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

arXiv:2608.07393v1 [cs.LG] 7 Aug 2026

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

Deepank Girish Nanyang Technological University deepank002@e.ntu.edu.sg

Yi Hao Chan Nanyang Technological University yihao001@e.ntu.edu.sg

Yubin Zheng Shanghai Jiao Tong University 1395115998@qq.com

Sukrit Gupta Indian Institute of Technology Ropar sukrit@iitrpr.ac.in

Jagath C. Rajapakse ∗ Nanyang Technological University ASJagath@ntu.edu.sg

Abstract 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 privacypreserving 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-theart methods in ASD and ADHD detection. Our results highlight its effectiveness in learning robust representations from multi-site datasets for reliable analysis.

1

Introduction

Recent advances in deep learning have significantly enhanced fMRI analysis, enabling improved characterization of brain states associated with neurodevelopmental conditions [19, 9]. These methods rely on large-scale datasets to achieve better generalizability and accuracy, but such data are often siloed across institutions due to strict privacy regulations [28]. Federated learning (FL) enables multiple imaging sites to collaboratively train models while keeping raw data local, thereby ∗

This research is supported by AcRF Tier-1 grant RG15/24 of Ministry of Education, Singapore.

Preprint.

mitigating privacy concerns through decentralized training. Recent studies in multi-site fMRI analysis have increasingly adopted the FL framework [13, 44]. However, existing FL approaches for fMRI analysis face three significant drawbacks: (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. In static FC, connectivity is computed over an entire scan, lasting several minutes. dFC extends static FC by capturing temporal variations in brain connectivity patterns at much faster timescales, on the order of seconds. Although dFC analysis is known to outperform static FC in detecting neurodevelopmental disorders [20, 47], it typically introduces high dimensionality. If not addressed, this may lead to overfitting and high communication costs in FL settings. Data harmonization (e.g., ComBat [15]) and deep generative models (e.g., variational autoencoders [23]) have been widely used to address heterogeneity across site distributions as they mitigate site-related differences. However, these methods may also remove biologically relevant variance correlated with the site (e.g., age-related connectivity differences across populations). This may induce spurious invariance, which can potentially degrade predictive performance by attenuating informative signals and individual identifiability [38]. To better understand site differences, prior work has decomposed them into measurement bias (scanner effects) and sampling bias (participant differences) [40]. However, this approach relies on a small traveling-subject dataset and estimates these biases using simple statistical assumptions. It is imperative to design a principled framework that explicitly separates the underlying sources of site variability. Doing so helps prevent these factors from entangling with disease-relevant signals in FC representations. To address these challenges, we introduce FedDOSE (Federated Decomposition Of Site Effects), a novel framework that performs a supervised decomposition of site-related variability via dFC. The overall architecture of the proposed framework is illustrated in Figure 1. Our approach 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 modularlevel 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. After alignment, we construct the global prototype using site-specific phenotypic embeddings. FedDOSE is extensively evaluated on three multi-site rs-fMRI datasets covering two neurodevelopmental disorders, namely ASD and ADHD, with a total of approximately 2000 subjects. We demonstrate that FedDOSE consistently outperforms other state-of-the-art FL methods in detecting both ASD and ADHD, while providing high interpretability, thereby enhancing its clinical utility.

2

Related Work

Substantial research has been conducted in FL since its introduction with FedAvg [22]. Although simple and based on gradient descent, the non-IID nature of client data led to poor generalisation and remains a challenge in FL. Methods such as FedProto [32] and FD [14] address this challenge through prototype learning and knowledge distillation. Prototype learning aggregates and exchanges class prototypes or averaged feature representations instead of model gradients. Knowledge distillation transfers knowledge from one model to another, enabling efficient representations with low communication cost. Another widely adopted strategy is personalized FL methods, which are known to adapt well to the characteristics of local data. pFedMe [30] is a popular personalized FL method that uses a Moreau envelope-based regularised loss function to decouple the optimization of personalized models from global model learning. Similarly, FPE [35] uses parameters from the trained global model to further adapt to local data. However, these models generalize well to local data but struggle to generalize globally and are also computationally demanding. Due to the privacy-preserving nature of medical data, FL has also gained traction in neuroimaging, particularly in fMRI data analysis. FedBrain [41] utilises a graph neural network–based FL framework 2

that incorporates brain connectome properties for disease prediction using multimodal imaging data. FedNI [25] performs node classification via a GCN-based population graph and leverages network inpainting for disorder detection. Domain adaptation has also been used to align distributions in multi-site fMRI data analysis [18]. Their method trains a domain discriminator at each site to mitigate domain shift. However, most existing approaches do not utilise non-imaging priors (e.g., age, gender, scanners) during aggregation to mitigate site heterogeneity.

3

Methods

Figure 1: Overview of the proposed FedDOSE framework. At the local level, the Modularity-Guided Tucker Decomposition (MGTKD) block extracts compressed modular-level spatiotemporal features from dFC matrices. These features are decomposed into disease, phenotype, and scanner subspaces to separate site differences. Spatial and temporal factors from the MGTKD block are coupled via attention matrix to construct prototypes. At the global level, site prototypes are aligned using OT and Procrustes analysis and then combined with phenotypic saliency scores to form the global prototype. 3.1

Functional Connectivity Representation

For each participant, fMRI data are represented as a sequence of dynamic FC matrices generated using a sliding window over time, indexed by t ∈ {1, 2, 3, . . . , T }. For each time window, an FC matrix Xt ∈ RN ×N is constructed, where N denotes the number of brain regions of interest (ROIs). Each entry represents the connectivity strength between a pair of ROIs. Xt is defined as the Pearson correlation of each node with all other nodes. We further construct an adjacency matrix Et from Xt using an orthogonal minimum spanning tree filtering scheme [6], which removes spurious edges while preserving informative weak connections, yielding a more reliable connectivity representation. 3.2

