Neural correlates of reinforcement learning models in substance use disorder: a systematic review

Background

Addiction can be conceptualised as a disorder of maladaptive learning, in which drug-seeking persists despite negative consequences, reflecting aberrant value updating and a shift from goal-directed to habitual control. Computational reinforcement learning models decompose behaviour into free parameters, including learning rate, inverse temperature, and stickiness, offering a more precise characterisation of decision-making deficits than behavioural measures alone. A key question is whether these parameters have a meaningful biological basis or remain abstract descriptors of behaviour; linking them to neural substrates would provide evidence that they map onto known brain systems, strengthening their validity as mechanistic markers of addiction. To the best of our knowledge, no review has been published that examines the extent to which reinforcement learning free parameters have been linked to their neural substrates in substance use disorder, or how task and state factors moderate these relationships. The objective of this review was to systematically evaluate the existing evidence.

Methods

A systematic search of PubMed identified human neuroimaging studies (fMRI, PET, or EEG) that fitted a computational reinforcement learning model to behaviour in a substance-using sample; both criteria were required for inclusion. Fifteen studies were identified, spanning alcohol, cocaine, nicotine, opioid, and methamphetamine use disorder samples and mixed substance use disorder samples, and were evaluated with respect to reward prediction error-related neural activity, group differences in free parameters, and parameter-imaging associations.

Results

Blunted ventral striatal reward prediction error signals were identified in five of the ten studies examining reward prediction error-related activity, occurring exclusively in non-reversal paradigms; the four probabilistic reversal learning studies reported intact striatal responses but found disrupted reward prediction error-related activity in prefrontal regions, including the dorsolateral prefrontal cortex, medial prefrontal cortex, and orbitofrontal cortex. Reduced learning rates were reported in four of eleven studies, again exclusively in non-reversal designs. Reduced inverse temperature was reported in five of twelve studies across both paradigm types, with two studies finding it sensitive to drug satiety state. Only four studies directly linked free parameters to neural substrates; learning rate was associated with anterior cingulate cortex and medial prefrontal cortex activation and frontostriatal connectivity, and inverse temperature with dorsolateral prefrontal cortex and insular engagement.

Conclusions

Reinforcement learning dysfunction in substance use disorder spans striatal value-coding and prefrontal regulatory systems, dissociating systematically by paradigm type. Tasks with stable contingencies appear more sensitive to deficits in striatal value representation, while probabilistic reversal learning designs more consistently reveal prefrontal impairments, suggesting that apparently conflicting findings in the literature reflect differential task sensitivity. Reduced inverse temperature across paradigm types, and its sensitivity to drug satiety state, implicates tonic dopamine tone in modulating value-guided choice. The scarcity of studies directly linking reinforcement learning parameters to neural substrates represents a critical gap; establishing the biological basis of these computational markers is essential for advancing precision psychiatry and, ultimately, improving treatment outcomes in addiction.

References

No references