Federated variational inference for incomplete multimodal learning

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder, with Mild Cognitive Impairment (MCI) often preceding its onset. Early prediction of MCI progression to AD is crucial for effective intervention. Although multimodal neuroimaging can improve prediction accuracy, incomplete datasets and strict privacy regulations hinder broad implementation in machine learning (ML) models. Current federated learning (FL) solutions in neuroimaging typically focus on unimodal data, thereby missing the synergistic advantages of multimodal approaches. We propose Federated Variational Inference for Cross-Model Learning (FedVICML), a novel FL framework that allows institutions with varying neuroimaging modalities to collaboratively train predictive models without centralized data sharing, thus keeping raw patient data local to each site. FedVICML uniquely combines FL with a new cross-model variational inference (VI) method, using two distinct Bayesian neural networks for multimodal and unimodal data respectively, and an extended VI-based mechanism for knowledge transfer. This decentralized, data-local approach enables robust learning with incomplete multimodal data by integrating FL and cross-model VI, without pooling raw data across sites. We apply FedVICML to simulated and real-world datasets from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), demonstrating improved performance and uncertainty quantification over conventional FL baselines for predicting MCI to AD progression. Our results highlight FedVICML’s promise for collaborative medical research in heterogeneous data environments.