Reward circuit local field potential modulations precede risk taking

Abstract

Risk-taking behaviour is a symptom of multiple neuropsychiatric disorders and often lacks effective treatments. Reward circuitry regions including the amygdala, orbitofrontal cortex, insula and anterior cingulate have been implicated in risk-taking, but electrophysiological activity predictive of risk taking in these regions is not well understood in humans. Identifying local field potential frequency signatures of risk-taking may provide therapeutic insight into disorders associated with risk-taking.

Eleven patients with medically refractory epilepsy who underwent stereotactic electroencephalography with electrodes in the amygdala, orbitofrontal cortex, insula and/or anterior cingulate participated in this experiment. Patients completed a gambling task where they wagered on a visible playing card being higher than a hidden card, betting $5 or $20 on this outcome, while local field potentials were recorded from implanted electrodes. We used linear regression models and cluster-based permutation testing to identify oscillatory power modulations associated with reward prediction error signal. We also computed a risk-taking value for each trial using card number and bet choice and similarly used linear regression and cluster-based permutation testing to identify power changes associated with risk-taking value. We then used two-way ANOVA with bet and risk level to identify power clusters predictive of risky decisions. We used linear mixed effects models to evaluate the relationship between reward prediction error and risky decision signals across trials.

Time-frequency clusters associated with reward prediction error were identified in the amygdala (two clusters: all P < 0.001) and orbitofrontal cortex (four clusters: all P < 0.001). Risky decisions were predicted by increased oscillatory power in theta-to-beta frequency range during card presentation in the orbitofrontal cortex (P = 0.00053; η²_bet = 0.15, η²_risk = 0.27, η²_bet*risk = 0.017) and by high beta power in the insula (P = 0.0003; η²_bet = 0.15, η²_risk = 0.20, η²_bet*risk = 0.0018). Subsequent analysis localized these signals to lateral orbitofrontal cortex and posterior insula respectively. The power within an insula cluster associated with risky decisions was associated with a theta-alpha reward prediction error signal in the orbitofrontal cortex (P = 0.023). In addition, an amygdala reward prediction error signal was associated with overall percentage of high bets (P = 0.0015) and a lateral orbitofrontal cortex risky decision signal was associated with high bets in risky scenarios (P = 0.028).

Our findings identify and help characterize reward circuitry activity predictive of risk-taking in humans. These findings identify oscillatory power signatures within these regions preceding risky decisions, which may serve as potential biomarkers to inform the development of novel treatment strategies such as closed loop neuromodulation for disorders of risk taking.

By studying human brain activity during a gambling task, Hughes and Qian et al. identify activity predictive of risk taking behaviour in areas traditionally associated with reward signalling. They also find that activity in the insula that precedes risky decisions is associated with reward signalling in the orbitofrontal cortex.

Introduction

Evaluating risks and benefits of available options is critical to our ability to make day-to-day decisions. Increased propensity for risk-taking can be present in many psychiatric and neurological diseases including substance use disorder,^1-3 bipolar disorder,⁴ frontotemporal dementia⁵ and Parkinson’s disease.^6-8 Deficits in the ability to evaluate and avoid risk can contribute to significant morbidity for patients with these diseases, leading to unfavourable choices impacting their quality of life and increasing overall risk of death.^9-12 Unfortunately, this pathologic tendency towards risk-taking is not well-managed by current treatment regimes.^13,14 Understanding neural activity changes associated with risk-taking decisions is critical for better understanding the underlying mechanisms that give rise to these pathologies and developing targeted treatments.

Several processes including reward valuation, loss sensitivity, reward and risk anticipation and counterfactual processing contribute to the decision to take risks.^15-18 Reward circuitry regions such as the orbitofrontal cortex (OFC), amygdala, insula and anterior cingulate cortex (ACC) play important roles in these processes to evaluate risk and reward.^19-24 Human intracranial recordings have found that gamma oscillatory power in reward regions encodes rewards and losses, with greater increases in gamma power in medial OFC and ACC following reward and in lateral OFC following loss.²⁵ Reward prediction error (RPE) represents the difference between expected and received reward, coded by dopaminergic neurons and may modulate risk-taking.²⁶ Gamma power also signals RPEs, with gamma power in OFC associated with positive RPEs and in insula associated with both positive and negative RPEs.^27,28 Intracranial EEG signals recorded from OFC have also been associated with anticipated reward and its subjective value.^29-31 In addition to information about reward received, representation of alternative outcomes (termed counterfactual processing) may contribute to risk-taking behaviour. Primate studies have shown that ACC neurons respond to both received and not-received potential outcomes.³² In humans, lesion and intracranial recording studies suggest that OFC is critical to the experience of regret, which depends on counterfactual processing.^31,33,34

Reward regions may also primarily encode risk, another variable critical to risk-taking decisions. Both primate and human studies have suggested that OFC neural firing encodes risk level.^29,34,35 Imaging studies have found that ACC and insula activity increase with greater risk,^36-39 and patients with amygdala lesions have impaired decision making in high-risk conditions.⁴⁰ While neural activity patterns associated with these risk and reward related processes that contribute to risk-taking have been extensively studied, specific signals predicting risk-taking and how these relate to reward signalling are less well understood. Neuroimaging studies suggest that the insula,^41,42 ACC,^43,44 OFC^43,45 and amygdala^46-48 modulate their activity during risky decision making. Scalp EEG has suggested that activity in prefrontal cortex predicts risk-taking behaviours.^49-51 A prior human intracranial EEG study found that baseline gamma activity in the ventromedial prefrontal cortex and dorsal anterior insula fluctuated with mood and predicted associated risk-taking decisions.⁵² However, the specific time-course and dynamics of these neural activity changes predicting risk-taking behaviour outside of the context of mood fluctuations and how this is associated with reward signalling remains unknown.

In the present study we recorded from intracranial electrodes implanted in patients with medically refractory epilepsy to investigate the neural dynamics that predict high-risk behaviour. Our primary analysis focused on OFC, ACC, insula and amygdala. Subjects participated in a card-based gambling task where they made the decision to place a higher or lower monetary wager on an outcome with varying levels of risk. We then evaluated neural signals of RPE and signals that occurred prior to risky decisions. Finally, we evaluated the relationship between reward and risk-taking signals. Evaluating how reward circuitry structures represent reward and predict subsequent risk-taking could provide insight into future therapeutic targets for disorders of risk-taking.

