequency bands per electrode (including 0). Download Table 5-1, DOCX file.
To better understand these power modulations, we examined the relative frequency of power modulations across frequencies and regions, as well as the sign of modulation (safe bet greater than gamble or vice versa; Fig. 6). We found that choice information was more predominant in high frequencies (ɣ–HFA) compared with low frequencies across all regions, suggesting a widespread involvement of higher-frequency activity in choice behavior (Fig. 6A). Choice-related activity was present on average in 1.29 ± 0.13 frequency bands per electrode, indicating that choice information could be represented in multiple frequencies in individual electrodes (Extended Data Table 5-1). HFA was significantly more likely to increase during safe bet trials compared with gamble trials in all regions examined except for limbic regions (hippocampus, Amy, and Ins; Fig. 6B,C; binomial test, p < 0.05). This pattern was especially prominent for prefrontal regions (LPFC, p < 0.0001; cingulate, p = 0.0012; and OFC, p = 0.0023). Taken together, these results indicate that differences in HFA between gamble and safe bet trials are widespread, appearing in most regions of interest, with safe bet trials associated with greater HFA power increases compared with gamble trials.
Choice-active electrodes showed power differences between safe bet and gamble trials during deliberation. A, Mean proportion of choice-active electrodes across subjects, grouped by both region and frequency. B, Proportion of electrodes separated by safe bet significantly greater than gamble (positive-going bars) and gamble greater than safe bet (negative bars). There were significantly more electrodes with greater power in safe bet greater than in gamble trials in several regions (binomial test; ***p < 0.001; **p < 0.01; *p < 0.05). C, Same data as in B, grouped by power band (vertical axes, proportion safe bet greater; horizontal axes, proportion gamble greater).
HFA encodes choice-related computations in a widespread but regionally specific pattern
Choice behavior depends on several underlying choice-related computations, including win probability and risk. Therefore, we next sought to examine the encoding of these choice-related computations across regions of interest. To do this, we examined the relationship between power modulations and choice-related computations (Saez et al., 2018). Given the role of high frequency in discriminating between choices shown above and because HFA is known to reflect coordinated spiking activity among local populations of neurons (Nir et al., 2007; Manning et al., 2009; Jia et al., 2013), we focused on HFA for this analysis. We selected five behavioral regressors that reflect choice-related information or potential sensorimotor confounds: win probability, gamble risk, risky side, choice side, and choice (as above; see Materials and Methods, Extended Data Fig. 7-1, and Extended Data Tables 7-1 and 7-2). We carried out GLM regressions across trials for all selected regressors for each time point of the HFA time course and used sum-of-F-stat permutation in combination with a stepwise regression analysis (see Materials and Methods) to identify whether HFA encoded any of the regressors of interest.
We found that HFA encoded choice-related computations (Fig. 7; Extended Data Table 7-3): win probability (31 ± 7% of all electrodes) and risk (31 ± 8%). HFA also encoded sensorimotor aspects of the task, including the location of the gamble bet on the screen (left/right; 33 ± 9%) and the chosen side (left/right; 33 ± 10%). A similar proportion of electrodes encoded these choice-related computations across regions. In contrast, choice (i.e., safe bet or gamble) was encoded in a significantly greater proportion of electrodes (48 ± 8% mean ± SD; Fig. 7A; paired t tests vs other regressors all p < 0.01) compared with the other four regressors (t test, p < 0.01). Individual electrodes often encoded multiple choice-related computations; to quantify this degree of multiplexing, we compared the number of encoded regressors per electrode (Fig. 7B). We found that 21.2% of electrodes (188/887) did not encode any regressors, 24.7% (219/887) encoded a single regressor, and 54.1% of electrodes (480/887) encoded two or more regressors (1.74 ± 1.3; mean ± SD regressors per electrode; Fig. 7C), suggesting a high degree of multiplexing that was similar across regions (ANOVA, p = 0.28), with 7/9 regions showing a mean between 1.50 and 2.00 encoded regressors per electrode (minimum in the hippocampus, 1.34 ± 1.11, and maximal in PC, 2.03 ± 1.55). Despite these similarities, there were differences in the extent to which individual regions represented each regressor (Fig. 7D). For example, risk was maximally encoded in OFC (47.2%), whereas the choice side was maximally encoded in PoG (51.2%). These results show that HFA encoding of choice-related computations were distributed but followed a more regional distribution and were less predominant than choice signals. To verify that our results were robust to our linear regression strategy and given the low correlation among regressors (Extended Data Fig. 7-1), we carried out a regression analysis using a simple GLM strategy (see Materials and Methods), including all regressors in a single model instead of stepwise. The results of the GLM analyses were highly correlated to those of the stepwise regression (Extended Data Fig. 7-2), indicating that the results were robust to the regression strategy and not heavily influenced by regressor collinearity.
