A unified approach for characterizing static/dynamic connectivity frequency profiles using filter banks

Ashkan Faghiri, Armin Iraji, Eswar Damaraju, Jessica A. Turner, Vince D. CalhounNetwork 2020 · DOI

TL;DR

This paper introduces Filter-Banked Connectivity (FBC), a novel method to analyze brain connectivity across its full frequency spectrum, unifying static and dynamic analyses. FBC overcomes limitations of traditional sliding-window approaches by using filter banks to capture high-frequency connectivity patterns, revealing distinct brain states and differences in connectivity profiles between individuals with schizophrenia and typical controls.

What Problem Does This Paper Address?

Functional connectivity (FC) and functional network connectivity (FNC) are crucial for understanding brain function. Traditionally, these are studied either as static (constant over time) or dynamic (changing over time). Dynamic FNC (dFNC) methods, like sliding-window approaches, often suffer from a low-pass filtering effect, meaning they miss high-frequency connectivity patterns. This limitation prevents a comprehensive understanding of brain dynamics across all frequencies and can lead to an incomplete picture, especially when comparing groups with potentially different connectivity profiles. The authors aim to bridge this gap by developing a method that can analyze connectivity across the entire frequency range in a unified manner.

Methods

The core innovation is Filter-Banked Connectivity (FBC). Unlike standard sliding-window methods (like SWPC) that inherently apply a low-pass filter, FBC uses a bank of filters to decompose the connectivity signal into multiple frequency sub-bands.

Here's a breakdown:

  • Standard Sliding Window (SWPC): Estimates connectivity over a moving window. The authors illustrate that this process, particularly with a rectangular window, acts as a low-pass filter, attenuating high-frequency information.
  • Filter Banks: FBC replaces the single low-pass filter in the SWPC system with an array of filters. These filters are designed to cover a wide range of frequencies, from very low (approaching static connectivity) to higher frequencies.
  • Frequency Tiling in Connectivity Domain: A key methodological point is that FBC performs frequency tiling directly in the connectivity domain, not in the activity domain (time series of brain signals). This is crucial because the relationship between activity frequencies and connectivity frequencies is complex and not fully understood; tiling in the activity domain can distort frequency information.
  • Unified Analysis: By analyzing connectivity across multiple frequency bands simultaneously, FBC allows for the examination of both static (low-frequency) and dynamic (higher-frequency) connectivity patterns within a single framework.
  • Clustering: The resulting multi-band connectivity estimates are then clustered (using k-means) to identify distinct brain network states.

Datasets

  • Toy Example: A simulated dataset was generated using a multivariate normal distribution to illustrate the benefits of FBC, particularly its ability to capture high-frequency connectivity missed by SWPC. This involved simulating time series with distinct low- and high-frequency connectivity states.
  • FBIRN Dataset: A real resting-state fMRI dataset was used, comprising 163 typical controls (TC) and 151 individuals with schizophrenia (SZ). This dataset has been used in previous studies, allowing for comparison.

Key Insights and Ideas

  • High-Frequency Connectivity is Real and Important: Traditional methods miss significant high-frequency connectivity patterns, which can be crucial for understanding brain dynamics and group differences.
  • FBC Unifies Static and Dynamic Analysis: The method naturally incorporates static connectivity (represented by the lowest frequency band) and dynamic connectivity across a broad spectrum, providing a more holistic view.
  • Frequency Profile of Connectivity Matters: Different brain states or conditions might be characterized by distinct frequency profiles of connectivity, not just the presence or absence of connections.
  • Schizophrenia and Connectivity Frequencies: Individuals with schizophrenia show a tendency to spend more time in states exhibiting higher frequency connectivity compared to typical controls, who favor lower frequency states.
  • SWPC's Bias: SWPC is inherently biased towards low-frequency connectivity, making it less sensitive to dynamic changes occurring at higher frequencies.

Key Results

  • Toy Example Success: FBC accurately estimated both low- and high-frequency connectivity states in simulations, outperforming SWPC when high-frequency patterns were present. SWPC struggled to distinguish states when their frequencies fell outside its limited low-pass window.
  • FBIRN Dataset Findings:
    • FBC identified eight distinct connectivity states, some of which were not captured by SWPC due to their higher frequency profiles.
  • Three states showed strong low-frequency profiles, similar to static connectivity.
  • Individuals with schizophrenia (SZ) spent more time in states with higher frequency connectivity (clusters 7 and 8), while typical controls (TC) spent more time in lower frequency states (clusters 2, 3, 4, and 6).
  • Static FNC results for SZ and TC showed similarities to specific low-frequency FBC clusters, reinforcing the link between static and low-frequency dynamic states.
  • SZ subjects showed weaker connections in auditory/visual/somatomotor networks in specific high-frequency states (clusters 7 and 8) compared to TCs.

Limitations and Future Work

  • Windowing for Mean/Std: The initial step of calculating mean and standard deviation using a window can be a limitation if these values change faster than the window can track. Future work could explore other instantaneous connectivity estimators.
  • Noise in Higher Frequencies: Higher frequencies might be more susceptible to noise contamination.
  • Assumption-Free Approach: While the authors aimed to avoid strong assumptions about connectivity frequencies, this means the specific frequency bands chosen for analysis are somewhat arbitrary and could be tailored based on prior knowledge or specific research questions.
  • Statistical Significance: The k-means clustering doesn't inherently distinguish between significant and non-significant connections. Future work could incorporate change-point detection algorithms or built-in statistical tests.
  • Parameter Choices: The choice of filter type (Chebyshev type 2 was used, but Butterworth and elliptic were also tested), window size, and number of clusters can influence results. While the study explored some of these, further investigation into optimal parameter selection is warranted.
  • Exploration of Oscillatory Patterns: The observation of opposite connectivity patterns in certain states suggests a potential for dynamic connectivity to oscillate. Further exploration of this phenomenon and its biological/clinical meaning is suggested.
  • Comparison with Other Estimators: The study primarily compared FBC with SWPC. A more comprehensive comparison with other novel dFNC estimators is needed.

Summary updated

Abstract

Static and dynamic functional network connectivity (FNC) are typically studied separately, which makes us unable to see the full spectrum of connectivity in each analysis. Here, we propose an approach called filter-banked connectivity (FBC) to estimate connectivity while preserving its full frequency range and subsequently examine both static and dynamic connectivity in one unified approach. First, we demonstrate that FBC can estimate connectivity across multiple frequencies missed by a sliding-window approach. Next, we use FBC to estimate FNC in a resting-state fMRI dataset including schizophrenia patients (SZ) and typical controls (TC). The FBC results are clustered into different network states. Some states showed weak low-frequency strength and as such were not captured in the window-based approach. Additionally, we found that SZs tend to spend more time in states exhibiting higher frequencies compared with TCs who spent more time in lower frequency states. Finally, we show that FBC enables us to analyze static and dynamic connectivity in a unified way. In summary, FBC offers a novel way to unify static and dynamic connectivity analyses and can provide additional information about the frequency profile of connectivity patterns.
BibTeX
@inproceedings{faghiri2020unified,
  author = {Ashkan Faghiri and Armin Iraji and Eswar Damaraju and Jessica A. Turner and Vince D. Calhoun},
  title = {{A unified approach for characterizing static/dynamic connectivity frequency profiles using filter banks}},
  booktitle = {Network},
  year = {2020},
  doi = {10.1162/netn\_a\_00155},
}
Figure