Poster F96, Tuesday, March 28, 8:00 – 10:00 am, Pacific Concourse
Task Evoked Dynamics in Whole Brain HMM Brain States
Andrew Quinn1, Eva Patai1,4, Diego Vidarre1,3, Anna Nobre1,2, Mark Woolrich1,3; 1Oxford Centre for Human Brain Activity, University of Oxford, 2Department of Experimental Psychology, University of Oxford, 3Oxford Centre for Functional MRI of the Brain,University of Oxford, 4Institute of Behavioural Neuroscience, University College London.
Estimation of whole brain dynamics is critical for understanding how network interactions subserve rapid cognition, yet to robustly perform such estimation requires time-series much longer than the time-scale of cognitive dynamics of interest. Here we show that Hidden Markov Model (HMM) states can characterise rapid dynamics and efficiently utilise the whole dataset to generate robust estimates of whole brain networks. This is illustrated in the context of a long-term memory paradigm involving spatial and contextual associations. MEG data were collected from 16 participants. The data were then filtered from 3-40Hz and projected into source space using a LCMV beamformer before parcellation into 44 nodes and multivariate leakage correction (Colclough et al 2015). Alpha band power envelopes were used to infer a Gaussian-HMM (Baker et al 2014;Vidaurre et al 2016) to identify transient brain states characterised by patterns of power and/or functional connectivity. Spatial maps of the relative amplitude for each HMM state were computed using the partial correlation between the state time-courses and the amplitude envelopes. Critically, the HMM decomposition is performed without any knowledge of the task conditions or timings within the dataset. To identify task-evoked changes in the HMM states, the state time-courses were epoched, and the average fractional occupancy of each state (i.e. the proportion of trials for which each state is active) was computed across participants for each point in time. The HMM approach robustly identifies dynamic brain networks across large datasets, retaining both a rich description of on-going dynamics and task-evoked responses.
Topic Area: METHODS: Neuroimaging