Source-linked AI summary
Federated Learning in Distributed Medical Databases: Meta-Analysis of Large-Scale Subcortical Brain Data
Santiago Silva, Boris Gutman, Eduardo Romero, Paul M Thompson, Andre Altmann, Marco Lorenzi
TL;DR
Distributed brain-imaging datasets offer opportunities for studying brain diseases, but privacy, legal, and institutional barriers restrict direct sharing. The paper proposes an end-to-end federated framework for standardized and multivariate analysis without sharing individual information, and evaluates it on synthetic and multi-centric disease data. The framework is presented as a validated approach for federated analysis of distributed biomedical data, while consistent federated correction remains sensitive to center-specific covariate handling.
Problem
Privacy, legal, and institutional constraints limit direct sharing of distributed biomedical data, while mass-univariate analysis models only one dependent variable at a time despite correlated high-dimensional features.
Method
The paper proposes an end-to-end federated framework combining data standardization, confounding-factor correction, and multivariate analysis, with an ENIGMA-compatible implementation using ADMM to reduce iterations.
Results
The framework was benchmarked on synthetic data and applied to subcortical thickness and shape features across multi-centric, multi-database cohorts including Alzheimer’s disease, MCI, Parkinson’s disease, and healthy individuals.
Takeaways & Limitations
The study presents a validated framework for federated analysis of distributed biomedical data and identifies potential extension to large-scale imaging-genetics studies.
Takeaways & Limitations
Local confounding correction can introduce bias because covariates may be handled differently across centers, so correction must also be performed consistently in a federated way.
Abstract
from arXiv · showhide
At this moment, databanks worldwide contain brain images of previously unimaginable numbers. Combined with developments in data science, these massive data provide the potential to better understand the genetic underpinnings of brain diseases. However, different datasets, which are stored at different institutions, cannot always be shared directly due to privacy and legal concerns, thus limiting the full exploitation of big data in the study of brain disorders. Here we propose a federated learning framework for securely accessing and meta-analyzing any biomedical data without sharing individual information. We illustrate our framework by investigating brain structural relationships across diseases and clinical cohorts. The framework is first tested on synthetic data and then applied to multi-centric, multi-database studies including ADNI, PPMI, MIRIAD and UK Biobank, showing the potential of the approach for further applications in distributed analysis of multi-centric cohorts
1 Introduction
Large, distributed neuroimaging datasets could advance brain-disease research, but privacy, legal, and logistical barriers limit direct data sharing. The paper proposes an end-to-end federated framework for standardized, multivariate analysis across centers, illustrated with ENIGMA-compatible pipelines and multi-centric disease data.
- Research gap: Privacy, legal, and bureaucratic barriers prevent some institutions from directly sharing large biomedical datasets.Federated learning shares model parameters rather than individual information, while meta-analysis combines statistical results across centers.
- Research gap: ENIGMA brings together neuroimaging groups worldwide, but its secure meta-analytic studies mostly use mass-univariate models.The consortium analyzes neuroimaging data from over 10,000 subjects.
- Research gap: Mass-univariate analysis models one dependent variable at a time, limiting its treatment of correlated biomedical features such as neighboring voxels or genetic variations.Multivariate methods have therefore been proposed to account for covariance in high-dimensional data.
- Research gap: Existing federated multivariate methods may require many communications across centers because they rely mainly on stochastic gradient descent.This creates potential computational and practical bottlenecks for multi-centric, high-dimensional data.
- Research gap: A prior eigenvalue-decomposition approach avoids iteration over centers but does not consistently federate data harmonization, including standardization and covariate adjustment.This limits its practical utility in real applications.
- Contribution: The proposed end-to-end framework federates data standardization, confounding-factor correction, and multivariate variability analysis using ADMM to reduce iterations.It is illustrated through an ENIGMA-compatible tool and evaluated on synthetic data and multi-centric, multi-database cohorts spanning Alzheimer’s disease, MCI, Parkinson’s disease, and healthy individuals.
2 Methods
The framework enables distributed multivariate analysis while keeping individual biomedical data within each center. It standardizes data, corrects confounding factors with shared parameters, and estimates global variability through federated PCA.
