FedDOSE: Federated Learning Framework Decomposing Site Effects for Modeling Brain Dynamic Functional Connectivity

FedDOSE (Federated Decomposition Of Site Effects) is a novel federated learning framework for modeling brain dynamic functional connectivity (dFC) from multi-site resting-state fMRI data. The paper states: We propose FedDOSE, a novel framework that explicitly decomposes site differences for analysis of dynamic FC (dFC). The framework addresses three significant drawbacks of existing FL approaches for fMRI analysis: "(1) they primarily focus on static FC; (2) they often ignore site-specific characteristics such as phenotypic factors (e.g., age and gender) and scanner variability; and (3) they continue to face practical challenges arising from data heterogeneity."

The method introduces three key innovations: "(1) we propose a Modularity-Guided Tucker Decomposition (TKD) (MGTKD) block to address the high dimensionality of dFC while also serves as a feature extractor for modular-level spatio-temporal patterns; (2) we decompose site differences into disease, phenotypic and scanner subspaces; (3) we construct class-specific prototypes from the factors obtained via the MGTKD block by capturing spatio-temporal coupling within each module and we align all site prototypes at the global level by correcting for distributional heterogeneity using OT and Procrustes analysis."

For the MGTKD block, the paper describes: "For each subject, we have a dFC tensor X ∈ RN ×N ×T. Since each temporal slice Xt lies on a non-linear space of symmetric positive-semi definite matrices, we first apply the matrix logarithm under Riemannian geometry to map the matrix to the tangent space at the identity matrix. The tensor is divided into modular sub-tensors using binary selection matrices, and TKD is applied to each: X̃ m ≈ Gm ×1 (Usm) ×2 (Usm) ×3 (Ut) where Usm ∈ RNm ×Rs is the module-specific spatial factor and Ut ∈ RT ×Rt is the shared temporal factor across all modules. The reconstruction loss is defined as Lrecon = log(X) − X̂ 2 F + PM m=1 Usm ⊤ Usm − IRs 2 F which ensures that each module’s spatial factor remains orthonormal during backpropagation."

For site effect decomposition, the paper states: we decompose the modular core tensors into (i) disease-related subspace, (ii) phenotype-preserving subspace, and (iii) scanner-related subspace using softmax-based gating to the aggregated modular core tensors. The disease subspace loss is Ldisease = LCE (zd, y) where y is the ground-truth disease label. The phenotype loss is Lpheno defined as LMSE (zb, a) where a is the subject-level phenotypic embedding. The scanner loss uses a gradient reversal layer: Lscanner = LMSE (zv, v) + LMSE (GRL(zd), v) + LMSE (GRL(zb), v) which prevents scanner-related variance from leaking into the disease and phenotypic subspaces.

For prototype learning, the paper describes: For each module m, the spatio-temporal coupling map F m is defined as: F m = Usm Σ Ut⊤ which captures how the spatial activity of each ROI in module m evolves over time. Class prototypes are computed as pck = Pomega i=1 PM ′ m=1 αim · Fim where α are attention logits. The prototype regularization loss is Lproto = 1 − PK k=1 PC c=1 ⟨pck, pg ⟩ ∥pck ∥F · ∥pg ∥F which encourages each local model to align its prototypes with the global prototype based on directional similarity while being scale-invariant. The total local objective is Ltotal = Ldisease + µLrecon + γ(Lpheno + Lscanner) + λLproto.

For global aggregation, the paper states: we first align local prototypes using generalized Procrustes analysis, where OT is used to infer cross-site mappings and define a shared reference space. The OT barycenter formulation is pref = arg min p PK k=1 W22 (p, pk) where pref denotes the reference distribution that minimizes the aggregate Wasserstein distance across all site prototype distributions. Then each site's prototype is aligned via rk = arg min r ∥pref − pk r∥F with the closed-form solution rk = Uk′ Vk′⊤ obtained via SVD. Finally, pg = PK k=1 wk · pk rk where wk denotes the phenotype-based saliency score for site k.

The framework was evaluated on three multi-site rs-fMRI datasets: ABIDE-I, ABIDE-II, and ADHD-200, with approximately 2000 subjects total. The paper reports: FedDOSE outperforms other FL methods in both site and global accuracy across all three datasets. Compared to FedGMKD (the strongest baseline), the gains are most pronounced on ABIDE-I and ABIDE-II, with an average improvement of 4% across all sites. On ADHD-200, the improvement is smaller at 3%. The paper also notes: The Centralised baseline serves as a privacy-violating upper bound for FL. FedDOSE narrows or closes this performance gap (in site accuracy) while maintaining high global accuracy.

Ablation studies on ABIDE-I showed a significant drop in performance when the phenotypic and scanner subspaces, along with their associated loss terms, are removed and Removing geometric alignment of local prototypes using Procrustes analysis and OT resulted in inferior performance compared to the full architecture. Sensitivity analysis found optimal hyperparameters: we set µ = 0.05, γ = 0.1, and λ = 0.2 as the default values.