Modularity-Guided Tucker Decomposition

Prior studies have applied TKD to fMRI datasets [11, 12]. However, these methods were primarily tailored to static FC-based studies involving small tensor representations. Applying such methods to large tensors like dFC matrices is computationally expensive and may also disrupt the spatial and temporal structure of the data [45]. Moreover, tensor decomposition approaches do not incorporate known properties of the connectome, such as the modular structure of FC and the presence of brain hubs. Modularity in FC refers to the organization of brain networks into groups of densely connected ROIs that have specific brain functions [29]. To address aforementioned challenges, we propose a learnable MGTKD block that captures compressed modular dFC features for downstream analyses. 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 [34], we first apply the matrix logarithm under Riemannian geometry to map the matrix to the tangent space at the identity matrix. We perform this step because our adopted TKD algorithm [33] assumes Euclidean structure and applying Euclidean operations on Xt can introduce spurious artifacts [24]. To reduce computational 3

complexity, we divide the large dFC tensor into smaller modular sub-tensors. For each brain module m ∈ {1, 2, 3, . . . , M ′ }, we define a binary selection matrix B m ∈ {0, 1}Nm ×N , where Nm denotes the number of ROIs assigned to module m. Using this, we extract the corresponding modular sub-tensor as:   X̃ m = log(X) ×1 (B m ⊤ ) ×2 (B m ⊤ ) (1) where ×n denotes tensor multiplication applied along the nth dimension of the tensor. We then apply TKD to each sub-tensor corresponding to module m. This step enables faster computation and provides good approximation at a much reduced size. It is given by: X̃ m ≈ Gm ×1 (Usm ) ×2 (Usm ) ×3 (Ut )

(2)

where Usm ∈ RNm ×Rs is the module-specific spatial factor and Ut ∈ RT ×Rt is the shared temporal factor across all modules. Rs and Rt denote the Tucker ranks controlling the compression. We cyclically apply Ut across all modules during the decomposition step to preserve global temporal dynamics. Gm ∈ RRs ×Rs ×Rt denotes the module-specific core tensor, capturing spatial and temporal patterns within module m. The core optimization objective of the MGTKD block is to learn robust feature representations for Gm . This is achieved by reconstructing the original dFC tensor as follows: ′

X̂ =

M X

Gm ×1 (B m Usm ) ×2 (B m Usm ) ×3 (Ut )

(3)

m=1

The reconstruction loss Lrecon is then formulated as: ′

Lrecon = log(X) − X̂

2

+ F

M X m=1

Usm ⊤ Usm − IRs

2

(4)

F

where ∥ · ∥F represents the Frobenius distance. The second term in Lrecon ensures that each module’s spatial factor remains orthonormal during backpropagation. This enforces the spatial patterns to be independent across columns and more interpretable. Modular core tensors from all modules are aggregated and then passed to the site-effect decomposition block (Section 3.3). The residual shared spatial and temporal factors are subsequently used for prototype construction (Section 3.4). 3.3

Decomposition of Site Effects

In this section, we decompose the modular core tensors into (i) disease-related subspace, (ii) phenotype-preserving subspace, and (iii) scanner-related subspace. We apply softmax-based gating to the aggregated modular core tensors to partition the variance in the core tensors into three orthogonal subspaces, which is given by: ′

[Md , Mb , Mv ] = Softmax(MLP(

M X

′

m

G )),

m=1

zd = MLP(vec(Md ⊙

M X

Gm ))

(5)

m=1

where ⊙ denotes the Hadamard product and vec(·) is the vectorization operator. zd represents the disease-space embedding obtained by flattening the disease-space mask Md and projecting it into a subspace representation. Similar steps are used to obtain the phenotype embedding zb and scanner embedding zv . We then define individual loss terms for each of the three subspaces to perform downstream tasks. The disease subspace loss term Ldisease is defined as Ldisease = LCE (zd , y), where y denotes the ground-truth disease labels and LCE represents the cross-entropy loss. Similarly, the phenotype subspace loss term Lpheno is defined as LMSE (zb , a), where a denotes the subject-level phenotypic embeddings and LMSE represents the mean squared error. In contrast to the other two losses, the scanner subspace loss term Lscanner is designed to capture scanner-specific information while enforcing invariance to the disease and phenotypic representations. This is performed using a gradient 4

reversal layer [8], which acts as an identity function during the forward pass and reverses the gradients during backpropagation. It is defined as: Lscanner = LMSE (zv , v) + LMSE (GRL(zd ), v) + LMSE (GRL(zb ), v)

(6)

where v denotes the subject-level scanner embeddings. This design choice prevents scanner-related variance from leaking into the disease and phenotypic subspaces. In this way, each source of site variation is encouraged to be captured in each distinct subspace. 3.4

Spatiotemporal Coupling for Prototype Learning

Each site computes its class prototypes in the disease subspace using the per-module spatial factors and temporal factor from the MGTKD block. We compute attention logits over all modules to weight each module’s contribution to the prototype as: α = Softmax(W ′ zd ) ′

where W ′ is a learnable modular attention matrix and α = (αm )M m=1 . The disease subspace embedding zd is transformed and reshaped into a coupling matrix Σ using a MLP. For each module m, the spatio-temporal coupling map F m is defined as: F m = Usm Σ Ut⊤

(7)

F m captures how the spatial activity of each ROI in module m evolves over time. For each site k and subject class c, spatio-temporal prototype pck is defined as: ′

pck =

Ω X M X

αim · Fim

(8)

i=1 m=1

where Ω denotes the total number of subjects belonging to class c at site k. The class prototypes are shared across all sites in our framework. Lastly, we introduce a prototype regularization term Lproto to use cosine similarity to measure the similarity between local and global prototypes. This encourages each local model to align its prototypes with the global prototype based on directional similarity while being scale-invariant. It is given by: Lproto = 1 −

K X C X k=1

⟨pck , pg ⟩ ∥pck ∥F · ∥pg ∥F c=1

(9)

where K denotes the total number of sites and C is the total number of classes. The total local objective Ltotal at the local level is defined as the combination of five loss functions described above: Ltotal = Ldisease + µLrecon + γ(Lpheno + Lscanner ) + λLproto

(10)

We control the contribution of Lrecon using µ and Lproto using λ as they operate on a different scales. We control the combination of Lpheno and Lscanner using γ, as disease detection is our primary task objective and is encouraged to be invariant to site-related variability. After prototype computation, the average phenotypic embedding across all subjects is computed. Together with the model parameters and class prototypes from each site, it is transmitted to the global module. 3.5