Materials and methods

Participants

We recruited 11 patients with medically refractory epilepsy who were implanted with stereotactic-EEG (SEEG) electrodes to localize their seizure focus. Electrode implant locations for each patient were clinically decided based on preoperative hypotheses of site of seizure onset. Following implantation, patients were admitted to the epilepsy monitoring unit where they were monitored for seizures and their anti-seizure medications were gradually weaned. All subjects provided written informed consent prior to participating in this study and this study was approved by the Institutional Review Board of Vanderbilt University Medical Center (Nashville, TN, IRB No. 211037). The day prior to participating in the behavioural task, a subset of patients completed several behavioural scales to measure psychiatric traits (Supplementary material, ‘Methods’ section).

Electrode localization

Each participant had multiple SEEG leads each containing 8–16 contacts (PMT Corp.). Contacts were localized by registering a post-operative CT scan to a pre-operative MRI scan using CRAnial Vault Explorer.⁵³ Patient-specific segmentations were performed with FreeSurfer.^54,55 The Desikan–Killiany Atlas^56,57 was utilized to identify brain regions along with patient-specific atlases developed for thalamic segmentation.⁵⁸ Recordings sites within medial and lateral OFC were combined for the OFC region in our primary analysis. The Brainnetome Atlas⁵⁹ was used for segmentation of the insula along anterior-posterior and dorsal-ventral axes into dorsal anterior (including dorsal agranular, dorsal dysgranular insula), ventral anterior (including ventral agranular, ventral dysgranular) and posterior (dorsal and ventral granular, hypergranular) regions (Supplementary Fig. 1). The location of bipolar recording sites was computed as the midpoint between the two electrode contacts in the bipolar pair. All automated localizations were manually verified.

Experimental design

Participants performed a gambling task while local field potentials were recorded from SEEG electrodes using the clinical recording system (Natus). The task was presented on a portable 13-inch tablet (Microsoft Surface) using Psychophysics Toolbox Version 3 core and MATLAB 2022 (MathWorks, Natick, MA).⁶⁰ This experiment was performed 1–2 days post-operatively while patients were in the hospital epilepsy monitoring unit. Digital pulses were sent to the EEG data acquisition system to synchronize task events with EEG recordings.⁶¹

Patients were instructed to maximize their winnings in a gambling task. This task has been previously used to assess risk-taking behaviours in Parkinson’s disease, depression, obsessive compulsive disorder and epilepsy patients.^62-64 Patients were shown a card from a deck made up of the cards 2, 4, 6, 8 and 10 and after a pause were prompted to place a wager as to whether their card was higher than a hidden card. They could bet either high ($20) or low ($5). Following response, the face down card was revealed, followed by feedback about the trial outcome (Fig. 1A). The placement of $5 and $20 on right or left sides of the screen during bet cue presentation indicating which button the subject should press to record the respective bet was randomized each trial to control for directional effects. Optimal strategy to maximizing earnings is to place high bets on higher card numbers. Before participation in study recordings, the patient participated in a practice block for at least 10 trials or until they felt comfortable with the task. After this, patients participated in two blocks of the gambling task of 75 trials each. Trials in which the patient took longer than 10 s (39 total trials) or less than 0.1 s (104 total trials) to respond were excluded from analysis, leaving 1507 trials included across all patients.

Figure 1.

Gambling task and behavioural data. (A) Behavioural task events and timing between events in seconds (mean ± standard deviation). (B) Mean response time (time from bet cue presentation to patient response) for each patient card number. Error bars represent standard error. (C) Mean per cent of trials on which subjects bet high ($20) for each patient card number. Asterisk indicates P < 0.05. Error bars represent standard error.

Electrophysiology recording and pre-processing

SEEG recordings were collected using the Natus clinical data acquisition system (Natus) at a sampling rate of 512 Hz and were referenced to a scalp electrode contact. SEEG pre-processing was performed using MATLAB and FieldTrip MATLAB toolbox.⁶⁵ Data was filtered with a bandpass filter from 1–256 Hz with padding and with 60 Hz, 120 Hz, 180 Hz and 240 Hz notch filters to remove line noise. Data were then bipolar re-referenced between adjacent recording sites along an electrode shaft. All traces were manually inspected for significant artifacts, which were excluded from further analysis. Contacts identified as the seizure onset zone were also excluded.

For each recording site, time-frequency representations of power between 2–255 Hz were computed using a Hanning window of 0.5 s that slides in steps of 0.05 s. Power was computed at 2 Hz frequency intervals between 2–60 Hz and at 15 Hz frequency intervals between 75–255 Hz. Oscillatory power was z-scored for each recording site across each block to allow comparison across subjects and recording sites. Processed SEEG data were segmented into epochs by alignment to the following behavioural events: (i) card presentation, lasting from time of patient card presentation to 1000 ms after; (ii) bet cue, lasting from time of the cue to bet until 2000 ms afterwards; and (iii) result, lasting from time of result reveal until 1000 ms after.

Statistical analysis

For behavioural analysis, a repeated measures ANOVA was used to identify relationships between card number and patient response times. Tukey’s post hoc honest significant difference test was then used to identify significant differences between response times for each card number. A repeated measures ANOVA was used to determine if response times differed in trials where patients bet high compared to trials where patients bet low, with patient identity as a random variable. Similarly, a repeated measures ANOVA accounting for individual patient variability followed by Tukey’s post hoc honest significant difference test was also used to determine the relationship between average per cent high bet and card number and to compare response times for different card numbers.

