FedPref Enables Federated Learning for Radiology Report Extraction

Extracting structured information from free-text radiology reports is crucial for downstream analysis and search, but it requires extensive labeled data. This data is often unevenly distributed across healthcare institutions, with smaller hospitals having fewer local examples, and direct data pooling is often impossible due to privacy concerns. This research introduces FedPref, a novel federated preference learning approach designed to overcome these challenges. FedPref allows multiple institutions to collaboratively train compact Qwen3-8B adapters. The process involves public language models proposing JSON extractions, local annotations ranking these proposals, and only model updates being shared between sites, never the raw patient data or annotations. The system demonstrated significant improvements on development data from simulated hospitals, boosting client-mean F1 scores by 2.49 points and worst-site F1 by 9.10 points compared to isolated training. While central training on pooled data achieved slightly higher F1, FedPref successfully enables collaboration and performance gains for institutions with disparate data volumes, all while maintaining data privacy.