Prototype Alignment via OT

Upon receiving the class prototypes, model parameters and phenotypic embedding, the global module performs a two-stage aggregation. In the first stage, it aggregates the local model parameters via PK 1 standard federated averaging [22], yielding the updated global model: fg = K k=1 fk In the second stage, aggregating prototypes using simple averaging is suboptimal. This is because Tucker decompositions are performed independently across sites, which leads to misalignment in the ordering of latent components. This introduces heterogeneity among prototypes. For instance, a spatial factor corresponding to the default mode network (DMN) in one site may correspond to a different functional network, such as the salience network (SN), in another site. To address this, we first align local prototypes using generalized Procrustes analysis [10], where OT is used to infer cross-site mappings and define a shared reference space. Procrustes-based methods have been 5

widely used in neuroscience, particularly in fMRI preprocessing, where they have been used to align functional and anatomical brain structures [1]. As it is reasonable to assume that local prototypes are drawn from different distributions, we employ an OT barycenter formulation [16] as it provides a principled approach for averaging the distributions of prototypes across sites: K X pref = arg min W22 (p, pk ) (11) p k=1 2 where W2 denotes the squared 2-Wasserstein distance function and p denotes the candidate reference prototypes. pref denotes the reference distribution that minimizes the aggregate Wasserstein distance

across all site prototype distributions. Once the reference matrix is obtained, each site’s prototype is aligned to it using the Procrustes problem. It is formulated as: rk = arg min ∥pref − pk r∥F (12) r

where rk is the rotational matrix for site k’s prototype. When multiplied by the site prototype, it aligns it as closely as possible to the reference prototype. r denotes the set of all possible rotation matrices and is the variable over which optimization is performed. The closed-form solution is obtained via singular value decomposition (SVD) of the cross-covariance matrix p⊤ k pref . The optimal rotation is ′ ′⊤ ′ ′ then given by rk = Uk Vk , where Uk and Vk are the left and right singular vectors obtained from the SVD. Finally, we can perform global prototype aggregation, defined as: K X pg = wk · pk rk (13) k=1

By aligning all site prototypes to a common reference coordinate system, we obtain geometrically and biologically meaningful representations. Here, wk denotes the phenotype-based saliency score for site k. It determines the contribution of each site to the global prototype. It is computed as the normalized inverse mean distance of its averaged phenotypic embedding to those of other sites. Finally, the global prototype pg is distributed back to all the sites for next round of end-to-end federated training.

4

Experiments and Results

4.1

Datasets and Preprocessing

We evaluated our framework on three public multi-site rs-fMRI datasets, demonstrating its generalizability and robustness. ABIDE-I [5]: The Autism Brain Imaging Data Exchange I (ABIDE-I) dataset contains 387 rs-fMRI scans from individuals diagnosed with ASD and 436 typically developing controls, collected from 20 sites. We selected three sites: NYU (88 NC / 73 ASD), UCLA (39 NC / 36 ASD), and UM (50 NC / 37 ASD). These sites were chosen as they are the largest in the dataset and have been used to benchmark model performance in other FL studies on rs-fMRI data. ABIDE-II [4]: Similar to ABIDE-I, the Autism Brain Imaging Data Exchange II (ABIDE-II) dataset comprises 107 individuals with ASD and 111 typically developing controls from three sites: BNI (29 NC / 29 ASD), EMC (27 NC / 27 ASD), and GU (55 NC / 51 ASD). ADHD-200 [2]: Data from the ADHD-200 dataset was used to further validate our findings. ADHD200 contains rs-fMRI scans from 301 subjects diagnosed with ADHD and 549 age-matched controls, collected from 6 sites. We selected NYU (212 NC / 184 ADHD), OHSU (125 NC / 112 ADHD), and PKU (116 NC / 78 ADHD), as the remaining sites suffered from low sample size (NI) and severe class imbalance (KKI and WUSTL). Following recommended practices [17, 42], we constructed dFC matrices by segmenting BOLD signals into overlapping windows of 60s length with a 1s stride. Power atlas [27] was used to define 264 ROIs. We used eight modules from the Power atlas for the MGTKD block. FC matrices were computed using Pearson correlation between the mean time series of each pair of ROIs. Resulting dFC matrices from each sliding window were used as model inputs. From each dataset, the phenotypic embedding is constructed by concatenating the subject’s age with an encoded representation of gender. Similarly, the scanner embedding is constructed by concatenating encoded site labels, encoded scanner-type information, and subject-specific temporal resolution. 6

4.2

Baselines

To assess the effectiveness of our framework, we implemented state-of-the-art FL models from both fMRI-specific and generic (non-fMRI) domains. These models include FedAvg [22], FedProto, TDPFed [37], GAFD [36], FedGST [21], FedAli [7], and FedGMKD [43]. FedAvg is the standard federated optimization method that constructs a global model by performing weighted parameter averaging across clients. In FedProto, class-level feature prototypes are exchanged instead of full model parameters. TDPFed uses TKD to design a tensorised local model and decouples personalized model optimisation from global model learning. GAFD introduced graph augmentation for static FC brain networks and applied knowledge distillation to improve multi-site aggregation. FedGST extracts spatiotemporal patterns using a GCN and dFC, with the temporal model trained on the client side and the spatial model trained on the server side. FedAli utilizes OT to enforce consistency between local and global representations, reducing client drift and improving convergence stability. Lastly, FedGMKD combines a knowledge distillation with differential aggregation to enable a personalized prototype-based FL framework. 4.3

Implementation Details

For all experiments, model training and evaluations were conducted using five seeds. All three datasets were divided into training, validation, and test sets at a 6:2:2 ratio. Gradient descent was done using the Adam optimizer with a learning rate of 0.005. The federated process ran for 40 rounds with 5 local epochs per round. Tuning was performed using grid search, where the tuned hyperparameters include µ ∈ {0.01, 0.05, 0.1}, γ ∈ {0.1, 0.3, 0.5}, and λ ∈ {0.1, 0.2, 0.3}. FedDOSE’s parameters were tuned using the validation set, and all reported results correspond to performance on the test set. We used a batch size of 8. All models were implemented using Python 3.10 and PyTorch 2.1 on an NVIDIA A100 GPU. We also compare FedDOSE against two standard learning approaches: Local and Centralised. In the Local setting, sites train independently on their own datasets. In the Centralised setting, a global model is trained on pooled data from all sites. Both settings are trained for 40 epochs. Default hyperparameters were used to train other baseline models. We evaluate performance using two metrics: site accuracy (Site Acc) and global accuracy (Global Acc). Site accuracy is the classification accuracy of each site’s locally adapted model on its own data after federated training. Global accuracy is the classification accuracy of the global model evaluated on each site’s data. 4.4

