Characterizing Whole Brain Temporal Variation of Functional Connectivity via Zero and First Order Derivatives of Sliding Window Correlations

Flor A. Espinoza, Victor M. Vergara, Eswar Damaraju, Kyle G. Henke, Ashkan Faghiri, Jessica A. Turner, Aysenil A. Belger, Judith M. Ford, Sarah McEwen, Daniel H. Mathalon, Bryon A. Mueller, Steven G. Potkin, Adrian Preda, Jatin G. Vaidya, Theo G.m. van Erp, Vince D. CalhounFrontiers in Neuroscience 2019 · DOI

TL;DR

This paper introduces a novel method, temporal variation of functional network connectivity (tvFNC), that enhances the analysis of dynamic brain connectivity by incorporating the rate of change (first derivative) of connectivity patterns. This approach reveals more nuanced differences in brain states, particularly between healthy controls and schizophrenia patients, than traditional methods.

What Problem Does This Paper Address?

Functional Magnetic Resonance Imaging (fMRI) allows us to study the brain's "connectome" – how different brain regions communicate. While static measures exist, the brain is inherently dynamic, with connectivity changing over time. Existing methods for analyzing these dynamic changes, like Dynamic Functional Network Connectivity (dFNC), use a "sliding window" approach to capture connectivity patterns. However, these methods primarily look at the connectivity patterns themselves (zero-order changes) and can struggle with noisy data or subtle differences between states. This can lead to similar states being indistinguishable or important group differences being missed. The authors aim to improve the sensitivity and accuracy of dynamic connectivity analysis by incorporating information about how quickly connectivity is changing.

Methods

The core innovation is the temporal variation of functional network connectivity (tvFNC) method. It builds upon the standard dFNC approach, which uses a sliding window to calculate functional connectivity (Pearson correlation) at multiple time points.

Here's how tvFNC extends dFNC:

  • Standard dFNC: Calculates windowed functional connectivity (wFNC) over time. These wFNC patterns are then clustered to identify recurring "states" of connectivity. This is essentially looking at the "zero-order" changes in connectivity.
  • tvFNC:
    1. Calculates Derivatives: In addition to the wFNC, tvFNC computes the first-order time derivatives of these wFNC vectors. This captures the instantaneous rate of change (increase or decrease) in connectivity between brain regions.
  • Concatenates Data: The wFNC data (zero-order derivatives) and the derivative data (first-order derivatives) are concatenated together.
  • Clustering: This combined dataset is then subjected to a clustering algorithm (k-means) to identify recurring states that now incorporate both the connectivity pattern and its rate of change.

This approach allows the method to differentiate between states that might look similar in terms of connectivity strength but differ in how rapidly that connectivity is evolving.

Datasets

  • Simulated Data: A simulated dataset was generated to validate the tvFNC method. This dataset was designed to mimic real fMRI data with a controlled number of subjects and time-course data, allowing for systematic evaluation of the method's performance under known conditions.
  • Resting-State fMRI Data: The study utilized resting-state fMRI data from the Functional Imaging Biomedical Informatics Research Network (FBIRN) Phase III study. This dataset included:
  • Healthy Controls (HC): 163 participants.
  • Schizophrenia (SZ) Patients: 151 participants.
  • The data was acquired across multiple sites using 3T Siemens and GE scanners.

Key Insights and Ideas

  • Derivatives Matter: The rate of change in brain connectivity (its derivatives) carries significant information that is not captured by looking at connectivity patterns alone.
  • Enhanced State Differentiation: Incorporating derivatives allows for a finer-grained distinction between connectivity states, especially for states that are otherwise very similar. This is crucial for understanding subtle dynamic shifts.
  • Increased Sensitivity to Group Differences: The tvFNC method proved more sensitive in detecting differences between healthy controls and schizophrenia patients, uncovering group distinctions in states and their derivatives that were not apparent with standard dFNC.
  • Complementary Information: The derivatives of connectivity states can provide complementary information, sometimes showing opposite trends (e.g., one state increasing connectivity while its derivative shows a decrease) that offer a more complete picture of dynamic brain function.

