Adaptive behavior relies on evaluating environmental features, particularly reward expectations. However, real-world unpredictability necessitates tracking higher-order variables like expected uncertainty. The temporal dynamics of these computations remain poorly understood due to limitations of non-invasive neuroimaging. This study uses intracranial EEG (iEEG) to investigate the neural evolution of reward and risk computations at high spatiotemporal resolution. Reward prediction error (RPE) signals, quantifying the discrepancy between observed and expected reward, also inform outcome variability (risk). Risk prediction error (RIPE) signals are thought to be computed when predictive associations change. A normative model suggests RPE should precede and predict RIPE computations. While previous scalp EEG studies suggest a mix of parallel and sequential computations, they lack the spatial resolution of iEEG. The study aims to test two hypotheses: that RPE precedes RIPE, and that RPE directly relates to RIPE computations. Specific brain regions, such as prefrontal cortex (PFC), anterior insula, and amygdala, are hypothesized to play key roles. The study uses iEEG recordings from epilepsy patients performing a card game designed to decouple reward and risk computations to investigate the temporal organization of reward and risk encoding across widespread brain networks.

Existing research demonstrates the crucial role of reward expectations in guiding human and animal actions. However, the inherent unpredictability of real-world settings necessitates the brain's ability to track and utilize higher-order variables such as expected uncertainty. The temporal dynamics underlying the brain's computation of these variables remain largely uncharacterized due to the limitations of non-invasive neuroimaging techniques, which often lack either the spatial or temporal resolution needed. Reward prediction error (RPE) signals have been shown to iteratively improve reward estimates and provide a measure of outcome variability, or risk. Analogously, risk prediction error (RIPE) signals are thought to be computed when associations between choices and outcomes change. A normative model suggests a temporal relationship where RPE should precede and predict RIPE signals. Previous studies, primarily utilizing scalp EEG, have begun to explore the neural correlates of reward and risk at finer timescales, providing evidence for both parallel and sequential computations. However, these studies have limitations in spatial resolution and their ability to probe deeper subcortical regions. This study builds upon previous work in human electrophysiology and aims to expand our understanding of the temporal dynamics of reward and risk processing by directly testing for interactions in the neural representations of these variables.

Ten patients undergoing chronic epilepsy evaluation participated. They performed a card game designed to decouple reward and risk computations. iEEG data were recorded from widespread cortical and subcortical regions. A normative model allowed for the calculation of expected value (EV), outcome (OUT), expected risk (E.Risk), RPE, and RIPE. Behavioral data were analyzed using mixed-effects regression models to assess task understanding and rule out learning effects. Multivariate decoding was used to analyze neural data, with single-trial decoding employed to unpack locally distributed multivariate representations of each computational variable. Bipolar referencing and ICA-based denoising improved signal-to-noise ratio. Statistical analyses used a non-parametric maximum cluster statistic approach to identify significant temporal clusters of decoding, controlling for family-wise error rate (FWE). Bootstrapping was used to estimate uncertainty around decoding accuracies. A generalized decoding analysis was conducted to test if the neural representation of one variable could predict another (e.g., RePE predicting RIPE). Leave-one-feature-out analysis examined feature importance in decoding.

Wide-spread outcome processing unfolded sequentially along the brain's anteroposterior axis, with the most posterior regions showing the earliest decoding. Expected value (EV) decoding occurred simultaneously across multiple regions, including vmPFC, amygdala, and hippocampus. Reward prediction error (RePE) decoding was significant in OFC and anterior insula, with mixed sequential and parallel processing. Risk prediction error (RIPE) decoding was significant in anterior insula and OFC, with similar mixed temporal organization. Anterior insula RePE representations predicted RIPE at a later time, suggesting a computational contribution. In contrast, OFC showed parallel decoding of outcome, RePE, and RIPE, but distinct sub-regions contributed to each decoding, indicating independent representations. Expected risk (E.Risk) was decoded from OFC but not anterior insula, contrasting with previous fMRI findings.

The findings reveal distinct modes of temporal organization for reward and risk computations. The sequential anteroposterior cascade for outcome representations suggests a hierarchical processing pathway, consistent with prior research on other cognitive processes. The simultaneous encoding of expected value across multiple regions may reflect its top-down computational nature. The mixed parallel and sequential processing of error signals highlights the brain's flexibility in integrating information. The anterior insula's role in leveraging RePE for RIPE computation demonstrates its contribution to constructing risk assessments. The parallel, yet independently represented, computations within OFC could reflect a specialization of sub-regions for distinct aspects of decision-making. Discrepancies with prior fMRI findings regarding E.Risk in anterior insula may stem from differences in the computational information captured by the two methodologies or variations in the anatomical location of recordings. The results advance our understanding of how the brain integrates reward and risk information across time and space during decision-making.

This study reveals a dynamic interplay of sequential and parallel processing in the brain's computation of reward and risk. The findings highlight the importance of considering temporal dynamics and distributed neural representations in understanding complex decision-making. Future research could explore the precise functional connectivity between regions, examine learning and belief updating, and investigate the generalizability of findings across larger and more diverse populations.

The study's relatively small sample size, based on a pseudo-population of recording sites from epilepsy patients, limits the generalizability of findings. Electrode placement, guided by clinical needs, may have missed relevant brain regions for certain computations. The focus on specific temporal windows for each variable may overlook representations at other time points within a trial. Future research should address these limitations by increasing sample size, improving anatomical coverage, and adopting methodologies to capture the full temporal extent of computational variables.