Home
This Title All WIREs
WIREs RSS Feed
How to cite this WIREs title:
WIREs Cogn Sci
Impact Factor: 2.881

Generative models for clinical applications in computational psychiatry

Full article on Wiley Online Library:   HTML PDF

Can't access this content? Tell your librarian.

Despite the success of modern neuroimaging techniques in furthering our understanding of cognitive and pathophysiological processes, translation of these advances into clinically relevant tools has been virtually absent until now. Neuromodeling represents a powerful framework for overcoming this translational deadlock, and the development of computational models to solve clinical problems has become a major scientific goal over the last decade, as reflected by the emergence of clinically oriented neuromodeling fields like Computational Psychiatry, Computational Neurology, and Computational Psychosomatics. Generative models of brain physiology and connectivity in the human brain play a key role in this endeavor, striving for computational assays that can be applied to neuroimaging data from individual patients for differential diagnosis and treatment prediction. In this review, we focus on dynamic causal modeling (DCM) and its use for Computational Psychiatry. DCM is a widely used generative modeling framework for functional magnetic resonance imaging (fMRI) and magneto‐/electroencephalography (M/EEG) data. This article reviews the basic concepts of DCM, revisits examples where it has proven valuable for addressing clinically relevant questions, and critically discusses methodological challenges and recent methodological advances. We conclude this review with a more general discussion of the promises and pitfalls of generative models in Computational Psychiatry and highlight the path that lies ahead of us.

This article is categorized under:

