Frequency modulation increases the specificity of time-resolved connectivity: A resting-state fMRI study

Ashkan Faghiri, Kun Yang, Andréia V. Faria, Koko Ishizuka, Akira Sawa, Tülay Adalı, Vince D. Calhoun

TL;DR

This paper introduces a novel method, SSB+SWPC, that uses single sideband modulation to improve the estimation of time-resolved functional connectivity (trFNC) from fMRI data. By mitigating the low-frequency information loss inherent in traditional sliding window Pearson correlation (SWPC), SSB+SWPC allows for shorter window sizes, leading to more sensitive detection of rapid connectivity changes and better overall performance in both simulations and real fMRI data analysis.

What Problem Does This Paper Address?

Analyzing the dynamic nature of brain activity is crucial for understanding brain function. Time-resolved functional network connectivity (trFNC) aims to capture how brain regions' interactions change over time. A common method for this is Sliding Window Pearson Correlation (SWPC). However, SWPC inherently applies a high-pass filter to the data. This means that to capture rapid changes in connectivity (requiring short windows), important low-frequency information, which is known to be significant in resting-state fMRI (rsfMRI), is filtered out. This creates a trade-off: short windows capture temporal dynamics but lose low-frequency information, while long windows preserve low frequencies but smooth out rapid changes. The authors aim to overcome this limitation, enabling the estimation of trFNC with short windows without sacrificing crucial low-frequency information.

Methods

The core innovation is the integration of Single Sideband Modulation (SSB) from communication theory into the SWPC framework, creating SSB+SWPC.

  • Standard SWPC: This method involves calculating Pearson correlation within a sliding window across time series. It can be conceptually broken down into a high-pass filter (removing the moving average) and a low-pass filter (applied after multiplying the filtered signals). The window size dictates the cutoff frequencies of these filters.
  • SSB+SWPC: This method first applies SSB to the time series before calculating the SWPC. SSB modulates the frequency content of the signal, effectively shifting low-frequency components to higher frequencies. This allows the subsequent SWPC's high-pass filter to operate on these shifted frequencies, preserving the original low-frequency information. The authors mathematically demonstrate that this modulation preserves the correlation estimation as long as the modulated sidebands don't overlap in a specific way.
  • Simulation Framework: The authors designed two sets of simulations:
  • To evaluate SSB+SWPC's performance in estimating time-resolved correlations against true simulated correlations, using Pearson correlation and RMSE as metrics.
  • To evaluate a pipeline using SSB+SWPC for estimating connectivity states (clustering trFNC) and compare it to a similar pipeline using SWPC.
  • Real fMRI Data Analysis: They applied both SSB+SWPC and SWPC to resting-state fMRI data from typical controls (TC) and individuals with first-episode psychosis (FEP). They used metrics like the difference between averaged trFNC and static FNC (sFNC), spectral properties, and clustering-based dwell time to compare the methods and group differences.

Datasets

  • Simulated Data: Two sets of simulated time series were generated to evaluate the methods under controlled conditions. These simulations involved varying correlation frequencies, amplitudes, and state lengths.
  • Real Resting-State fMRI Data: Data from 183 subjects (94 TC, 89 FEP) were analyzed. This dataset has been previously published and analyzed in other works.

Key Insights and Ideas

  • Low-frequency information is critical: The paper strongly emphasizes the importance of low-frequency signals in rsfMRI for capturing brain connectivity dynamics.
  • SSB as a frequency shifter: The core idea of using SSB is to "hide" low-frequency information in higher frequencies, making it invisible to SWPC's high-pass filter while still being recoverable.
  • Trade-off resolution: SSB+SWPC effectively decouples the window size from the need to preserve low-frequency information, allowing for shorter windows to capture rapid dynamics without sacrificing spectral content.
  • Improved sensitivity: The method is argued to be more sensitive to temporal variations in trFNC.
  • Real-world applicability: The study demonstrates the method's utility in analyzing clinical populations (FEP vs. TC), suggesting it can reveal subtle differences in brain connectivity.

Key Results

  • Simulation Performance: SSB+SWPC consistently outperformed SWPC in simulations across various parameters, showing higher Pearson correlation and lower RMSE between estimated and true connectivity. The improvement was particularly noticeable with small window sizes.
  • Real Data Analysis:
    • SSB+SWPC showed a better ability to capture the zero-frequency content (time-averaged FNC) of connectivity compared to SWPC, as indicated by lower mean squared error.
  • The spectral analysis of trFNC estimated by SSB+SWPC was more similar to the ideal low-pass filter transfer function than that of SWPC.
  • In the FEP vs. TC comparison, SSB+SWPC revealed more distinct differences in connectivity states and dwell times between groups. Notably, FEP individuals showed weaker connectivity patterns, while TC individuals spent more time in a state with negative connectivity between subcortical and cortical regions.

Limitations and Future Work

  • Additional Parameter: SSB+SWPC introduces an additional parameter (modulation frequency, $f_m$) compared to SWPC, increasing complexity. Users need to select $f_m$ carefully to avoid aliasing.
  • Band-limited Assumption: The method assumes signals are band-limited. While up-sampling can mitigate this, it introduces another parameter and complexity. The authors suggest exploring this further.
  • Standard Error: The paper focuses on the bias reduction of the estimator, not its standard error. Deriving the standard error for SSB+SWPC is noted as a complex undertaking for future work.
  • Future Directions: The authors suggest exploring the impact of up-sampling more deeply and investigating the relationship between the power law exponent of activity and connectivity time series. They also propose further research into the theoretical underpinnings of the standard error for SSB+SWPC.

Summary updated

Abstract

Representing data using time-resolved networks is valuable for analyzing functional data of the human brain. One commonly used method for constructing time-resolved networks from data is sliding window Pearson correlation (SWPC). One major limitation of SWPC is that it applies a high-pass filter to the activity time series. Therefore, if we select a short window (desirable to estimate rapid changes in connectivity), we will remove important low-frequency information. Here, we propose an approach based on single sideband modulation (SSB) in communication theory. This allows us to select shorter windows to capture rapid changes in the time-resolved functional network connectivity (trFNC). We use simulation and real resting-state functional magnetic resonance imaging (fMRI) data to demonstrate the superior performance of SSB+SWPC compared to SWPC. We also compare the recurring trFNC patterns between individuals with the first episode of psychosis (FEP) and typical controls (TC) and show that FEPs stay more in states that show weaker connectivity across the whole brain. A result exclusive to SSB+SWPC is that TCs stay more in a state with negative connectivity between subcortical and cortical regions. Based on all the results, we argue that SSB+SWPC is more sensitive for capturing temporal variation in trFNC.
BibTeX
@inproceedings{faghiri2024frequency,
  author = {Ashkan Faghiri and Kun Yang and Andréia V. Faria and Koko Ishizuka and Akira Sawa and Tülay Adalı and Vince D. Calhoun},
  title = {{Frequency modulation increases the specificity of time-resolved connectivity: A resting-state fMRI study}},
  booktitle = {Network},
  year = {2024},
  doi = {10.1162/netn\_a\_00372},
}
Figure