Explicitly Nonlinear Dynamic Functional Network Connectivity In Resting-State fMRI Data

Sara Motlaghian, Ayşenil Belger, Juan Bustillo, Ashkan Faghiri, James M. Ford, Armin Iraji, Kelvin O. Lim, Daniel H. Mathalon, Robyn L. Miller, Bryon A. Mueller, Daniel S. OʼLeary, Godfrey D. Pearlson, Steven G. Potkin, Adrian Preda, T.g. van Erp, Vince D. CalhounbioRxiv 2022 · DOI

TL;DR

This paper introduces a novel method to analyze "explicitly nonlinear" dynamic functional network connectivity (dFNC) in resting-state fMRI data, going beyond traditional linear correlation analyses. By isolating and quantifying these nonlinear relationships, the study reveals distinct brain states and significant differences between healthy controls and individuals with schizophrenia, suggesting nonlinear connectivity patterns are crucial for understanding brain function and dysfunction.

What Problem Does This Paper Address?

Functional connectivity (FC) and its dynamic counterpart (dFNC) are powerful tools for understanding brain function by examining how different brain regions or networks interact over time. However, most existing dFNC research primarily focuses on linear correlations between brain activity time series. This approach overlooks potentially important nonlinear relationships that might capture unique information about brain dynamics. The authors argue that these nonlinear effects are not only present but may be widespread and crucial for a complete understanding of brain function, especially in the context of neurological disorders like schizophrenia, where altered connectivity is a known factor.

Methods

The core of this paper is the development of a method to quantify "explicitly nonlinear" (EN) dynamic functional network connectivity. The approach involves:

  1. Independent Component Analysis (ICA): First, resting-state fMRI data is decomposed into intrinsic connectivity networks (ICNs) using group ICA.
  2. Sliding Window Analysis: Time courses of these ICNs are then analyzed using a sliding window approach to capture dynamic changes.
  3. Linear Correlation Removal: For each window, the linear correlation between pairs of ICN time courses is calculated and then removed.
  4. Normalized Mutual Information (NMI): The remaining "residual" dependency, after linear effects are accounted for, is then measured using Normalized Mutual Information (NMI). NMI quantifies the statistical dependence between two variables, capturing both linear and nonlinear relationships. By applying it to the residuals, the method specifically targets the explicitly nonlinear component of the dependency.
  5. State Detection: K-means clustering is applied to the windowed NMI matrices to identify distinct dynamic states of EN dFNC.

This method is novel in its explicit focus on isolating and quantifying nonlinear dependencies after accounting for linear ones, using NMI on residuals.

Datasets

The study utilized the fBIRN dataset, which includes resting-state fMRI (rsfMRI) data from:

  • 163 healthy controls (HC)
  • 151 patients with schizophrenia (SZ)

The data was collected across seven sites using 3-Tesla scanners.

Key Insights and Ideas

  • Nonlinearities are Widespread and Informative: The study strongly suggests that explicitly nonlinear relationships are not just noise but represent a significant and widespread aspect of brain connectivity, often more so than linear relationships.
  • Complementary Information: Linear and nonlinear dFNC capture different, complementary information about brain states. Neither approach alone provides a complete picture.
  • Schizophrenia and Nonlinear Connectivity: Significant differences in both the presence and co-occurrence of linear and nonlinear states were observed between healthy controls and individuals with schizophrenia, highlighting the potential role of altered nonlinear dynamics in the disorder.
  • Dynamic States: The method successfully identified distinct, structured dynamic states for both linear and nonlinear dFNC, providing a more nuanced view of brain activity over time.

Key Results

  • Three Distinct States: Both linear and explicitly nonlinear dFNC analyses revealed three distinct, highly structured functional states.
  • Group Differences in States: Significant differences were found in the fraction occupancy (time spent in a state) and dwell time (duration in a state) for both linear and nonlinear states between HC and SZ.
  • For instance, in linear states, SZ spent significantly longer in state 3 (low correlation) compared to HC.
  • In nonlinear states, HC tended to spend more time in state 3 (higher EN dependency in VIS and SM), while SZ spent longer in state 2 (higher EN dependency in DM).
  • Nonlinear Relationships More Widespread: Nonlinear relationships were found to be more widespread than linear ones.
  • Visual Network Differences: Healthy controls showed a higher level of explicitly nonlinear dependencies in transient connectivity within the visual network compared to schizophrenia patients.
  • Co-occurrence Matters: Significant differences in the co-occurrence of linear and nonlinear states were observed between HC and SZ, suggesting this interplay is important for understanding the disorder.

Limitations and Future Work

  • NMI Symmetry: The authors acknowledge that NMI is not inherently symmetric with respect to the order of input time courses. They addressed this by averaging results, but it remains a consideration.
  • Positivity/Negativity Interpretation: Unlike linear correlation, NMI does not directly provide information about the "positivity" or "negativity" of a relationship, which is a limitation for interpretation.
  • Window Size Impact: The choice of sliding window size is crucial and can affect the estimation of nonlinear dependencies. While simulations showed good performance with 50 time points, further investigation into the impact of window size and exploring alternative approaches like filter banks are suggested for future work.
  • Further Investigation: The authors suggest exploring the impact of window size and using filter bank approaches to cover a broader range of frequencies for future research. They also noted no significant relationship was found between EN dwell time and SZ symptoms in this study, which could be a direction for future exploration.

Summary updated

Abstract

ABSTRACT Most dynamic functional connectivity in fMRI data is focused on linear correlations, and to our knowledge, no study has studied whole brain explicitly nonlinear dynamic relationships within the data. While some approaches have attempted to study overall connectivity more generally using flexible models, we are particularly interested in whether the non-linear relationships, above and beyond linear, are capturing unique information. This study thus proposes an approach to assess the explicitly nonlinear dynamic functional network connectivity derived from the relationship among independent component analysis time courses. Linear relationships were removed at each time point to evaluate, typically ignored, explicitly nonlinear dFNC using normalized mutual information. Simulations showed the proposed method accurately estimated NMI over time, even within relatively short windows of data. Results on fMRI data included 151 schizophrenia patients, and 163 healthy controls showed three unique, highly structured, mostly long-range, functional states that also showed significant group differences. This analysis identifies a higher level of explicitly nonlinear dependencies in transient connectivity within the visual network in healthy controls compared to schizophrenia patients. In particular, nonlinear relationships tend to be more widespread than linear ones. We also find highly significant differences in the relative co-occurrence of linear and explicitly nonlinear states in HC and SZ, suggesting these may be an important aspect of the disorder. Overall, this work suggests that quantifying nonlinear dependencies of dynamic functional connectivity may provide a complementary and potentially valuable tool for studying brain function by exposing relevant variation that is typically ignored.
BibTeX
@misc{motlaghian2022explicitly,
  author = {Sara Motlaghian and Ayşenil Belger and Juan Bustillo and Ashkan Faghiri and James M. Ford and Armin Iraji and Kelvin O. Lim and Daniel H. Mathalon and Robyn L. Miller and Bryon A. Mueller and Daniel S. OʼLeary and Godfrey D. Pearlson and Steven G. Potkin and Adrian Preda and T.G. van Erp and Vince D. Calhoun},
  title = {{Explicitly Nonlinear Dynamic Functional Network Connectivity In Resting-State fMRI Data}},
  year = {2022},
  doi = {10.1101/2022.06.22.497262},
  howpublished = {bioRxiv},
}
Figure