New Framework for Privacy-Preserving, Multi-Institutional LLM Adaptation in Medicine

What happened

Researchers introduced Fed-MedLoRA and Fed-MedLoRA+, a parameter-efficient federated framework for collaborative LLM adaptation across healthcare institutions. The framework transmits only low-rank adapters to reduce communication overhead and includes a privacy-preserving variant using Gaussian perturbation. Evaluated on five patient cohorts totaling 42,198 entities and 41,570 relations, the methods improved extraction performance and generalized better to heterogeneous cohorts compared to zero-shot/fine-tuned LLMs and domain-specific BERT models. A case study using clinical notes from the Yale New Haven Health System demonstrated strong performance under low-resource new-site deployment.

Why it matters

This addresses privacy and governance constraints preventing multi-institutional data sharing, making federated LLM adaptation feasible, scalable, and effective for multi-institutional clinical deployment.