HFA showed widespread but regionally specific encoding of choice-related computations. A, Average proportion of encoding electrodes per regressor (each dot corresponds to one region, averaged across subjects); mean ± SEM indicated by error bars. Using a stepwise regression approach, we examined encoding of win probability, risk, risky choice side (L/R), choice side (L/R), and choice (gamble/safe bet). See Extended Data Table 7-1 for regressor definitions and Extended Data Figure 7-1 for collinearity results among regressors. See Extended Data Table 7-2 for electrode counts per patient and regressor included in analysis. Encoding for choice was significantly greater than all other regressors (paired t test; ***p < 0.005; **p < 0.01). B, The number of significant regressors encoded per electrode (regressors multiplexing): histogram showing the number of electrodes that encode 0, 1, 2, 3, 4, or 5 regressors out of 5 regressors of interest. White dotted line indicates overall mean (1.74). C, Multiplexing: proportion of electrodes per ROI with 0, 1, 2, 3, 4, or 5 significant regressors, gray stacked bars, shaded as in A. There was no difference in the mean number of significant regressors across regions (ANOVA, p = 0.28). D, The radar plot shows proportion of significant electrodes per regressor (5 colored lines) per region (9 vertices); see Extended Data Table 7-3. Stepwise results were similar to those obtained using a GLM approach (see Extended Data Fig. 7-2). See Extended Data Figure 7-3 for timing regressor results.
Figure 7-1
Cross correlations among regressors of interest. Cross correlation matrix among five behavioral regressors of interest studied. All-off diagonal R2 values were less than or equal to 0.01, with the exception of win probability and choice (R2 = 0.34). Download Figure 7-1, TIF file.
Figure 7-2
Comparison of stepwise and GLM regression approaches. GLM (x-axis) and stepwise (y-axis) regression approaches resulted in a similar proportion of encoding electrodes per regressor and subject (R2 = 0.68, p < 10−6). Download Figure 7-2, TIF file.
Figure 7-3
Temporal progression of encoding or choice-related computations and choice. Boxplot representing the timing of maximum peak encoding (maximum T stat) across electrodes, for all regressors in our analysis (win probability, risk, risky side [L/R], choice side [L/R] and choice [gamble/safe bet]). Win probability (mean = -0.57 s pre-button press; median = 0.5 s) and risk (mean = 0.44 s pre-button press; median = 0.2 s) information arose earliest, followed by risky side (mean = 0.33 s pre-button press; median = 0.2 s) and choice (mean = 0.32 s pre-button press; median = 0.05) side, and finally gamble index (mean = 0.03 s pre-button press; median = 0.35 s post-button press). Download Figure 7-3, TIF file.
Table 7-1
Behavioral regressors. Formulas and descriptions of selected behavioral regressors. Download Table 7-1, DOCX file.
Table 7-2
Electrode counts for stepwise regression analyses. Electrode counts per patients and region of interest included in Stepwise and GLM regression analyses. Download Table 7-2, DOCX file.
Table 7-3
Percentage of encoding electrodes per region and regressor. Percentage of encoding electrodes per behavior regressor and ROI as determined by stepwise regression model. Download Table 7-3, DOCX file.
