Identifying the Relationship Structure Among Multiple Datasets Using Independent Vector Analysis: Application to Multi-Task fMRI Data

Isabell Lehmann, Tanuj Hasija, Ben Gabrielson, Mohammad A. B. S. Akhonda, Vince D. Calhoun, Tülay AdaliIEEE Access 2024 · DOI

TL;DR

This paper introduces a novel, robust method for uncovering the underlying relationships between multiple datasets, going beyond simple grouping to reveal a detailed structure. By combining Independent Vector Analysis (IVA) with bootstrap-based hypothesis testing and hierarchical clustering, the method effectively handles complex, non-Gaussian data and has been successfully applied to multi-task fMRI data to identify brain regions associated with schizophrenia.

What Problem Does This Paper Address?

Identifying relationships among multiple datasets is crucial for summarizing information and making group inferences, particularly in fields like medical applications. However, existing methods for subgroup identification often rely on user-defined thresholds or strong distributional assumptions, limiting their robustness and interpretability. Furthermore, many approaches focus solely on identifying groups rather than understanding the intricate relationships between these groups. This paper aims to bridge this gap by providing a data-driven method that can effectively learn and reveal the relationship structure among datasets without subjective choices, even when dealing with complex, non-Gaussian data like fMRI.

Methods

The proposed method is a three-step process:

  1. Estimation of Latent Sources (IVA): The core of this step is Independent Vector Analysis (IVA), specifically the IVA-L-SOS variant. IVA is a blind source separation technique that extends Independent Component Analysis (ICA) to multiple datasets. It aims to find underlying "source components" that are statistically independent. IVA-L-SOS is chosen because it can handle non-Gaussian data and incorporates higher-order statistics, making the estimated components more interpretable. This step generates "Source Component Vectors" (SCVs) which represent the relationships between datasets.
  2. Identification of Common and Structured SCVs: This step focuses on identifying which SCVs contain meaningful information about dataset relationships. This is achieved by analyzing the covariance matrices of the SCVs. Structured SCVs are those whose covariance matrices exhibit a block-diagonal structure, indicating dependencies within specific subsets of datasets. The paper introduces a bootstrap-based hypothesis testing approach to robustly estimate the number of eigenvalues greater than one in these covariance matrices, which is key to distinguishing between common (single block) and structured (multiple blocks) SCVs. This bootstrap method avoids relying on asymptotic assumptions often made in traditional model-order selection.
  3. Identification of Relationship Structure using Structured SCVs: The structured SCVs identified in Step 2 are then used to build a feature matrix. Specifically, the leading eigenvectors corresponding to eigenvalues greater than one in the structured SCV covariance matrices are concatenated. This feature matrix serves as input for a hierarchical clustering algorithm. Hierarchical clustering is chosen because it naturally reveals the relationship structure through a dendrogram, showing how datasets group together and their relative distances, without requiring prior knowledge of the number of clusters.

Datasets

The paper uses two types of datasets:

  • Simulated Datasets: These were generated to evaluate the performance of the proposed method, particularly the bootstrap technique for estimating the number of blocks per SCV. The simulations involved generating SCV covariance matrices with varying structures (common and structured) and correlation coefficients, from which SCVs with Laplacian-distributed sources were created.
  • Real Multi-Task fMRI Data: The method was applied to fMRI data from the MIND Clinical Imaging Consortium (MCIC) collection. This dataset includes data from 271 subjects (patients with schizophrenia and healthy controls) performing three different tasks: Auditory Oddball (AOD), Sensory Motor (SM), and Sternberg Item Recognition Paradigm (SIRP). The data was preprocessed using standard fMRI analysis pipelines.

Key Insights and Ideas

  • Beyond Grouping to Structure: The paper emphasizes that simply identifying subgroups is insufficient; understanding the relationships between these subgroups is crucial for a deeper understanding of the data.
  • Robustness to Non-Gaussianity: By using IVA-L-SOS, the method effectively handles the non-Gaussian nature of fMRI data, which is a significant advantage over methods relying on Gaussian assumptions.
  • Eliminating User-Defined Thresholds: The bootstrap-based hypothesis testing for estimating the number of eigenvalues greater than one removes the need for arbitrary thresholds, leading to a more objective and reproducible analysis.
  • Leveraging All Structured SCVs: The method's strength lies in fusing information from all structured SCVs, rather than analyzing them individually, which enhances the robustness and comprehensiveness of the identified relationship structure.
  • Interpretable Components: IVA's ability to extract interpretable latent components (fMRI activation maps) allows for direct biological interpretation of the findings, linking them to known brain regions and functions.

Key Results

  • Simulation Performance: The proposed bootstrap technique for estimating the number of blocks per SCV demonstrated superior performance compared to existing methods (EV and GD) in simulations, accurately estimating the true number of blocks across various correlation values. The method achieved an AMI of 1 for simulated data, indicating perfect identification of groupings.
  • fMRI Data Analysis:
    • The method successfully identified structured SCVs that revealed meaningful relationships between the fMRI tasks. For instance, it grouped auditory tasks (AOD, SM) together and visual tasks (SIRP) together, aligning with prior knowledge.
  • Within the SIRP tasks, finer groupings were identified for encoding and probe phases, reflecting the task's structure.
  • The estimated fMRI activation maps (source components) revealed interpretable brain regions. Notably, it showed stronger deactivations of the Default Mode Network (DMN) in patients with schizophrenia compared to controls, and stronger activations in auditory regions for controls.
  • The identified relationship structure of the task datasets was consistent with existing work but extended it by providing a more detailed structural understanding.

Limitations and Future Work

  • Identity Covariance Matrices: The paper acknowledges a limitation that the proposed method does not identify SCVs with identity covariance matrices, which correspond to completely uncorrelated sources. Future work could involve adapting the bootstrap test to address this.
  • Interpretability: While the method provides interpretable components, the authors note that interpretation is not always straightforward, and a good guideline is to explore a range of dimension reduction values (R) and compare components with established brain atlases (like Brodmann areas).
  • Generalizability: The authors suggest that the method is not limited to task datasets and is applicable to more general problems, such as identifying subgroups of subjects in neuroimaging and other fields. Future work could explore its direct implications for clinical significance and treatment strategies by applying it to diverse datasets, particularly in precision medicine.
  • Adaptability: The authors mention that Step 1 (source estimation) can be replaced by other source separation methods if the data has different source distributions.

Summary updated

BibTeX
@article{lehmann2024identifying,
  author = {Isabell Lehmann and Tanuj Hasija and Ben Gabrielson and Mohammad A. B. S. Akhonda and Vince D. Calhoun and Tülay Adali},
  title = {{Identifying the Relationship Structure Among Multiple Datasets Using Independent Vector Analysis: Application to Multi-Task fMRI Data}},
  journal = {IEEE Access},
  year = {2024},
  doi = {10.1109/access.2024.3435526},
}
Figure