Exploring model-based and model-free reinforcement learning in obsessive-compulsive disorder
Year: 2020
Type: dissertation
Abstract: Obsessive-compulsive disorder (OCD) is a common, chronic and disabling neuropsychiatric condition for which current treatments are ineffective in a large proportion of cases. The gold-standard instrument to assess the severity of OCD symptoms is the Yale-Brown Obsessive-Compulsive Scale (Y-BOCS), which was recently revised (Y-BOCS-II). However, its construct validity has been reported has moderate and its criterion-related validity for the diagnosis of OCD has never been tested. In the first chapter of this dissertation, I tested, for the first time, criterion-related validity of the Y-BOCS-II and demonstrated that a cut-off of 13 (total score) attains the best balance between sensitivity and specificity for the diagnosis of OCD. However, I confirmed that its divergent validity is far from excellent. This last finding led me to search for other potential markers of OCD. Several abnormalities have been demonstrated in OCD patients in studies using neuropsychological and neuroimaging approaches, but we still lack a consistent marker for the disorder which is able to discriminate patients with OCD from healthy subjects or from patients with other mental disorders, which is sensitive to treatment induced changes, and which can be mapped to brain circuits or function. An approach which has been followed over the last decade is considering OCD as a disorder of action learning systems of the brain. Sequential decision tasks have recently emerged as an influential and sophisticated tool to investigate action learning in humans through the reinforcement learning (RL) framework. According to the RL framework, actions can be learned in two different ways: model-based control works by learning a model of the dynamics of the environment and later using that model to plan future behavioral trajectories, while model-free control works by storing the estimated value of recently taken actions and updating these estimates by trial-and-error. Sequential decision tasks have been used to assess associations between dysfunction in RL control systems and certain behavioral disorders, such as OCD, where an unbalance between model-based and model-free RL has been hypothesized. In fact, using the most commonly applied sequential decision task, the two-step task, evidence has been produced suggesting that OCD patients have a deficit in model-based learning. However, in this specific paradigm, subjects typically receive detailed information about task structure prior to performing the task. Thus, it remains unclear how different RL systems contribute when subjects learn exclusively from experience, and how explicit information about task structure modifies RL strategy. To address these questions, I created a sequential decision task requiring minimal prior instruction, the reduced two-step task. I assessed performance both prior to and after delivering explicit information on task structure, in healthy volunteers, patients with OCD and patients with other mood and anxiety disorders. Initially model-free control dominated, with model-based control emerging only in a minority of subjects after significant task experience, and not at all in patients with OCD, who had instead a tendency to increase their use of model-free control. Once explicit information about task structure was provided, a dramatic increase in the use of model-based RL was observed,similarly across healthy volunteers and both patient groups, including OCD. The debriefing also significantly decreased the use of model-free RL in healthy volunteers and in patients with mood and anxiety disorders, but not in OCD patients. Additionally, after instructions, model-free action value updates were influenced more by state values and less by trial outcomes, in all groups, and subject choices became more perseverative in healthy subjects, consistent with changes in exploration strategy. These results help in clarifying the RL profile for patients with OCD, with unspecific findings of deficient model-based control, and more specific findings of enhanced model-free control, in both cases prior to information about task structure. Finally, as the literature is not yet consensual on how model-free and modelbased RL systems interact in human brain circuits, I developed a functional magnetic resonance imaging (fMRI) protocol to assess uninstructed and instructed sequential action choice. Preliminary results in healthy subjects suggest that the fMRI version of the reduced two-step task allows to separate predominantly model-free control (before instructions) from predominantly model-based control (after instructions), in the same subject, task structure and environment. Across all sessions, choice events were associated with increases blood-oxygen-level-dependent (BOLD) activity in the left precentral gyrus and reward events were associated with increased BOLD activity in the ventral striatum. I found that explicit knowledge about task structure modifies blood-oxygen-level-dependent (BOLD) activity in the paracingulate cortex (medial prefrontal cortex) during the transition from the first- to the second-step of the task. Future directions include using multivariate pattern analysis techniques to explore how the brain represents state space in sequential decision tasks and applying the current fMRI protocol in clinical populations.
Author: Rodrigues, Pedro Manuel Ferreira de Castro
Language: English
Subfield: Clinical Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7113598323
