VC-RDAgent: An efficient rare disease diagnosis agent via virtual case construction informed by hybrid statistical-metric and hyperbolic-semantic prioritization

Abstract

While Large Language Models (LLMs) have shown promise in clinical decision support, current Retrieval-Augmented Generation (RAG) paradigms face a fundamental bottleneck in rare disease diagnosis: the scarcity, privacy restrictions, and extreme heterogeneity of real-world patient records. This reliance on sparse or inaccessible data leads to a severe "retrieval mismatch," where the lack of high-quality reference cases causes diagnostic performance to degrade sharply. To break this deadlock, we propose VC-RDAgent, a privacy-preserving and offline-capable framework that decouples diagnostic reasoning from sensitive real-world records by synthesizing virtual standardized cases. The system is powered by VC-Ranker, a multi-dimensional engine that integrates statistical-metric measures with hyperbolic-semantic embeddings to capture deep hierarchical ontology relationships. This approach allows for the dynamic generation of high-fidelity virtual references directly from authoritative knowledge bases. Extensive benchmarking across four diverse datasets demonstrates that VC-RDAgent effectively functions as a "performance equalizer." It boosts average Top-1 hit rates by 8.7% to 85.9% over zero-case baselines, enabling lightweight open-source models to rival frontier commercial systems. Notably, VC-Ranker alone achieved an aggregate Top-10 hit rate of 0.819, outperforming prior state-of-the-art methods by 6%. By eliminating the dependency on real-time web retrieval and private case sharing, VC-RDAgent provides a scalable, robust, and clinically deployable solution to shorten the diagnostic odyssey, which is made accessible through an intuitive, chat-based web application https://rarellm.service.bio-it.tech/rdagent/.