Differential reinforcement learning responses to positive and negative information in unmedicated individuals with depression

Abstract

Major depressive disorder (MDD) is characterized by behavioral and neural abnormalities in processing both rewarding and aversive stimuli, which may impact motivational and affective symptoms. Learning paradigms have been used to assess reinforcement encoding abnormalities in MDD and their association with dysfunctional incentive-based behavior, but how the valence and context of information modulate this learning is not well understood. To address these gaps, we examined responses to positive and negative reinforcement across multiple temporal phases of information processing. While undergoing functional magnetic resonance imaging (fMRI), 47 participants (23 unmedicated, predominantly medication-naïve participants with MDD and 24 demographically-matched HC participants) completed a probabilistic, feedback-based reinforcement learning task that allowed us to separate neural activation during motor response (choice) from reinforcement feedback and monetary outcome across two independent conditions: pursuing gains and avoiding losses. In the gain condition, MDD participants showed overall blunted learning responses (prediction error) in the dorsal striatum when receiving monetary outcome, and reduced responses in ventral striatum for positive, but not negative, prediction error. The MDD group showed enhanced sensitivity to negative information, and symptom severity was associated with better behavioral performance in the loss condition. These findings suggest that striatal responses during learning are abnormal in individuals with MDD but vary with the valence of information.

Keywords: Learning; Major Depressive Disorder; Putamen; Reinforcement; Reward; Ventral Striatum.

Fig. 1

Reinforcement Learning Paradigm. Each participant completed a probabilistic reward-based learning task in two contexts, (A) incentivized by gaining money, and (B) incentivized by avoiding losing money. Condition order and stimuli were counterbalanced. (C) There was a 70/30 probabilistic contingency that linked the choice to verbal feedback. Though expected value was higher with “correct” feedback and lower with “incorrect” feedback, there was a (D) 50% uncertainty of reward receipt magnitude.

Fig. 2

Behavioral Results. (A) Participants’ reaction time to choices became shorter as they progressed across blocks (10 trials/block) in both conditions. (B) Likewise, participants improved performance as reflected in the percentage of time they chose the correct (optimal) shape, or the one most associated with a rewarding outcome. (C) Model-based analyses indicate there is a group effect greater than 0 based on negative feedback PE valence (top panel; grp=group, con=condition). (D) A correlation between anhedonic depression severity and loss performance was observed in the MDD group (r = 0.42, p = 0.04). All group labels are consistent with colors in the top row legend.

Fig. 3

Learning Signal Differences for HCs > MDD. (A) Regions of interest defined by automated meta-analysis (neurosynth.org) using terms “ventral” (teal) and “dorsal” (yellow) striatum. (B) Whole-brain corrected results (FWE corrected p < 0.005) were examined as they differed for HC > MDD for prediction error learning signals during presentation monetary outcome in the gain condition and (C) results for HC > MDD represent positive PE across feedback and outcome in the gain condition. No significant results were found for the loss condition or negative PE in the striatum.