Computational modelling of reversal learning in individuals with OCD versus healthy controls

Computational psychiatry provides mathematical tools to decode cognitive mechanisms underlying mental disorders. This dissertation investigates decision-making in obsessive-compulsive disorder (OCD) using a reversal learning task (RLT), aiming to identify latent behavioural markers that differentiate individuals with OCD from healthy controls.

Trial-by-trial empirical data from 142 subjects (80 OCD, 62 Healthy) was analysed. A computational pipeline was specifically developed, featuring seven cognitive models categorized into three families: Associative Learning, Bayesian Inference, and Heuristics.

Latent parameters were estimated using a just-in-time compiled optimization engine performing multi-start maximum a posteriori fitting, tested via surrogate data simulations.Model-agnostic analysis revealed no significant differences in overarching task performance between the groups. In a model comparison via the Bayesian information criterion (BIC), two strategies emerged for both groups: the winning model for most subjects was a win-stay lose-shift heuristic, while Bayesian models won when comparing average model quality across the sample.

The OCD group exhibited higher variance in baseline inattention. Within a subgroup that excluded inattentive subjects, exploratory parameter analysis suggested a potential increase in choice stochasticity in OCD during associative learning. This work exemplifies how abstract psychiatric traits could be translated into measurable computational variables.