How computational psychiatry can advance the understanding and treatment of obsessive‐compulsive disorder

The behavioral repertoires of patients with obsessive‐compulsive disorder (OCD) often appear puzzling and irrational. For example, an OCD patient who just locked a door might repeatedly return and check that it is locked. Similarly, a patient might continue washing and rewashing his hands, waiting for a vague “just‐right” feeling before deciding to stop.

Numerous models have been proposed to explain such symptoms. Prominent theories argue that compulsions are driven by an attempt to reduce potential threat or anxiety. Such theories stem from patients’ reports of obsessional preoccupations with catastrophic, even if improbable, scenarios. Other equally compelling theories argue that compulsions do not relate to attaining any instrumental goal but rather to difficulty stopping a repetitive, habitual behavior ¹ . The latter accounts rely primarily on patients’ habit‐like performance on neuropsychological tasks, but their role in real‐life symptoms and experiences is less well studied. Since these (and other) theories refer to different assumptions and methods, they are rarely formally evaluated against one another, let alone integrated. Furthermore, how such theoretical debates can constructively contribute to understanding and improving pharmacological and psychosocial treatments for OCD remains unclear.

One way to overcome these impasses is to specify a mechanism tying together symptoms, performance in neurocognitive tasks, and the mode of action of existing treatments. This is an overarching goal of the field of computational psychiatry ² . A computational model of OCD might first ask ² , ³ : what computations are normally performed by the brain to solve the everyday problems of deciding when to stop handwashing or checking that a door is locked? One class of models, relying on principles of Bayesian inference, highlights a prominent role of expectations and predictions ³ , ⁴ . For example, when locking your door, you rely not only on sensory information (seeing, hearing, and feeling a click), but also on a prediction that locking the door determines that it is locked and will remain that way unless someone unlocks it. This necessity to infer the actual consequences of an action from its expected outcomes is even more evident in the case of handwashing. Given that we have no reliable sensory evidence for the absence (or presence) of germs, we nevertheless infer that our hands are clean and disinfected from the mere fact that we have just washed them.

The consequences of an inability to rely on such “top‐down” predictions is likely to include an exaggerated need to repeatedly verify that the goals of such actions have actually been attained. Furthermore, it can naturally lead to an experience of the world as unstable and unpredictable, thereby also explaining OCD patients’ excessive preoccupation with catastrophic scenarios. This mechanistic perspective also allows linking such symptoms and experiences to patients’ behavior in neurocognitive tasks requiring the integration of predictions and sensory evidence ³ , ⁵ , ⁶ .

In addition to explaining how people in general integrate predictions and sensory information to plan and infer the consequences of their behavior, a Bayesian framework also offers insight into why people sometimes persist in doing what they are used to, regardless of consequences ⁴ , and why OCD patients seem more prone to this ¹ , ³ . The basic idea is that people rely on habits especially when they cannot reliably predict the outcomes of an action ⁴ . Thus, the repetitive, habitual nature of some compulsions (and of some behaviors that patients exhibit in neurocognitive tasks) might reflect a compensatory mechanism, allowing patients to avoid uncertainty and indecision ³ .

This mechanistic, computational perspective allows us to integrate different, ostensibly inconsistent, explanations of OCD. Compulsions can be both attempts to reduce overestimated threat, and expressions of inflexible habits. Both proximal causes stem from the same core impairment (unreliable predictive models), and differentiating them becomes a question of context (e.g., some contexts encourage habit formation more than others), rather than a theoretical stance.

This perspective also has important treatment implications. In principle, it can allow clinicians to go beyond the classical question of what works for whom, to ask what works for whom, for what, and when. For example, a recent study suggested that selective serotonin reuptake inhibitors (SSRIs) reduce patients' difficulties in maintaining a predictive model of their actions and outcomes ⁷ . A computational model can explain how this helps reduce obsessions and compulsions. However, after sufficient time and repetitions, some compulsions may reach a tipping point rendering them so deeply ingrained that they are no longer maintained by this core impairment alone. Such compulsions may also be less sensitive to cognitive interventions that aspire to convince a patient that no harm will accrue if a compulsive act is not executed. Since habit‐based and non‐habit‐based compulsions may co‐occur within the same patient, such interventions might alleviate some symptoms but not others. A behavioral “outcome devaluation” test ¹ can help with differentiation: for example, an urge to check a door persisting even when seeing that it is locked might imply a habit‐based compulsion. This dynamic conceptualization of compulsions also highlights the importance of early interventions aimed at preventing the conversion of goal‐directed compulsions into habitual compulsions. Overall, these considerations serve to highlight a need for more research examining how effective different therapeutic interventions are for compulsions arising out of different putative proximal causes.

A computational approach also has the benefit of allowing researchers to perform computer simulations that can examine how a manipulation of key factors might affect specific pathological dynamics; this, in turn, can suggest a focus for novel, targeted interventions. For example, stopping compulsions completely can be intolerable for many patients. Simulations can be used to examine whether nudging a patient to occasionally avoid a compulsion ³ , or to perform it in a constantly changing manner, helps reduce the emergence of habitual dominance and improves behavioral flexibility. Similarly, the clinical practice of stressing the harm caused by certain compulsive behaviors is also supported by simulations ³ . Thus, computational simulations can efficiently reveal effects and mechanisms for various potential interventions. These predictions can then be examined in controlled experimental environments (e.g., by introducing different micro‐interventions in simple decision‐making tasks) and subsequently converted into personalized in vivo interventions, paving the way for a precision psychiatry approach to the management of OCD.

More generally, a computational psychiatry perspective helps promote greater integration of the perspectives of clinicians and basic researchers, allowing the common clinical intuition that symptoms can change across time and context to be tested using well‐specified, falsifiable models ² . Ultimately, computational models aspire to advance diagnosis and treatment.