We assigned each trial a ‘risk-taking value’, calculated by the formula:

$$Risk-takingvalue=\left(1-{\displaystyle \frac{Cardnumber}{10}}\right)\times \left({\displaystyle \frac{Bet}{20}}\right)$$ (1)

As in a previous publication using a similar task,^34,66 we calculated RPE for each trial by subtracting the expected reward value from the monetary gain or loss outcome of that trial, or:

$$RPE=Monetarygainorloss\text{–}Expectedoutcome$$ (2)

$$RPE=Monetarygainorloss\text{–}Bet(Probabilit{y}_{win}\text{–}Probabilit{y}_{loss})$$ (3)

$$\begin{array}{rl}& RPE=Monetarygainorloss-Bet\\ & \times \left({\displaystyle \frac{Cardnumber-2}{10}-{\displaystyle \frac{10-Cardnumber}{10}}}\right)\end{array}$$ (4)

For each recording site, we performed linear regressions assessing the effect of the behavioural measure on power at each time and frequency point. To evaluate risk-taking we regressed risk-taking value against power within card presentation and bet cue epochs. To evaluate reward and loss-associated power changes we separately regressed (i) wins versus other outcomes (losses and ties) and (ii) losses versus other outcomes (wins and ties) all against power within the result epoch.³⁴ To evaluate RPE we regressed RPE value against power within the result epoch. We also performed similar RPE regressions with only win trials and only loss trials. To verify that RPE clusters were associated with true reward prediction errors with both expectation and received outcome signal components, we also performed regression analysis with expected outcomes (‘expectation’) and received outcomes (‘outcome’; $20, $5, $0, −$5, −$20) as independent predictors of power: Power ∼ (Outcome) + (Expectation), then performed cluster based permutation testing on the outcome and expectation regression estimates testing to identify outcome and expectation-associated time-frequency clusters. Group level statistics for each brain region were calculated using two-sided, one-sample Student’s t-tests on recording site regression estimates at each time-frequency point for all recording sites within a region of interest. We used cluster-based permutation testing to identify time-frequency clusters in which power was significantly associated with risk-taking value or RPE (Supplementary material, ‘Methods’ section).⁵²  P-values were Bonferroni-corrected by the number of clusters identified in a region and by number of brain regions analysed. Corrected P-values are denoted P_corr. Statistical significance was set at P_corr < 0.05. Clusters were excluded from analysis if they did not span at least 100 ms and 5 Hz (for clusters with frequencies below 60 Hz) or 30 Hz (for clusters with frequencies above 60 Hz).

To identify oscillatory power specifically predictive of high-risk behaviour, trials were separated into ‘high-risk’ (presented card of ≤6, i.e. ≤ 50% winning) and ‘low risk’ (a presented card of 8 or greater, i.e. > 50% winning) (Supplementary Table 1). We averaged z-scored power within the time and frequency boundaries of each identified risk-taking cluster. For each recording site, we averaged this power across trials of the same type: high-bet high-risk, low-bet high-risk, high-bet low-risk and low-bet low-risk. We used two-way repeated measures ANOVAs to identify associations between cluster power, bet and risk. Post hoc Tukey’s honest significant difference was then used to evaluate differences between trial type pairs. We classified a cluster as predicting high-risk behaviour (‘risky decision’ cluster) if average power for high-bet high-risk trials was significantly different from average power during all other risk-bet conditions.

To control for the effect of subject and recording site on the signals of interest, we also fitted linear mixed effects models⁵² to evaluate the effect of risk-taking or RPE value on average power within cluster boundaries: [Power ∼ (Risk-taking value) + (1 | Recording site) + (1 | Subject)]; or [Power ∼ (RPE value) + (1 | Recording site) + (1 | Subject)]. To identify individual recording sites significantly involved in risk-taking and RPE signalling, we conducted cluster-based permutation tests on individual recording site regressions (Supplementary material, ‘Methods’ section).⁶⁷ We identified a recording site as a significant contributor to a risky decision or RPE signal if it had a significant time-frequency cluster within the same epoch that met the same size criteria as the group-level analysis.

To evaluate whether our primary regions of interest contributed unique information to predicting risky decisions, we additionally performed regression using power from pairs of recording sites within the two regions that significantly predicted risk-taking decisions (Risk-taking value ∼ Region 1 power + Region 2 power). Linear regressions were performed at each time-frequency point for all pairs of recording sites for each subject that had both regions sampled. We then used cluster-based permutation testing to separately evaluate the significance of regression estimates of each predictor region and selected the most significant cluster for each region.

To evaluate whether risky decision and RPE signals were related, we calculated average z-scored power within each risky decision and RPE cluster’s time and frequency boundaries at each recording site for the five patients with leads in OFC, ACC, amygdala and insula. We then constructed a linear mixed effects model for each risky decision cluster, with power of all RPE signals and trial outcome (win, loss or draw) as predictors for risky decision power [Risky decision cluster power ∼ (RPE Cluster 1 power)(Trial outcome) + … + (RPE final cluster power)(Trial outcome) + (1 | Subject)].

To determine whether risky decision and RPE neurophysiology signals within our primary regions of interest were associated with behavioural measures, we evaluated the relationship between power in risky decision or RPE clusters in the primary regions of interest and behavioural measure scores (Supplementary material, ‘Methods’ section). For each risky decision signal, we calculated the average z-scored power within time-frequency cluster boundaries during high-bet high-risk trials for each subject. For each RPE signal, we calculated the average z-scored power during high-risk trials in which subjects won, which we define as ‘unexpected reward.’ We used Pearson correlation to determine if cluster power was associated with any behavioural metrics. P-values were corrected for multiple comparisons using the Benjamini–Hochberg false discovery rate. We additionally analysed whether risky decision cluster power was predictive of the total percentage of high bets that subjects made overall or during high-risk trials using a similar method to determine whether our identified signals were associated with overall risk-taking tendency as measured by our task.

Results

Behavioural results

Demographic and clinical information is shown in Supplementary Table 2. Average response time was 1.69 ± 1.33 s (Fig. 1A). Response time was not significantly associated with card number [F(41,503) = 2.45, P = 0.061] (Fig. 1B). Patients were more likely to place a high bet when the card number was greater, indicating that they used appropriate strategy in the task [F(4,50) = 8.17, P = 0.0001] (Fig. 1C). There was no significant association between response times and bet choice [F(11,506) = 3.5, P = 0.09]. There was also no significant relationship between task performance and behavioural measures (Supplementary Tables 3 and 4).

