Pith. sign in

REVIEW 4 major objections 5 minor 34 references

A prompt-based BERT model extracts fine-grained Crohn's disease findings from Hebrew radiology reports, outperforming standard fine-tuning and cutting inference cost by a factor of five.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · deepseek-v4-flash

2026-08-05 11:06 UTC pith:3TZYDYVA

load-bearing objection Practical Hebrew radiology extraction pipeline with a real 5x speedup, but the raw-text leakage of test reports into pretraining makes the headline F1/kappa optimistic until a leakage analysis is run. the 4 major comments →

arxiv 2509.04519 v1 pith:3TZYDYVA submitted 2025-09-03 cs.CL

Hierarchical Section Matching Prediction (HSMP) BERT for Fine-Grained Extraction of Structured Data from Hebrew Free-Text Radiology Reports in Crohn's Disease

classification cs.CL
keywords Hebrew radiology reportsstructured information extractionprompt-based learningBERTsection matching predictionCrohn's diseasehierarchical inferencelow-resource clinical NLP
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that a BERT model adapted to Hebrew radiology text can extract fine-grained organ–finding labels from free-text reports well enough for population-level Crohn's disease research, using only a few hundred expert-annotated examples. The method adds a pretraining objective that teaches the model whether a report's Findings and Impression sections belong together, then uses hierarchical prompting: scan-level, organ-level, finding-level. On 24 organ–finding combinations with sufficient positive cases, HSMP-BERT reaches mean F1 0.83 and Cohen's kappa 0.65, far above a zero-shot baseline (F1 0.49) and standard fine-tuning (F1 0.30). Hierarchical inference cuts runtime about fivefold, which is what makes applying the model to nearly ten thousand reports practical. A sympathetic reader would care because unstructured Hebrew radiology reports have been a bottleneck for cohort studies; this work argues the bottleneck can be removed with small annotation effort.

Core claim

The central discovery is that a section-aware pretraining objective plus prompt-based hierarchical inference yields a model that, after light supervision on 512 reports, extracts 24 organ–finding labels from Hebrew radiology reports with a mean F1 of 0.83 ± 0.08 and Cohen's kappa of 0.65 ± 0.17, outperforming both a zero-shot prompted model (F1 0.49) and standard fine-tuning (F1 0.30, p < 10^-7). The same model applied to all 9,683 reports recovers clinically expected associations—ileal wall thickening co-occurring with stenosis and pre-stenotic dilatation—and reveals age- and sex-stratified patterns of inflammatory findings. The paper presents this as evidence that structured extraction in

What carries the argument

The load-bearing mechanism is the Section Matching Prediction (SMP) pretraining objective: the model is trained to classify whether a Findings section and an Impression section come from the same radiology report, analogous to next-sentence prediction. This teaches the model the logical coherence between the detailed observations and the interpretive summary, which is exactly the relation the downstream extraction prompts exploit. At inference, a hand-designed hierarchical tree routes queries from 'is the scan normal?' to 'is there an abnormality in the ileum?' to 'is there wall thickening in the ileum?', skipping lower-level prompts when a parent answer is negative, which reduces BERT calls

Load-bearing premise

The reported test accuracy assumes that including the 512 annotated reports in the 9,683-report pretraining corpus does not let the model memorize the test text; with no leakage check, the absolute numbers may be optimistic.

What would settle it

Retrain HSMP-BERT with the 512 annotated reports excluded from the MLM and SMP pretraining corpus, evaluate on the same 24 targets, and compare mean F1 and kappa; a drop greater than about 0.05 in mean F1 would confirm the pretraining overlap inflates the reported results.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Fine-grained structured extraction from Hebrew radiology reports is achievable with a few hundred annotated examples, lowering the barrier to cohort curation and disease surveillance.
  • The hierarchical inference scheme cuts runtime roughly fivefold with no loss in F1, making whole-corpus analysis of ~10,000 reports practical.
  • Zero-shot SMP already outperforms standard fine-tuning, indicating the section-matching objective captures much of the task's structure without labeled data.
  • Applying the model to the full cohort surfaces the expected Crohn's co-occurrence pattern of ileal wall thickening with stenosis and pre-stenotic dilatation, supporting its use for population-level phenotyping.
  • The per-label F1/kappa table across 24 targets offers a reproducible benchmark for future Hebrew radiology extraction systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reported F1/kappa are upper bounds: the 512 annotated reports, test split included, are a subset of the 9,683 reports used for pretraining, so the model saw the raw test text before evaluation; a fresh test set outside the pretraining corpus is needed to know true deployment performance.
  • SMP pretraining needs only paired Findings/Impression sections and no labels, so the same recipe could transfer to other sectioned clinical documents (pathology, discharge summaries) and other low-resource languages by swapping the base language model.
  • The 5x speedup relies on a manually designed prompt tree; learning the hierarchy from label co-occurrence could improve efficiency and accuracy, especially for the rarest findings where F1 drops to 0.63.
  • Collapsing 'organ not visible' and 'organ resected' into negative labels may mask surgically relevant information; a model that distinguishes these categories could add clinical value in follow-up and resection planning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The manuscript presents HSMP-BERT, a Hebrew BERT (HeRo) adapted to radiology reports via masked language modeling and a novel Section Matching Prediction (SMP) objective, then prompted hierarchically to classify 24 organ–finding labels from MRE/CTE reports. On a 512-report annotated subset, it reports mean F1 0.83±0.08 and Cohen's κ 0.65±0.17, surpassing a zero-shot SMP baseline and standard fine-tuning, with a 5.1× inference speedup from hierarchical routing. The model is then applied to the full corpus to estimate prevalence and associations by age, sex, and anatomical site.