Performance comparison with the baselines

Table 1: Comparison of classification performance among existing methods on the ADHD-200 dataset across sites NYU, OHSU, and PKU. Results are reported as mean ± standard deviation. * indicates statistically different significance (p-value < 0.05) with FedDOSE. Models

NYU

OHSU

PKU

Site Acc

Global Acc

Site Acc

Global Acc

Site Acc

Global Acc

Local Centralised

0.58 ± 0.07 N/A

N/A 0.62 ± 0.05

0.60 ± 0.05 N/A

N/A 0.64 ± 0.06

0.62 ± 0.06 N/A

N/A 0.66 ± 0.06

FedAvg [22] FedProto [32] TDPFed [37] GAFD [36] FedGST [21] FedAli [7] FedGMKD [43] FedDOSE

0.54 ± 0.06* 0.56 ± 0.08* 0.57 ± 0.07 0.57 ± 0.09 0.59 ± 0.08 0.60 ± 0.08 0.61 ± 0.07 0.64 ± 0.07

0.51 ± 0.05* N/A 0.53 ± 0.06* 0.54 ± 0.06 0.55 ± 0.06 0.57 ± 0.04 0.57 ± 0.05 0.60 ± 0.06

0.55 ± 0.08* 0.56 ± 0.10* 0.58 ± 0.09 0.59 ± 0.07 0.60 ± 0.09 0.61 ± 0.06 0.62 ± 0.06 0.65 ± 0.07

0.52 ± 0.06* N/A 0.54 ± 0.08* 0.55 ± 0.08 0.57 ± 0.06 0.58 ± 0.07 0.59 ± 0.07 0.62 ± 0.06

0.53 ± 0.08* 0.55 ± 0.09* 0.56 ± 0.08* 0.58 ± 0.07* 0.61 ± 0.06* 0.63 ± 0.05 0.64 ± 0.05 0.68 ± 0.06

0.50 ± 0.07* N/A 0.53 ± 0.07* 0.56 ± 0.07* 0.58 ± 0.04* 0.60 ± 0.06 0.61 ± 0.04 0.65 ± 0.05

FedDOSE outperforms other FL methods in both site and global accuracy across all three datasets. Some of the observed improvements were statistically significant based on the Mann–Whitney U test. The experimental results on ADHD-200, ABIDE-I, and ABIDE-II are summarized in Tables 1, 2, and 4 (see Appendix), respectively. FedDOSE consistently achieves the highest site accuracy. Compared to FedGMKD, the strongest baseline, the gains are most pronounced on ABIDE-I and ABIDE-II, 7

Table 2: Comparison of classification performance among existing methods on the ABIDE-I dataset across sites NYU, UCLA, and UM. Results are reported as mean ± standard deviation. * indicates statistically different significance (p-value < 0.05) with FedDOSE. Models

NYU

UCLA

UM

Site Acc

Global Acc

Site Acc

Global Acc

Site Acc

Global Acc

Local Centralised

0.64 ± 0.05 N/A

N/A 0.69 ± 0.06

0.68 ± 0.06 N/A

N/A 0.74 ± 0.05

0.70 ± 0.05 N/A

N/A 0.75 ± 0.04

FedAvg [22] FedProto [32] TDPFed [37] GAFD [36] FedGST [21] FedAli [7] FedGMKD [43] FedDOSE

0.56 ± 0.07* 0.58 ± 0.08* 0.59 ± 0.07* 0.61 ± 0.06* 0.64 ± 0.08* 0.65 ± 0.06 0.67 ± 0.05 0.71 ± 0.06

0.52 ± 0.08* N/A 0.55 ± 0.09* 0.58 ± 0.06* 0.61 ± 0.07 0.62 ± 0.05 0.63 ± 0.05 0.67 ± 0.05

0.64 ± 0.09* 0.62 ± 0.09* 0.65 ± 0.08* 0.68 ± 0.07* 0.69 ± 0.08 0.71 ± 0.08 0.71 ± 0.07 0.75 ± 0.06

0.59 ± 0.07* N/A 0.61 ± 0.08* 0.62 ± 0.06* 0.64 ± 0.07* 0.67 ± 0.08 0.68 ± 0.05 0.71 ± 0.05

0.59 ± 0.09* 0.62 ± 0.08* 0.65 ± 0.07* 0.67 ± 0.07* 0.68 ± 0.07* 0.71 ± 0.08 0.72 ± 0.06 0.77 ± 0.07

0.56 ± 0.08* N/A 0.59 ± 0.06* 0.64 ± 0.07* 0.66 ± 0.06* 0.68 ± 0.07 0.69 ± 0.05 0.73 ± 0.06

with an average improvement of 4% across all sites. On ADHD-200, the improvement is smaller at 3%, but follows the same trend. This suggests that FedDOSE’s global alignment mechanism and decomposition of site differences are particularly beneficial for larger datasets with greater inter-site heterogeneity. 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. On the ABIDE-I dataset, FedDOSE achieves a residual performance gap of about 2% across all three sites. On ADHD-200, this gap further reduces to below 2%, with FedDOSE nearly matching the centralised upper bound at site PKU. We observe similar trends in ABIDE-II, where FedDOSE nearly matches the performance of the centralised baseline at site EMC. These results indicate its potential as a reliable FL method. Another consistent trend across all three datasets is that vanilla aggregation and simple prototype-based methods achieve the lowest performance. Brain network–modeled methods show modest improvements, while prototype-based personalization methods with alignment strategies and knowledge distillation achieve comparatively higher performance levels. These results suggest that the gains in FedDOSE’s performance are likely systematic rather than dataset-specific. 4.5

Ablation Studies

Table 3: Ablation study of FedDOSE on the ABIDE-I dataset across sites NYU, UCLA, and UM. Results are reported as mean ± standard deviation. * indicates statistically different significance (p-value < 0.05) with FedDOSE. Models

NYU Site Acc

w/o Lpheno , Lscanner w/o Lrecon w/o Procrustes w/o Lpheno w/o OT w/o Lproto FedDOSE