Finally, we sought to analyze whether timing differences in encoding existed between regressors or regions of interest. To do this, we further examined the results of our time-resolved linear regression (see Materials and Methods) and examined the timing of encoding, captured by the time (relative to button press) reflecting maximum association between HFA and the regressor of interest for each electrode and regressor (Extended Data Fig. 7-3). These analyses showed no difference in peak timing across regions (Kruskal–Wallis test, p = 0.0865; df = 8), indicating that there are no overall differences in the timing of encoding among regions. However, we observed a significant difference in peak timing across regressors (Kruskal–Wallis test, p = 1.792 × 10−6; df = 4). In terms of relative timing, win probability and risk regressors peaked earliest (mean 0.57 and 0.44 s before button press, respectively), followed by risky side presentation (L/R; 0.33 s before button press) and chosen side (L/R; 0.32 s before button press). Finally, activity reflecting gamble choice [gamble index [gamble/safe bet)] appears last (0.03 s before button press; Extended Data Fig. 7-3). Post hoc tests (Dunn's test) showed the main significant difference was between gamble index (i.e., the patients' decision to gamble or not) and all other regressors (all p < 0.05), with gamble information being represented later than the other.
Discussion
Value-based decision–making depends on the coordinated activation of multiple brain areas. How this distributed neural activity supports choices in the human brain is not fully understood. Here, we sought to examine the extent of involvement of several brain regions in encoding choice-related computations and choices (safe bet vs gamble) and the nature of the neural activity underlying this representation. Our results provide several novel insights in this area: first, we observed region-specific power modulations in neural activity during choice–behavior, which were distinct in known cognitive subcircuits (prefrontal, frontoparietal, and limbic; Figs. 2–4). Second, choice information was primarily represented in HFA and not in lower-frequency oscillations, and this representation was present in multiple brain regions (Figs. 5, 6). Finally, choice-related computations (win probability, gamble risk, risky side, choice side) were also represented in HFA, showed greater regional modularity than the choice signal, and appeared earlier than choice information (Fig. 7).
Low-frequency power modulations during deliberation do not encode choices
We observed a rich pattern of bidirectional power modulations across a broad range of frequency bands and brain regions during the deliberation period (1 s prior to choice), including widespread modulation in the lower-frequency bands (δ/θ/ɑ/β; Fig. 3A). Clustering analyses showed that various sets of regions (prefrontal, frontoparietal, and limbic) display similar patterns of low-frequency power modulation (Fig. 4). Importantly, these low-frequency modulations are not related to choice output (Fig. 5B), which raises the question of their functional significance. Although our task design does not allow for any strong inference on the exact role of these power modulations, one possibility is that common patterns of low-frequency modulation reflect the reversible establishment of functional subcircuits in service of specific cognitive functions. In support of this notion, some of their features (i.e., magnitude and directionality of power modulations; Fig. 3A) are consistent with previously described roles of individual frequency bands in specific cognitive processes. For example, δ–θ increases were widespread in limbic and prefrontal electrodes, an observation consistent with the role of θ-band oscillations in the hippocampus and prefrontal cortex during goal-directed behavior (Cohen and Donner, 2013; Weber et al., 2024), navigation, and memory formation (Ekstrom et al., 2005; Buzsáki and Moser, 2013), suggesting they may underlie goal-oriented behavior. Oscillations in the α frequency band, known to be associated with attentional processes (Spaak et al., 2014; Samaha and Postle, 2015), were also modulated in frontolimbic regions (except for the Ins) and may underlie attentional processing. Conversely, the most prominent power modulation in frontoparietal regions (but not others) was a widespread β power decrease, likely reflecting β desynchronization associated with premotor processes (Kühn et al., 2004).
Interestingly, these patterns coexist, suggesting that frequency-specific modulations may offer a clue to the nature of neural processing in that specific region. For example, PrG shows both δ–θ and β modulations (Fig. 3) that may support both goal-oriented and motor functions. This is consistent with the notion that low-frequency power modulations reflect functional communication across regions, an idea reminiscent of “spectral fingerprinting” in which frequency-specific activations reflect interactions among regions for specific cognitive processes (Siegel et al., 2012). Under this framework, our observed region-specific patterns of low-frequency modulation would reflect the engagement of cognitive subcircuits (including limbic, frontoparietal, and prefrontal) in service of task demands, including goal-oriented behavior (δ–θ; Cohen and Donner, 2013; Weber et al., 2024), attention (α; Helfrich et al., 2018), and motor output (β; Khanna and Carmena, 2015).
