Conference on Cognitive Computational Neuroscience

Learning to maximise reward, and minimise punishment, is the theoretical premise of reinforcement learning (RL). Although both might seem computationally similar, their contrasting affect modulates learning behaviour. Humans tend to integrate incoming feedback with a positive valence more than feedback with a negative valence. In this study, we investigate how RL-based valence biases might be related to mental health profiles characterised by three novel psychiatric dimensions (Anxious Depression (AD), Compulsivity (C) and Substance Abuse (SA)), while accounting for context-dependent value encoding. This positivity bias is hypothesised to be further altered by mental health states. Additionally, we posit context-dependent value encoding as a relevant modulator valence biased behaviour. While we find a correlation between our SA dimension and value encoding transfer between experimental phases, the correlation between our psychiatric dimensions and both valence bias and context dependency remain elusive. We propose these behavioural traits might lack temporal robustness. This is supported by our findings showing thorough replicability but poor interindividual reliability. In contrast questionnaire derived results show good reliability. In conclusion, we present a rationale for a reappraisal of task elicited mental health studies without reliability analysis. Furthermore, we advocate for systematic state versus trait analysis in task-related mental health approaches.