Computational neuroimaging strategies for single patient predictions

Abstract

Neuroimaging increasingly exploits machine learning techniques in an attempt to achieve clinically relevant single-subject predictions. An alternative to machine learning, which tries to establish predictive links between features of the observed data and clinical variables, is the deployment of computational models for inferring on the (patho)physiological and cognitive mechanisms that generate behavioural and neuroimaging responses. This paper discusses the rationale behind a computational approach to neuroimaging-based single-subject inference, focusing on its potential for characterising disease mechanisms in individual subjects and mapping these characterisations to clinical predictions. Following an overview of two main approaches - Bayesian model selection and generative embedding - which can link computational models to individual predictions, we review how these methods accommodate heterogeneity in psychiatric and neurological spectrum disorders, help avoid erroneous interpretations of neuroimaging data, and establish a link between a mechanistic, model-based approach and the statistical perspectives afforded by machine learning.

Variational Bayesian inversion for hierarchical unsupervised generative embedding (HUGE)

“…The approach described in this paper -hierarchical unsupervised generative embedding (HUGE) -unifies two important streams of development in neuroimaging: (i) hierarchical models for empirical Bayesian analyses of multi-subject fMRI data (Friston et al, 2016;Lindquist et al, 2017;Sanyal et al, 2012), and (ii) combining generative models of single-subject fMRI data with (un)supervised learning for clinical predictions Brodersen et al, 2011;Stephan et al, 2017). An early version of HUGE was based on computationally demanding MCMC sampling .…”

“…Its unsupervised variant was introduced as a strategy to address a central problem in computational psychiatry: the need to stratify heterogeneous spectrum disorders into pathophysiologically more homogenous subgroups and thus enhance the predictive validity of diagnoses (Stephan et al, 2017). HUGE unifies the original two-step procedure of generative embedding into the inversion of a single hierarchical model.…”

“…This is particularly helpful for dealing with heterogeneous clinical populations that are thought to consist of numerous (but typically poorly known) subgroups Stephan et al, 2017). By contrast, even in the presence of prior evidence for the existence of multiple subgroups in the population, classical generative embedding would use the same priors for the inversion of the DCM of all subjects.…”

“…This type of analysis has the advantage of inferring putative mechanisms underlying neurophysiological and cognitive processes from neuroimaging and behavioral measurements. Such mechanistic accounts are not only highly beneficial for understanding the healthy human brain (for examples, see Piray et al, 2017;Rae et al, 2015;van Leeuwen et al, 2011) but also for identifying possible disease mechanisms in psychiatric disorders and for guiding differential diagnosis in individual patients (for review, see Stephan et al, 2017). Ultimately, this might help to overcome the lack of predictive validity of current symptom-based diagnostic schemes (DSM-5 or ICD-11) of psychiatric disorders ).…”

Neuroplasticity and the brain connectome: what can Jean Talairach’s reflections bring to modern psychosurgery?

“…32 Moreover, as discussed above, structural connectivity and effective connectivity may be coupled with a computational description of behavior to infer (patho)physiological mechanisms underlying clinical features. However, an appealing challenge would be to characterize these mechanisms at the individual level and use them to guide therapeutic intervention (see Stephan et al for a comprehensive review of the computational approach to neuroimaging-based single-subject inference 80 ). Such an approach would dissect the heterogeneity of psychiatric diseases through the definition of subgroups characterized by specific network impairments and thus provide the foundation for targeted surgery.…”