An information theory framework for capturing multi-connectivity via spatial network encoding reveals reduced population count (Hamming weight) localized to auditory, visual, and motor networks

Biozid Bostami, Noah Lewis, Oktay Agcaoglu, Rogers F. Silva, Jessica A. Turner, Theo G.m. van Erp, Judith M. Ford, Vince D. CalhounbioRxiv 2025 · DOI

TL;DR

This paper introduces a novel voxel-centric framework to analyze how individual brain voxels contribute to multiple functional networks simultaneously. By encoding these contributions as binary "words" and using information theory metrics like Hamming weight, they reveal altered network interactions in schizophrenia, particularly reduced contributions in higher-order cognitive networks and increased activity in sensory networks.

What Problem Does This Paper Address?

Traditional neuroimaging analyses of brain function often rely on "functional connectivity" (FC), which typically measures pairwise correlations between brain regions. While useful, this approach can oversimplify the brain's complex, interconnected nature by overlooking how individual voxels might participate in multiple networks at once. This "multi-way interplay" is crucial for understanding intricate brain processes and how disruptions in these interactions might underlie neurological and psychiatric disorders like schizophrenia. The paper aims to bridge this gap by developing a method that captures these multi-network contributions at a finer, voxel-level resolution.

Methods

The core of the proposed method is a voxel-centric framework that moves beyond pairwise connections.

  1. Independent Component Analysis (ICA): They first used spatially constrained ICA (via the NeuroMark pipeline) to identify 53 distinct functional brain networks (ICNs) from fMRI data.
  2. Binarization: For each subject and each network, voxel contributions were binarized. Voxels with a significant contribution (z-score > 1) were marked as "1" (present), and others as "0" (absent). This creates a binary spatial map for each network.
  3. Voxel Encoding (Binary "Words"): Each voxel's participation across all 53 networks is represented as a 53-bit binary word. This effectively encodes how a voxel is embedded within the multi-network architecture.
  4. Hamming Weight: This metric quantifies the number of networks a voxel significantly contributes to. It's calculated by summing the binary values in a voxel's 53-bit word. A higher Hamming weight indicates greater multi-network participation for that voxel.
  5. Dice Similarity Coefficient (DSC): At the network level, DSC measures the spatial overlap between two networks by quantifying the proportion of shared voxels. This helps understand how integrated or segregated different networks are.

Datasets

The study utilized resting-state fMRI data from:

  • 151 patients with schizophrenia (SZ)
  • 160 age- and gender-matched healthy controls (HC)

This data was obtained as part of the fBIRN project.

Key Insights and Ideas

  • Voxels as Multi-Network Mediators: The framework highlights that individual voxels aren't confined to single networks but act as crucial mediators of communication across multiple networks.
  • Information Theory for Brain Architecture: Using Hamming weight, a concept from information theory, provides a computationally efficient way to quantify this multi-network participation at the voxel level.
  • Schizophrenia's Dual Nature: The findings suggest schizophrenia involves a "dual nature" of brain alterations: reduced integration in higher-order cognitive networks (like DMN, FPN) and increased or aberrant integration in sensory/perceptual networks (like VIS, AUD, SM).
  • Age-Related Network Integration: Network overlap (measured by DSC) increases with age, suggesting a progressive integration of functional networks throughout life.

Key Results

  • Schizophrenia vs. Controls:
    • Individuals with schizophrenia showed significantly reduced Hamming weights in higher-order cognitive networks (e.g., Default Mode Network - DMN, Frontoparietal Network - FPN). This indicates fewer voxels contributing significantly to these networks.
  • Conversely, increased Hamming weights were observed in sensory and perceptual networks (Visual - VIS, Auditory - AUD, Sensorimotor - SM).
  • Dice Similarity: * Schizophrenia groups showed higher overlap (increased DSC) between VIS, AUD, and SM networks.
  • Healthy controls showed higher overlap between DMN and Cerebellar (CB) networks.
  • Age Effect: A significant positive linear relationship was found between age and Dice Similarity, meaning older individuals tend to have greater network overlap.

Limitations and Future Work

  • Loss of Continuous Information: The binarization step, while simplifying analysis, might lose valuable information present in the continuous spectrum of functional connectivity strengths. A voxel with moderate connectivity to multiple networks is treated the same as one with very strong connectivity.
  • Resting-State Limitations: The study relies on resting-state fMRI, which captures intrinsic connectivity but doesn't account for task-related network dynamics or temporal fluctuations in connectivity.
  • Clinical Correlations: The absence of correlations with clinical or cognitive measures limits the understanding of the behavioral relevance of the identified voxel-level contributions.
  • Cross-Sectional Design: The study's cross-sectional nature prevents analysis of connectivity progression over time or its response to treatment. Factors like medication status, illness duration, and comorbidities were not controlled.

Future Directions: * Incorporate task-based fMRI to examine network alterations during specific cognitive or motor tasks. * Utilize dynamic functional connectivity (dFC) to capture time-varying connectivity patterns. * Develop probabilistic frameworks (e.g., using mutual information) to account for intersubject variability. * Integrate other imaging modalities (e.g., DTI, MEG) to link structural and temporal dynamics. * Extend the framework to other neuropsychiatric disorders. * Conduct longitudinal studies and include clinical data for a more comprehensive understanding.

Summary updated

Abstract

1. Abstract The human brain operates as a complex system where functional networks evolve and interact across spatially distributed regions. In traditional neuroimaging analyses, functional connectivity (FC), based on pairwise correlations or statistical dependencies of temporal fluctuations in the BOLD signal, has been a primary method for exploring interactions between brain regions and decoding brain function. However, traditional FC methods often overlook the intricate, multi-way interplay among brain elements that emerge from the brain’s densely interconnected nature. To overcome these limitations, we introduce a novel voxel-centric framework that captures the multi-way interactions between voxels and networks identified via high-model order independent component analysis. This framework posits that individual voxels serve as critical mediators of multi-network communication, reflecting the brain’s complex functional architecture. By encoding voxel contributions from brain networks into binary representations and quantifying the population count at each voxel via Hamming weights, the proposed method prioritizes high-contribution voxels that facilitate inter-network interactions. This approach provides new insights into the brain’s functional organization, revealing previously unrecognized patterns of voxel-to-network entanglement. Specifically, in the context of schizophrenia, our method enables the identification of spatial patterns that may underpin the cognitive and perceptual disturbances characteristic of the disorder. This enhanced understanding could improve diagnostic precision and help tailor interventions that target specific dysfunctional networks, offering a pathway to more effective treatments and better patient outcomes in schizophrenia.
BibTeX
@misc{bostami2025information,
  author = {Biozid Bostami and Noah Lewis and Oktay Agcaoglu and Rogers F. Silva and Jessica A. Turner and Theo G.M. van Erp and Judith M. Ford and Vince D. Calhoun},
  title = {{An information theory framework for capturing multi-connectivity via spatial network encoding reveals reduced population count (Hamming weight) localized to auditory, visual, and motor networks}},
  year = {2025},
  doi = {10.1101/2025.05.22.655511},
  howpublished = {bioRxiv},
}
Figure