Reward prediction error signals

Recording sites are shown in Fig. 2. Our primary analysis focused on our a priori regions of interest OFC, ACC, amygdala and insula (Fig. 2B–E, Table 1 and Supplementary Table 5). We found that all four of these regions had power changes following rewarded and/or punished outcomes (Supplementary Figs 2 and 3 and Supplementary Tables 6 and 7). OFC beta-to-high-gamma activity increased in rewarded trials and decreased during punishments (Supplementary Fig. 2A and E). Similarly, amygdala theta-to-gamma power increased during rewards and decreased during loss trials (Supplementary Fig. 2B and F). Clusters unique to either rewarded or punished outcomes were also identified in both the ACC and the insula (Supplementary Fig. 2C, D, G and H). Furthermore, OFC, ACC and amygdala all had power changes associated with RPE (Fig. 3). There were three OFC time-frequency clusters in which power increased with increasing RPE. These spanned beta-gamma range (14–52 Hz) frequencies from 0.35–1 s, gamma from 0.7–0.9 (75–135 Hz) and 0.45–0.6 s (56–75 Hz) after result reveal (Fig. 3M, N and P). There was also one OFC theta (6–10 Hz) range time-frequency cluster in which power decreased with increasing RPE from 0.75–1 s (P_corr = 0.0013) (Fig. 3A and O). There was one amygdala gamma (34–60 Hz) cluster from 0.2–0.95 s in which power increased with increasing RPE and one theta (2–8 Hz) cluster from 0.75–1 s in which power decreased with increasing RPE (both P_corr = 0.0017; Fig. 3B, Q and R). An alpha-beta (12–18 Hz) cluster from 0.85–1 s in ACC had increasing power with increasing RPE (P_corr = 0.0015) (Fig. 3C and S). The insula did not have any significant RPE signals (Fig. 3D). Overall, 23.9% (22/92) of OFC recording sites, 44.2% (27/61) of amygdala sites and 21.1% (4/19) of ACC sites were significantly involved in RPE signalling (Supplementary Table 8). We also found this relationship between RPE and cluster power remained significant when using a linear mixed effects model to control for subject and recording site (Supplementary Table 9). OFC (Fig. 3A, E and I) and amygdala (Fig. 3B, F and J) RPE signals had both overlapping expectation and outcome components; however, the ACC lacked an overlapping expectation signal (Fig. 3C, G and K) and thus may not be a true RPE signal (Supplementary Tables 12 and 13). We performed a secondary analysis to evaluate whether medial and lateral OFC and dorsal anterior, ventral anterior and posterior insula contribute differently to RPE signalling given prior literature on the roles of these regions^27,34,52,67 (Fig. 2F and G and Supplementary Fig. 1E and F). We found lateral OFC clusters in beta and gamma ranges overlapping with those from the combined OFC analysis (all P_corr = 0.0020) and a medial OFC theta cluster (P_corr = 0.0018) in which power increased with increasing RPE (Fig. 4A, B and P–T). Power in a theta-alpha cluster in lateral OFC also inversely correlated with RPE (P_corr = 0.0020) (Fig. 4A and I). There was a high gamma cluster in posterior insula in which power correlated with our computed RPE value (P_corr = 0.040) (Fig. 4C and K); however, this did not have a significant association with either expectation or outcome separately and therefore may not be a true RPE. There were no significant RPE clusters in the dorsal or ventral anterior insula (Fig. 4D and E).

Figure 2.

Visualization of recording site locations. (A) Visualization of recording sites for each brain region across patients, colour-coded by assigned region. (B–E) Visualization of recording sites within (B) orbitofrontal cortex, (C) insula (D) amygdala and (E) anterior cingulate, colour coded by patient. (F–I) Detailed visualization of recordings sites within different regions including (F) insula subdivisions, (G) orbitofrontal cortex subdivisions, (H) temporal regions and (I) frontal regions. ACC = anterior cingulate; OFC = orbitofrontal cortex; Pt = patient.

Table 1.

Number of patients and bipolar recording site pairs for each included region

	Number of contributing patients	Number of channel pairs
Orbitofrontal cortex	10	92
Anterior cingulate cortex	6	19
Insula	10	45
Amygdala	10	61
Lateral orbitofrontal cortex	10	72
Medial orbitofrontal cortex	7	20
Dorsal anterior insula	8	19
Ventral anterior insula	6	8
Posterior insula	7	18
Superior temporal gyrus	11	92
Middle temporal gyrus	11	115
Inferior temporal gyrus	6	25
Hippocampus	7	62
Temporal pole	4	22
Fusiform gyrus	6	10
Inferior frontal gyrus	9	55
Middle frontal gyrus	4	27
Superior frontal gyrus	6	30
Precentral gyrus	5	20
Supramarginal gyrus	4	26
Putamen	5	16
Pulvinar of the thalamus	3	15
Ventral thalamus	5	18

Figure 3.

Primary region of interest reward prediction error signals. (A–D) Spectrograms showing average linear regression estimates for the relationships between power and reward prediction error (RPE), (E–H) expected outcome (‘expectation’) and (I–L) received outcome (‘outcome’) for (A, E and I) orbitofrontal cortex, (B, F and J) amygdala, (C, G and K) anterior cingulate and (D, H and L) insula. White outlines indicate time-frequency clusters within which power was significantly associated with RPE, expectation or outcome (P_corr < 0.05 for all clusters). (M and N) Scatter plots showing average power within each RPE time-frequency cluster for each RPE value and subject. This relationship is shown for (M) orbitofrontal cortex beta-gamma cluster from 0.35–1 s after result reveal, (N) orbitofrontal cortex high gamma cluster from 0.7–0.9 s after result reveal, (O) orbitofrontal cortex theta-alpha cluster from 0.75–1 s after result reveal, (P) orbitofrontal cortex gamma cluster from 0.45–0.6 s after result reveal, (Q) amygdala low gamma cluster from 0.2–0.95 s after result reveal, (R) amygdala delta-theta cluster from 0.75–1 s after result reveal and (S) anterior cingulate low beta cluster from 0.85–1 s after result reveal. Scatter points are colour-coded by subject, with black X representing mean power across all subjects at each RPE value. Blue line represents the linear regression model estimate for all subject data points with blue shading representing confidence interval, corrected for the total number of clusters detected and number of regions analysed. Red asterisk indicates that average cluster power during high-bet high-risk trials was significantly correlated with task performance metrics. ACC = anterior cingulate cortex; OFC = orbitofrontal cortex; Pt = patient.

