Cortical Similarities in Psychiatric and Mood Disorders Identified in Federated VBM Analysis via COINSTAC

Kelly Rootes-Murdy, Sandeep Panta, Ross Kelly, Javier Romero, Yann Quidé, Murray J. Cairns, Carmel Loughland, Vaughan J. Carr, Stanley V. Catts, Assen Jablensky, Melissa J. Green, Frans Henskens, Dylan Kiltschewskij, Patricia T. Michie, Bryan Mowry, Christos Pantelis, Paul E. Rasser, William R. Reay, Ulrich Schall, Rodney J. Scott, Oliver J. Watkeys, Gloria Roberts, Philip B. Mitchell, Janice M. Fullerton, Bronwyn J. Overs, Masataka Kikuchi, Ryota Hashimoto, Junya Matsumoto, Masaki Fukunaga, Perminder S. Sachdev, Henry Brodaty, Wei Wen, Jiyang Jiang, Negar Fani, Timothy D. Ely, Adriana Lorio, Jennifer S. Stevens, Kerry J. Ressler, Tanja Jovanović, Sanne J.h. van Rooij, Sergey Plis, Anand D. Sarwate, Vince D. CalhounmedRxiv 2023 · DOI

Abstract

Psychiatric disorders such as schizophrenia (SZ), major depressive disorder (MDD), and bipolar disorder (BP) exhibit significant overlap across genetics, neurochemistry, and symptom profiles. While structural neuroimaging has identified various gray matter (GM) alterations, direct comparisons across disorders are often limited by the logistical and privacy-related challenges of pooling large-scale datasets. This study utilizes the Collaborative Informatics and Neuroimaging Suite Toolkit for Anonymous Computation (COINSTAC) to perform a federated voxel-based morphometry (VBM) analysis. By analyzing T1-weighted images from 3,287 individuals across eight diagnostic groups and seven international sites, we identified significant overlap in GM patterns—specifically within the bilateral insula—among individuals with schizophrenia, MDD, and autism spectrum disorder (ASD). These results suggest a shared neuroanatomical vulnerability across traditional diagnostic boundaries.

Introduction: Breaking Down Data Silos

In modern psychiatry, diagnostic boundaries are increasingly recognized as "leaky." Patients across the spectrum of schizophrenia, bipolar disorder, and major depression frequently share symptoms such as cognitive impairment and emotional dysregulation. Despite these clinical overlaps, neuroimaging research has historically treated these conditions as isolated categories.

To identify common neuroanatomical threads, researchers require massive, heterogeneous datasets. However, pooling raw MRI scans presents significant ethical and logistical hurdles; medical data is highly sensitive, and transferring terabytes of files between global institutions is often prohibitive. This "data silo" problem frequently results in underpowered studies. This paper addresses these challenges by employing a federated learning approach to identify transdiagnostic patterns of gray matter loss.

Methodology: Federated Voxel-Based Morphometry

The researchers employed Federated Voxel-Based Morphometry (VBM) using the COINSTAC platform. COINSTAC is an open-source application designed for decentralized analysis, allowing researchers to analyze neuroimaging data without ever moving the raw files from their original locations.

The Federated Workflow

  1. Local Computation: Each of the seven participating sites performed initial processing and regression models on their local servers.
  2. Privacy-Preserving Aggregation: Only summary statistics (such as gradients or averages) were transmitted to a central cloud aggregator.
  3. Global Synthesis: The aggregator combined these summaries to generate a global model, which was then shared back with the participating sites.

The study utilized a General Linear Model (GLM) to associate specific diagnoses with voxel-wise GM concentration, while controlling for confounding variables including age, sex, site, and the APOE gene—a known marker for neurodegeneration.

Datasets and Cohorts

The analysis included T1-weighted MRI scans from N = 3,287 individuals. The data was sourced from several major international consortia and databases, including:

  • ASRB (Australian Schizophrenia Research Bank)
  • COCORO (Cognitive Genetics Collaborative Research Organization)
  • Emory University (Grady Trauma Project)
  • IGP (Imaging Genetics in Psychosis)
  • MAS/OATS (Studies on aging and twins)
  • Sydney (Bipolar Kids and Sibs)

The diagnostic breakdown encompassed Schizophrenia (SZ), Bipolar Disorder (BP), Major Depressive Disorder (MDD), PTSD, Autism Spectrum Disorder (ASD), and Mild Cognitive Impairment (MCI), alongside a control group of nearly 2,000 healthy individuals.

Key Insights: A Transdiagnostic Continuum

The Insula as a Universal Hub

The most significant finding was the identification of gray matter reductions in the bilateral insula across nearly every disorder studied (SZ, BP, MDD, ASD, and MCI). The insula plays a critical role in emotional regulation and "salience"—the brain's ability to determine which internal and external stimuli deserve attention. This suggests that the insula may serve as a primary neuroanatomical marker for general psychiatric vulnerability.

Schizophrenia and the Neuroanatomical Spectrum

Schizophrenia exhibited the most widespread and severe global reductions in gray matter. Interestingly, the GM alteration patterns in MDD and "Spectrum" disorders were highly correlated with those found in SZ (with correlation coefficients ranging from r = 0.72 to 0.79). This suggests that these disorders may exist on a neuroanatomical continuum rather than being entirely distinct biological entities.

