Decentralized multi-site VBM analysis during adolescence shows structural changes linked to age, body mass index, and smoking: A COINSTAC analysis

Harshvardhan Gazula, Bharath Holla, Zuo Zhang, Jiayuan Xu, Eric Verner, Ross Kelly, Günter Schumann, Vince D. CalhounbioRxiv 2019 · DOI

TL;DR

This paper introduces COINSTAC, a novel framework for decentralized neuroimaging analysis, enabling researchers to study brain structure across multiple sites without pooling sensitive data. They demonstrate its utility by analyzing structural brain changes related to age, BMI, and smoking in adolescents, revealing significant associations and highlighting the power of large-scale, privacy-preserving collaborative research.

What Problem Does This Paper Address?

Collecting neuroimaging data is expensive and computationally intensive. Furthermore, pooling data from multiple institutions raises significant privacy concerns. This creates a barrier to conducting large-scale studies that could yield more robust and reliable findings. Researchers need methods to analyze distributed datasets without compromising individual privacy or requiring data centralization. This paper tackles this challenge by presenting and validating a decentralized analysis framework, COINSTAC, specifically for voxel-based morphometry (VBM) of structural MRI data.

Methods

The core of this work lies in the COINSTAC (Collaborative Informatics and Neuroimaging Suite Toolkit for Anonymous Computation) platform. COINSTAC enables decentralized analysis by implementing a privacy-preserving message-passing infrastructure. For the specific analysis, they employ decentralized voxel-based morphometry (VBM). VBM is a statistical technique used to compare brain structure (specifically gray matter concentration) across different groups by analyzing MRI data voxel by voxel.

The decentralized VBM approach builds upon a decentralized regression algorithm. This algorithm fits a linear model to relate covariates (like age, BMI, smoking) to brain responses (gray matter volume at each voxel) across multiple sites. Crucially, each site computes its local contribution to the regression, and only aggregated statistics are shared with a central aggregator, preserving data privacy. The paper also details how statistical measures like t-values and R-squared are computed in this decentralized manner.

Datasets

The study utilizes data from two large, multi-site neuroimaging projects:

  • IMAGEN: A longitudinal study of adolescent brain development and mental health in Europe, with participants recruited from eight sites across England, Ireland, France, and Germany. Data acquired at age 14 was used.
  • cVEDA: A multi-site, collaborative cohort study in India focused on externalizing disorders and addiction, covering an age span of 5-24 years. MRI data from 6 sites was used.

These datasets represent real-world, sensitive neuroimaging data from different geographical locations and scanner types, making them ideal for testing the decentralized framework.

Key Insights and Ideas

  • Privacy-Preserving Large-Scale Analysis: The most significant takeaway is the successful demonstration that complex neuroimaging analyses like VBM can be performed across multiple sites without pooling raw data, thus safeguarding participant privacy.
  • Feasibility of Decentralized VBM: The paper validates that decentralized VBM using COINSTAC yields results comparable to traditional, centralized analyses, proving its practical utility.
  • Adolescent Brain Structure and Lifestyle Factors: The study reveals interesting associations between structural brain changes in adolescents and factors like smoking and BMI, providing insights into developmental trajectories.
  • User Experience Matters: Feedback from users of the COINSTAC platform highlighted the value of examining voxel-wise effects and the efficiency of decentralized computations, even with data-privacy concerns.

Key Results

The decentralized VBM analysis identified several significant associations:

  • Smoking: Non-smokers showed higher gray matter concentration in regions including the right anterior cingulate cortex, bilateral thalamus, brainstem, and left dorsal prefrontal cortex, with an increase in precuneus.
  • Body Mass Index (BMI): Increases in gray matter were observed in bilateral putamen, bilateral hypothalamus, and lateral and medial temporal lobes.
  • Age: Widespread decreases in gray matter were noted, particularly in frontal regions, with bilateral increases in the hippocampus.

The paper also notes that the decentralized analysis for the two models (age + BMI and age + smoking) took approximately 30-45 minutes to complete, demonstrating computational efficiency.

Limitations and Future Work

  • Limitations: The paper acknowledges that the causal relationship between smoking and certain brain regions (like the orbitofrontal cortex) is still debatable. Additionally, while the study found an increase in hypothalamic volume with BMI in adolescents, existing literature often reports atrophy, suggesting a potential non-linear effect specific to this age group or chronic obesity.
  • Future Work: The authors express optimism that this successful exercise will foster more collaborations. They also suggest that the insights gained from user feedback will help improve the user experience of the COINSTAC platform. The potential for using decentralized algorithms in large-scale neuroimaging studies is strongly encouraged.

Summary updated

Abstract

Abstract In the recent past, there has been an upward trend in developing frameworks that enable neuroimaging researchers to address challenging questions by leveraging data across multiple sites all over the world. One such framework, Collaborative Informatics and Neuroimaging Suite Toolkit for Anonymous Computation (COINSTAC), provides a platform to analyze neuroimaging data stored locally across multiple organizations without the need for pooling the data at any point during the analysis. In this paper, we perform a decentralized voxel-based morphometry analysis of structural magnetic resonance imaging data across two different sites to understand the structural changes in the brain as linked to age, body mass index and smoking. Results produced by the decentralized analysis are contrasted with similar findings in literature. This work showcases the potential benefits of performing multi-voxel and multivariate analyses of large-scale neuroimaging data located at multiple sites.
BibTeX
@misc{gazula2019decentralized,
  author = {Harshvardhan Gazula and Bharath Holla and Zuo Zhang and Jiayuan Xu and Eric Verner and Ross Kelly and Günter Schumann and Vince D. Calhoun},
  title = {{Decentralized multi-site VBM analysis during adolescence shows structural changes linked to age, body mass index, and smoking: A COINSTAC analysis}},
  year = {2019},
  doi = {10.1101/846386},
  howpublished = {bioRxiv},
}
Figure