Significance. If the results are validated, the combination of self-supervised section matching and hierarchical prompt routing is a practical contribution for low-resource clinical NLP, particularly for Hebrew radiology text. The paper has real strengths: external radiologist labels, a multilabel-stratified split, multiple metrics, repeated-measures ANOVA with paired post-hoc tests across seeds, and a concrete efficiency analysis with measured runtime, BERT calls, and token counts. However, the overlap between the pretraining corpus and the annotated test set, together with the unreported quantitative validation of the population-level predictions, means that the headline absolute performance numbers and the downstream clinical associations are not yet established. The strengths are sufficient to merit a revision rather than rejection.

major comments (4)
  1. [§2.1, §2.2, §2.2.1] The 512 annotated reports—including the held-out test reports—are a subset of the 9,683-report corpus used for domain-adaptive MLM and SMP pretraining. Because SMP is explicitly trained on Findings–Impression pairs from these reports, the model has seen the raw text of test reports before evaluation. The paper provides no leakage analysis and no ablation with pretraining restricted to the 9,171 non-annotated reports. This makes the absolute F1 0.83±0.08, κ 0.65±0.17, and the zero-shot results potentially optimistic, and it compromises the claimed validation of population-level predictions against the annotated subset. Please rerun pretraining excluding the annotated reports, or provide a quantitative leakage check (e.g., performance stratified by nearest-neighbor overlap between test reports and pretraining data).
  2. [§3, Table 3, §2.1] The paper gives two inconsistent cohort sizes: §2.1 and the abstract state 9,683 reports from 8,093 unique patients, while §3 and Table 3 state 9,658 reports from 7,389 patients. Since the population-level prevalences, Figure 5, and the 5.1× speedup figures are computed on the corpus, this discrepancy must be resolved and the correct denominators reported consistently throughout.
  3. [§2.3, §4, Figure 5, Figure S1] The population-level associations ('revealed associations', prevalence patterns, and pairwise correlations) are derived entirely from the model's own predictions. The paper claims that 'aggregate patterns' were validated against expert annotations, but no quantitative agreement metrics (e.g., prevalence calibration, rank-order correlation, correlation-structure agreement) are reported. Without such validation, these associations should be described as exploratory model outputs, not as validated clinical findings.
  4. [Abstract; Table 4] The abstract states that HSMP-BERT 'outperformed' the baselines without mentioning the SMP-Finetune arm. Table 4 and the post-hoc tests show that HSMP-BERT and SMP-Finetune have indistinguishable accuracy, F1, κ, and balanced accuracy (p > 0.80). The demonstrated advantage of the hierarchical component is inference speed, not accuracy. Please state this plainly so readers do not infer an accuracy gain over the strongest baseline.
minor comments (5)
  1. [§3, Table 2, Table S1] There are numeric inconsistencies: the text says 'near-perfect NPV of 0.9' for rectum wall thickness, but Table 2 reports NPV 0.99; the text says rectal wall enhancement prevalence is 7%, but Table S1 reports 4.4%; the abstract says a 33% test split, while §2.3 says 34%.
  2. [§3, Table 3] The sentence '1.77 vs. 0.35 seconds per single radiology report' needs to specify which number corresponds to flat vs. hierarchical inference. As written, the ordering is ambiguous and should be matched to the total runtimes in Table 3.
  3. [Figure 5] There are minor discrepancies between the text and the figure caption regarding which findings are more common in adults vs. pediatric patients (e.g., stenosis and inflammation). Please align the text and caption.
  4. [Table S3] The table title contains a typo: 'SMP/BERT' should be 'SMP-BERT' or 'SMP BERT'.
  5. [§2.3–§2.4] Section 2.4 appears after §2.3 but is titled 'Evaluation Framework'; this suggests a numbering or heading error. Also, the model configurations are described with inconsistent names ('Fine-tuned HSMP-BERT' vs. 'SMP Finetune').