Figure 4.

Secondary analysis of orbitofrontal cortex and insula subregion reward prediction error signals. Spectrograms showing average linear regression estimates for the relationships between power and (A–E) reward prediction error (RPE), (F–J) expectation, (K–O) outcome for (A, F and K) lateral orbitofrontal cortex, (B, G and L) medial orbitofrontal cortex, (C, H and M) posterior insula, (D, I and N) dorsal anterior insula and (E, J and O) ventral anterior insula. White outlines indicate time-frequency clusters in which power was significantly associated with RPE (P_corr < 0.05 for all clusters). (P–U) Scatter plots showing average power within each RPE time-frequency cluster for each RPE value and subject. This relationship is shown for (P) lateral orbitofrontal cortex beta cluster from 0.25–1 s after result reveal, (Q) lateral orbitofrontal cortex low gamma cluster from 0.5–0.95 s after result reveal, (R) lateral orbitofrontal cortex gamma cluster from 0.45–0.9 s after result reveal, (S) lateral orbitofrontal cortex theta-alpha cluster from 0.75–1 s after result reveal, (T) medial orbitofrontal cortex delta-alpha cluster from 0–0.6 s after result reveal and (U) posterior insula high gamma cluster from 0.7–0.9 s after result reveal. Data points are colour-coded by subject, with black X representing mean power across all subjects at each RPE value. The blue lines represent the linear regression model estimate for all subject data points with shading representing confidence interval, corrected for the total number of clusters detected and number of regions analysed). OFC = orbitofrontal cortex; Pt = patient.

We additionally performed exploratory analysis evaluating outcome signals in all brain regions sampled by at least 10 recording sites and three patients. Power changes associated with both rewarded and punished outcomes were found in the hippocampus, putamen and temporal pole, as well as in the middle temporal, superior temporal, inferior parietal, inferior frontal and middle frontal gyri. (Supplementary Figs 5 and 6 and Supplementary Tables 6 and 7). Clusters positively correlated with RPE were identified in hippocampus, temporal pole, superior temporal gyrus, middle temporal gyrus, inferior frontal gyrus and inferior parietal gyrus (all P_corr < 0.01; Supplementary Table 10 and Supplementary Fig. 4). We also identified clusters in hippocampus, superior temporal gyrus, middle temporal gyrus, inferior temporal gyrus, middle frontal gyrus and pulvinar nucleus of the thalamus where power was inversely correlated with RPE (all P_corr < 0.01). Inferior, middle and superior temporal gyri, temporal pole, inferior parietal gyrus and pulvinar all had both expectation and outcome clusters overlapping with RPE clusters (Supplementary Fig. 4). We also evaluated power changes associated with RPE for wins and losses separately to evaluate for separate reward and loss prediction errors (Supplementary Tables 16 and 17). In the OFC and inferior frontal gyrus, we found both reward and loss prediction errors overlapping with our identified overall RPE clusters, whereas in regions such as amygdala, insula, middle frontal, temporal pole, hippocampus, inferior parietal, pulvinar and inferior, middle and superior temporal gyrus, we found unique clusters specifically associated with either reward or loss prediction errors only (Supplementary Figs 7–10), indicating that some RPE signals span the full positive to negative range, whereas others are specific to reward or loss.

Risk-related signals

To identify power modulations associated with risk-taking, we first utilized linear regressions between power at each time-frequency point and risk-taking value and used cluster-based permutation testing to assess group level statistical significance. To understand how identified clusters were associated with specific aspects of risk-taking, we performed a secondary ANOVA to evaluate the impact of bet choice, risk level and their interaction on power within each cluster. We classified clusters as predicting risky decisions if power in the high-bet high-risk condition was significantly different from all other conditions. Several clusters were significantly associated with risk-taking value but did not meet risky decision criteria on secondary testing. In our primary regions of interest, we found clusters in OFC and amygdala in which power was both inversely correlated with bet choice and positively correlated with risk level (Fig. 5, Supplementary Table 11 and Supplementary Fig. 11). Both OFC and amygdala had clusters in which power was correlated with bet choice and OFC also had two clusters that were correlated with risk level alone. Finally, we found clusters associated with bet-risk interaction in OFC, amygdala and insula.

Figure 5.

Primary region of interest risk-related signals. (A–H) Spectrograms showing average linear regression estimates for the relationship between power and risk-taking value for (A and B) orbitofrontal cortex, (C and D) insula, (E and F) anterior cingulate and (G and H) amygdala during card presentation (A, C, E and G) and bet cue (B, D, F and H). Outlined regions are significant time-frequency clusters, with outlines colour-coded by signal type as determined by secondary ANOVA testing (P_corr < 0.01 for all clusters). Red = risky decision signals; purple = signals that only had a significant effect of risk alone; white = signals that were significant for the interaction of bet and risk and/or had multiple significant effects but did not meet risky decision criteria (P < 0.05). (I and L) Average z-score power within risky decision cluster plotted over time for high (red) and low (blue) bets during high-risk trials. Shaded bars indicate standard error. Grey box indicates time range of significant clusters. (J and M) Box plots of the mean power within risky decision signal clusters for the four bet and risk conditions. (K and N) Box plots of the mean power within risky decision clusters prior to high bets based on patient card number. Significant difference between groups (P < 0.05) is indicated with line and asterisk above. Black X indicates mean, whereas horizontal line within box indicates median. ACC = anterior cingulate cortex; OFC = orbitofrontal cortex.

In our secondary analysis we found that clusters in lateral OFC and posterior insula were inversely correlated with bet choice and positively correlated with risk (Fig. 6, Supplementary Table 11 and Supplementary Fig. 11). There were clusters correlated with risk alone in lateral OFC and dorsal anterior insula. Lateral OFC had a cluster in which power was correlated with bet choice alone. We also found clusters associated with bet-risk interaction in both medial and lateral OFC.

Figure 6.