Key Results

  • Simulation Validation: The tvFNC method, when applied to simulated data, demonstrated a clearer identification of similar clusters with different temporal behaviors compared to the standard dFNC approach. The Elbow criterion for cluster selection also showed a smoother decay for tvFNC, suggesting better data representation.
  • Resting-State fMRI Findings:
    • Five distinct dynamic connectivity states were identified in the resting-state fMRI data, consistent with previous dFNC studies.
  • tvFNC revealed additional patterns in the connectivity derivatives, complementing the identified states.
  • Significant group differences in connectivity dynamism were observed between HC and SZ participants. Notably, SZ individuals spent more time in a state characterized by weak connectivity (State 5), while HC individuals spent more time in states with stronger within- and between-connectivity (States 2 and 3).
  • tvFNC uncovered group differences in the derivatives of connectivity states, particularly in states 1-3, highlighting differences in the rate of change of connectivity between HC and SZ groups.

Limitations and Future Work

  • Limitations:
    • Pearson Correlation: The study relies on Pearson correlation, a lower-order statistic. Future work could explore higher-order statistics like mutual information for a more comprehensive analysis.
  • Window Length Selection: The choice of window length for calculating connectivity is critical. The study used a fixed window size (22 TRs), and future research should investigate how different window sizes impact state derivative changes.
  • Scan Duration: The resting-state fMRI data was collected for a relatively short duration (5.4 minutes). Longer scan durations might reveal more critical time points where FNC states reach a more stable or "stationary" behavior.
  • Future Work:
    • Investigate the use of higher-order statistics for connectivity analysis.
  • Explore the impact of different window sizes on tvFNC results.
  • Analyze longer resting-state fMRI datasets to capture more prolonged dynamic states.

Summary updated

Abstract

Brain functional connectivity has been shown to change over time during resting state fMRI experiments. Close examination of temporal changes have revealed a small set of whole-brain connectivity patterns called dynamic states. Dynamic functional network connectivity (dFNC) studies have demonstrated that it is possible to replicate the dynamic states across several resting state experiments. However, estimation of states and their temporal dynamicity still suffers from noisy and imperfect estimations. In regular dFNC implementations, states are estimated by comparing connectivity patterns through the data without considering time, in other words only zero order changes are examined. In this work we propose a method that includes first order variations of dFNC in the searching scheme of dynamic connectivity patterns. Our approach, referred to as temporal variation of functional network connectivity (tvFNC), estimates the derivative of dFNC, and then searches for reoccurring patterns of concurrent dFNC states and their derivatives. The tvFNC method is first validated using a simulated dataset and then applied to a resting-state fMRI sample including healthy controls (HC) and schizophrenia (SZ) patients and compared to the standard dFNC approach. Our dynamic approach reveals extra patterns in the connectivity derivatives complementing the already reported state patterns. State derivatives consist of additional information about increment and decrement of connectivity among brain networks not observed by the original dFNC method. The tvFNC shows more sensitivity than regular dFNC by uncovering additional FNC differences between the HC and SZ groups in each state. In summary, the tvFNC method provides a new and enhanced approach to examine time-varying functional connectivity.
BibTeX
@inproceedings{espinoza2019characterizing,
  author = {Flor A. Espinoza and Victor M. Vergara and Eswar Damaraju and Kyle G. Henke and Ashkan Faghiri and Jessica A. Turner and Aysenil A. Belger and Judith M. Ford and Sarah McEwen and Daniel H. Mathalon and Bryon A. Mueller and Steven G. Potkin and Adrian Preda and Jatin G. Vaidya and Theo G.M. van Erp and Vince D. Calhoun},
  title = {{Characterizing Whole Brain Temporal Variation of Functional Connectivity via Zero and First Order Derivatives of Sliding Window Correlations}},
  booktitle = {Frontiers in Neuroscience},
  year = {2019},
  doi = {10.3389/fnins.2019.00634},
}
Figure