This thesis deals with modeling temporal changes in functional brain connectivity derived from functional magnetic resonance imaging (fMRI). These changes, observed in both task and rest settings, have been coined dynamic functional connectivity (dFC), and are often clustered into a discrete set of so-called dFC states. In the five included research papers, we analyse these repeating patterns of connectivity using Bayesian machine learning methods and relate these to cognitive traits and disease status in different resting-state datasets. In dFC state models, we are faced with many parameter choices, which we in this thesis have tackled using a predictive likelihood framework allowing for quantitative model comparison. Furthermore, this can also be used to assess the relative plausibility of a set of candidate models. We applied this framework to the Wishart mixture model, a probabilistic extension of the sliding-window k-means approach used in many dFC studies. Here, we show that the predictive likelihood can be used to quantify the support for dFC given different window lengths. Furthermore, in another paper we show that the predictive likelihood can be used to choose both the number of states and the model structure in a hidden Markov model (HMM) applied to a highly sampled single subject's resting-state fMRI data. Another way to investigate the relevance of dFC models is to relate them to subject specific cognitive traits or disease status. The former was investigated in a large cohort of healthy subjects' resting-state fMRI data and we found almost no association between the temporal characteristics of the dFC models and the higher order cognitive traits. In another paper we investigated different HMMs ability to distinguish between patients with schizophrenia and healthy controls based on resting-state fMRI data. We found that the simplest characterizations using static FC were adequate for the classiffcation task. Our ndings underline the importance of quantitative evaluation of dFC models and furthermore shows that we need better models that can account for subject variability and noise confounds.
|Number of pages||100|
|Publication status||Published - 2018|
|Series||DTU Compute PHD-2018|