Secondary analysis of orbitofrontal cortex and insula subregion risk-related signals. (A–E) Spectrograms showing average linear regression estimates for the relationship between power and risk-taking value for (A) lateral orbitofrontal cortex, (B) medial orbitofrontal cortex, (C) posterior insula, (D) dorsal anterior insula and (E) ventral anterior insula during card presentation (left) and bet cue (right). Outlined regions are significant time-frequency clusters, with outlines colour-coded by signal type as determined by secondary ANOVA testing (P_corr < 0.01 for all clusters). Red = risky decision signals; purple = signals that only had a significant effect of risk alone; white =signals that were significant for the interaction of bet and risk and/or had multiple significant effects but did not meet risky decision criteria (P < 0.05). (F, I, L, O, R and U) Average z-score power within risky decision cluster frequency range plotted over time for high bet (red) and low bet (blue) trials during high-risk conditions. Shaded bars indicate standard error. Grey box indicates time range of significant clusters. Red asterisk indicates that average cluster power during high-bet high-risk trials was significantly correlated with task metrics. (G, J, M, P, S and W) Box plots of the mean power within risky decision clusters for each bet and risk combination. (H, K, N, Q, T and X) Box plot of the mean power within risky decision clusters prior to high bets for each card number. Significant difference between groups (P < 0.05) is indicated with line and asterisk above. Black X indicates mean, whereas horizontal line within box indicates median. OFC = orbitofrontal cortex.

Risky decision signals

For translational significance, we were particularly interested in identifying power modulations predicting high-risk taking behaviour. We defined clusters as predicting risky decisions if power in high bet high-risk trials was significantly different from all other bet/risk combinations. In the OFC, there was a delta-beta (2–22 Hz) frequency cluster from 0–0.55 s following card presentation in which increased power predicted subsequent risky decision, (P_corr = 0.0021 for cluster; F(1,91) = 33.84, P = 8.8E-08, η² = 0.27 for risk; F(1,91) = 15.59, P = 1.6E-04, η² = 0.15 for bet; F(1,91) = 1.60, P = 0.21, η² = 0.017 for the interaction bet and risk) (Fig. 5A and I–K and Supplementary Fig. 12A). In the insula, there was a high-beta (26–32 Hz) time-frequency cluster 0.7–0.9 s following subject card presentation in which increased power predicted risky decisions [P_corr = 0.0013 for cluster; F(1,44) = 10.72, P = 2.07 × 10⁻³, η² = 0.20 for risk; F(1,44) = 7.86, P = 7.48 × 10⁻³, η² = 0.15 for bet; F(1,44) = 0.078, P = 0.78, η² = 0.0018 for the interaction of bet and risk; Fig. 5C and L–N and Supplementary Fig. 12B). Overall, 32.6% (30/92) of OFC recording sites and 17.8% (8/45) of insula recording sites were significantly involved in risky decision signalling (Supplementary Table 8). Furthermore, when using a linear mixed effects model controlling for subject and recording site, the relationship between risk-taking value and the power of these clusters remained significant (Supplementary Table 14).

To evaluate whether the OFC and insula clusters from the primary analysis contributed unique information to predicting risky-decisions, we utilized a linear regression model utilizing both OFC and insula power at each time-frequency point as predictors of risk-taking value. This analysis identified a 2–46 Hz cluster in OFC following subject card presentation (P_corr = 0.00077; Supplementary Fig. 13A). In the insula, we identified a beta-gamma cluster following subject card presentation (P_corr = 0.00067; Supplementary Fig. 13B). Both of these clusters overlapped with the identified signals from single region analysis, suggesting that our findings represent unique contributions of OFC and insula power in predicting risk-taking behaviour.

In our secondary subregion analysis, we found two lateral OFC clusters following card presentation in which increased power predicted risky decisions and which overlapped with the OFC cluster from our primary analysis. One was in the delta-alpha range and the other beta (both cluster P_corr = 0.0026) (Fig. 6A and F–K and see Supplementary Table 15 for ANOVA and pairwise comparison statistics). We also found one delta-alpha cluster in medial OFC following bet cue in which increased power predicted risky decisions (cluster P_corr = 0.0034; Fig. 6B and R–T). Finally, we found two other alpha-beta frequency lateral OFC clusters following bet cue in which decreased power predicted risky decisions (both cluster P_corr = 0.0044; Fig. 6A and L–Q). In posterior insula we found a beta frequency cluster following subject card presentation that overlapped with the insula cluster from our primary analysis and in which increased power similarly predicted risky decisions (cluster P_corr = 0.0018; Fig. 6C and U–X). There were no risky decision signals in dorsal or ventral anterior insula (Fig. 6D and E).

We performed exploratory analysis examining all sampled brain regions for power changes preceding risky decisions. We identified clusters in which increased power predicted risky decisions in hippocampus, middle temporal gyrus and supramarginal gyrus (all P_corr < 0.05). We also found clusters in which decreased power predicted risky decisions in hippocampus, middle temporal gyrus, precentral gyrus, middle frontal gyrus and supramarginal gyrus (all P_corr < 0.05). (Supplementary Fig. 14 and see Supplementary Table 15 for ANOVA and pairwise comparison statistics).

Relationship between risk and reward signals and behavioural measures

We evaluated whether the risky decision and RPE signals that we identified were associated with behavioural and task performance measures. Power in an amygdala delta-theta cluster (Fig. 3B and J) during unexpected rewards was correlated with the percentage of high bets made across all trials (R = 0.88, P_corr = 0.0015) (Fig. 7A). Additionally, increased power in the lateral OFC beta range cluster (Fig. 6A and O–Q) following the bet cue that predicted risky decisions was correlated with a greater percentage of high bets made during high-risk trials (R = 0.74, P_corr = 0.028; Fig. 7B). There were no significant relationships with behavioural scales (Supplementary Table 4).

Figure 7.

Relationships of power in risky decision and reward prediction error clusters to behavioural measures. (A) Power during unexpected wins in amygdala reward prediction error (RPE) delta-theta cluster (indicated by red asterisk in Fig. 3) was correlated with percentage of high bets made overall (R = 0.88, P_corr = 0.0015). Black line represents the linear regression model estimate for all data points with grey shading representing the 95% confidence interval. (B) Power preceding high bets during high-risk conditions in lateral orbitofrontal cortex beta range risky decision cluster (indicated by red asterisk in Fig. 6) was correlated with percentage of high bets made during high-risk trials (R = 0.74, P_corr = 0.028). OFC = orbitofrontal cortex.

