PaperlayerView exact revision on arXiv

Source-linked AI summary

FedPref: Federated Preference Learning for Structured Radiology Report Extraction

Flint Xiaofeng Fan, Cheston Tan, Yew-Soon Ong, Roger Wattenhofer

arXiv:2608.16971v1cs.AIcs.LG

TL;DR

Radiology reports encode findings and locations in free text, while structured downstream use requires fixed-schema relations and institutions may lack enough local labels to train effectively. FedPref uses locally ranked, teacher-generated JSON alternatives to train federated adapters without sharing reports or annotations, improving client-mean and worst-site F1 over isolated training, especially at the least-data sites.

  • Problem

    Institutions have unequal case volumes and prevalence, limiting local evidence for fixed-schema radiology extraction when reports and annotations cannot be pooled.

  • Method

    FedPref has frozen teachers generate JSON alternatives, local annotations rank them, and sites exchange only LoRA adapter states aggregated by weighted FedAvg.

  • Results

    FedPref improves client-mean F1 by +2.49 points and worst-site F1 by +9.10 points over isolated training, with largest gains at the smallest sites.

  • Takeaways & Limitations

    FedPref offers a practical middle ground between isolated and pooled training while keeping reports and task-specific judgments under institutional control.

  • Takeaways & Limitations

    The protocol provides no formal privacy guarantee because updates lack secure aggregation or differential privacy, and leakage attacks were not evaluated.

Abstract

from arXiv · show

Radiology reports describe findings and locations in free text, but downstream search and analysis require these relations in a fixed schema. Learning this extraction requires labels that are unevenly distributed across institutions: smaller hospitals have less local evidence, and pooling data may be infeasible. We introduce FedPref: frozen public language models propose alternative JSON extractions, local annotations rank them, and 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 development data from six simulated hospitals with unequal data volume and disease prevalence, FedPref improves client-mean F1 by 2.49 points and worst-site F1 by 9.10 points compared with training each site in isolation, with the largest gains at the sites holding the least data. Central training on the pooled preference-pair union is 2.66 points higher on client-mean F1. On a locked, 400-report manually validated gold test set, FedPref reaches 68.68 F1 and pooled training 71.67, preserving that same ordering. FedPref thus lets institutions with unequal, unpooled data benefit from collaboration without ever sharing reports or annotations.

1 Introduction

FedPref addresses structured radiology extraction across institutions with unequal case distributions by using local preference supervision and federated adapter aggregation without pooling reports or annotations. Its design uses heterogeneous teacher candidates to create contrastive JSON preferences and evaluates whether aggregation improves over isolated training while retaining pooled-training performance.

  • Motivation: Radiology extraction must jointly identify each finding and its anatomical location, because assigning a disease to the wrong anatomy is an error.The benchmark covers nine diseases and their locations using MIMIC-CXR and Chest ImaGenome.
  • Motivation: Unequal case volumes and disease prevalence leave smaller hospitals with less evidence, motivating federated aggregation as an alternative to centralized or isolated training.The schema and annotation rule remain fixed while patient distributions differ across sites.
  • Method: FedPref uses four frozen teachers to propose JSON extractions that each site projects into a common schema and ranks as chosen c+ or rejected c− using local annotations.Reports, annotations, candidates, and preference pairs remain local while adapters are trained collaboratively.
  • Method: Heterogeneous cross-model candidates provide more useful preference contrast than repeated self-sampling from the Qwen3-8B target model, whose outputs often collapse to the same structure.This feasibility diagnostic is separate from the operational training pipeline and motivates the four-teacher pool.
  • Evaluation: The evaluation compares Local, FedPref, and Central after one chosen-response SFT epoch followed by fixed-reference DPO, testing client-mean, worst-site, and pooled-performance retention.Local trains sites independently, FedPref aggregates site adapters, and Central trains on the union of retained preference pairs.

2 Related Work

Prior work supplies report-level and anatomy-linked supervision for structured radiology extraction, while federated learning and preference optimization offer complementary tools for siloed medical data. FedPref combines these ideas by using local annotations to rank heterogeneous teacher outputs and sharing only LoRA model states.

