Structural connectivity-informed dynamic estimation (STRiDE): Multimodal connectivity constrained ICA

Mahshid Fouladivanda, Armin Iraji, Lei Wu, Jiayu Chen, Pablo Andrés-Camazón, Theodorus G.m. van Erp, Ayşenil Belger, Godfrey D. Pearlson, Tülay Adalı, Vince D. CalhounbioRxiv 2025 · DOI

TL;DR

This paper introduces STRiDE, a novel multimodal approach that integrates structural and functional brain connectivity to better understand dynamic brain states. By using structural information as a guide, STRiDE enhances the sensitivity and robustness of detecting brain connectivity differences, particularly in disorders like schizophrenia.

What Problem Does This Paper Address?

Understanding how the brain's functional connectivity changes over time (dynamic functional connectivity or dFNC) is crucial for grasping cognitive processes and identifying disruptions in neurological disorders. However, previous methods often relied on unimodal data (either functional or structural alone) or used structural information in a limited, deterministic way. This can hinder accurate estimation of dynamic brain states, especially in conditions like schizophrenia where both structural and functional disruptions are present. The authors aim to bridge this gap by developing a more sophisticated multimodal approach that leverages the complementary information from both structural and functional brain networks.

Methods

The core of this work is the Structural Connectivity-Informed Dynamic Estimation (STRiDE) model. It's a data-driven, multimodal approach that builds upon Independent Component Analysis (ICA). Here's a breakdown of the key steps:

  1. Structural Connectivity Bases Estimation: They first derive "structural bases" from individual structural connectivity data using group ICA (GICA). These bases represent common, coherent patterns of white matter pathways across subjects, acting as a low-dimensional representation of structural architecture.
  2. Dynamic Functional Connectivity (dFNC) Estimation: Standard dFNC is computed from resting-state fMRI data using a sliding window approach.
  3. Multimodal Connectivity Constrained ICA: This is the novel part. STRiDE uses the estimated structural bases as "references" or "priors" to constrain the ICA decomposition of dFNC. This means the model is guided by the underlying structural architecture when identifying dynamic functional states. The optimization process aims to maximize statistical independence of the functional components while ensuring they are similar to the structural bases.
  4. State Vector Analysis: After identifying the dynamic states (STRiDEs), the authors analyze their temporal dynamics using metrics like frequency, dwell time, occupancy rate, and transition probability.

Datasets

The study uses two main types of data:

  • Simulated Data: A synthetic dataset was generated to evaluate the model's performance and compare it to single-modal approaches. This allowed for controlled introduction of group differences and noise.
  • Real Data: Resting-state fMRI (rs-fMRI) and diffusion MRI (dMRI) data from the Functional Biomedical Informatics Research Network (FBIRN) dataset were used. This dataset includes 311 participants: 162 healthy controls (HC) and 149 individuals with schizophrenia (SZ).

Key Insights and Ideas

  • Structural Priors Enhance Sensitivity: Integrating structural connectivity as priors significantly improves the model's ability to detect group differences (e.g., between HC and SZ) compared to using functional data alone. This is crucial for identifying subtle disruptions.
  • Robustness to Noise: The multimodal approach (STRiDE) is more robust to noise than single-modal methods, consistently identifying more significant connections with less variability. This is a critical advantage for real-world neuroimaging data.
  • Hierarchical Brain Organization: The identified STRiDEs reflect a spectrum of brain network configurations, from localized unimodal segregation to distributed trans-modal integration, mirroring known principles of cortical organization and suggesting a hierarchical structure.
  • State-Specific Dysconnectivity in Schizophrenia: Schizophrenia-related disruptions are not uniform but appear to be state-dependent, with different STRiDEs capturing distinct facets of the disorder. This provides a more nuanced understanding than traditional single-modality approaches.
  • Link Between Structure and Dynamics: The strong correspondence between estimated STRiDEs and structural bases validates the model's effectiveness in leveraging structural information to inform dynamic functional states.

