MedLatentDx: Latent Multi-Agent Communication for Cross-Hospital Rare-Disease Diagnosis

MedLatentDx is a newly introduced latent multi-agent communication framework designed to facilitate cross-hospital rare-disease diagnosis while preserving patient privacy. It enables hospital agents to send compact latent key-value (KV) blocks rather than identifiable clinical text, utilizing latent KV distillation for same-backbone agents and cross-family latent alignment for different backbone models. The framework demonstrates improved diagnostic performance on the CrossRare-Bench benchmark while minimizing the risk of reconstructing sensitive clinical information, making it significant for practitioners focused on secure collaborative AI in healthcare.