Recent neuroimaging research has demonstrated that the brain has a highly adaptable functional network structure (Hampshire & Sharp, 2015; Matthews & Hampshire, 2016). Different networks are specialised to different functions(Damoiseaux et al., 2006; Dosenbach, Fair, Cohen, Schlaggar, & Petersen, 2008; Dosenbach et al., 2006; Rosazza & Minati, 2011), but the diversity of tasks that we face in everyday life can tap these functions in different combinations. This leads to the observations of ‘dynamic network states’. Simply put, the vast array of possible tasks that we can perform is reflected by the diverse conjunctions of network configurations that the brain can transiently express. These conjunctions are visible to fMRI as transient patterns of correlated activity across brain regions (Cole et al., 2013).
In a recent study (see Figure) we demonstrated that the dynamic network states that are expressed when different types of working-memory task are being performed are highly consistent across cohorts and individuals. We trained multivariate machines to classify working-memory domains (e.g., object, number or location) and stages (encode, maintain or recall). Machine trained on data from one cohort were able to classify these aspects of working memory with near-ceiling (90%) accuracy when applied to connectivity graphs from a completely independent cohort.
The high reliability of this approach holds great promise for clinical research and diagnostics because it contrasts favourably with the noisy/un-reliable contrasts from classic neuroimaging, which primarily focused on activity within individual brain regions or connections. It also presents the challenge of how measures of complex states may be simplified/summarised into tractable diagnostic markers, e.g., for sub-classifying patients or determining the impact of interventions on macroscopic brain dynamics.
This project will develop and validate machine-learning methods that can quantify/sub-classify abnormalities in patients’ dynamic functional network profiles. This will be with a focus on selective attention and response inhibition; these are key aspects of cognition that are affected in many clinical populations, e.g., psychiatric groups like ADHD, substance abusers and pathological gamblers, and neurological populations such as traumatic brain injury and stroke patients.
The supervisors collectively have >1000 fMRI datasets for fMRI tasks that fractionate aspects of selective attention (Hampshire & Owen, 2006) and response inhibition (Aron & Poldrack, 2006; Aron, Robbins, & Poldrack, 2004; Erika-Florence, Leech, & Hampshire, 2014). These include healthy controls, older adults, patients and pharmacological trials. To date, the data have been analysed with simple univariate statistics. The student will use cutting-edge machine-learning methods to characterise the dynamic connectivity states that are evident under different task conditions in normal healthy adults. They will develop univariate distance measures for these states and use them as a framework for characterising the network abnormalities that are evident in clinical populations with cognitive-control problems. They will determine how pharmacological treatments that are used to treat these populations modulate dynamic-network states.
The resultant measures have the potential for far reaching impact in clinical research and assessment, e.g., as a framework for stratifying/sub-classify patients based on underlying network disruption, measuring intervention effects on the brain and guiding individualised treatments.