  • Neuroscience > Computation
  • Neuroscience > Clinical Neuroscience
Taxonomy for different disciplines in the computational neurosciences and their relation to clinical questions. (Reprinted with permission from Stephan, Siemerkus, Bishop, & Haker, . Copyright 2017 Hogrefe AG)
[ Normal View | Magnified View ]
Regression DCM (rDCM) as a novel variant of DCM for inferring whole‐brain effective connectivity patterns from fMRI data. (a) Endogenous connectivity architecture (A matrix) among the 66 brain regions from the parcellation reported by Hagmann et al. () as well as the driving inputs, mimicking the effects of visual stimulation in the right and the left visual field (left), as well as an actual “observation” of the endogenous connectivity (right). L = left hemisphere; R = right hemisphere; A = anterior; P = posterior; LVF = left visual field; RVF = right visual field. (b) Parameter recovery of rDCM in terms of the root mean squared error (RMSE) and (c) the number of sign errors (SE) for various combinations of the signal‐to‐noise ratio (SNR) and the repetition time (TR) of the synthetic fMRI data. Results are shown when restricting the analysis to parameter estimates with a non‐negligible effect size (i.e., the 95% Bayesian credible interval of the posterior not containing zero). (Reprinted with permission from Frässle et al. (). Copyright 2017 Elsevier)
[ Normal View | Magnified View ]
Development of computational assays based on DCM for M/EEG for model‐based pathophysiological phenotyping. A conductance‐based DCM with ligand‐gated sodium, calcium, and chloride channels, as well as voltage‐gated potassium and calcium channels was constructed. Shown are the posterior estimates of two ionotropic (AMPA, NMDA) and one potassium channel for a large cohort of 94 healthy controls (dark grey ellipsoids). These serve as a multivariate reference distribution against which a single patient (red ellipsoid), suffering from a mutation affecting the potassium channel gene KCNJ2, could be compared. This patient is placed at the edge of the multivariate distribution, suggesting that DCM could identify the synaptic channel abnormality with high sensitivity and specificity. (Reprinted with permission from Gilbert et al. (). Copyright 2016 Elsevier)
[ Normal View | Magnified View ]
Examples demonstrating generative embedding based on DCM as a formal tool for differential diagnosis and dissection of spectrum disorders. (a) Classification accuracy of the supervised generative embedding approach for various measures. Input features were either based on measures of BOLD activity (light grey), functional connectivity (dark grey), or effective connectivity (blue). For all measures, the balanced accuracy and its 95% posterior probability interval is shown, as well as chance level (50%). Generative embedding based on the posterior means of the model parameters of a plausible DCM significantly outperformed more conventional classification approaches that operated on regional BOLD activity or measures of functional connectivity. Furthermore, balanced accuracy was markedly reduced for biologically unlikely models. (b) Representation of aphasic patients (red) and healthy controls (grey) in the reduced voxel space—that is, the space spanned by the BOLD activity in the three peaks of the most discriminative activation clusters (left), as well as in the reduced generative score space—that is, the space spanned by the three individually most discriminative effective connectivity parameters (right). (c) Results of the unsupervised generative embedding approach based on the variational Bayesian inversion of a Gaussian mixture model, operating on the posterior parameter estimates of a three‐region DCM. Results suggested highest model evidence for the number of clusters being equal to three. (d) Different effective connectivity profiles for the three distinct subgroups. (e) Clusters of the schizophrenic patients differed significantly in the negative symptom severity scores on the Positive and Negative Syndrome Scale (PANSS). (Reprinted with permission from Brodersen et al. (). Copyright 2011 PLOS and Brodersen et al. (). Copyright 2014 Elsevier)
[ Normal View | Magnified View ]
Bayesian model selection (BMS) as a principled framework for differential diagnosis. Given measurements (e.g., clinical observations), the relative plausibility of a set of competing hypotheses (models) of how the observations have been generated, can be evaluated in terms of the posterior model probability. (Reprinted with permission from Stephan, Schlagenhauf, et al. (). Copyright 2017 Elsevier)
[ Normal View | Magnified View ]
Example demonstrating Bayesian model selection (BMS) in DCM as a formal tool for differential diagnosis. Subjects with different forms of grapheme–color synesthesia were analyzed, namely projector synesthesia (left) and associator synesthesia (right). (a) Two alternative hypotheses of the putative effective connectivity underlying synesthesia, formulated as a bottom‐up DCM and (b) top‐down DCM. (d) BMS results with shaded areas representing the posterior probability distribution of the winning model. Results suggest that no (strong) evidence was found for either model when comparing DCMs across the entire populations (grey). However, dividing subjects into projectors (red) and associators (blue) based on their synesthetic experience, the two competing models mapped almost perfectly onto the different subgroups. (c) Posterior densities of modulatory parameters for projectors and (e) associators. (Reprinted with permission from van Leeuwen et al. (). Copyright 2011 Society for Neuroscience)
[ Normal View | Magnified View ]
Schematic summary of key prospective endeavors in Computational Psychiatry and the necessary methodological building blocks. Ultimately, Computational Psychiatry strives to enable generative models of brain activity (and behavior) as computational assays for differential diagnosis and dissection of spectrum disorders in routine clinical practice. (Reprinted with permission from Stephan and Mathys (). Copyright 2014 Elsevier)
[ Normal View | Magnified View ]
Graphical summary of the generative model of DCM for EEG/MEG, representing a single source by a neural mass model based on the Jansen‐and‐Rit model (). The neural mass model comprises three interacting neuronal subpopulations (left). In DCM for EEG/MEG, these subpopulations are taken to mimic excitatory spiny stellate cells in granular layer IV, inhibitory interneurons in supragranular layers II and III, and excitatory deep pyramidal cells in infra‐granular layers V and VI. Subpopulations are interconnected via intrinsic (i.e., within‐source) connections γ1,2,3,4. Dynamics of the neuronal states are described by a set of differential equations (right). The model effectively yields a prediction of the depolarization of pyramidal cells (which is assumed to underlie the measured M/EEG signals) by first transforming average density of presynaptic inputs into an average postsynaptic membrane potential (i.e., convolution), which is then converted into an estimate of the average rate of action potentials fired by each neuronal subpopulation. (Reprinted with permission from David et al. (). Copyright 2006 Elsevier and Moran, Pinotsis, and Friston (). Copyright 2013 Frontiers)
[ Normal View | Magnified View ]
Graphical summary of the generative model of DCM for fMRI, comprising the neuronal and hemodynamic model, as well as the (static) nonlinear BOLD signal change equation. The neuronal state equation is cast as a bilinear differential equation, describing the dynamics of neuronal states as a function of the endogenous connectivity (A matrix), the modulatory influences (B matrix) and driving inputs (C matrix). The neuronal states then enter a cascade of differential equations, which make up the hemodynamic model and describe how neuronal dynamics lead to changes in cerebral blood flow, which, in turn, affect venous blood volume and deoxyhemoglobin content. These two quantities then enter a static BOLD signal observation equation that yields a prediction of BOLD signal time courses. A more comprehensive description is provided elsewhere (Daunizeau, David, & Stephan, ; Friston et al., ; Kahan & Foltynie, ; Stephan et al., ). (Reprinted with permission from Stephan et al. (). Copyright 2015 Elsevier)
[ Normal View | Magnified View ]

Browse by Topic

Neuroscience > Computation
Neuroscience > Clinical Neuroscience

Access to this WIREs title is by subscription only.

Recommend to Your
Librarian Now!

The latest WIREs articles in your inbox

Sign Up for Article Alerts