The relatively new field “Computational Psychiatry” aims to develop and apply state-of-the-art computational methods to improve the mechanistic understanding, classification, diagnosis, outcome prediction, and treatment of psychiatric disorders. As the computational methodology might seem daunting at first, this symposium will provide an accessible overview and introduction to current methods, and present exemplary recent results of clinical relevance.
Computational Psychiatry relies both, on data-driven, multivariate methods from machine learning, and on explicit mathematical models of neural and behavioral data. The combination of computational modelling and specifically tailored experimental paradigms allows for a mechanistic understanding of pathophysiology and therapy effects, via the individual estimation of model parameters and dynamic variables, which are often not directly accessible. In combination with machine learning methods, this yields meaningful, compressed representations of complex datasets, which distill the relevant structure and dynamics, yielding a solid base to better inform (patho-)physiological understanding and clinical decision-making.
Kai Ueltzhöffer will give a short overview over computational modeling from single neurons, via perception, learning, and cognition, to action and behavior. Isabel Berwian will present a study using computational modelling of effort-based decision making in depression. Decision time in the model could be used for prediction of relapse after discontinuation of medication. Christoph Korn will present behavioral and fMRI studies, which explore decision making under risk, approach-avoidance-conflicts and social conflicts from a computational perspective. Georgia Koppe will show how state-of-the-art methods from “deep learning” can be applied to analyze the dynamic structure of complex psychiatric time-series data.
16:00 Uhr
Computational models in psychiatry – from neurons to cognition
Kai Ueltzhöffer, Heidelberg (Germany)
Details anzeigen
Autor:in:
Kai Ueltzhöffer, Heidelberg (Germany)
Computational models explicitly describe the distribution or dynamics of an experimentally accessible variable – e.g. reaction time distributions, functional MRI time courses, or event-related potentials. They link these observables to an underlying process, which can be described using parameters and latent variables, which are not directly accessible by measurement, but which can be inferred by inverting the computational model. In this way, they can yield a summary of the experimental data, which is informed by the structure or mechanism formalized in the model. This structure can be based on neurophysiological findings, e.g. by using neurophysiologically plausible spiking neurons and architectures to implement basic functions, such as perceptual decision making. But it can also be informed by psychological or cognitive models of learning and perception.
Current approaches in computational psychiatry try to use these models in at least two ways: 1. To yield a meaningful compression of the observed data, to feed into machine-learning algorithms, with the aim of better predicting clinical trajectories and treatment responses. 2. To get a mechanistic understanding of disease mechanism, which even might bridge the modeling scales, i.e. from neural substrates to cognitive processes. This talk will give a quick overview over such approaches and show exemplarily, how spiking attractor network models, describing neural activity, can be connected to drift-diffusion models, used to model evidence accumulation and decision making, and how this approach was used to analyze behavioral and functional imaging data. A short outlook will be given on similar approaches, which might be used in future research to connect computational models at different levels of description.
16:10 Uhr
Computational modelling of effort and reward decisions and its application to clinical questions in major depressive disorder
Isabel Berwian, Zürich (Switzerland)
Details anzeigen
Autor:in:
Isabel Berwian, Zürich (Switzerland)
The field of "Computational Psychiatry" attempts to reveal mechanisms underlying psychopathology using computational models of brain and behavior with the goal of applying them to problems in treatment selection in the clinical setting. In this talk, I will illustrate this approach focusing on relapse prediction after antidepressant discontinuation. One in three patients with Major Depressive Disorder relapse after antidepressant discontinuation even when remitted. There are no established predictors of relapse. Such predictors could improve the decision to discontinue and reduce relapse risk. We examined potential predictors of relapse after antidepressant discontinuation in the AIDA study. This was a naturalistic longitudinal study in which remitted patients on antidepressants discontinued their medications and were followed up for six months to assess relapse. Here, we report on a physical effort task in which individuals decided between high effort/high reward and low effort/low reward options. Overall, patients invested less effort for reward than healthy controls. A generative computational model that captured the decision-making process involved in the task by means of combining a value-based with a drift-diffusion model revealed that this behavior was driven by increased effort sensitivity and effort avoidance but no difference in reward sensitivity was found. Patients who discontinued and went on to relapse showed a reduction in the vigor during the execution of effortful behavior, whereas no such change was seen in patients who discontinued and remained well. Furthermore, patients who went on to relapse had increased decision times before discontinuation which was captured by an increased decision boundary in the model. Future relapsers hence needed more evidence before they committed to a decision. Decision times also predicted relapse better than chance when using leave-one cross validation in the main sample and in a separate validation sample.
16:30 Uhr
Alterations in resting state dynamics in PTSD assessed via recurrent neural networks
Georgia Koppe, Mannheim (Germany)
Details anzeigen
Autor:in:
Georgia Koppe, Mannheim (Germany)
The brain is a highly interconnected and highly nonlinear dynamical system and cognitive functions have often been described as being implemented in the system dynamics. Alterations in the dynamics have in turn been suggested as the root of cognitive dysfunction observed in various psychiatric disorders. For instance, increased depth of attractor basins could be a cause of pathological obsessive thoughts or ruminations as observed in obsessive compulsive disorder or depression, while increased system noise may result in a high distractibility as seen for instance in attention deficit hyperactivity disorder.
This talk will demonstrate in an illustrative and intuitive way how we can extract network dynamics via state-of-the-art statistical latent variable models based on piecewise linear recurrent neural networks. We will have a look at how the parameters of such networks allow us to derive new measures to directly assess the level of nervous system noise, the nonlinear connectivity between brain areas, the transition from ordered to more chaotic behavior, as well as attractors and limit cycles. The methods will be applied to functional magnetic imaging recordings during resting state of a sample of patients with posttraumatic stress disorder with and without the dissociative subtype (PTSD-DS and PTSD+DS, respectively) and healthy controls (HCs). Our findings provide evidence for alterations in network dynamics between both PTSD subtypes with dissociation being linked to more ordered activity as well as reduced nervous system noise. Reduced noise levels and ordered activity could reflect one underlying mechanisms of a system to become less sensitive to external perturbations, as observed for instance during dissociative experiences.