FedPref: Federated Preference Learning for Structured Radiology Report Extraction

FedPref is a federated preference learning framework using frozen language models to propose JSON extractions, local hospitals rank them, and collaboratively train Qwen3-8B adapters without sharing data.

Central training on pooled data outperforms by 2.66 points in client-mean F1, but FedPref yields significant gains for data-poor sites while preserving privacy.

Yes, it suits institutions with unequal data, enabling cross-border collaboration without violating data privacy regulations.