Key Results

  • Simulation Study: STRiDE demonstrated significantly enhanced sensitivity in detecting group differences and improved robustness to noise compared to single-modal functional analysis.
  • Real Data Analysis (Schizophrenia vs. Healthy Controls):
    • Significant group differences were found in connectivity patterns across multiple STRiDEs, highlighting widespread alterations in sensory, cognitive control, default mode, and subcortical domains.
  • STRiDEs showed significant associations with cognitive scores (attention vigilance, working memory, visual learning, CMINDS composite) and positive symptoms in schizophrenia.
  • Specifically, STRiDE 3 was identified as a symptom-relevant state, showing connectivity changes linked to positive symptoms.
  • Temporal dynamics analysis revealed that STRiDE1 is the most stable and dominant state, while others are more transient, reflecting a hierarchical organization.
  • Differences in dwell time and occupancy rate were observed between HC and SZ groups, with STRIDE1 showing decreased dwell time in SZ.

Limitations and Future Work

  • Unexplained Variance: While STRiDE captures dynamic functional connectivity well, some variance remains unexplained, suggesting that other factors (e.g., genetics, environmental influences) might contribute to brain dynamics.
  • Future Directions:
    • Incorporating other modalities: Future work could integrate electrophysiological data (e.g., EEG, MEG) for a more comprehensive view.
  • Advanced modeling: Applying more sophisticated approaches like dynamic causal modeling (DCM) could explore directional interactions between brain regions.
  • Longitudinal and personalized studies: The authors suggest that future longitudinal and multimodal studies could further inform personalized intervention strategies.

Summary updated

Abstract

Abstract Understanding the dynamic intrinsic interactions between brain regions has been advanced by functional magnetic resonance imaging (fMRI), particularly through the connectivity analysis used to characterize reoccurring patterns in the brain known as brain states. However, previous studies have primarily focused on unimodal models, which can hinder optimal dynamic state estimation, especially in cognitive disorders, cases where there may be both structure and function disruption. To better estimate intrinsic brain interactions, it is important to account for the factors shaping this estimation not only in terms of time resolved variability of the connectivity but also regarding the underlying physical pathways between brain regions. However current approaches mostly use deterministic weighting of structural connectivity. To address this, we propose a flexible multimodal connectivity constrained independent component analysis (ICA) model, termed structural connectivity-informed dynamic state estimation (STRiDE), that enhances stability and sensitivity by leveraging white matter structural connectivity and dynamic functional connectivity information. Using this model, we decompose brain interactions into independent, reoccurring multimodal patterns or structural-functional states, guided by maximally independent structural connectivity priors derived from the group level data. We first evaluate our proposed model using a simulation pipeline, showing the approach works as design and improves sensitivity to group differences and enhances robustness to noise. Next, we applied the proposed multimodal model to real dataset including a cohort of subjects with schizophrenia (SZ) and healthy controls (HC). Results demonstrated its potential to enhance group-differences in both connectivity domain and temporal dynamics parameters. Specifically results highlighted disruption within and between sensory and trans-modal domains, through a SZ vs HC comparison. Symptoms severity and cognitive scores statistical analysis specifies their significant association with default mode domain, offering insights into the disrupted functional and neural mechanisms underlying schizophrenia. In addition, temporal interplay of the estimated STRiDEs reveals that visual-related STRiDE is significantly impacted in SZ, regarding the speed of processing score, underscoring the link between visual system and speed of processing. In sum, the STRiDE approach provides a flexible way to link structural and functional connectivity at the network level and represents a general approach for studying multimodal dynamic patterns and leveraging these to study the typical and disordered brain.
BibTeX
@misc{fouladivanda2025structural,
  author = {Mahshid Fouladivanda and Armin Iraji and Lei Wu and Jiayu Chen and Pablo Andrés-Camazón and Theodorus G.M. van Erp and Ayşenil Belger and Godfrey D. Pearlson and Tülay Adalı and Vince D. Calhoun},
  title = {{Structural connectivity-informed dynamic estimation (STRiDE): Multimodal connectivity constrained ICA}},
  year = {2025},
  doi = {10.1101/2025.06.18.660402},
  howpublished = {bioRxiv},
}
Figure