Adaptive Multimodal Fusion in Vertical Federated Learning for Decentralized Glaucoma Screening

Background/Objectives: Early and accurate detection of glaucoma is vital for preventing irreversible vision loss, yet traditional diagnostic approaches relying solely on unimodal retinal imaging are limited by data sparsity and constrained context. Furthermore, real-world clinical data are often fragmented across institutions under strict privacy regulations, posing significant challenges for centralized machine learning methods. Methods: To address these barriers, this study proposes a novel Quality Aware Vertical Federated Learning (QAVFL) framework for decentralized multimodal glaucoma detection. The proposed system dynamically integrates clinical text, retinal fundus images, and biomedical signal data through modality-specific encoders, followed by a Fusion Attention Module (FAM) that adaptively weighs the reliability and contribution of each modality. Unlike conventional early fusion or horizontal federated learning methods, QAVFL operates in vertically partitioned environments and employs secure aggregation mechanisms incorporating homomorphic encryption and differential privacy to preserve patient confidentiality. Results: Extensive experiments conducted under heterogeneous non-IID settings demonstrate that QAVFL achieves an accuracy of 98.6%, a recall of 98.6%, an F1-score of 97.0%, and an AUC of 0.992, outperforming unimodal and early fusion baselines with statistically significant improvements (p < 0.01). Conclusions: The findings validate the effectiveness of dynamic multimodal fusion under privacy-preserving decentralized learning and highlight the scalability and clinical applicability of QAVFL for robust glaucoma screening across fragmented healthcare environments.