Impaired Probabilistic Inference in Mental Disorders. Schizophrenia and Autism
Lead Research Organisation:
University of Edinburgh
Department Name: Sch of Informatics
Abstract
In recent years it has been argued that the positive symptoms of autism (perceptual hyper- sensitivity) and schizophrenia (delusions and hallucinations) might be a result of an impaired probabilistic inference where low-level sensory information dominates over high-level prior expectations. These hypotheses are a direct result of recent advances in cognitive sciences:
perception (and reasoning) have been formalized in terms of a general Bayesian inference that combines ambiguous low-level sensory information with high-level prior expectations of what might have caused the sensory experience. Accumulating evidence suggests that mentally healthy people perform as optimal Bayesian observers, while people with schizophrenia and autism perform differently (e.g., they are less susceptible to certain perceptual illusions).
In my PhD I am planning to thoroughly test this rather general hypothesis. First, via behavioral tasks (perceptual and probabilistic reasoning) I will try to determine whether different performance in people with the aforementioned disorders arises due to impaired updating of their internal beliefs (how well these beliefs correspond to the actual environment) or due impaired integration of these top-down beliefs with bottom-up information. This can be determined by modeling patients performance in statistical learning tasks (which are used to induce specific top-down expectations). The models have already been developed and tested in previous experiments. This project will be done in collaboration with the Royal Edinburgh Hospital. We have already acquired ethics approval and will start recruiting participants in November. Second, depending on the results of the project described, I will attempt to test different algorithmic implementations of any impairments that we find. So far, predictive coding and belief propagation models have passed 'proof of a concept test', but they have not yet been adapted to be used for modelling actual data. This will require designing additional experiments in order to tease out the relevant aspects of cognitive functioning. The adapted versions of models and the design of the experiment should be done in the second year and the experiment itself in the third year.
perception (and reasoning) have been formalized in terms of a general Bayesian inference that combines ambiguous low-level sensory information with high-level prior expectations of what might have caused the sensory experience. Accumulating evidence suggests that mentally healthy people perform as optimal Bayesian observers, while people with schizophrenia and autism perform differently (e.g., they are less susceptible to certain perceptual illusions).
In my PhD I am planning to thoroughly test this rather general hypothesis. First, via behavioral tasks (perceptual and probabilistic reasoning) I will try to determine whether different performance in people with the aforementioned disorders arises due to impaired updating of their internal beliefs (how well these beliefs correspond to the actual environment) or due impaired integration of these top-down beliefs with bottom-up information. This can be determined by modeling patients performance in statistical learning tasks (which are used to induce specific top-down expectations). The models have already been developed and tested in previous experiments. This project will be done in collaboration with the Royal Edinburgh Hospital. We have already acquired ethics approval and will start recruiting participants in November. Second, depending on the results of the project described, I will attempt to test different algorithmic implementations of any impairments that we find. So far, predictive coding and belief propagation models have passed 'proof of a concept test', but they have not yet been adapted to be used for modelling actual data. This will require designing additional experiments in order to tease out the relevant aspects of cognitive functioning. The adapted versions of models and the design of the experiment should be done in the second year and the experiment itself in the third year.
Organisations
People |
ORCID iD |
| Peggy Series (Primary Supervisor) | |
| Frank Karvelis (Student) |
Publications
Valton V
(2019)
Acquisition of visual priors and induced hallucinations in chronic schizophrenia.
in Brain : a journal of neurology
Studentship Projects
| Project Reference | Relationship | Related To | Start | End | Student Name |
|---|---|---|---|---|---|
| EP/N509644/1 | 01/10/2016 | 30/09/2021 | |||
| 1789263 | Studentship | EP/N509644/1 | 01/09/2016 | 29/02/2020 | Frank Karvelis |
| Description | Modern psychiatry is far from precise. Current diagnostic categories are based on surface symptomatology, often rely on trial-and-error treatment and focus on managing symptoms rather than treating the underlying illness. That is primarily a result of not having a sufficient understanding of how the observed mental dysfunctions arise from the complex computations performed by the brain. Computational psychiatry steps in to fill this gap: mathematical models of the neural computations are used together with experimental data to study how disruption at neural levels propagate upward to produce dysfunction at the level of behavior. A widespread hypothesis in the field of computational psychiatry is that schizophrenia (SCZ) and autistic spectrum disorder (ASD) are related to impairments in Bayesian inference performed by the brain; that is, how accumulated prior knowledge ('prior') is combined with incoming ambiguous sensory information ('likelihood'). More precisely, it has been proposed that either priors are 'weaker' or that likelihoods are 'enhanced', resulting in percepts being systematically closer to the sensory input. Such impairment could explain many of the ASD and SCZ symptoms, such as sensory hypersensitivities & abnormalities in ASD and hallucinations & delusions in SCZ. However, very few studies have directly tested this hypothesis and even fewer have fitted computational models to the data instead of inferring the underlying cognitive processes from the behavioral data analysis alone. Due to these and other methodological limitations, most of the central questions surrounding the hypothesis remain unanswered: (1) do the impairments result from 'weaker' priors or 'enhanced' likelihoods? (2) how do SCZ and ASD differ in the underlying impairments in Bayesian inference? (3) if priors are indeed implicated, is there a problem with the acquisition or simply utilization of priors? These are some of the main questions that I seek to address in my PhD research. In our first study (Karvelis et al., 2018, eLife) we investigated how people from a normal population but along autistic and schizotypy traits spectra acquire and use perceptual priors. In a sample of 83 people we found that schizotypy traits showed no effects on the perceptual inference under our experimental conditions. However, ASD traits were associated with decreased perceptual effects. By applying computational models to analyze such effects, it was found that these effects were driven not by any differences in the acquired priors as proposed by the original Bayesian accounts of autism, but by increased precision in sensory representations. More recent Bayesian accounts of autism are in line with such finding, although our results are the first ones to provide direct experimental evidence. In our second study (Valton & Karvelis et al., 2019, Brain) we investigated patients with schizophrenia and again found no impairments in the acquisition of perceptual priors; however, when faced with complete sensory ambiguity, patients were less likely to rely on such priors. This finding is a small but valuable piece in the puzzle of how exactly Bayesian inference is impaired in schizophrenia. |
| Exploitation Route | Our findings can be seen as a valuable piece in a complicated puzzle of understanding the computational character of autism and schizophrenia. However, it is an enormous task that requires a lot of experimental work and patience. We hope that our contribution will make this process faster and smoother for all the academics working in this field. Once we have a better handle on the computational basis of these disorders, we hope that our work will be translated into new tools in the clinical practice: be it medication, therapy or perhaps even new forms of tools. In addition, we hope that our work shines more light not only on brain disorders, but also on the general principles of brain functioning, informing academics doing basic research. |
| Sectors | Communities and Social Services/Policy,Education,Healthcare,Pharmaceuticals and Medical Biotechnology |
| URL | https://elifesciences.org/articles/34115 |