832 Multiple Computations of Reward-Based Learning are Differentially Encoded in the Human Limbic Network

Abstract

INTRODUCTION: Adaptive decision making requires the capacity to anticipate and detect rewarding or aversive outcomes. Prior animal and neuroimaging studies have implicated the limbic network as an important contributor to this process. However, how different limbic regions differ in supporting distinct phases of reward-related computations are not well understood. METHODS: Fourteen human participants undergoing electrographic monitoring were included based on implantation with a depth electrode targeting at least one of the following limbic regions: anterior cingulate, hippocampus, amygdala, orbitofrontal cortex, or anterior insula. Participants completed a reinforcement learning task known as the monetary incentive delay (MID) paradigm, which is a widely used computer-based task utilized in neuroimaging research to assay anticipatory and outcome signals to monetary reward and loss. A reinforcement model was constructed based on task behavioral data. Computations of this model were regressed against electrophysiologic signals. RESULTS: We found that an expectant value signal was encoded during both reward anticipation and outcome while prediction error was solely encoded in the outcome phase. Broadband gamma activity (25-180 Hz) positively scaled with expectation value and prediction error, while low-frequency activity (1-12 Hz) negatively scaled with prediction error. Limbic regions including the anterior insula, hippocampus, orbitofrontal and anterior cingulate cortices exhibited differential encoding schemes for reward valence. Finally, there was a difference in the distribution of individual channels across limbic regions encoding expectation value, prediction error, or both. CONCLUSIONS: These findings indicate spatially distributed limbic regions with their own unique functions support complex computations of reinforcement learning.