VBA: a probabilistic treatment of nonlinear models for neurobiological and behavioural data.

Daunizeau J, Adam V, Rigoux L. · 2014

Read VBA: a probabilistic treatment of nonlinear models for neurobiological and behavioural data. by Daunizeau J, Adam V, Rigoux L. (2014) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc3900378.txt.

Introduction

Jean Daunizeau, Vincent Adam, and Lionel Rigoux present a modeling toolbox in PLoS Computational Biology, with the record dated 2014. Brain and behavior models have grown rich and nonlinear, with unknown settings that must be fit to data. Without a formal stats frame, such model studies are hard to repeat or refute.

Their toolbox, VBA, gives one probabilistic answer to three shared jobs. It can simulate data from a model, fit settings and pick between rival models, and tune study design to sharpen the next test. The tools are built to be generic, fast, and robust across model types. The authors aim at a quantitative science of mind and brain where models earn their keep against data, patient by patient where illness varies widely.

Catalog record

CREATORDaunizeau J, Adam V, Rigoux L.
YEAR2014
COLLECTIONPAPERS
FORMATTXT
FILE SIZE69 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC3900378

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

VBA: A Probabilistic Treatment of Nonlinear Models for Neurobiological and Behavioural Data By Jean Daunizeau, Vincent Adam, Lionel Rigoux, Andreas Prlic Published in: PLOS Computational Biology DOI: 10.1371/journal.pcbi.1003441 ABSTRACT This work is in line with an on-going effort tending toward a computational (quantitative and refutable) understanding of human neuro-cognitive processes. Many sophisticated models for behavioural and neurobiological data have flourished during the past decade. Most of these models are partly unspecified (i.e. they have unknown parameters) and nonlinear. This makes them difficult to peer with a formal statistical data analysis framework. In turn, this compromises the reproducibility of model-based empirical studies. This work exposes a software toolbox that provides generic, efficient and robust probabilistic solutions to the three problems of model-based analysis of empirical data: (i) data simulation, (ii) parameter estimation/model selection, and (iii) experimental design optimization. "This is a PLOS Computational Biology Software Article" INTRODUCTION Spectrum diseases in psychiatry, such as schizophrenia or depression, display profound heterogeneity with regard to the underlying pathophysiological mechanisms, requiring the development of models that can infer subject-specific mechanisms from neurophysiological and/or behavioural data [1]–[3]. Developing quantitative approaches that can do this will require merging expert knowledge on neurobiology, biophysical generation of neuroimaging signals, cognitive psychology, and statistical data analysis, to mention but a few. Recall that most cerebral mechanisms can be described at two levels of abstraction: - The cognitive or functional level is concerned with the information processing that is needed to explain behavioural measurements (e.g., choices, reaction times) or subjective reports (e.g., emotions, thoughts). - The neurobiological or physiological level is related to the neurobiological substrate of the system. Imaging neuroscience or neuroimaging (e.g., electroencephalography/magnetoencephalography or EEG/MEG, functional Magnetic Resonance Imaging or fMRI) is capable of observing (non-invasively) certain biophysical characteristics of this biological substrate. These typically map to two classes of models, i.e. (i) formal models of perception, learning and decision making that predict behavioural responses, and (ii) biophysically realistic models that describe how electrophysiological activity propagate through neural networks. The issue with such models is that they are based upon mechanisms that are usually both hidden (they are not directly accessible from experimental data) and nonlinear (this is the curse of realism). As a consequence, one requires sophisticated statistical approaches that can deal efficiently with parameter estimation and model selection (given experimental data). If only, these are necessary to capture the inter-individual variability of neurophysiological and behavioural responses. More generally, such schemes would embed the models into the data analysis and act as a “mathematical microscope” that is capable of unravelling mechanisms, which are hidden deep within experimental data [4]. This article describes a (matlab) software toolbox that is designed to perform such model-based analyses of neuroimaging and behavioural data. Essentially, it consists of a probabilistic model inversion scheme that borrows from disciplines such as inverse problems, statistical physics and machine learning. More precisely, the toolbox implements a variational Bayesian approach (VBA) that can deal with a very general class of generative models, which we describe below. In brief, VBA address the following issues: - performing efficient and robust parameter estimation on nonlinear models; - providing quantitative diagnostics of model fitting (including summary statistics that can be used for model comparison); - optimizing the experimental design in the aim of maximizing the statistical power of model-based data analysis; - assessing the results reproducibility at the group level (e.g., between-groups and between-conditions model comparisons). In addition, the toolbox gathers a growing number of established and novel biophysical and cognitive models, which capture a broad range of phenomena. Among these are deterministic and stochastic variants of dynamic causal models for fMRI data (see, e.g. [5] for a recent review) and bayesian models for human learning and decision making [6]–[7]. Importantly, VBA includes diagnostic analyses that can be used directly to refine such models, i.e. to account for yet unpredicted features of experimental data (cf. Volterra decompositions). We will demonstrate this below. Our ambition is twofold: (i) to disseminate models and methods that serve experimental purposes, and (ii) to provide a flexible platform, which modellers can easily contribute to. This paper is organized as follows. In the next section, we will describe the design and implementation of the toolbox. In particular, we will describe the class of generative models that the VBA toolbox can handle, expose the main aspects of its algorithmic treatment, and summarize the organization of the code. In section “Results”, we will demonstrate its capabilities through analyses of deposited test data. In section “Availability and future directions”, we will discuss limitations and on-going developments of this work. DESIGN AND IMPLEMENTATION In brief, the toolbox furnishes solutions to the three canonical problems of model-based data analysis, which relate to the experimental cycle (cf. Figure 1). One starts with a set of competing hypotheses, e.g.: is the incentive value discounted (or not) by the amount of cognitive effort that is required to obtain the reward? This question will eventually be framed in terms of a model comparison (e.g., model 1 -with effort discounting- versus model 2 -without effort discounting-). First, one has to be able to predict behavioural and/or biological signals from the candidate models. This simply means simulating, e.g. people's choices under model 1 and 2. Second, one may want to optimize the experimental design, in the aim of best discriminating the candidate models. In our example, this naturally involves manipulating both the reward that is at stake and the amount of effort. However, there might be an optimal manipulation of these two factors, such that models 1 and 2 yield radically different predictions. Third, one proceeds to parameter estimation and/or model selection, given the acquired experimental data. Here, statistical inference follows the variational Bayesian treatment of the candidate models (see below). [Figure: The experimental cycle.The experimental cycle summarizes the interaction between modelling, experimental work and statistical data analysis. One starts with new competing hypotheses about a system of interest. These are then embodied into a set of candidate models that are to be compared with each other given empirical data. One then designs an experiment that is maximally discriminative with respect to the candidate models. Data acquisition and analysis then proceed, the conclusion of which serves to generate a new set of competing hypotheses, etc… Adapted from [14].] Data analysis results can now serve to identify a new set of competing hypotheses, which then triggers a new experimental cycle… The toolbox is designed so that data simulation, parameter estimation, model selection and design optimization can be handled automatically, given standardized information about the model (see below). Stochastic nonlinear state-space models Any Bayesian data analysis relies upon a generative model, i.e. a probabilistic description of the mechanisms by which observed data are generated. Such descriptions can go from simple static linear models to nonlinear dynamical models in continuous time. This toolbox does not invert any generative model: it has been developed to deal with stochastic nonlinear state-space models. As we will see below, this class of generative models grandfathers most models used in the literature. It is defined by a joint probability distribution over the following set of variables: - y: the data time-series - x: the hidden states time-series - : the initial conditions of the system - u:

Read the complete text