Relationship between risk and reward signals

Finally, to evaluate the relationship between risk and reward signalling we used linear mixed effects models to determine whether the identified RPE signals were associated with each identified risky decision signal within our primary regions of interest. We found that the interaction between power in the OFC theta-alpha RPE signal cluster and trial wins was inversely correlated with insula risky decision signal power (P_corr = 0.023, adjusted R² = 0.0047), suggesting that outcome specific RPE power in OFC is associated with power modulations in insula that predict risky decisions.

Discussion

In this study we identify and characterize risk-taking and reward related oscillatory power in multiple brain regions. In our primary regions of interest analysis we identified RPE signals in amygdala and OFC. We also identified power changes associated with risky decisions in the OFC and insula that occur prior to a response being made, suggesting that these signals may be predictive of a subsequent risky decision. We found that lateral OFC risky decision and amygdala RPE signals were associated with overall risk-taking propensity as measured by task performance. We also found that a theta OFC RPE signal was correlated with insula risky decision signalling. In exploratory analysis we found additional power changes associated with risk-taking behaviour in limited frontal and temporal regions and with RPE in widespread brain regions. Together these findings deepen our understanding of how reward circuitry signalling may influence risk-taking.

Reward prediction error signals

We identified several oscillatory power changes associated with RPE in our primary reward-associated regions of interest. In ACC, amygdala and OFC there were clusters in which power increased with increasing computed RPE. In OFC and amygdala oscillatory activity was negatively correlated with expectation and positively correlated with outcome, consistent with what is expected for an RPE signal. In the ACC RPE cluster power was not inversely correlated with expectation, suggesting that this may not be a true RPE signal. In OFC and amygdala there were also separate clusters in which power decreased with increasing RPE. This aligns with prior literature which has identified roles of the ACC, amygdala and OFC in RPE signalling. Several prior studies have found an important role for beta power in reward signalling.⁶⁸ In ACC specifically, beta power has been associated with reward anticipation and receipt.⁶⁹ We similarly identified an outcome related ACC beta power signal that did not have an overlapping expectation component, suggesting that ACC beta power may not be an RPE signal but may signal outcome alone. OFC has previously been shown to encode several reward characteristics that may influence risk-taking including reward probability and uncertainty.²⁹ Gamma power in medial and lateral OFC has been associated with reward receipt, subjective value and outcome processing.^25,34,70 Our findings support a direct role of both medial and lateral the OFC in RPE signalling, as we find lateral OFC power in beta to gamma frequencies that increases with greater positive RPE and a low frequency signal in medial OFC that also correlated with RPE. A previous study found OFC high gamma power increases associated with RPE.³⁴ We similarly found gamma range 75–135 Hz and 56–75 Hz clusters correlated with RPE; however, we additionally identify a robust lower frequency 14–52 Hz cluster associated with RPE. Nonetheless, our findings are similar to other studies that have found both positive and negative prediction errors are positively correlated with broadband gamma activity in medial and lateral orbitofrontal cortex,²⁷ and with alpha event related potentials in ventromedial prefrontal cortex.⁷¹

The amygdala is well-known to be strongly associated with reward outcome and reward evaluation^72,73 and prior functional MRI research has suggested that it has activity changes with both positive and negative RPEs.⁷⁴ Amygdala gamma power has been shown to increase with both reward and loss receipt.³⁰ We found amygdala gamma power increase with positive RPEs and low frequency power increase with negative RPEs, supporting a role for amygdala in signalling both positive and negative RPEs. Several previous studies have identified RPE signals in the insula using both functional MRI and intracranial EEG,^{25,27,28,72,75} with negative RPE in particular coded by gamma power in anterior insula.²⁷ Our findings suggest that positive RPE may be encoded by gamma power in posterior insula. Lack of signal detection within the anterior insula may be due to a low number of recording sites in this area in our study.

In our exploratory analysis we found power changes in widespread frontal, temporal and parietal regions associated with RPE. These findings may be consistent with a broader network of reward signalling that involves vision, action and memory circuits as previously proposed,^76,77 which could suggest that unexpected positive outcomes may similarly have a modulatory role throughout the brain to impact and guide future actions. Supporting the existence of network-level RPE signalling, prior literature suggests frontal cortex subsystem interactions in both reward and punishment prediction error signalling.⁷⁸

Risky decision signals

Neural activity associated with risk-taking has previously been demonstrated in reward regions including amygdala, OFC and insula, and our study reinforces and expands upon these findings. We identified gamma power modulations in the amygdala that were positively associated with the bet-risk interaction and with bet and risk but that did not predict risky decisions specifically. Amygdala activity is known to be important in associating stimuli with subjective value and in biasing decisions based on prior trials.⁷⁹ OFC gamma power has been shown to play a role in multiple processes that may contribute to risk-taking including reward anticipation,^29,30 risk level processing and anticipation,^29,34 and coding recent outcomes.^31,34 We found two theta-beta clusters in OFC that were associated with risk level, supporting a role for OFC in coding this property. We also found beta and gamma power changes that were associated with the interaction of bet and risk but did not predict risky decisions, supporting a role for OFC in encoding multiple processes contributing to risk-taking. While each frequency band may be associated with multiple neural processes, gamma power has been linked with neural firing,⁸⁰ while beta power has been suggested to help maintain neural equilibrium or to play a role in clearing and updating information.⁸¹ Gamma power in ventromedial prefrontal cortex, which overlaps with medial OFC, has been shown to fluctuate with mood and correlate with risky decisions associated with mood fluctuations.⁵² We found that theta to beta power in both lateral and medial OFC predicted risky decisions, suggesting that OFC power changes can independently predict high-risk behaviours outside of the context of mood fluctuations. Our findings overall support a role for OFC in multiple processes contributing to risk-taking and identify an OFC oscillatory power signature that specifically predicts high-risk decisions, which may be useful in developing novel treatments for disorders with pathological risky behaviour in the future.