  • Structured extraction: MIMIC-CXR and Chest ImaGenome provide report-level and anatomy-linked supervision for structured extraction, alongside Sabour et al.’s centralized formulation.These resources and formulation establish the supervision and task setting used here.
  • Federated learning: Federated learning coordinates models without centralizing raw institutional records, while secure aggregation or differential privacy provide formal privacy mechanisms.Data locality alone does not guarantee privacy; FedAvg combines local states and LoRA reduces communicated trainable parameters.
  • Preference optimization: DPO learns from response rankings, and federated variants distribute feedback-driven, personalized, or medical-dialogue alignment.FedPref connects preference optimization with federated learning for medical extraction.
  • FedPref: FedPref uses local disease-location annotations to rank heterogeneous teacher outputs and aggregates only the resulting LoRA states.This design connects preference learning, heterogeneous teachers, and parameter-efficient federated coordination.

3 Method

FedPref builds local preference pairs from multiple frozen teachers, then uses fixed-reference preference optimization to train and aggregate LoRA adapters across sites. Centralized and local training provide matched controls, while reports, annotations, candidates, and preference pairs remain site-local.

  • Fixed-reference preference optimization: Round-0 supervised fine-tuning establishes the schema and fixed reference, after which DPO rounds t = 1, 2, 3 refine preferences while retaining the round-0 root as πref.The fixed SFT reference keeps likelihood changes comparable across rounds.
  • Local preference construction: Four frozen teachers—Qwen3-14B, Qwen2.5-14B-Instruct, Llama-3.1-8B-Instruct, and Mistral-7B-Instruct-v0.3—each generate one deterministic completion for the excluded Qwen3-8B target.Teachers use identical task messages through native chat templates with reasoning disabled.
  • Local preference construction: Each site ranks eligible teacher-generated JSON structures by annotation agreement, prioritizing Location+Disease F1 and requiring structural differences, a score margin of at least 0.10, and higher F loc+dis1 for the chosen response.Failed ontology projections are discarded, and annotations determine ordering but never enter teacher prompts or candidates.
  • Aggregation and matched controls: At each round, FedPref sites train one local epoch from the preceding global LoRA state, and pair-weighted aggregation makes site influence proportional to retained training evidence.Central trains on the exact pair union, while Local retains one independent root per site.
  • Aggregation and matched controls: Reports, annotations, candidates, and preference pairs remain at their simulated sites; only LoRA states cross the boundary.Central and Local are matched controls: Central trains one adapter for three epochs on the union, whereas Local trains six isolated adapters for three local epochs.

4 Experimental Setup

The experiments use patient-disjoint MIMIC-CXR and Chest ImaGenome cohorts, partitioned across six heterogeneous simulated hospitals. Models are trained with teacher-generated preference pairs and evaluated primarily by micro-F1 on positive disease-location labels after canonical JSON projection.

  • Task and cohorts: 4,800 training, 600 development, 100 diagnostic gold-standard, and 400 locked-test reports form patient-disjoint cohorts across nine diseases and 18 locations.The automatically derived cohorts exclude gold-standard patients before construction; 15 locations have positive support.
  • Simulated federation: Six sites contain 384–1,344 training reports and 297–717 retained pairs, with disease-prevalence divergence of 8.83 versus 1.89 points for training pairs and 10.14 versus 4.14 for development under IID allocation.A capacity-constrained multilabel Dirichlet allocator uses α = 0.5 while preserving site sizes; both prevalence comparisons have p = 0.0005.
  • Preference data and optimization: 19,200 training and 2,400 development teacher responses yield 2,961 training and 376 development preference pairs for Qwen3-8B LoRA adaptation.LoRA uses rank 16 and α = 32; each of three seeds trains the SFT root and three DPO rounds with three DPO exposures per pair in every arm.
  • Evaluation: The primary endpoint is micro-F1 over positive disease-location labels (Location+Disease F1) after canonical projection, requiring both finding and anatomy to be correct.Secondary measures include Disease F1, projection success, false positives on reports with no targets, and held-out preference accuracy; global F1 weights larger sites more, whereas client mean weights sites equally.

5 Results

FedPref improves federated extraction performance over home-routed Local, especially at smaller sites, while Central remains stronger overall on development and locked gold-standard test data.

