Vertical Federated Learning in Medical AI: Concepts, Advances, and Challenges

Abstract

The growing need for data-driven medical intelligence is constrained by strict privacy regulations and fragmented healthcare data distributed across multiple institutions. Vertical Federated Learning (VFL) offers a secure solution by enabling col-laborative model training on feature-partitioned datasets, where different organizations hold complementary information about the same patients. This review presents a comprehensive overview of VFL with a dedicated focus on healthcare applications. We explore the fundamental principles, training protocols, and key design challenges of VFL, followed by a detailed review of recent advancements in communication efficiency, model performance, privacy preservation, and fair-ness. To unify these perspectives, we introduce the Federated Optimization Framework (VFLow), a conceptual framework that captures the optimization of trade-offs across privacy, utility, efficiency, and equity. We further examine real-world applications in disease prediction, diagnostics, clinical trials, personalized treatment, and public health. Unlike previous surveys, this work provides a domain-specific synthesis tailored to the practical constraints and ethical considerations of medical AI. Finally, we outline open research challenges and future directions to guide the development of trustworthy and scalable VFL systems for healthcare.