Symposium 7 – What do computational models tell us about cognition in depression and psychosis?

This symposium discussed how computational models could be used to explore decision-making and prediction learning in psychiatric disorders. By including computational modelling in our analyses we can identify and quantify key parameters, which are not visible to traditional analyses, and can often be successfully linked to the function of underlying neural systems or to clinically relevant outcomes.

For reference, the Transcontinental Computational Psychiatry Workgroup organises a monthly webinar to foster discussions between those involved in computational psychiatry –

Dr Thomas Akam began by describing how it is necessary to develop ways of exploring rich, human-like decision-making in animals (specifically rodents in this case), which could then be used to inform psychiatric disorders that display impairments in decision-making abilities.

Mice readily learn multi-step dynamic decision tasks and are able to use a model to inform active plans that satisfy our needs, such as planning our action to get from A (where we are currently) to B (where the food reward is located). Moreover, mouse two-step task behaviour looks consistent with model-based reward learning observed in humans, but this may reflect a latent state inference rather than planning.

When exploring human two-step learning, we typically provide them with a lot of instruction prior to task completion, but what happens if we do not give them any instruction? Before receiving instructions patients performed pretty poorly, but once they were told the model structure they showed increased use of the model and updated their understanding. Thus, uninstructed human two-step tasks dissociates learning and instruction effects of model-based choice.

Professor Paul Fletcher used a meta-level view of psychosis as a worked example of the use of coding/processing to explore the challenges in neuropsychiatry. He began by looking at the inherent difficulties in modelling a noisy, ever-changing world to explore how the brain interacts with the world, and in some cases may deviate from reality.

Our personal perception of the world must be a process of inference which is optimised by prior knowledge of our environment. Thus, we can use predictions to enhance our perceptions and understanding of our environment, but what if we create perceptions that aren’t there? – These can be represented as delusions/hallucinations. Our understanding of predictions in psychosis are mixed within the literature with research outlining how delusions/hallucinations in psychosis may be a hyper-strengthening of top-down signals, with a greater tendency to exercise top-down processing in those who experience hallucinations. In contrast however, some research has suggested there to be a weakening of top-down mechanisms in psychosis. There is clearly still more work required in order to understand this model of psychosis.

It is therefore possible to see that predictive processing is proving increasingly widespread and influential. It offers a powerful framework for exploring neuropsychiatric illness, uniting multiple levels of description enabling us to bridge the explanatory gap, and formulating testable hypotheses at different levels. But, as ever, more work is required in order to further understand the model of psychosis.

Dr Quentin Huys asked why we would want to apply ‘maths’ to questions referring to emotions and how we could apply Bayes rules to everyday clinical practice? We understand that computations are processes that store and order/extract statistics which infer meaning and assist in making predictions about the world. However, we can observe illnesses that come from aspects of the brain where this goes awry. The application of computational analyses therefore provides an important analysis tool that can be applied to clinical questions.

We know that in depression, antidepressants are useful in preventing relapse into symptomatology, but some patients relapse despite taking antidepressants. Thus, Quentin asked, “Who can take antidepressants without relapse? And, who will relapse despite taking antidepressants?” Very little research has explored individual vulnerability in this topic. As such, sadness reactivity was explored using prefrontal asymmetry in alpha power using EEG, where it was observed that only people who relapse experienced sadness reactivity.

The aspects of decision-making that seem to influence patients were explored by asking patients to choose between either receiving one point with a certain number of button presses in a paradigm, or pressing the button many more times and receiving more points. It was found that patients didn’t choose the higher options despite receiving more points for them, whereas controls would choose the option with more points. As such, it was concluded that patients were more sensitive to effort.

Dr Claire Gillan began by asking, “if we’re going to develop these models of complex analysis, then how can we actually use them effectively?” We understand that DSM disorder classifications are an imperfect ‘ground-truth’ for research, with patients being heterogeneous within disorders and also similar across disorders with a high prevalence of co-morbidity. Typically we would aim to link cognitive and neural factors to symptom dimensions, but patients require a complex combination of these factors to meet the criterion for diagnosis.

Claire took a dimensional approach to leverage the efficiency of large-scale data collection among ‘healthy’ individuals to focus on the precise psychiatric phenotype associated with a deficit in goal-directed control in a model-based planning task. She concluded that evidence supported the conclusion that ‘Compulsive Behaviour and Intrusive Thought’ was a symptom dimension associated with deficits in goal-directed control linking features from symptoms of OCD, addiction and eating disorders. Thus, this dimension goes beyond the behaviour previously associated with compulsivity, suggesting a trans-diagnostic ‘compulsivity’ dimension maps onto goal-directed deficits more strongly than a diagnosis of OCD. This same factor analysis structure has since been replicated and proven its robustness (for example, Roualt et al 2018, Patzelt et al., 2019).

Following this, Claire explored whether this approach would be meaningful for ‘real’ patients too. In a paper currently under revision, comparisons were made between patients diagnosed with OCD and general anxiety disorder (clinical control cohort), to separate obsessions from compulsions and explore whether goal-directed planning maps compulsivity. In short, higher compulsivity indicated poorer goal-directed planning, but did they map onto diagnosis? No, in fact they did this really poorly, with the obsessionality dimension showing no association whatsoever. Thus, the effect of compulsivity was predictive of the deficits observed in patients. Perhaps dimensions may be a better ground-truth for research than diagnosis.

Leave a Reply

Your email address will not be published. Required fields are marked *