The REM effect: How sleep stages influence learning strategies in complex reward-based decision-making

Abstract

Model-free (MF) and model-based (MB) learning strategies are complementary decision-making processes used in evaluating choices with potential rewards. Disorders involving compulsive behaviours (e.g., substance use, gambling) are suggested to emerge from an overreliance on MF learning, though the reasons for this bias remain unclear. Sleep disruptions, common in these disorders, could be a contributing factor, however no study has examined the impact of sleep and/or sleep loss on an individual’s engagement of each strategy. Thus, this study examined the influence of sleep on MF/MB learning in healthy adults. Participants (n = 67, M_age = 26.21yrs, SD = 5.82yrs, females = 65.67%) completed a two-stage reinforcement learning paradigm following a week of either sleep restriction (5-hr time in bed/night) or well-rested sleep (9-hr/night). Using mixed-effect logistic regressions and comprehensive computational modelling, we found no differences in MF and MB learning based on sleep condition (all p = > 0.05). However, regressions showed less REM sleep was associated with increased use of MB learning, whilst greater levels of REM sleep were associated with increased use of MF learning. Computational modelling supported this, revealing negative associations between the MB parameter estimate and REM sleep percentage (τ = -0.22, p = 0.02). This suggests the amount of REM sleep prior to learning may potentially play a role in determining which strategy will dominate. In particular, individuals with less REM sleep may be less willing or able to assess the relative costs and benefits of each strategy. Future research should explore this relationship further.