A Communication-Efficient Personalized Federated Learning Framework for Multi-Site Brain Disease Diagnosis

Abstract

Functional magnetic resonance imaging (fMRI) enables data-driven diagnosis of neuropsychiatric disorders, but the brain-connectivity features it produces are extremely highdimensional relative to the number of subjects available at any single hospital, and patient-privacy regulations prevent institutions from pooling raw scans into a shared database. Federated learning (FL) allows multiple sites to jointly train a diagnostic model without exchanging raw data, yet standard FL algorithms such as FedAvg still suffer from two problems in this setting: (i) transmitting full-dimensional connectivity weight vectors over many communication rounds is expensive, and (ii) heterogeneity in scanner hardware and patient populations across sites causes both covariate shift and concept drift, which a single global model cannot capture. This paper proposes a lightweight federated framework that combines sparse, L1-regularized clientside feature selection, a marginal-and-conditional domain-alignment step, and a personalized global/local mixture-of-experts predictor, wrapped with differentialprivacy noise and homomorphicencryption protection during transmission. Because the public ABIDE neuroimaging repository could not be accessed in the present computing environment, the framework is evaluated on a controlled synthetic multi-site benchmark that reproduces the sample sizes and heterogeneity profile of six widely used ABIDE sites (NYU, UM, USM, UCLA, KKI, YALE). Experimental results show that the proposed framework reduces the per-round communication payload by 79.5% relative to raw-feature FedAvg and improves average six-site test accuracy from 0.614 to 0.678 (+6.4 percentage points), with the personalized mixture-of-experts component contributing the largest and most consistent gain across ablations.