ProMoE-FL Enhances Multimodal Federated Learning with Missing Data

This research addresses a significant challenge in multimodal federated learning: dealing with missing data modalities. Existing approaches often rely on public datasets or simplistic feature synthesis, which can be inadequate. The authors propose ProMoE-FL, a novel framework that leverages a Prototype-conditioned Mixture-of-Experts architecture to robustly synthesize missing features. ProMoE-FL establishes a global prototype bank that is aware of client-specific characteristics, capturing clinically relevant modality priors across different institutions. This prototype bank then conditions a Mixture of Experts, which uses modality indices to enable dynamic, direction-aware routing of experts. This mechanism allows for the intelligent synthesis of missing features. Extensive evaluations on four public chest X-ray datasets (MIMIC-CXR, NIH Open-I, PadChest, and CheXpert) demonstrate that ProMoE-FL consistently surpasses current state-of-the-art methods in both homogeneous and more complex heterogeneous settings.