  • Candidate contrast: Repeated Qwen3-8B sampling produced 71 identical parsed sets, 3.05 points of headroom, and 20.0% usable-pair coverage across 100 automatically labeled reports.Candidate contrast was measured by oracle headroom and the fraction of reports satisfying the pair margin.
  • Federated aggregation: 69.11 ± 0.17 client-mean F1 and 68.12±0.22 worst-site F1 were reached by FedPref at round 3, up from 54.93 ± 0.47 and 53.67 ± 0.18.FedPref overtook Local on client mean at round 2, while round 3 maximized the prespecified selection criterion.
  • Federated aggregation: +2.49 points in client-mean F1 and +9.10 points in worst-site F1 were FedPref’s improvements over home-routed Local at the selected round.The corresponding 95% confidence intervals were +1.24 to +3.79 and +1.01 to +11.87.
  • Central comparison: 2.66 points higher client-mean F1 was achieved by Central than FedPref on development data.FedPref and Local had similar held-out preference accuracy despite different extraction F1.
  • Site-level effects: +9.46 and +4.47 points were FedPref’s largest site-level gains, occurring at S4 and S5, the two smallest training cohorts.S0 and S2 favored Local, but gains otherwise concentrated at sites with less local evidence.
  • Locked gold-standard generalization: 71.67±0.23 Location+Disease F1 was recorded by Central versus 68.68 ± 0.35 by FedPref on the locked 400-report gold test.The primary-endpoint comparison yielded +2.98 points for Central minus FedPref.

6 Discussion and Conclusion

FedPref uses local preference judgments over candidate structured annotations while aggregating adapters across institutions, keeping reports and task-specific judgments local. In controlled federation, it improves extraction most at data-poor sites and offers a middle ground between isolated and pooled training.

  • Local preference interface: FedPref turns existing institutional annotations into local rankings of candidate structures under a shared extraction rule.Cross-model candidates provide local supervision, while adapter aggregation transfers it into a shared model.
  • Federated design: Adapter aggregation accommodates institutions with shared schemas but differing data volume, case mix, and governance.Reports and task-specific judgments remain under local institutional control.
  • Results: Federated aggregation improves extraction most at sites with the least evidence.Central training retains an advantage under matched preference evidence and exposure, defining the observed pooled–federated gap.
  • Conclusion: A locked gold test preserves the same ordering, positioning FedPref between isolated and pooled training for structured radiology extraction.The method is presented as a practical middle ground that enables collaboration without centralizing institutional reports or judgments.

A Scientific Supplement

This supplement documents the task contract, cohort construction, candidate projection, federated optimization, and supporting results underlying the main paper. It reports frozen experiment artifacts without additional model selection.

  • Scope: The supplement records the task contract, cohort construction, candidate projection, and federated optimization used in the main paper.These components define the experimental setup and learning procedure.
  • Scope: It also includes supporting results associated with the main paper.The supplement is intended to document the results used in the paper.
  • Reporting: All tables report frozen experiment artifacts, with no additional model selection performed for the supplement.The reported tables reflect fixed artifacts rather than newly selected models.

A.1 Scope and Limitations

FedPref is evaluated only in a controlled six-site simulation, so its generalizability to independent institutions and external domain shift remains untested. The protocol also has evaluation ambiguity, aggregation-method limitations, and no formal privacy guarantee.

  • Generalizability: The six-site simulation varies capacity and disease prevalence but does not test institution-specific reporting styles, annotation policies, clinical workflows, or external domain shift.Independent institutional cohorts are required to assess these questions.
  • Evaluation: Eight locked-test reports form three normalized-text duplicate groups, including one four-report group with inconsistent targets.This introduces a small source of evaluation ambiguity.
  • Comparison design: FedPref’s comparison with Local estimates cross-site aggregation within a shared training pipeline, while Central optimizes directly on the six-shard corpus.Central, FedPref, and Local use the same retained preference-pair corpus and matched per-pair training exposure; the supplied passage ends before fully describing Central’s setup.
  • Aggregation and privacy: FedPref uses pair-count-weighted, factor-wise averaging of complete LoRA states without comparison to exact low-rank aggregation methods such as FedEx-LoRA.Factor-wise averaging need not equal averaging the clients’ effective updates.
  • Aggregation and privacy: The protocol keeps reports and annotations local but provides no formal privacy guarantee because updates lack secure aggregation or differential privacy, and leakage attacks were not evaluated.The supplied passage also notes limitations involving the six-client, all-participation protocol, but its text is truncated.