Global Acc

UCLA Site Acc

Global Acc

UM Site Acc

Global Acc

0.60 ± 0.06* 0.55 ± 0.05* 0.61 ± 0.06* 0.53 ± 0.06* 0.61 ± 0.06* 0.55 ± 0.05* 0.62 ± 0.06* 0.57 ± 0.04* 0.63 ± 0.07* 0.56 ± 0.05* 0.64 ± 0.06* 0.56 ± 0.04* 0.65 ± 0.05* 0.59 ± 0.04* 0.67 ± 0.05* 0.60 ± 0.04* 0.68 ± 0.05* 0.61 ± 0.06* 0.65 ± 0.06* 0.60 ± 0.05* 0.66 ± 0.06* 0.58 ± 0.06* 0.69 ± 0.07* 0.62 ± 0.05* 0.67 ± 0.07 0.61 ± 0.05* 0.70 ± 0.07 0.64 ± 0.06* 0.72 ± 0.07 0.65 ± 0.05* 0.69 ± 0.07 0.64 ± 0.06 0.71 ± 0.06 0.66 ± 0.07 0.74 ± 0.08 0.68 ± 0.07 0.71 ± 0.06 0.67 ± 0.05 0.75 ± 0.06 0.71 ± 0.05 0.77 ± 0.07 0.73 ± 0.06

To validate the effectiveness of our proposed method, we conducted ablation studies on the ABIDE-I dataset. The results, summarized in Table 3, show a significant drop in performance when the phenotypic and scanner subspaces, along with their associated loss terms, are removed. This highlights the importance of decomposing site differences into learned subspaces. Otherwise, such confounded site factors may interfere with disease-related signals and degrade predictive performance. Furthermore, removing the reconstruction loss from the MGTKD block and applying a direct TKD to the large dFC tensor led to a noticeable performance drop. Removing geometric alignment of local prototypes using Procrustes analysis and OT resulted in inferior performance compared to the full 8

architecture. This confirms the effectiveness of our prototype alignment strategy in addressing site heterogeneity. Our ablation experiments demonstrate the importance of each component of FedDOSE for distinguishing ASD subjects from typically developing controls. 4.6

Sensitivity Analysis

Figure 2: FedDOSE mean global accuracy on the ABIDE-I dataset for different hyperparameter pairs: (λ, γ) with µ fixed (Left), (µ, λ) with γ fixed (Middle), and (γ, µ) with λ fixed (Right). The red circle highlights the highest accuracy obtained in each plot. We conducted a sensitivity analysis on the ABIDE-I dataset to evaluate the effects of varying the hyperparameters µ, γ, and λ within the FedDOSE framework. Evaluation was conducted by averaging the global accuracy across three seeds. We fixed one hyperparameter at its optimal value and varied the other two across multiple combinations. Our findings indicate that λ and γ are optimal at 0.2 and 0.1, respectively. Across all sites, performance with respect to λ typically peaked at 0.2 before declining. In contrast, global accuracy consistently decreased with increasing γ. µ showed an increasing trend up to 0.05, after which it dropped for higher values. At site UM, µ was highest at 0.01, which differed from the other two sites. Model performance generally degraded at extreme hyperparameter values. Based on this empirical evaluation, we set µ = 0.05, γ = 0.1, and λ = 0.2 as the default values for downstream analyses. Overall, the analysis indicates that FedDOSE is robust across all sites, maintaining stable accuracy across different settings.

5

Discussion and Conclusion

This study presents several key findings. FedDOSE demonstrates strong performance across two ASD datasets and one ADHD dataset, outperforming seven state-of-the-art FL frameworks. This can be attributed to FedDOSE’s supervised decomposition of site differences into disease, phenotype, and scanner subspaces. The integration of TKD and brain modularity for capturing spatiotemporal patterns in dFC matrices was another key factor contributing to FedDOSE’s superior performance. Dividing the dFC tensor into modular subtensors and introducing a reconstruction regularization term helped preserve meaningful spatial network topology. The cyclic temporal factor further enabled the modular core tensor to learn robust representations. 9

Due to distributional heterogeneity across site prototypes, we first align them by solving the Procrustes problem using a reference matrix constructed via OT. We then perform a weighted aggregation of the aligned prototypes using site-specific phenotypic saliency scores. This resulted in high global accuracy as observed in both baseline and ablation results. Through hyperparameter sensitivity analysis, we show that FedDOSE is robust across various settings. Our biomarker analysis (Section A.1) also revealed that FedDOSE can capture reliable biomarkers, highlighting the interpretability of our framework. 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. We would also like to clarify that our primary goal is to develop a framework that can be trained on multi-site datasets, rather than designing a systems-level architecture. Future work could focus on developing a client–server architecture and studying communication costs, training efficiency, and scalability. Future work could also explore other neurological disorders and extend the approach to benchmarks beyond fMRI data.

