Swarm-GestaltMatcher: distributed Gestalt learning to enhance facial phenotyping for rare genetic syndromes

Deep learning-based facial phenotyping represents a major paradigm shift in the diagnosis of rare and ultra-rare genetic disorders. By capturing disease-specific craniofacial "gestalts" that are often subtle, overlapping, but overlooked in routine clinical practice, these technologies surpass the traditional limits of dysmorphology assessment. Despite this, data scarcity and stringent privacy policies constraint centralized model training and its clinical translation. Swarm learning, a decentralized paradigm that combines edge-based training with blockchain-mediated parameter synchronization, offers a potential solution by enabling collaborative model development without sharing raw patient data. However, it remains uncertain whether a decentralized model can achieve performance comparable to the centralized model, particularly in the context of rare disease diagnosis. Using two complementary datasets, GMDB, and AIDY, we evaluate the feasibility, performance, and clinical relevance of swarm learning against centralized and institution-specific local models. Model performance is assessed across in-distribution, cross-institutional, and ultra-rare disorder scenarios (not part of model training), with additional analyses of calibration and epistemic uncertainty. Our results show that swarm-trained models consistently match the accuracy of centralized training and outperformed local models. Importantly, swarm learning preserves sensitivity to low-prevalence and ultra-rare syndromes despite extreme data scarcity across sites, while exhibiting more conservative and reliable uncertainty calibration. Although swarm learning has previously been applied to well-characterized diseases, this study represents the first application of a swarm model in real-world settings involving a large and diverse set of disease classifications. Taken together, swarm learning emerges as a scalable, equitable, and trustworthy framework that shortens the diagnostic odyssey and advances precision medicine for rare disease diagnosis in routine clinical practice.