FedPref: Federated Preference Learning Boosts Radiology Report Extraction in Data-Scarce Hospitals
There's this new machine learning method called FedPref that really improves how we extract structured information from radiology reports, even when different hospitals have varying amounts of data. These reports are usually written in a messy way, so we need a standard format to make searching and analyzing them easier, which can be tough when the data isn’t evenly spread. FedPref uses established public language models to propose different JSON formats, ranks them using local notes, and collaborates to train smaller Qwen3-8B models, only sharing updates. In tests with six simulated hospitals facing different data levels, FedPref increased the average F1 score by 2.49 points and the lowest site’s F1 by 9.10 points, especially helping those with less data.
Key facts
- Radiology reports describe findings and locations in free text, requiring a fixed schema for downstream search and analysis.
- Labels for extraction are unevenly distributed across institutions; smaller hospitals have less local evidence.
- Pooling data may be infeasible due to privacy or logistical constraints.
- FedPref uses frozen public language models to propose alternative JSON extractions.
- Local annotations rank the proposed extractions.
- Sites collaboratively train compact Qwen3-8B adapters while sharing only model updates.
- A heterogeneous teacher pool provides cross-model contrast when repeated single-model samples collapse.
- On six simulated hospitals, FedPref improved client-mean F1 by 2.49 points and worst-site F1 by 9.10 points versus isolated training.
Entities
—