10

References [1] Angela Andreella, Riccardo De Santis, Anna Vesely, and Livio Finos. Procrustes-based distances for exploring between-matrices similarity. Statistical Methods & Applications, 32(3):867–882, 2023. [2] Matthew RG Brown, Gagan S Sidhu, Russell Greiner, Nasimeh Asgarian, Meysam Bastani, Peter H Silverstone, Andrew J Greenshaw, and Serdar M Dursun. Adhd-200 global competition: diagnosing adhd using personal characteristic data can outperform resting state fmri measurements. Frontiers in systems neuroscience, 6:69, 2012. [3] Robert M Califf. Biomarker definitions and their applications. Experimental biology and medicine, 243(3):213–221, 2018. [4] Adriana Di Martino, David O’connor, Bosi Chen, Kaat Alaerts, Jeffrey S Anderson, Michal Assaf, Joshua H Balsters, Leslie Baxter, Anita Beggiato, Sylvie Bernaerts, et al. Enhancing studies of the connectome in autism using the autism brain imaging data exchange ii. Scientific data, 4(1):1–15, 2017. [5] Adriana Di Martino, Chao-Gan Yan, Qingyang Li, Erin Denio, Francisco X Castellanos, Kaat Alaerts, Jeffrey S Anderson, Michal Assaf, Susan Y Bookheimer, Mirella Dapretto, et al. The autism brain imaging data exchange: towards a large-scale evaluation of the intrinsic brain architecture in autism. Molecular psychiatry, 19(6):659–667, 2014. [6] Stavros I Dimitriadis, Christos Salis, Ioannis Tarnanas, and David E Linden. Topological filtering of dynamic functional brain networks unfolds informative chronnectomics: a novel data-driven thresholding scheme based on orthogonal minimal spanning trees (omsts). Frontiers in neuroinformatics, 11:28, 2017. [7] Sannara Ek, Kaile Wang, François Portet, Philippe Lalanda, and Jiannong Cao. Fedali: Personalized federated learning alignment with prototype layers for generalized mobile services. arXiv preprint arXiv:2411.10595, 2024. [8] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180–1189. PMLR, 2015. [9] Deepank Girish, Yi Hao Chan, Sukrit Gupta, Jing Xia, and Jagath C Rajapakse. Encoding and decoding of brain dynamic functional connectivity for adhd diagnosis. IEEE Journal of Biomedical and Health Informatics, 2026. [10] John C Gower. Generalized procrustes analysis. Psychometrika, 40(1):33–51, 1975. [11] Yue Han, Qiu-Hua Lin, Li-Dan Kuang, Xiao-Feng Gong, Fengyu Cong, and Vince D Calhoun. Tucker decomposition for extracting shared and individual spatial maps from multi-subject resting-state fmri data. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1110–1114. IEEE, 2021. [12] Yue Han, Qiu-Hua Lin, Li-Dan Kuang, Bin-Hua Zhao, Xiao-Feng Gong, Fengyu Cong, Yu-Ping Wang, and Vince D Calhoun. A core tensor sparsity enhancement method for solving tucker-2 model of multi-subject fmri data. Biomedical Signal Processing and Control, 95:106471, 2024. [13] Zhi-An Huang, Yao Hu, Rui Liu, Xiaoming Xue, Zexuan Zhu, Linqi Song, and Kay Chen Tan. Federated multi-task learning for joint diagnosis of multiple mental disorders on mri scans. IEEE Transactions on Biomedical Engineering, 70(4):1137–1149, 2022. [14] Eunjeong Jeong, Seungeun Oh, Hyesung Kim, Jihong Park, Mehdi Bennis, and Seong-Lyun Kim. Communication-efficient on-device machine learning: Federated distillation and augmentation under non-iid private data. arXiv preprint arXiv:1811.11479, 2018. [15] W Evan Johnson, Cheng Li, and Ariel Rabinovic. Adjusting batch effects in microarray expression data using empirical bayes methods. Biostatistics, 8(1):118–127, 2007. 11

[16] Alexander Kolesov, Petr Mokrov, Igor Udovichenko, Milena Gazdieva, Gudmund Pammer, Anastasis Kratsios, Evgeny Burnaev, and Alexander Korotin. Energy-guided continuous entropic barycenter estimation for general costs. Advances in Neural Information Processing Systems, 37:107513–107546, 2024. [17] Nora Leonardi and Dimitri Van De Ville. On spurious and real fluctuations of dynamic functional connectivity during rest. Neuroimage, 104:430–436, 2015. [18] Xiaoxiao Li, Yufeng Gu, Nicha Dvornek, Lawrence H Staib, Pamela Ventola, and James S Duncan. Multi-site fmri analysis using privacy-preserving federated learning and domain adaptation: Abide results. Medical image analysis, 65:101765, 2020. [19] Xiaoxiao Li, Yuan Zhou, Nicha Dvornek, Muhan Zhang, Siyuan Gao, Juntang Zhuang, Dustin Scheinost, Lawrence H Staib, Pamela Ventola, and James S Duncan. Braingnn: Interpretable brain graph neural network for fmri analysis. Medical image analysis, 74:102233, 2021. [20] Lekai Luo, Lizhou Chen, Yuxia Wang, Qian Li, Ning He, Yuanyuan Li, Wanfang You, Yaxuan Wang, Fenghua Long, Lanting Guo, et al. Patterns of brain dynamic functional connectivity are linked with attention-deficit/hyperactivity disorder-related behavioral and cognitive dimensions. Psychological medicine, 53(14):6666–6677, 2023. [21] Junbin Mao, Hanhe Lin, Xu Tian, Yi Pan, and Jin Liu. Fedgst: Federated graph spatio-temporal framework for brain functional disease prediction. In 2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 1356–1361. IEEE, 2023. [22] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273–1282. Pmlr, 2017. [23] Daniel Moyer, Greg Ver Steeg, Chantal MW Tax, and Paul M Thompson. Scanner invariant representations for diffusion mri harmonization. Magnetic resonance in medicine, 84(4):2174– 2189, 2020. [24] Bernard Ng, Gael Varoquaux, Jean Baptiste Poline, Michael Greicius, and Bertrand Thirion. Transport on riemannian manifold for connectivity-based brain decoding. IEEE transactions on medical imaging, 35(1):208–216, 2015. [25] Liang Peng, Nan Wang, Nicha Dvornek, Xiaofeng Zhu, and Xiaoxiao Li. Fedni: Federated graph learning with network inpainting for population-based disease prediction. IEEE Transactions on Medical Imaging, 42(7):2032–2043, 2022. [26] Gabriel Peyré and Marco Cuturi. Computational optimal transport: With applications to data science. Now Foundations and Trends, 2019. [27] Jonathan D Power, Alexander L Cohen, Steven M Nelson, Gagan S Wig, Kelly Anne Barnes, Jessica A Church, Alecia C Vogel, Timothy O Laumann, Fran M Miezin, Bradley L Schlaggar, et al. Functional network organization of the human brain. Neuron, 72(4):665–678, 2011. [28] Micah J Sheller, G Anthony Reina, Brandon Edwards, Jason Martin, and Spyridon Bakas. Multi-institutional deep learning modeling without sharing patient data: A feasibility study on brain tumor segmentation. In International MICCAI Brainlesion Workshop, pages 92–104. Springer, 2018. [29] Olaf Sporns and Richard F Betzel. Modular brain networks. Annual review of psychology, 67(1):613–640, 2016. [30] Canh T Dinh, Nguyen Tran, and Josh Nguyen. Personalized federated learning with moreau envelopes. Advances in neural information processing systems, 33:21394–21405, 2020. [31] Ken Takeda, Masaru Sasaki, Kota Abe, and Masafumi Oizumi. Unsupervised alignment in neuroscience: Introducing a toolbox for gromov–wasserstein optimal transport. Journal of Neuroscience Methods, 419:110443, 2025. 12

