Reward and punishment learning deficits among bipolar disorder subtypes

Reward sensitivity is an essential dimension related to mood fluctuations in bipolar disorder (BD), but there is currently a debate around hypersensitivity or hyposensitivity hypotheses to reward in BD during remission, probably related to a heterogeneous population within the BD spectrum and a lack of reward bias evaluation. Here, we examine reward maximization vs. punishment avoidance learning within the BD spectrum during remission. Patients with BD-I (n = 45), BD-II (n = 34) and matched (n = 30) healthy controls (HC) were included. They performed an instrumental learning task designed to dissociate reward-based from punishment-based reinforcement learning. Computational modeling was used to identify the mechanisms underlying reinforcement learning performances. Behavioral results showed a significant reward learning deficit across BD subtypes compared to HC, captured at the computational level by a lower sensitivity to rewards compared to punishments in both BD subtypes. Computational modeling also revealed a higher choice randomness in BD-II compared to BD-I that reflected a tendency of BD-I to perform better during punishment avoidance learning than BD-II. Our patients were not naive to antipsychotic treatment and were not euthymic (but in syndromic remission) according to the International Society for Bipolar Disorder definition. Our results are consistent with the reward hyposensitivity theory in BD. Computational modeling suggests distinct underlying mechanisms that produce similar observable behaviors, making it a useful tool for distinguishing how symptoms interact in BD versus other disorders. In the long run, a better understanding of these processes could contribute to better prevention and management of BD. •Participants with bipolar disorder showed impaired reward-based learning performance compared with control participants•Participants with bipolar disorder showed a hyposensitivity to reward compared to punishments•Patients with the BD-II subtype showed a high degree of choice variability at the computational level compared to the BD-I