Interpreting and Quantifying Whole-brain Dynamic Functional Connectivity

Functional connectivity (FC) is usually defined as the correlation between spatially distributed signals measured by Blood Oxygen Level Dependent (BOLD) fMRI, and it provides valuable insights into the large-scale functional organization of the human brain. During recent years, dynamic functional connectivity (dFC), measuring the time-varying property of FC, has emerged as a power tool to extract more information about brain function than the traditional FC analysis. All pairwise dFC estimation between brain regions across the entire brain can be combined to generate a dynamic description of brainwide functional architecture, which is termed as whole-brain dFC analysis. The aim of this thesis is to investigate the behavioral and cognitive implications of whole-brain dFC, and evaluate the efficacy of different dFC representations to identify underlying cognitive processes. Using a multitask dataset, in which subjects engaged and transitioned between four mental states (rest, working memory, visual search, and mental calculation), we show that the whole-brain dFC patterns generated by sliding window approach encode information regarding subject identity and ongoing cognition. We decompose the group-level dFC patterns into a subject-specific FC signature pattern and a task-evoked FC signature pattern, and we show that removing subject-specific FC signature pattern can increase visibility of cognitively relevant mental states in a group-level analysis. Given the prior evidence that subjects with good performance have more compact and stable task-evoked FC patterns, we further hypothesize that dFC patterns may carry fine-grained information that can be used to relate to short-term task engagement levels. We use k-means algorithm as a vector quantization tool to extract two engagement-specific FC patterns representing active engagement and passive engagement. Then, we derive three engagement markers from whole-brain dFC patterns, i.e. dissimilarity between dFC patterns and engagement-specific FC patterns, and brainwide integration level. Those engagement markers are evaluated against windowed task performance using a linear mixed effects model. Significant relationships are observed between engagement markers and windowed task performance for the working memory task only, partially confirming our hypothesis. We also evaluate the efficacy of four dFC representations — namely sliding window correlation, sliding window correlation with L1 regularization, multiplication of temporal derivatives, and dynamic conditional correlation — by comparing the discrimination power of dFC estimates to separate different tasks. We find that moving averaged dynamic conditional correlation produces the best overall results, especially for short window length (WL ≤ 9sec), while all four methods offer comparable performance for commonly-used window length (WL ≥ 30sec).