Abstract:Numerous dynamic functional connectivity (dFC) methods have been proposed to study time-resolved network reorganization in rest and task fMRI. However, a comprehensive comparison of their performance is lacking. In this study, we compared the efficacy of seven dFC methods (and their enhanced versions) to track transient network reconfiguration using simulation data. The seven methods include flexible least squares (FLS), dynamic conditional correlation (DCC), general linear Kalman filter (GLKF), multiplication of temporal derivatives (MTD), sliding-window functional connectivity with L1-regularization (SWFC), hidden Markov models (HMM), and hidden semi-Markov models (HSMM). Multiple datasets of non-fMRI-BOLD and fMRI-BOLD signals with predefined covariance structures, signal-to-noise ratio levels, and sojourn time distributions were simulated. We adopted inter-subject analysis to eliminate the effects of signals of non-interest, resulting in enhanced methods: ISSWFC, ISMTD, ISDCC, ISFLS, ISKF, ISHMM, and ISHSMM. Efficacy was defined as the spatiotemporal association between simulated and estimated data. We found that all enhanced dFC methods outperformed their original versions. Efficacies depend on several factors, such as considering the neurovascular effect in simulated data, the covariance structure between two time series, state sojourn distribution, and signal-to-noise ratio levels. These results highlight the importance of selecting appropriate dFC methods in fMRI study.

What problem does this paper attempt to address?

The problem this paper attempts to address is: systematically evaluating the performance of various dynamic functional connectivity (dFC) methods on simulated data to determine their effectiveness and applicability in tracking transient network reorganization. Specifically, the researchers compared seven dFC methods and their enhanced versions (by incorporating inter-subject analysis (ISA) to improve signal-to-noise ratio) on different types of simulated datasets. These methods include: 1. **Flexible Least Squares (FLS)** 2. **Dynamic Conditional Correlation (DCC)** 3. **General Linear Kalman Filter (GLKF)** 4. **Multiplication of Temporal Derivatives (MTD)** 5. **Sliding-Window Functional Connectivity with L1-regularization (SWFC)** 6. **Hidden Markov Models (HMM)** 7. **Hidden Semi-Markov Models (HSMM)** The researchers generated two types of simulated datasets: 1. **Non-fMRI-BOLD signals**: Without considering neurovascular effects, used to evaluate the methods' performance on pure time series data. 2. **fMRI-BOLD signals**: Considering neurovascular effects, generated by convolving the blood-oxygen-level-dependent (BOLD) response function, used to evaluate the methods' performance on actual fMRI data. These datasets included different covariance structures, signal-to-noise ratio levels, and state dwell time distributions. Through these simulated data, the researchers aimed to comprehensively evaluate the effectiveness of different dFC methods and explore factors affecting their performance, such as neurovascular effects, covariance structure, state dwell time distribution, and signal-to-noise ratio levels. Ultimately, the research results emphasize the importance of selecting appropriate dFC methods in fMRI studies.

A systematic evaluation of dynamic functional connectivity methods using simulation data