Spatial Dynamic Functional Connectivity Analysis Identifies Distinctive Biomarkers in Schizophrenia

Suchita Bhinge, Qunfang Long, Vince D. Calhoun, Tülay AdalıFrontiers in Neuroscience 2019 · DOI

Abstract

Dynamic functional network connectivity (dFNC) analysis is a widely-used method to study associations between dynamic functional correlations and cognitive abilities. Traditional methods analyze time-varying associations of different spatial networks while assuming that the spatial network itself is stationary. However, there has been very little work focused on voxelwise spatial variability. Exploiting the variability across both the temporal and spatial domains provides a more promising direction to obtain reliable dynamic functional patterns. In this work, we implement subsequent extraction of exemplars and dynamics using a constrained independent vector analysis (SED-cIVA), a data-driven method that efficiently estimates spatial and temporal dynamics from large-scale resting-state fMRI data. Our results indicate that for resting-state fMRI data, spatial dFNC (sdFNC) patterns were able to better classify patients and controls, and yield more distinguishing features compared with temporal dFNC (tdFNC) patterns. We also observe that patients with schizophrenia tend to switch to or stay in a state corresponding to a hyperconnected brain network.

Introduction

Functional Magnetic Resonance Imaging (fMRI) is a cornerstone of modern neuroscience, allowing researchers to study Dynamic Functional Network Connectivity (dFNC)—the way different brain regions coordinate over time. Traditionally, dFNC models have operated under the assumption of "spatial stationarity," meaning that while the strength of connections between regions fluctuates, the physical boundaries and shapes of the networks themselves remain fixed.

However, recent evidence suggests that the spatial maps of functional networks are also dynamic. This study addresses the computational challenges of capturing these spatio-temporal shifts and investigates whether spatial dynamic functional network connectivity (sdFNC) can serve as a more effective biomarker for schizophrenia than traditional temporal methods.

Methodology

The authors propose a novel framework termed SED-cIVA (Subsequent Extraction of Exemplar and Dynamic components using constrained Independent Vector Analysis). The process involves two primary stages:

  1. Exemplar Extraction: Group Independent Component Analysis (GICA) is used to identify 17 stable "exemplar" networks across the study population.
  2. Dynamic Estimation: A parameter-tuned constrained IVA is employed to allow these spatial maps to adapt within sliding time windows, capturing how a network’s spatial footprint expands or shifts over time.
The SED-cIVA technique and post-analysis workflow.

To evaluate the diagnostic utility of these features, the researchers applied a Naïve Bayes classifier to the COBRE dataset, comparing the performance of sdFNC against temporal dFNC (tdFNC). Additionally, k-means clustering was used to identify recurring "brain states."

Results and Discussion

The findings demonstrate a clear superiority of spatial features in identifying schizophrenia. While tdFNC features performed near the level of random chance, sdFNC patterns provided highly discriminative signals for classification.

Analysis of state dynamics revealed that patients with schizophrenia spend a significantly higher fraction of time and have longer dwell times in a specific "hyperconnected" state (State 2). This state is characterized by abnormally high positive correlations between the visual, frontal, and anterior Default Mode Network (DMN) components.

Example of spatial dynamics and network activation maps.

Crucially, these sdFNC patterns were linked to clinical symptoms; the hyperconnectivity observed between the DMN and frontal regions correlated with paranoia traits and impulse control abnormalities in the patient group. In contrast, healthy controls spent more time in a more balanced connectivity state (State 1).

Conclusion

This study highlights that the spatial variability of functional networks is a rich source of information often overlooked in traditional connectivity studies. By utilizing the SED-cIVA method, researchers were able to identify distinctive spatial biomarkers that outperform temporal ones in classifying schizophrenia. Future work may involve applying more complex machine learning architectures to these spatial features to further enhance diagnostic precision.

Summary updated

Abstract

Dynamic functional network connectivity (dFNC) analysis is a widely-used to study associations between dynamic functional correlations and cognitive abilities. Traditional methods analyze time-varying association of different spatial networks while assuming that the spatial network itself is stationary. However, there has been very little work focused on voxelwise spatial variability. Exploiting the variability across both the temporal and spatial domains provide a more promising direction to obtain reliable dynamic functional patterns. However, methods for extracting time-varying spatio-temporal patterns from large-scale functional magnetic resonance imaging (fMRI) data present some challenges, such as degradation in performance with respect to increase in size of the data, estimation of the number of dynamic components, and the potential sensitivity of the resulting dFNCs to selection of the networks. In this work, we implement subsequent extraction of exemplars and dynamics using a constrained independent vector analysis, a data-driven method that efficiently estimates spatial and temporal dynamics from large-scale resting-state fMRI data. We explore the benefits of analyzing spatial dFNC (sdFNC) patterns over temporal dFNC (tdFNC) patterns in the context of differentiating healthy controls and patients with schizophrenia. Our results indicate that for resting-state fMRI data, sdFNC patterns were able to better classify patients and controls, and yield more distinguishing features compared with tdFNC patterns. We also estimate structured patterns of connectivity/states using sdFNC patterns, an area that has not been studied so far, and observe that sdFNC was able to successfully capture distinct information from healthy controls and patients with schizophrenia. In addition, sdFNC patterns were also able to identify functional patterns that associate with signs of paranoia and abnormalities in the patients group. We also observe that patients with schizophrenia tend to switch to or stay in a state corresponding to a hyperconnected brain network.
BibTeX
@inproceedings{bhinge2019spatial,
  author = {Suchita Bhinge and Qunfang Long and Vince D. Calhoun and Tülay Adalı},
  title = {{Spatial Dynamic Functional Connectivity Analysis Identifies Distinctive Biomarkers in Schizophrenia}},
  booktitle = {Frontiers in Neuroscience},
  year = {2019},
  doi = {10.3389/fnins.2019.01006},
}
Figure