FedPref: Federated Preference Learning for Structured Radiology Report Extraction

Let’s craft.

Word count: We’ll approximate.

I’ll write and then count.

Draft:

Hook paragraph:

Extracting structured information from radiology reports is a common bottleneck for research and clinical decision support. Most approaches rely on large labeled corpora, but hospitals differ greatly in the amount and type of data they can share. A new method called FedPref tackles this problem by combining frozen language models, local preference annotations, and federated learning of small adapters.

Word count? Let’s count later.

Now What You Need to Know (2-3 paragraphs). We’ll write two paragraphs.

Paragraph1: explain frozen public language models propose JSON extractions, local annotators rank them.

Paragraph2: describe heterogeneous teacher pool and compact Qwen3-8B adapters, sharing only model updates.

Now Why It Matters (1-2 paragraphs). Write about addressing data imbalance, privacy, enabling smaller hospitals.

Key Details: bullet list 4-6 items.

What’s Next: one paragraph about future work, scaling, real-world evaluation.

Now let’s write and count.

I’ll write text then count words manually approx.

Text:

Extracting structured information from radiology reports is a common bottleneck for research and clinical decision support. Most approaches rely on large labeled corpora, but hospitals differ greatly in the amount and type of data they can share. A new method called FedPref tackles this problem by combining frozen language models, local preference annotations, and federated learning of small adapters.

What You Need to Know

FedPref starts with a pool of publicly available large language models that are kept frozen. Each model receives a radiology report and proposes a JSON‑style extraction of findings and anatomic locations. Because the models are frozen, their outputs are consistent across sites and do not require any local fine‑tuning.

At each participating hospital, radiologists or trained annotators view the proposed JSONs and rank them according to correctness and completeness. These rankings act as preference signals that are used locally to update a compact adapter—specifically a low‑rank modification of a Qwen3‑8B backbone. Only the adapter weights are sent to a central server, where they are averaged and redistributed, preserving the raw report text on site.

Why It Matters

Many smaller hospitals lack enough labeled examples to train a reliable extraction model on their own, yet pooling raw reports raises privacy and legal concerns. By learning from preference rankings instead of full annotations, FedPref reduces the labeling burden while still capturing site‑specific nuances. The federated scheme ensures that no patient text leaves the hospital, addressing a key barrier to multi‑institution collaboration.

Empirical results on a simulated benchmark with six hospitals showing varied data volume and disease prevalence indicate that FedPref yields higher extraction fidelity than training adapters in isolation or using a single global model. The improvement is most pronounced for the lowest‑resource sites, where the gain approaches double digits in F1 score.

Key Details

• Frozen teacher pool: multiple public LLMs generate candidate JSON extractions.
• Local preference labeling: annotators rank candidates; rankings are converted to a loss for adapter updates.
• Adapter architecture: low‑rank (LoRA) style modules attached to Qwen3‑8B, keeping total trainable parameters under 1 % of the base model.
• Federated averaging: adapter weight updates are aggregated across sites each round; no raw text is transmitted.
• Heterogeneous teacher contrast: when a single model’s samples become repetitive, alternative models provide diverse proposals to prevent collapse.
• Evaluation metric: structured extraction measured with slot‑filling F1 on a held‑out set from each simulated hospital.

What’s Next

Future work includes testing FedPref on real multi‑institution radiology datasets, exploring different base models beyond Qwen3‑8B, and investigating how the preference signal can be enriched with uncertainty estimates or semi‑automatic ranking to further reduce annotation cost.

Now need word count 400-600. Let’s approximate.

We’ll count words roughly.

I’ll count each paragraph.

Hook paragraph: “Extracting(1) structured2 information3 from4 radiology5 reports6 is7 a8 common9 bottleneck10 for11 research12 and13 clinical14 decision15 support.16 Most17 approaches18 rely19 on20 large21 labeled22 corpora,23 but24 hospitals25 differ26 greatly27 in28 the29 amount30 and31 type32 of33 data34 they35 can36 share.37 A38 new39 method40 called41 FedPref42 tackles43 this44 problem45 by46 combining47 frozen48 language49 models,50 local51 preference52 annotations,53 and54 federated55 learning56 of57 small58 adapters59.”

So hook ~59 words.

What You Need to Know paragraph1: “FedPref1 starts2 with3 a4 pool5 of6 publicly7 available8 large9 language10 models11 that12 are13 kept14 frozen.15 Each16 model17 receives18 a19 radiology20 report21 and22 proposes23 a24 JSON‑style25 extraction26 of27 findings28 and29 anatomic30 locations.31 Because32 the33 models34 are35 frozen,36 their37 outputs38 are39 consistent40 across41 sites42 and43 do44 not45 require46 any47 local48 fine‑tuning49.”

Count: Let’s count words: I’ll count again: FedPref(1) starts2 with3 a4 pool5 of6 publicly7 available8 large9 language10 models11 that12 are13 kept14 frozen15. Each16 model17 receives18 a19 radiology20 report21 and22 proposes23 a24 JSON‑style25 extraction26 of27 findings28 and29 anatomic30 locations31. Because32 the33 models34 are35 frozen,36 their37 outputs38 are39 consistent40 across41 sites42 and43 do44 not45 require46 any47 local48 fine‑tuning49.

So 49 words.

Paragraph2: “At1 each2 participating3 hospital,4 radiologists5 or6 trained7 annotators8 view9 the10 proposed11 JSONs12 and13 rank14 them15 according16 to17 correctness18 and19 completeness.20 These21 rankings22 act23 as24 preference25 signals26 that27 are28 used29 locally30 to31 update32 a33 compact34 adapter—specifically35 a36 low‑rank37 modification38 of39 a40 Qwen3‑8B41 backbone.42 Only43 the44 adapter45 weights46 are47 sent48 to49 a50 central51 server,52 where53 they54 are55 averaged56 and57 redistributed,58 preserving59 the60 raw61 report62 text63 on64 site65.”

Count words: Let’s count: At1 each2 participating3 hospital,4 radiologists5 or6 trained7 annotators8 view9 the10 proposed11 JSONs12 and13 rank14 them15 according16 to17 correctness18 and19 completeness20. These21 rankings22 act23 as24 preference25 signals26 that27 are28 used29 locally30 to31 update32 a33 compact34 adapter—specifically35 a36 low‑rank37 modification38 of39 a40 Qwen3

📌 Source: Arxiv Ai