[32] Yue Tan, Guodong Long, Lu Liu, Tianyi Zhou, Qinghua Lu, Jing Jiang, and Chengqi Zhang. Fedproto: Federated prototype learning across heterogeneous clients. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pages 8432–8440, 2022. [33] Ledyard R Tucker. Some mathematical notes on three-mode factor analysis. Psychometrika, 31(3):279–311, 1966. [34] Manasij Venkatesh, Joseph Jaja, and Luiz Pessoa. Comparing functional connectivity matrices: A geometry-aware approach applied to participant identification. NeuroImage, 207:116398, 2020. [35] Kangkang Wang, Rajiv Mathews, Chloé Kiddon, Hubert Eichner, Françoise Beaufays, and Daniel Ramage. Federated evaluation of on-device personalization. arXiv preprint arXiv:1910.10252, 2019. [36] Qianqian Wang, Junhao Zhang, Long Li, Lishan Qiao, Pew-Thian Yap, and Mingxia Liu. Graph augmentation guided federated knowledge distillation for multisite functional mri analysis. Pattern Recognition, 164:111559, 2025. [37] Qing Wang, Jing Jin, Xiaofeng Liu, Huixuan Zong, Yunfeng Shao, and Yinchuan Li. Tensor decomposition based personalized federated learning. arXiv preprint arXiv:2208.12959, 2022. [38] Yu-Wei Wang, Xiao Chen, and Chao-Gan Yan. Comprehensive evaluation of harmonization on functional brain imaging for multisite data-fusion. NeuroImage, 274:120089, 2023. [39] Prany Wantzen, Patrice Clochon, Franck Doidy, Fabrice Wallois, Mahdi Mahmoudzadeh, Pierre Desaunay, Mille Christian, Jean-Marc Guilé, Fabian Guénolé, Francis Eustache, et al. Eeg resting-state functional connectivity: evidence for an imbalance of external/internal information integration in autism. Journal of Neurodevelopmental Disorders, 14(1):47, 2022. [40] Ayumu Yamashita, Noriaki Yahata, Takashi Itahashi, Giuseppe Lisi, Takashi Yamada, Naho Ichikawa, Masahiro Takamura, Yujiro Yoshihara, Akira Kunimatsu, Naohiro Okada, et al. Harmonization of resting-state functional mri data across multiple imaging sites via the separation of site differences into sampling bias and measurement bias. PLoS biology, 17(4):e3000042, 2019. [41] Yi Yang, Han Xie, Hejie Cui, and Carl Yang. Fedbrain: federated training of graph neural networks for connectome-based brain imaging analysis. In PACIFIC SYMPOSIUM ON BIOCOMPUTING 2024, pages 214–225. World Scientific, 2023. [42] Han Zhang, Xiaobo Chen, Yu Zhang, and Dinggang Shen. Test-retest reliability of “high-order” functional connectivity in young healthy adults. Frontiers in neuroscience, 11:439, 2017. [43] Jianqiao Zhang, Caifeng Shan, and Jungong Han. Fedgmkd: An efficient prototype federated learning framework through knowledge distillation and discrepancy-aware aggregation. Advances in Neural Information Processing Systems, 37:118326–118356, 2024. [44] Junhao Zhang, Qianqian Wang, Xiaochuan Wang, Lishan Qiao, and Mingxia Liu. Preserving specificity in federated graph learning for fmri-based neurological disorder identification. Neural Networks, 169:584–596, 2024. [45] Chi Zhou, Qiu-Hua Lin, Yue Han, Jia-Yang Song, Xiao-Feng Gong, and Vince D Calhoun. Cyclic coupled tucker decomposition of large fmri datasets. In ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8162– 8166. IEEE, 2026. [46] Rui Zhou, Chenhao Sun, Mingxiang Sun, Yudi Ruan, Weikai Li, and Xin Gao. Altered intra-and inter-network connectivity in autism spectrum disorder. Aging (Albany NY), 16(11):10004, 2024. [47] Zhihao Zhu, Hongwei Wang, Hui Bi, Jidong Lv, Xiaotong Zhang, Suhong Wang, and Ling Zou. Dynamic functional connectivity changes of resting-state brain network in attentiondeficit/hyperactivity disorder. Behavioural Brain Research, 437:114121, 2023.

13

A

Technical Appendices and Supplementary Material

Table 4: Comparison of classification performance among existing methods on the ABIDE-II dataset across sites BNI, EMC, and GU. Results are reported as mean ± standard deviation. * indicates statistical significance (p-value < 0.05). Included here due to space constraints. Models

BNI

EMC

GU

Site Acc

Global Acc

Site Acc

Global Acc

Site Acc

Global Acc

Local Centralised

0.63 ± 0.05 N/A

N/A 0.69 ± 0.04

0.59 ± 0.04 N/A

N/A 0.64 ± 0.04

0.62 ± 0.05 N/A

N/A 0.67 ± 0.06

FedAvg [22] FedProto [32] TDPFed [37] GAFD [36] FedGST [21] FedAli [7] FedGMKD [43] FedDOSE

0.53 ± 0.09* 0.56 ± 0.08* 0.58 ± 0.09* 0.59 ± 0.07* 0.62 ± 0.08* 0.65 ± 0.08 0.67 ± 0.07 0.70 ± 0.05

0.51 ± 0.08* N/A 0.53 ± 0.08* 0.57 ± 0.08* 0.60 ± 0.07* 0.61 ± 0.06* 0.63 ± 0.06 0.67 ± 0.06

0.54 ± 0.11* 0.54 ± 0.09* 0.56 ± 0.09* 0.58 ± 0.08* 0.60 ± 0.08* 0.63 ± 0.08 0.65 ± 0.06 0.69 ± 0.06

0.51 ± 0.09* N/A 0.52 ± 0.08* 0.56 ± 0.09* 0.57 ± 0.08* 0.59 ± 0.06 0.61 ± 0.07 0.63 ± 0.05

0.54 ± 0.09* 0.56 ± 0.09* 0.57 ± 0.07* 0.59 ± 0.06* 0.61 ± 0.06* 0.62 ± 0.07 0.64 ± 0.06 0.68 ± 0.05