HFA encodes choice
In contrast to the region-specific low–frequency power modulation, we observed a much greater degree of similarity in higher-frequency (ɣ/HFA) activity, which additionally was the only frequency band to show bidirectional (both increases and decreases) modulation on all regions. More importantly, ɣ/HFA activity was directly related to subject choices and reflected whether patients chose the safe or risky choice in all regions examined, indicating widespread representation of choice information (Fig. 6A). Frontoparietal and prefrontal regions were overall more engaged in encoding choice than limbic ones (Fig. 5C), likely because of their central roles in reward-based evaluation and in the transformation of abstract value-based choices (e.g., in OFC) into motor outputs (e.g., in the premotor cortex), which progressively engages prefrontal to parietal regions through cortico-striato-thalamocortical loops (Haber and Knutson, 2010). In addition, ɣ/HFA was consistently higher in safe bet than in gamble choice trials (Fig. 6B) in frontoparietal regions and most significant in LPFC. This is consistent with the notion that LPFC activation underlies the exertion of cognitive control (MacDonald et al., 2000), which in our gambling task would translate to choice of a lower, but safe, reward (McClure et al., 2004; Kable and Glimcher, 2007; Smith et al., 2019).
More generally, our data provide evidence that value-based decision–making is a distributed process engaging multiple brain areas in the human brain, complementing evidence from primate single unit recordings (Wallis and Miller, 2003; Kennerley et al., 2009; Hunt and Hayden, 2017; Smith et al., 2019; Spitmaan et al., 2020) and rodent (Steinmetz et al., 2019; Ottenheimer et al., 2022) and human imaging studies (Vickery et al., 2011). This distributed model of decision-making (Cisek, 2012; Hunt and Hayden, 2017) extends “action-plan” models (Gold and Shadlen, 2000; Kable and Glimcher, 2009; Cisek and Kalaska, 2010) by proposing that decision-making is a parallel, distributed process that weighs information across all representational levels, including those of good-based valuation. In addition, it provides novel evidence for the involvement of specific neurophysiological mechanisms involved in this process, with local ɣ/HFA reflecting choice-related activity and lower frequencies likely reflecting other task demands such as sensorimotor processing.
Computations underlying choice are regionally specific yet broadly represented
In addition to representing the final choice, ɣ/HFA activity also reflected other choice-related computations. Decisions under uncertainty involve computations related to the utility and probability of the presented choices [win probability (Hampton and O’Doherty, 2007), risk (Schultz et al., 2008)] as well as sensorimotor aspects (i.e., spatial locations, motor plans). We show that these are reflected in ɣ/HFA in different brain areas (Fig. 7) but with a lower prevalence than choice encoding (31–33%, compared with 48% for choice encoding; Fig. 7A), which also appeared later in time. Individual choice-related computations were not equally represented across brain areas; for example, risk information was significantly more represented in OFC than in any other region studied, consistent with previous fMRI (Tobler et al., 2007; Schultz et al., 2008) and electrophysiological (Saez et al., 2018) evidence involving OFC in risk processing. Motor-associated activity (left/right choice) was primarily represented in motor and premotor areas (PrG, PoG). The location of the risky side presentation (left/right) was predominantly represented in Ins, Amy, and CG regions, consistent with their roles in affective valuation, integration of interoceptive signals, anticipation, and conflict or error monitoring (Preuschoff et al., 2008; Engelmann and Tamir, 2009; Rudorf et al., 2012; Smith et al., 2019; Yih et al., 2019). These observations fit into existing models of information processing during decision-making in which economic choice progresses from abstract valuation in goods space to concrete actions in action space. More specifically, the uneven representation of computation across brain areas is partially consistent with good-based models which predict functional separation of valuation (e.g., risk in prefrontal regions) and spatiomotor (e.g., choice side in PrG/PoG) information (Padoa-Schioppa, 2011). Our timing analyses further support this notion, since abstract information related to goods space activation (risk, win probability) arises earliest (∼0.5 s before choice), followed by information related to action space (i.e., risky side and choice side, ∼0.3 s before choice) and finally by the choice output (safe bet or gamble) at the time of choice. This timing pattern is consistent with a progressive transformation of reward information from goods space to action space, before generating a motor action plan reflecting the final choice (Cisek and Kalaska, 2010; Padoa-Schioppa, 2011). Despite these timing differences across regressors, there were no overall differences in timing across anatomical regions, suggesting that choice-related computations arise simultaneously across all implicated regions, which may reflect widespread information elaboration throughout cortico-striato-thalamocortical loops (Haber and Knutson, 2010; Haber, 2016). Together, these results are consistent with distributed consensus models that predict that choice representations compete simultaneously and flexibly right until the moment that the action is executed (Cisek, 2012; Hunt and Hayden, 2017), with early choice-related computations coalescing through goods-to-action-space transformation and contributing all the while to a highly widespread choice signal.