Biomarker analysis revealed connections involving the Sensorimotor Network (SM) and Dorsal Attention Network appear salient in sites NYU and UM, but are notably absent in UCLA and connections between the SM and Limbic Network (Li.) appear salient in NYU and UCLA, but are absent in UM, suggesting these FC patterns may be age-specific. Cross-site generalization testing showed FedDOSE consistently outperformed state-of-the-art models in terms of generalization accuracy, demonstrating FedDOSE effectively identifies and decomposes site-specific signatures while consistently capturing invariant ASD features.

The paper concludes: FedDOSE demonstrates strong performance across two ASD datasets and one ADHD dataset, outperforming seven state-of-the-art FL frameworks and notes One possible limitation of this study is that FedDOSE may introduce computational overhead in the federated training process. Nevertheless, the performance gains of the framework justify its added complexity.

Improvements to AI Systems:

1. Domain-Invariant Feature Decomposition for Multi-Site Medical Data: Implement a modular decomposition framework that explicitly separates data into disease-relevant, phenotype-preserving, and scanner-related subspaces. This allows AI systems to learn robust, generalizable features across heterogeneous data sources (e.g., different hospitals, scanners, demographics) without requiring data sharing, improving cross-site diagnostic accuracy.

2. Riemannian Geometry-Aware Tensor Processing: Use matrix logarithm mapping under Riemannian geometry to handle non-linear, symmetric positive semi-definite matrices (e.g., dynamic functional connectivity). This enables AI systems to process high-dimensional spatio-temporal data more accurately, reducing distortion from Euclidean approximations and improving feature extraction for time-varying brain states.

3. Modularity-Guided Tucker Decomposition (MGTKD) for High-Dimensional Tensors: Integrate modular sub-tensor decomposition with orthonormal constraints to reduce dimensionality while preserving modular-level spatio-temporal patterns. This allows AI systems to efficiently analyze large-scale dynamic data (e.g., fMRI, EEG) with lower computational cost and better interpretability of localized patterns.

4. Prototype Alignment via Optimal Transport and Procrustes Analysis: Construct class-specific prototypes from decomposed factors and align them across sites using Wasserstein barycenters and generalized Procrustes analysis. This improves AI systems’ ability to learn consistent, scale-invariant representations across distributed clients, enhancing federated learning performance under non-IID data distributions.

5. Adversarial Subspace Disentanglement with Gradient Reversal: Apply gradient reversal layers to prevent scanner-related variance from leaking into disease and phenotype subspaces. This enables AI systems to become more robust to nuisance variables, reducing overfitting to site-specific artifacts and improving generalization to unseen data sources.

6. Phenotype-Conditioned Global Aggregation: Use phenotype-based saliency scores to weight prototype contributions during global model aggregation. This allows AI systems to prioritize informative sites while down-weighting noisy or less relevant ones, leading to more accurate and fair global models in federated settings.

7. Biomarker Discovery via Modular Spatio-Temporal Coupling: Leverage the spatio-temporal coupling maps (e.g., F m = U s Σ U t T) to identify salient connectivity patterns across modules. This enables AI systems to generate interpretable, module-level biomarkers (e.g., specific brain network interactions) that are invariant across sites, aiding in clinical diagnosis and hypothesis generation.

8. Privacy-Preserving Upper-Bound Performance: By narrowing the gap to centralized models while maintaining privacy, this framework allows AI systems to achieve near-centralized accuracy in federated settings, making them viable for sensitive medical data where data cannot leave local institutions.

What the Improved AI System Can Do:

• Accurately classify neurological and psychiatric disorders (e.g., ASD, ADHD) from multi-site fMRI data with 3–4% higher accuracy than state-of-the-art federated methods.

• Generalize to unseen sites or scanners without retraining, maintaining high diagnostic performance despite domain shifts.

• Provide interpretable, site-invariant biomarkers of brain network dysfunction, revealing age- or phenotype-specific connectivity patterns.

• Operate in privacy-constrained environments (e.g., hospitals, research consortia) without sharing raw patient data, while approaching the performance of centralized models.

• Handle large-scale, high-dimensional dynamic connectivity data efficiently, enabling real-time or near-real-time analysis in clinical settings.

Functional Magnetic Resonance Imaging (fMRI) data are often pooled into collaborative multi-site consortia, as deep learning models for analyses require large datasets to generalize well. While Federated Learning (FL) offers a privacy-preserving paradigm for collaborative training, standard approaches continue to struggle with statistical heterogeneity. In particular, site differences pose a key challenge in multi-site data settings. Additionally, existing FL approaches for fMRI rely on static Functional Connectivity (FC), omitting dynamic information in brain networks. To address this, we propose FedDOSE, a novel framework that explicitly decomposes site differences for analysis of dynamic FC (dFC). FedDOSE introduces a Modularity-Guided Tucker Decomposition block to encode high-dimensional dFC tensors and capture modular-level spatio-temporal patterns efficiently. Class-specific prototypes are generated across all sites and subsequently aligned at the global level by using a combination of Optimal Transport (OT) barycenter formulation and Procrustes analysis. Extensive experiments for diagnosing Autism Spectrum Disorder (ASD) and Attention-Deficit Hyperactivity Disorder (ADHD) on three multi-site resting-state fMRI datasets: ABIDE-I, ABIDE-II, and ADHD-200, demonstrate that FedDOSE outperforms state-of-the-art methods in ASD and ADHD detection. Our results highlight its effectiveness in learning robust representations from multi-site datasets for reliable analysis.