Direct linkage detection with multimodal IVA fusion reveals markers of age, sex, cognition, and schizophrenia in large neuroimaging studies

Rogers F. Silva, Eswar Damaraju, Xinhui Li, Peter Kochunov, Ayşenil Belger, Judith M. Ford, Daniel H. Mathalon, Bryon A. Mueller, Steven G. Potkin, Adrian Preda, Jessica A. Turner, Theo G.m. van Erp, Tülay Adalı, Vince D. CalhounbioRxiv 2021 · DOI

TL;DR

This paper introduces a novel multimodal independent vector analysis (MMIVA) framework to fuse data from structural, functional, and diffusion MRI. This approach successfully identifies linked brain patterns associated with age, sex, cognition, and schizophrenia, offering a more comprehensive understanding of neurobiological variations.

What Problem Does This Paper Address?

Analyzing neuroimaging data from multiple modalities (like structural MRI, fMRI, and diffusion MRI) holds immense potential for understanding complex brain functions and identifying biomarkers for neurological and psychiatric conditions. However, effectively integrating these diverse data types to find shared underlying patterns has been a significant challenge. Traditional methods often analyze each modality in isolation, missing crucial cross-modal relationships. This paper aims to bridge this gap by developing a data-driven fusion method that can directly identify and characterize these linked, multimodal brain signatures.

Methods

The core of this work is the Multimodal Independent Vector Analysis (MMIVA) model, implemented within a broader framework called Multidataset Independent Subspace Analysis (MISA). MMIVA is an extension of Independent Vector Analysis (IVA), which itself builds upon Independent Component Analysis (ICA).

  • Standard ICA: Assumes linear mixing of independent sources in a single dataset and aims to recover these sources by estimating a single unmixing matrix.
  • Independent Vector Analysis (IVA): Extends ICA to multiple datasets (modalities in this case), assuming that while sources within each modality might be mixed, the corresponding sources across modalities are linked (dependent). It jointly estimates multiple unmixing matrices to find these linked sources.
  • MMIVA (as implemented in MISA): This paper utilizes a flexible MISA implementation of MMIVA. A key innovation is the initialization of the unmixing matrices using Multimodal Group Principal Component Analysis (MGPCA). MGPCA finds common directions of maximal variation across modalities, providing a robust starting point for the MMIVA estimation. The MMIVA model then refines these estimates, assuming source distributions follow a multivariate Kotz distribution, which allows for capturing higher-order statistics and non-Gaussianity.

Datasets

The study utilized two large, independent, multi-site neuroimaging datasets:

  1. UK Biobank: A prospective epidemiological study with a large imaging database. The analysis included data from 3497 subjects.
  2. Patient Datasets: Pooled data from four studies investigating schizophrenia, schizo-affective disorder, bipolar disorder, and age-matched healthy controls. These studies included:
  3. Center of Biomedical Research Excellence in brain function and mental illness (COBRE)
    • Function Biomedical Informatics Network (fBIRN)
    • Maryland Psychiatric Research Center (MPRC)
    • Bipolar and Schizophrenia Network for Intermediate Phenotypes (BSNIP)

The paper provides detailed demographic information for these datasets in Table 1.

Key Insights and Ideas

  • Multimodal Linkages are Key: The study strongly suggests that significant neurobiological phenomena (like aging, sex differences, or disease markers) manifest as coordinated patterns across different brain imaging modalities. Simply analyzing each modality in isolation misses this crucial interconnectedness.
  • MMIVA Captures Shared and Modality-Specific Information: The MMIVA approach successfully disentangles information that is shared across structural, functional, and diffusion MRI, as well as information that is unique to each modality within a linked pattern. This allows for a more nuanced understanding of how different aspects of brain structure and function contribute to observed phenotypes.
  • Brain-Age is Not Monolithic: The brain-age delta modeling revealed that different components (shared SVD features, partialled sMRI, and partialled fMRI) contribute uniquely to predicting brain age, highlighting the diverse neurobiological processes underlying aging.
  • Replication Across Datasets: The identification of similar age-related patterns in the UK Biobank and patient datasets (e.g., SCV 8 in UKB and SCV 19 in patients) provides strong evidence for the robustness and generalizability of the findings.

Key Results

  • Identification of Linked Subject Profiles (SCVs): The MMIVA model successfully identified 30 linked multimodal sources (SCVs) that captured significant associations with age, sex, cognitive performance (time to answer), and schizophrenia.
  • Age-Related Decline: Five SCVs were significantly associated with age, showing patterns of decline in gray matter density (sMRI), reduced functional connectivity (fMRI ALFF), and altered white matter integrity (dMRI FA) in regions like the caudate, thalamus, and cingulate cortex.
  • Sex Differences: Five SCVs showed significant sex effects, with distinct patterns of source intensity for males and females in various brain regions.
  • Schizophrenia Biomarkers: Two SCVs were associated with schizophrenia, showing reduced gray matter intensity in the temporal lobe and altered white matter integrity, consistent with previous findings.
  • Cognitive Performance: SCV 22 showed a significant association with time to answer in a prospective memory test, with faster responders exhibiting higher source intensities.
  • Brain-Age Delta: The MMIVA-derived features achieved a mean absolute error of 2.70 years in predicting brain age, demonstrating the model's ability to capture age-related brain changes.

Limitations and Future Work

  • Confound Variables: The brain-age delta analysis did not include all potential confound variables, such as head size, scanner table position, and scan-date-related slow drifts. Future work could incorporate these to further refine the models.
  • Modality-Specific Contributions: While the study highlights the value of multimodal analysis, further investigation into the precise contributions of each modality within specific SCVs could be beneficial.
  • Generalizability: While two large datasets were used, applying the MMIVA model to even more diverse populations and datasets could further validate its robustness.
  • Clinical Applications: The identified SCVs represent potential biomarkers. Future research could explore their clinical utility in diagnosis, prognosis, and treatment monitoring for various neurological and psychiatric conditions.

Summary updated

Abstract

Abstract With the increasing availability of large-scale multimodal neuroimaging datasets, it is necessary to develop data fusion methods which can extract cross-modal features. A general framework, multidataset independent subspace analysis (MISA), has been developed to encompass multiple blind source separation approaches and identify linked cross-modal sources in multiple datasets. In this work we utilized the multimodal independent vector analysis model in MISA to directly identify meaningful linked features across three neuroimaging modalities — structural magnetic resonance imaging (MRI), resting state functional MRI and diffusion MRI — in two large independent datasets, one comprising of control subjects and the other including patients with schizophrenia. Results show several linked subject profiles (the sources/components) that capture age-associated decline, schizophrenia-related biomarkers, sex effects, and cognitive performance. For sources associated with age, both shared and modality-specific brain-age deltas were evaluated for association with non-imaging variables. In addition, each set of linked sources reveals a corresponding set of multi-tissue spatial patterns that can be studied jointly.
BibTeX
@misc{silva2021direct,
  author = {Rogers F. Silva and Eswar Damaraju and Xinhui Li and Peter Kochunov and Ayşenil Belger and Judith M. Ford and Daniel H. Mathalon and Bryon A. Mueller and Steven G. Potkin and Adrian Preda and Jessica A. Turner and Theo G.M. van Erp and Tülay Adalı and Vince D. Calhoun},
  title = {{Direct linkage detection with multimodal IVA fusion reveals markers of age, sex, cognition, and schizophrenia in large neuroimaging studies}},
  year = {2021},
  doi = {10.1101/2021.12.13.472507},
  howpublished = {bioRxiv},
}
Figure