We also identified an insula high beta power increase that predicted subsequent risky choices. Previous functional MRI studies have shown that the insula is involved in multiple processes associated with risk-taking. Posterior insula activity increases during visualization of risky scenarios⁸² and may also be involved in delayed gratification,⁸³ and anterior insula is involved in processing of risk level.³⁶ Insula activity also increases during reward anticipation.⁸⁴ Human intracranial studies have suggested that mood fluctuations induce anterior insula gamma power which increases subjective value of losses and modulates risk-taking.⁵² Our findings build on this prior work, suggesting that posterior insula high beta power independently predicts high-risk decisions outside of risky decisions induced by mood fluctuations and identifies specific time limited power modulations during decision making associated with this process.

In our exploratory whole brain analysis, we identified power changes in fronto-temporal regions including hippocampus, middle temporal gyrus, middle frontal gyrus and precentral gyrus that predicted risk-taking. Previous imaging studies have demonstrated temporal lobe, middle frontal gyrus and parietal lobe involvement in risk-taking.^46,85-88 Our findings suggest that power modulations in this frontotemporal network can predict a risky decision.

Risk and reward signal association with behavioural measures

To associate these electrophysiological signatures with patient behaviour, we looked at if these signatures were associated with overall task performance. Overall, the behavioural performance of subjects in our study was similar to that in other studies utilizing this task in Parkinson’s disease, depression, obsessive compulsive disorders and epilepsy patients, with the percentage of high bets increasing with card number, indicating appropriate strategy and understanding of task.^62-64 While a similar study in Parkinson’s disease patients showed a statistically significant slowing in response times for cards with high uncertainty,⁶² the response time difference between different card numbers trended towards but did not reach statistical significance in our study, similar to prior studies using this task in epilepsy patients and in patients with depression/obsessive compulsive disorder.^63,64 This may reflect differences among Parkinson’s disease and epilepsy, depression and obsessive compulsive disorder patient populations, or may be due to other factors contributing to differences in task performance.

We found that increased power within the lateral OFC beta cluster predictive of risky decisions was correlated with greater percentage of high bets in high-risk scenarios, meaning that subjects with higher lateral OFC beta power were overall more likely to choose high-risk choices. Additionally, increased amygdala power during trials with high positive RPE was correlated with greater percentage of high bets across all trials. RPEs have previously been shown to modulate future behaviour, with greater RPE signaling contributing to a greater likelihood of pursuing future rewards.⁸⁹ Our findings suggest that amygdala may play an important role in this process. We also evaluated the relationship of behavioural scale scores to risky decision and RPE power modulations identified in our study but did not find any significant relationships. However, these behavioural scales were also not associated with task performance, suggesting that they may not be associated with risk-taking behaviour. Additionally, the behavioural measures were obtained in only a subset of patients and this analysis was therefore not well-powered to detect associations between these measures and average oscillatory power. Future studies with larger sample sizes and more detailed and risk taking targeted behavioural scales would be necessary to determine if such a relationship exists.

Risky decision signal association with reward prediction error

We also evaluated whether risk and reward signals were related. We found that the insula risky decision signal was inversely correlated with OFC RPE signal power during rewarded trials. Although the insula and OFC specifically have not previously been implicated in this interaction between RPE and risk-taking decisions, this relationship has been suggested in other brain regions using different methodological techniques. For example, prior event-related potential studies using scalp EEG have demonstrated that frontal feedback related negativity signal amplitudes may be associated with the level of risk taken, with reduced amplitudes in high-risk choices. Imaging studies suggest that risk prediction error signalling in the inferior frontal gyrus may be more pronounced in risk averse patients.⁹⁰ In the nucleus accumbens, a relationship between dopaminergic cells signalling unfavourable outcomes has been shown to play a role in real-time risky decision making.⁹¹ However, to our knowledge, the relationship as described here between OFC post-outcome RPE signal and pre-decision insula risky decision signal has not before been described. Future research may be necessary to better examine the elements of these relationships between risky decision signals and RPE signals, as they may be modulated by learning, magnitude of risk and reward or other factors.

Overall, we identify power changes in reward regions that predict risk-taking and find that risk-taking and reward signals in insula and OFC are related. These findings overall increase the understanding of neural activity predicting high-risk behaviours and have implications for developing neuromodulation approaches for disorders with pathological risk-taking. Neuromodulation has been used in humans for movement disorders and epilepsy, but also OCD, depression and has been postulated to even improve memory by altering neuronal network firing patterns.⁹² Identification of neural signatures predicting risky decisions may enable to future development of closed loop neuromodulation approaches to treat pathologic risk-taking in conditions such as substance use and bipolar disorder.

Limitations

This is a single-institution study with a specialized population of individuals with medically refractory epilepsy, limiting generalisability. Epilepsy may alter neural activity patterns, though we excluded electrode contacts within the seizure onset zone from analysis to decrease risk of this confounder. Furthermore, despite a homogenous patient population, the average Behavior Inhibition Scale-11 score was 66.9 ± 10.0, which is comparable to a previously reported adult average of 62.3 ± 10.3,⁹³ suggesting similar impulsivity to the general population. A limitation of our study is that our task has not been performed in healthy subjects and expected task behaviour has not been characterized in a healthy population. However, the task has been previously used in Parkinson’s disease,⁶² epilepsy,⁶³ depression⁶⁴ and obsessive compulsive disorder patients⁶⁴ with similar behavioural performance to our study population, with the exception of slower reaction times for more uncertain outcomes in Parkinson’s disease but not other disease group patients. Additionally, in a very similar task, behaviour was not different between epilepsy patients and healthy controls.³⁴ Additionally, cluster-based permutation is not suitable to identify effect latencies on a millisecond scale; therefore timing and frequency of reported signals are approximate.⁹⁴ Additionally, though these findings are consistent with and expand upon prior literature, this task is imperfect for evaluating reward prediction errors. Probability of a win is an indirect measurement of reward expectation, although reward probability associated with each stimulus was directly provided. However, the described methodology of calculating RPE from the difference between expectation and received outcome has been used in prior literature.^34,66

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Funding

National Institutes of Health NINDS K12 NS080223.

Competing interests

The authors report no competing interests.

Supplementary material

Supplementary material is available at Brain  online.