Fair-FedMOE: Group-Fair One-Shot Federated Learning via Prototype-Guided Experts for Medical Imaging Analysis

Group fairness can ensure equitable performance across different demographic subgroups for medical image analysis. However, the current fine-tuned foundation models (FMs) exhibit significant subgroup disparity. One-shot federated learning (OFL) can potentially mitigate this by leveraging cross-institutional data diversity within a single communication round. However, heterogeneous distributions across medical institutions may cause OFL local models to diverge severely, resulting in parameter conflicts that amplify disparity upon aggregation. To address these challenges, we propose Fair-FedMOE, a group-fair OFL framework for medical FMs. During local training, Fairness-aware Expert Routing leverages learnable prototypes to route samples to group-specific experts, enabling subgroup-specialized learning to capture group-specific features without inter-group interference. During model aggregation, Prototype-guided Differential Aggregation computes personalized weights based on prototype similarity and applies differentiated aggregation strategies to filter conflicting updates. We propose RES-AUC, a Rawlsian justice-inspired metric based on worst-group performance that remains stable as groups increase. Extensive experiments on retinal and chest X-ray datasets with multiple FMs demonstrate consistent fairness gains without sacrificing accuracy.

Medical AI tools are widely used to diagnose diseases from images such as retinal scans. But these tools often work less well for certain patient groups, for example older women or younger men, which can lead to missed or wrong diagnoses. One way to reduce this gap is to train models using data from multiple hospitals together. But hospitals cannot simply share patient records due to privacy rules. Federated learning allows hospitals to collaborate without sharing raw data. Standard federated approaches still require many rounds of communication between hospitals and a central server, which is expensive and often blocked by data regulations. Fair-FedMOE completes this collaboration in a single communication round. Each hospital trains dedicated modules for different demographic groups, ensuring that minority groups are not overshadowed by majority group signals. When aggregating models across hospitals, conflicting updates are filtered out to prevent degradation in model quality. We also propose RES-AUC, a fairness metric that stays stable as the number of patient subgroups grows, fixing a known weakness of existing metrics. Tests on retinal and chest X-ray data show that Fair-FedMOE improves fairness across patient groups without hurting overall accuracy.