Neurophysiologically, the role of ɣ/HFA in encoding choice and choice-related computations suggests that local neuronal activation and spiking underlies value-based decision–making processes, consistent with observations from animal models (Buzsáki et al., 2012; Buzsáki and Wang, 2012; Rich and Wallis, 2016; Parvizi and Kastner, 2018). HFA is highly correlated with both fMRI BOLD signal (Nir et al., 2007; Leszczyński et al., 2020) and single unit activity (Buzsáki et al., 2012; Buzsáki and Wang, 2012; Leszczyński et al., 2020). This, together with our choice of a simple task highly similar to others used in risk–reward studies in model animals (O’Neill and Schultz, 2010; Soltani and Izquierdo, 2019), helps resolve discrepancies between electrophysiological observations in animal models and noninvasive observations in human studies. Concretely, our evidence for distributed representation of choice-related information is consistent with observation from animal models that indicated a distributed basis for decision-making processes (Steinmetz et al., 2019; Shin et al., 2021; Ottenheimer et al., 2022) and less consistent with human fMRI studies that show more discrete activation, with a single or few brain regions encoding choice-related computations (Delgado et al., 2005; Kable and Glimcher, 2007; Tobler et al., 2007; Preuschoff et al., 2008). We postulate that this difference is due to methodological and analytical differences between electrophysiological methods and common fMRI approaches, which detect most salient activations in detriment of more distributed signals. Consistent with this idea, multivoxel studies also support the notion of distributed reward activations in the human brain (Vickery et al., 2011; Zhang et al., 2017; Kahnt, 2018). Our observations therefore provide a neurophysiologically detailed depiction of the neural basis of risk–reward choices in the human brain and can serve as ground to launch into more sophisticated exploration of the neural basis of complex human behavior.
In summary, we leveraged invasive iEEG recordings to obtain new insights into the regional versus distributed nature of neurophysiological activity underlying decision-making behavior and into the role of low- and high-frequency neural activity in decision-making processes. Specifically, we show that (1) task performance is associated with multiregional patterns of low-frequency power modulation likely to reflect coordination of multiple brain regions in service of reward-oriented (δ/θ) and motor (β) processes, but do not reflect choice information; (2) conversely, ɣ/HFA modulation is observed similarly across all regions and encodes choice information, with greater activity in safe bet versus gamble choice trials; and (3) ɣ/HFA also reflects other choice-related computations (risk, win probability), although in a weaker and more regionally specific manner, compared with widely distributed choice signals.
Overall, our results support the notion that decision-making processes engage both low-frequency activity, potentially reflecting region-specific cognitive demands of the task (i.e., attention, motor control), and highly distributed ɣ/HFA reflecting both regional encoding of choice-related computations (i.e., risk, win probability) and widespread encoding of the final choice. These observations demonstrate the distributed nature of choice encoding in the human brain and provide novel insights into the regional and global neurophysiological substrates supporting human decision-making, including the prominent role high-frequency neural activity in supporting reward-guided decisions.
Footnotes
We thank L. Nuñez, C. Meikle, and C. Foreman for their help with data collection. We especially thank the patients for their willingness to participate in this research. The project described was supported by the National Institute of Mental Health through Grant Numbers K01MH108815 and 1R01MH124763 and the National Center for Advancing Translational Sciences through Grant Number UL1TR001860 and linked award TL1TR001861.
The authors declare no competing financial interests.
- Correspondence should be addressed to Ignacio Saez at ignacio.saez{at}mssm.edu.
