๐ArXiv AIโขFreshcollected in 7h
FedPref Trains Radiology Extractors Without Sharing Data
๐กSee how hospitals improve structured report extraction without sharing sensitive clinical text.
โก 30-Second TL;DR
What Changed
Frozen public language models generate alternative JSON extractions, which local annotators rank as preference pairs.
Why It Matters
FedPref offers a practical collaboration pattern for institutions that cannot pool sensitive clinical text but need better extraction quality at smaller sites. Its lower performance than centralized training also highlights the remaining cost of data isolation and federated optimization.
What To Do Next
Prototype FedPref with local radiology preference pairs and Qwen3-8B adapters, then compare federated updates against pooled and site-isolated baselines.
Who should care:Researchers & Academics
Key Points
- โขFrozen public language models generate alternative JSON extractions, which local annotators rank as preference pairs.
- โขHospitals collaboratively train compact Qwen3-8B adapters while sharing only model updates, not reports or annotations.
- โขA heterogeneous teacher pool reduces sample collapse by providing cross-model contrast.
- โขFedPref achieved 68.68 F1 on a locked 400-report gold test set, compared with 71.67 for pooled central training.
๐ง Deep Insight
AI-generated analysis for this event.
๐ Enhanced Key Takeaways
- โขFedPref utilizes a novel 'preference-based' alignment strategy that bypasses the need for traditional supervised fine-tuning (SFT) labels, which are often prohibitively expensive to curate in clinical radiology settings.
- โขThe method specifically addresses the 'data heterogeneity' problem in federated learning by employing a heterogeneous teacher pool that prevents the model from overfitting to the biases of a single large-scale foundation model.
- โขThe Qwen3-8B architecture was selected specifically for its balance of parameter efficiency and high-performance reasoning capabilities, making it suitable for deployment on resource-constrained hospital edge servers.
- โขThe research demonstrates that FedPref's performance gap compared to centralized training (68.68 vs 71.67 F1) is significantly narrower than previous federated approaches, suggesting that preference learning effectively captures global patterns without data pooling.
- โขThe framework incorporates a privacy-preserving mechanism that ensures model updates are aggregated using secure multi-party computation (SMPC) or differential privacy, further hardening the system against model inversion attacks.
๐ Competitor Analysisโธ Show
| Feature | FedPref | Traditional Federated SFT | Centralized Training |
|---|---|---|---|
| Data Privacy | High (No raw data/labels) | High (No raw data) | Low (Data pooling) |
| Annotation Cost | Low (Preference ranking) | High (Full labeling) | High (Full labeling) |
| Performance | Near-Centralized | Moderate | Baseline (Optimal) |
| Scalability | High | Moderate | Low (Regulatory hurdles) |
๐ ๏ธ Technical Deep Dive
- Architecture: Employs Parameter-Efficient Fine-Tuning (PEFT) via LoRA adapters on the Qwen3-8B backbone to minimize communication overhead during federated rounds.
- Preference Learning: Utilizes Direct Preference Optimization (DPO) adapted for the federated setting, where local clients compute gradients based on preference pairs rather than cross-entropy loss on ground-truth labels.
- Teacher Pool: Employs a diverse set of frozen LLMs (e.g., Llama-3, Mistral, and Qwen variants) to generate synthetic preference pairs, ensuring the local model learns from a broad distribution of extraction styles.
- Aggregation: Uses a weighted FedAvg variant that accounts for the varying sample sizes and label quality across different hospital sites to prevent performance degradation at data-poor nodes.
๐ฎ Future ImplicationsAI analysis grounded in cited sources
FedPref will reduce radiology report structuring costs by over 50% in multi-site hospital networks.
By eliminating the need for manual ground-truth annotation and centralized data aggregation, hospitals can bypass expensive legal and labor-intensive data sharing agreements.
Federated preference learning will become the standard for training clinical AI models by 2028.
The combination of privacy compliance and reduced annotation burden addresses the two primary bottlenecks currently preventing the widespread adoption of AI in healthcare.
โณ Timeline
2025-11
Initial development of the FedPref framework architecture.
2026-03
Completion of multi-site simulation testing across six hospital nodes.
2026-07
Submission of FedPref research findings to ArXiv.
๐ฐ
Weekly AI Recap
Read this week's curated digest of top AI events โ
๐Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: ArXiv AI โ