0.52 ± 0.08* N/A 0.53 ± 0.07* 0.56 ± 0.07* 0.58 ± 0.06* 0.60 ± 0.05 0.60 ± 0.05 0.64 ± 0.05

A.1

Biomarker Analysis

Figure 3: Chord diagrams representing site-specific salient FC connections of the ABIDE-I dataset. In each chord diagram, only the top 10 connections were visualized to reduce cluttering. (a) NYU (b) UCLA (c) UM. Brain network labels as follows:- DMN: Default Mode Network; FPCN: Frontoparietal control network; VAN: Ventral Attention Network; Unce.: Uncertain; Li.: Limbic Network; SM: Sensorimotor Network; Visual: Visual network; DAN: Dorsal Attention Network In brain imaging analysis, a biomarker generally refers to a brain region or connectivity pattern associated with the presence of a disease or condition of interest, or with a specific disease subtype [3]. Existing FL studies in fMRI analysis typically provide limited evaluation of framework interpretability and biomarker identification. To assess the effectiveness of FedDOSE, we performed a biomarker analysis at the modular level. Saliency scores for FedDOSE were derived by computing the outer product of the two modular spatial factors from the MGTKD block. We expected such a cross product to capture rich modular interactions between the primary and secondary patterns encoded by the two spatial factors. Each chord diagram in Figure 3 represents the top 10 salient connections. In Figure 3, connections involving the Sensorimotor Network (SM) and Dorsal Attention Network appear salient in sites NYU and UM, but are notably absent in UCLA. Similarly, connections between the SM and Limbic Network (Li.) appear salient in NYU and UCLA, but are absent in UM. This could suggest that these FC patterns may be age-specific, as UCLA generally recruits younger cohorts, whereas subjects from UM tend to be older. These observations align with the existing literature on ASD biomarkers [46, 39]. One key factor contributing to FedDOSE’s ability to produce reliable biomarkers is the orthogonality constraint in the Lrecon loss. By enforcing independence among spatial patterns, it enables their cross-product interactions to capture diverse and salient relationships. 14

A.2

Cross-Site Generalization Performance of FedDOSE

Table 5: Comparison study of FedDOSE through external validation on sites from the ABIDE-I dataset. Results are reported as mean ± standard deviation. * indicates statistical significance (p-value < 0.05). Models FedAvg [22] FedProto [32] TDPFed [37] GAFD [36] FedGST [21] FedAli [7] FedGMKD [43] FedDOSE

NYU

UCLA

UM

Site Acc

Genr. Acc

Site Acc

Genr. Acc

Site Acc

Genr. Acc

0.56 ± 0.07* 0.58 ± 0.08* 0.59 ± 0.07* 0.61 ± 0.06* 0.64 ± 0.08* 0.65 ± 0.06 0.67 ± 0.05 0.71 ± 0.06

0.49 ± 0.09* 0.52 ± 0.08* 0.55 ± 0.08* 0.56 ± 0.08 0.58 ± 0.09 0.59 ± 0.07 0.60 ± 0.06 0.62 ± 0.07

0.64 ± 0.09* 0.62 ± 0.09* 0.65 ± 0.08* 0.68 ± 0.07* 0.69 ± 0.08 0.71 ± 0.08 0.71 ± 0.07 0.75 ± 0.06

0.55 ± 0.07* 0.57 ± 0.09* 0.59 ± 0.09* 0.59 ± 0.06* 0.60 ± 0.07* 0.64 ± 0.06 0.65 ± 0.06 0.68 ± 0.07

0.59 ± 0.09* 0.62 ± 0.08* 0.65 ± 0.07* 0.67 ± 0.07* 0.68 ± 0.07* 0.71 ± 0.08 0.72 ± 0.06 0.77 ± 0.07

0.53 ± 0.10* 0.56 ± 0.09* 0.61 ± 0.09* 0.60 ± 0.09* 0.62 ± 0.08* 0.66 ± 0.08 0.66 ± 0.06 0.69 ± 0.07

To assess the reliability of FedDOSE, we evaluated each site’s locally adapted model on the test data from the other two sites after federated training and reported the averaged classification accuracy. We define this metric as generalization accuracy (Genr. Acc.) and used the ABIDE-I dataset for this study. From Table 5, we find that FedDOSE consistently outperformed state-of-the-art models in terms of generalization accuracy. This demonstrates that FedDOSE effectively identifies and decomposes site-specific signatures while consistently capturing invariant ASD features. This further highlights its reliability and strong generalization capability across multiple sites. A.3

Sensitivity Analysis of OT

Table 6: Evaluation of FedDOSE under different OT formulation strategies on the ABIDE-I dataset. Results are reported as mean ± standard deviation. * indicates statistical significance (p-value < 0.05). Formulation

NYU Site Acc

UCLA

Global Acc

Site Acc

Global Acc

UM Site Acc

Global Acc

None 0.62 ± 0.06* 0.56 ± 0.06* 0.65 ± 0.06* 0.59 ± 0.05* 0.69 ± 0.06* 0.61 ± 0.06* Kantorovich [26] 0.67 ± 0.08 0.60 ± 0.06 0.72 ± 0.07 0.67 ± 0.06 0.73 ± 0.07 0.66 ± 0.07 Gromov–Wasserstein [31] 0.68 ± 0.06 0.60 ± 0.06 0.74 ± 0.07 0.67 ± 0.07 0.75 ± 0.06 0.68 ± 0.06 OT-barycenter [16] 0.71 ± 0.06 0.62 ± 0.07 0.75 ± 0.06 0.68 ± 0.07 0.77 ± 0.07 0.69 ± 0.07

In this section, we investigate the impact of different OT formulation strategies for constructing the reference matrix from multiple local prototypes for the Procustes analysis. We compare against a baseline strategy where the reference matrix is computed using simple averaging. In Table 6, we observe that the OT-barycenter formulation consistently outperforms other OT strategies across all sites in the ABIDE-I dataset. FedDOSE is typically stable across different OT formulations. This indicates that, while important, the OT component does not induce spurious improvements in performance. The differences between the OT formulations in our case can be attributed to whether they are based on point-to-point alignment or distribution-level alignment. Therefore, selecting an appropriate formulation is crucial for optimizing model performance.

15

Related documents

• High-Resolution Dynamic Functional Connectivity Generation with Graph-Variate Flow Matching#1132238
• Connectomic Alterations Following Acute Ischemic Stroke in the MCA Territory#871265
• Could Model Partitioning Make Federated Learning More Sustainable?#458254
• $f$-FUM: Federated Unlearning via min--max and $f$-divergence#1116203
• A conserved node degree-based backbone and flexible hub organization of brain connectome during naturalistic movie watching#1052813