- Biomedical data remain partitioned across centers, which share model parameters and run local feature-extraction pipelines instead of sharing individual information.
- The framework has three stages: data standardization, confounding-factor correction, and variability analysis.
- Data standardization: Distributed standardization centers each feature to zero mean and unit standard deviation using statistics intended to reflect the whole study.
- Correction from confounding factors: ADMM estimates a shared parameter matrix for linear confounder correction, avoiding direct sharing of local data or covariates.
- Correction from confounding factors: After convergence, centers subtract the shared covariate effect from their standardized structural data to obtain corrected observations.
- Federated PCA: Federated PCA approximates global covariance by combining local covariance information and sharing reduced principal components, with components selected at an 80% explained-variability threshold.
3 Experiments
The framework was evaluated on synthetic data and applied to multi-database neuroimaging cohorts using federated preprocessing and principal-component analysis. The resulting components captured disease-related subcortical variability across centers, while suggesting residual age effects after correction.
- Synthetic experiments: Synthetic matrices were split across centers, corrupted with Gaussian noise, and analyzed across 200 folds to assess convergence.Noise was set to 20% of ∥X∥, and convergence was analyzed for each scenario.
- Real-data cohort: The real-data study included 455 controls, 181 non-progressive MCI participants, 208 progressive MCI participants, 234 AD participants, and 232 PD participants.T1-weighted baseline MRI scans were processed with ENIGMA Shape Analysis using radial distance and log-Jacobian features.
- Federated preprocessing: Each database was modeled as an independent center, with sex, age, and age^2 used to correct the vertex-wise shape data.ADMM convergence was ensured through 10 iterations before federated variability analysis.
- Synthetic experiments: Figure 2 compares centralized and federated matrices using MSE, column estimates, and principal components for 100 centers.The synthetic setting used N = 2400, Nfeatures = 50,000, and dim(y) = 20.
- Real-data results: Federated principal components separated AD from controls across centers and displayed progressive versus stable MCI projections from ADNI.The projections used the first four components estimated from data obtained from four centers.
- Real-data results: Components 1 to 3 showed healthy-to-AD variability across centers, while healthy ADNI participants occupied an intermediate position.The intermediate position may reflect residual age effects after correction, and PD subjects were generally similar to healthy individuals for modeled subcortical information.
4 Conclusions
The study concludes that it proposed, tested, and validated a consistent framework for federated analysis of distributed biomedical data, with future extensions to imaging-genetics studies.
- 4 Conclusions: The framework was proposed, tested, and validated for federated analysis of distributed biomedical data.Further development is intended for large-scale imaging-genetics data, including ENIGMA meta-studies.
A Funding
The work received support from French, UK, European Union, regional, and NIH funding sources.
- A Funding: Support included the UCA-JEDI and French National Research Agency project ANR-15-IDEX-01, plus Medical Research Council and NIH grants.Additional funding came from the European Union Horizon 2020 programme and the Region Sud.
B The Alzheimer’s Disease Neuroimaging Initiative (ADNI)
The study used data from the Alzheimer’s Disease Neuroimaging Initiative, whose investigators contributed to ADNI’s design, implementation, or data provision but not this report’s analysis or writing.
- B The Alzheimer’s Disease Neuroimaging Initiative (ADNI): ADNI investigators supported database design, implementation, or data provision, while the report’s analysis and writing were conducted separately.The passage directs readers to the complete ADNI investigator acknowledgement list.
C The Parkinson’s Progression Markers Initiative (PPMI)
The article used data from the Parkinson’s Progression Markers Initiative (PPMI) database, a public-private partnership funded by the Michael J. Fox Foundation and other funding partners.
- The article’s data were obtained from the Parkinson’s Progression Markers Initiative (PPMI) database.
- PPMI is described as a public-private partnership.
- PPMI is funded by the Michael J. Fox Foundation for Parkinson’s Research and additional funding partners.
D UK Biobank
The research used the UK Biobank Resource under Application Number 20576, with Nicholas Ayache identified as principal investigator.
- The research used the UK Biobank Resource under Application Number 20576.
- Nicholas Ayache is identified as the principal investigator for the UK Biobank application.