PTSD as a Unique Outlier

In contrast to the other psychiatric conditions, PTSD showed minimal overlap with the common GM patterns identified. Its alterations appeared unique, potentially reflecting specific stress-response mechanisms rather than the broader heritable vulnerability seen in schizophrenia or MDD.

Neural Resilience in Relatives

The study also examined unaffected relatives of individuals with psychosis. Interestingly, these individuals showed increased gray matter in certain regions, such as the temporal poles and cerebellum. This may indicate a "compensatory" or protective mechanism that allows these individuals to remain healthy despite a shared genetic risk.

Results and Stability

  • Spatial Correlations: The highest similarity was found between Schizophrenia and the "Spectrum" group (r = 0.79).
  • Effect Sizes: SZ demonstrated the largest effect size for GM reduction across the entire cortex.
  • Regional Specificity: Beyond the insula, shared reductions were notably prominent in the medial prefrontal cortex and parahippocampal gyrus.
  • Robustness: Post-hoc analyses confirmed that these correlations remained stable even after removing individuals with MCI, ensuring that age-related cognitive decline was not the primary driver of the results.

Limitations and Future Directions

While the study provides a powerful look at shared brain structures, it is limited by its cross-sectional design, which captures only a snapshot in time. It remains unclear whether these GM changes precede the onset of illness or result from chronic disease progression and medication use (e.g., antipsychotics). Additionally, while the total sample size is large, some sub-groups (such as MDD) were relatively small, limiting specific head-to-head comparisons.

Moving forward, the authors suggest that the COINSTAC framework could be used to develop federated machine learning models. Such models could eventually classify or predict psychiatric outcomes across diverse global populations while maintaining strict patient privacy.

Summary updated

Abstract

Abstract Psychiatric disorders such as schizophrenia, major depressive disorder, and bipolar disorder have areas of significant overlap across multiple domains including genetics, neurochemistry, symptom profiles, and regional gray matter alterations. Various structural neuroimaging studies have identified a combination of shared and disorder-specific patterns of gray matter (GM) deficits across these different disorders, though few direct comparisons have been conducted. Given the overlap in symptom presentations and GM alterations, these disorders may have a common etiology or neuroanatomical basis that may relate to a certain vulnerability for mental illness. Pooling large data or heterogeneous data can ensure representation of several participant factors, providing more accurate results. However, ensuring large enough datasets at any one site can be cumbersome, costly, and may take many years of data collection. Large scale collaborative research is already facilitated by current data repositories and neuroinformatics consortia such as the Enhacing NeuroImaging Genetics through Meta-Analysis (ENIGMA) consortium, institutionally supported databases, and data archives. However, these data sharing methodologies can still suffer from significant barriers. Federated approaches can augment these approaches and mitigate some of these barriers by enabling access or more sophisticated, shareable and scaled up analyses of large-scale data which may not be shareable and can easily be scaled up with the number of sites. In the current study, we examined GM alterations using Collaborative Informatics and Neuroimaging Suite Toolkit for Anonymous Computation (COINSTAC). Briefly, COINSTAC ( https://coinstac.trendscenter.org ) is an open-source decentralized analysis application that provides a venue for analyses of neuroimaging datasets without sharing individual level data, while maintaining granular control of privacy. Through federated analysis, we examined T1-weighted images (N = 3,287) from eight psychiatric diagnostic groups across seven sites. We identified significant overlap in the GM patterns of individuals with schizophrenia, major depressive disorder, and autism spectrum disorder. These results show cortical and subcortical regions, specifically the bilateral insula, that may indicate a possible shared vulnerability to psychiatric disorders.
BibTeX
@misc{rootesmurdy2023cortical,
  author = {Kelly Rootes-Murdy and Sandeep Panta and Ross Kelly and Javier Romero and Yann Quidé and Murray J. Cairns and Carmel Loughland and Vaughan J. Carr and Stanley V. Catts and Assen Jablensky and Melissa J. Green and Frans Henskens and Dylan Kiltschewskij and Patricia T. Michie and Bryan Mowry and Christos Pantelis and Paul E. Rasser and William R. Reay and Ulrich Schall and Rodney J. Scott and Oliver J. Watkeys and Murray J. Cairns and Gloria Roberts and Philip B. Mitchell and Janice M. Fullerton and Bronwyn J. Overs and Masataka Kikuchi and Ryota Hashimoto and Junya Matsumoto and Masaki Fukunaga and Perminder S. Sachdev and Henry Brodaty and Wei Wen and Jiyang Jiang and Negar Fani and Timothy D. Ely and Adriana Lorio and Jennifer S. Stevens and Kerry J. Ressler and Tanja Jovanović and Sanne J.H. van Rooij and Sergey Plis and Anand D. Sarwate and Vince D. Calhoun},
  title = {{Cortical Similarities in Psychiatric and Mood Disorders Identified in Federated VBM Analysis via COINSTAC}},
  year = {2023},
  doi = {10.1101/2023.09.28.23296219},
  howpublished = {medRxiv},
}
Figure