Computational Neuroscience: Psychiatric Disorders that emerge during Development
Brainy Poster Presentation
It is quite well-established that many psychiatric disorders emerge during development, as the brain rapidly grows and changes. This applies not only to typical neurodevelopmental disorders like (ADHD) but actually to a lot of other psychiatric disorders too — see the colourful graph in Figure 1!
Using computational modelling we can characterise these interactions and ask why certain computations go wrong in patients.
Let’s focus on the cognitive functions of reward and social learning and how they inform our understanding of psychiatric disorders like Anxiety and Depression or Schizophrenia and Psychosis (which you can see in pink, blue and turquoise on Fig. 1)
One of the most popular computational approaches is Reinforcement Learning (RL). In one very prominent study such a RL model was trained on a behavioral task (see Fig.2), where participants had to learn from social feedback (approval or disapproval) whether a person likes them. The resulting changes in self-esteem (plotted on the graph in Fig. 3) depended on social prediction errors, so the difference between received and expected social feedback. As a result of this prediction error mechanism, self-esteem increased most after approval from someone with the lowest probability of approving (see the green bar at 15%)
Interestingly, when the model parameters were combined with symptoms of psychiatric anxiety and depression (see Fig.4), it was possible to identify dopaminergic brain regions (see the yellow patches on the brain scan in Fig.5) that responded more strongly to the prediction errors in those with high vulnerability on the questionnaires (indicated by the darker shade of turquoise in the plot).
The implication of this is that over-weighted dopaminergic prediction errors during social learning result in unhealthy self-esteem updates which can make someone more vulnerable to psychiatric disorders.
Another group of researchers used a Bayesian modelling approach and neuroimaging to link abnormal belief formation (during delusions & hallucinations) in Schizophrenia and Psychosis patients to reward learning and dopamine dysfunction.
Based on a reward learning task in which patients had to choose between neutral and financially rewarding stimuli the probability of choosing each was estimated using the formula in Figure 6. Whereby Q denotes the expected reward of choosing stimulus A or B and are updated with t (time) after every trial.
Figure 7 shows a surprising signal change in the dopaminergic midbrain for the psychosis group: Augmented in response to neutral stimuli (green bar) but attenuated for rewards.
Importantly the same abnormal midbrain activity was found with a causal inference task (see Fig. 8) in which fulfilment of previously learned associations evoked a stronger dopamine response (green bar) than its violation.
This suggests that impaired dopamine signalling makes abnormal prior beliefs persist because new evidence against them is suppressed and not updated appropriately.
Despite these incredible insights, challenges remain as computational models rely on simplifications that don’t always represent the pathological complexity of psychiatric disorders. For instance, ambiguities like comorbidity (whereby symptoms are classified under several psychiatric disorders) are disregarded by models which treat diagnoses as mutually exclusive. At the same time, attempts to make increasingly complex models can result in overfitting them with too many parameters. This can lead to ungeneralizable predictions tailored only to one specific and often noisy data set. Another crucial issue that remains is that research findings from experimental settings are not always useful in the psychiatric clinic where they are needed to benefit the patients.
Nevertheless, computational approaches are powerful tools for psychiatry to tease apart the interaction between behavioural and neural as well as cognitive mechanisms, like reward and social learning, while a developmental perspective can help us trace these trajectories over time and age. But we need even more robust measures of longitudinal development and rigorous hypothesis testing on complex data sets in order to determine which computational models provide the best explanations and clinically useful predictions.