Intracranial electrophysiological evidence for a novel neuro-computational mechanism of cognitive flexibility in humans

Flexibility in a dynamic and uncertain world is critical for individuals to survive and thrive. An appropriately flexible agent should know when to exploit a known resource or explore for a potentially better one. Previous research indicates that prefrontal brain regions including the orbitofrontal cortex (OFC) and medial prefrontal cortex (mPFC), as well as the anterior insula are critical for value computation and decision-making. However, how these brain regions encode and transfer computed values to resolve the exploration-exploitation dilemma still remains unknown. We studied human cognitive flexibility and its neuro-computational mechanisms with an explore-exploit restless three-armed bandit task while recording intracranial EEG data from mPFC, OFC, dorsal medial prefrontal cortex (dmPFC), and anterior-Insula (AIn). We found that neural oscillations in the alpha, beta and high-gamma band in mPFC and dmPFC reflected feedback processing, while OFC and AIn encoded information in both decision and feedback stages. Interestingly, AIn and OFC were more sensitive to exploit behaviors after positive feedback (i.e., reward), while dmPFC and mPFC only encoded explore behaviors after negative feedback (i.e., non-reward). Alpha-band oscillatory bursting in OFC encoded the explore/exploit latent state before choice. To examine how these prefrontal brain regions compute state-values in exploit/explore decisions, we constructed a foraging based reinforcement learning model. In high-gamma band, AIn encoded the model-computed value of exploitation before choice started, and OFC encoded exploit/explore just before the feedback stage. Finally, we conducted decoding analysis in the three frequency bands across patients by implementing cross-validation and found that feedback (reward and non-reward) in the last trial predicted explore/exploit behaviors in the next trial within OFC and mPFC above chance level. Our findings may shed light on the neuro-computational mechanisms underlying cognitive flexibility and show translational relevance for brain stimulation for treating mental disorders.