A.2 Task Contract and Cohort Support

The task requires exactly one JSON object containing supported diseases and disease–location findings, constrained by explicit inclusion, exclusion, vocabulary, and inference rules. Cohort evaluation excludes gold-standard patients from automatic training and development data and supports primary Location+Disease evaluation over 15 regions.

  • Output contract: Each report must produce exactly one JSON object with only diseases and findings keys.Diseases is a list of supported diseases; findings is a list of disease–location objects.
  • Output contract: Findings include positive, explicitly proposed, or hedged abnormalities but exclude findings that are only negated or absent.Localized findings use exactly one disease and one location from frozen vocabularies.
  • Output contract: Every localized disease must also appear in the disease list, while unsupported inference and unsupplied vocabulary are disallowed.A supported disease may lack a location when the report provides none.
  • Cohort integrity: 15 regions support primary Location+Disease evaluation because mediastinum, upper mediastinum, and trachea have zero positive support.Gold-standard patients were excluded from automatically labeled training and development data.

A.3 Candidate Generation, Projection, and Pair Construction

FedPref generates one greedy, schema-oriented completion per report from each frozen teacher, then projects teacher text into canonical JSON without accessing reference annotations. Preference pairs retain only materially different, higher-quality candidates, yielding 61.69% of training reports and 62.67% of development reports.

  • Generation: Each frozen teacher generated one completion per report using its native chat template with greedy decoding and a 768-token output limit.Generation used BF16 weights without quantization or retrieval augmentation; thinking was disabled for Qwen3 models.
  • Label-independent projection: Projection extracted the first brace-balanced JSON object, rejected duplicate keys and nonstandard constants, and required both top-level lists.It operated without opening reference annotations; extra fields were dropped, while unknown vocabulary and malformed findings were discarded without synonym or fuzzy matching.
  • Preference-pair eligibility: 61.69% of training reports and 62.67% of development reports formed eligible preference-pair corpora after filtering.Each report contributed at most one pair, requiring differing canonical structures, a composite-score margin of at least 0.10, and strictly higher Location+Disease F1 for the chosen candidate.
  • Preference-pair eligibility: 0.426 and 0.414 were the mean composite-score margins for retained training and development pairs, respectively.The corresponding median margins were 0.400 and 0.382.

A.4 Federation Construction and Non-IID Audit · A.5 Optimization Details · A.6 Supporting Results

The study constructs capacity-constrained, heterogeneous federated sites, standardizes adapter optimization and exposure, and evaluates supporting development trajectories. FedPref’s client-mean and worst-site performance improves over communication rounds, with its worst-site advantage reaching 9.10 points by round 3.

  • A.4 Federation Construction and Non-IID Audit: Site capacities were fixed at 28%, 22%, 18%, 14%, 10%, and 8% of each cohort before pair scoring.A multilabel Dirichlet profile with α = 0.5 and skew weight 0.7 allocated intact report rows.
  • A.4 Federation Construction and Non-IID Audit: The heterogeneity audit used nine disease indicators plus a normal-report indicator, for L = 10.The null distribution used 1,999 capacity-preserving permutations of intact multilabel report rows with a plus-one Monte Carlo correction.
  • A.5 Optimization Details: All arms used LoRA rank 16 with α = 32, zero dropout, no trainable bias, BF16 training, and sequence length 2,048.Optimization also used fused AdamW, cosine decay, warmup, zero weight decay, gradient checkpointing, and fail-closed truncation.
  • A.4 Federation Construction and Non-IID Audit: Disease prevalence by site was reported among retained training pairs, whose columns need not sum to 100% because reports may contain multiple diseases.The prevalence table covers retained training pairs and is distinct from the allocation procedure.
  • A.5 Optimization Details: Each of three DPO rounds used one epoch at 5 × 10−6 with β = 0.1, while SFT used one epoch at 2 × 10−4.Physical batch size was one, with gradient accumulation 24 for Central and four per client.
  • A.5 Optimization Details: Central, FedPref, and Local received three DPO exposures per pair, while FedPref clients used the previous global policy and differed from Local only by aggregation.The SFT root remained the fixed DPO reference, and Central opened the exact six-shard union.
  • A.6 Supporting Results: 9.10 points was FedPref’s worst-site advantage over Local at round 3, increasing from 0.36 points at round 0 as FedPref’s client mean passed Local at round 2.The reported trajectory was monotonic across prespecified communication rounds and was interpreted as accumulating cross-site evidence through aggregation.
Loading 2608.16971v1…