Circularity Check

1 steps flagged

Test reports are inside the MLM/SMP pretraining corpus, so the headline test F1/kappa are leakage-inflated; otherwise the core evaluation is against external radiologist labels and not circular.

specific steps
  1. fitted input called prediction [§2.1, §2.2, §2.2.1, §2.4]
    "Given the scarcity of publicly available Hebrew medical corpora, we adapted HeRo to the radiology domain through additional pretraining on all 9,683 MRE and CTE reports described in Section 2.1. ... Multilabel-stratified split (66% for training and validation; 34% for held-out evaluation). ... During SMP training, we construct input pairs as: xin = [CLS] xF_i [SEP] xI_i [EOS]. ... Each input took the form: xin = [CLS] xF_i [SEP] p_j [EOS]"

    The annotated 512-report set—including the held-out test reports—is a subset of the 9,683 reports used for domain-adaptive MLM and SMP pretraining. SMP is trained on the real Impression sections (xI_i) of those exact reports. At test time, the prompt p_j (e.g., 'There is wall thickening in the ileum.') is semantically the same kind of text as an Impression, and the input format is identical ([CLS] Findings [SEP] prompt). A model can therefore answer a test item by retrieving the Impression content it saw during pretraining rather than by generalizing from the annotated split. The reported absolute F1 and kappa are thus not clean out-of-sample predictions: the raw test inputs were already fitted during pretraining. This is test-contamination/leakage rather than label-level identity, so rela

full rationale

The core F1/kappa evaluation is against external radiologist annotations on a held-out split, and the comparison with SMP zero-shot and standard fine-tuning is an empirical result with independent content, so the paper is not fundamentally circular. The main circularity-adjacent flaw is the leakage described above: the held-out test reports appear verbatim in the MLM/SMP pretraining corpus, and the SMP pretraining task uses the Impression sections of those reports, which are the same semantic type as the test-time prompts. This makes the absolute test numbers optimistic. The population-level 'revealed' associations are computed from the model's own predictions, and the promised quantitative consistency check against expert annotations is not reported, so those downstream findings are self-referential. Reference [26] is a self-citation used to motivate the Findings–Impression structure, but it is not load-bearing because the paper provides its own SMP-vs-SFT experiments. Overall score 3: minor self-citation plus a real leakage/self-referentiality concern, but the central supervised evaluation retains independent content.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claim depends on the evaluation being unbiased; the most consequential unpaid assumptions are that test reports can appear in the pretraining corpus without leakage, that the 2/9 recoding does not distort prevalence, and that the hand-built prompt tree preserves accuracy. These are domain assumptions, not standard math.

free parameters (3)
  • minimum positive count for label selection = 15
    Only organ-finding pairs with at least 15 positives were evaluated (Section 2.1.1); this post hoc filter determines which 24 of 90 labels enter the average and improves apparent performance.
  • decision threshold for binary labels = not reported
    The paper reports F1/kappa but does not state the probability threshold used to convert model outputs into positive/negative predictions; threshold choice materially affects these metrics.
  • hierarchical prompt tree design = manually authored
    The scan, organ, and finding level prompts (Section 2.2.3) are hand-designed; the tree structure and wording are free choices that affect both speed and accuracy.
axioms (4)
  • domain assumption The 9,683-report pretraining corpus contains the 512 annotated reports, including the test set
    Section 2.1 says annotation is a random subset of the 9,683 reports, and Section 2.2.1 pretrains on all 9,683 reports; the paper silently assumes this transductive setup does not inflate test scores.
  • domain assumption Radiology reports have discrete Findings and Impression sections with logical consistency
    SMP construct and example in Section 2.2.1 rely on this; authors acknowledge in Section 4.1 that deviation degrades performance.
  • domain assumption Labels 2 (organ not visible) and 9 (organ resected) can be treated as negatives
    Section 2.1.1 states this simplification; it raises NPV and accuracy mechanically and may hide misses on organs frequently not visible.
  • domain assumption Adjudicated radiologist labels are ground truth
    No inter-rater reliability is reported; the evaluation treats the adjudicated labels as error-free.

pith-pipeline@v1.4.0-alltime-deepseek-medium · 14253 in / 12986 out tokens · 126777 ms · 2026-08-05T11:06:00.645704+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Hierarchical Section Matching Prediction (HSMP) BERT for Fine-Grained Extraction of Structured Data from Hebrew Free-Text Radiology Reports in Crohn's Disease." pith.science (2026). https://pith.science/paper/3TZYDYVA

@misc{pith2026250904519,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Section Matching Prediction (HSMP) BERT for Fine-Grained Extraction of Structured Data from Hebrew Free-Text Radiology Reports in Crohn's Disease},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3TZYDYVA}},
  note         = {Machine review of arXiv:2509.04519}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Extracting structured clinical information from radiology reports is challenging, especially in low-resource languages. This is pronounced in Crohn's disease, with sparsely represented multi-organ findings. We developed Hierarchical Structured Matching Prediction BERT (HSMP-BERT), a prompt-based model for extraction from Hebrew radiology text. In an administrative database study, we analyzed 9,683 reports from Crohn's patients imaged 2010-2023 across Israeli providers. A subset of 512 reports was radiologist-annotated for findings across six gastrointestinal organs and 15 pathologies, yielding 90 structured labels per subject. Multilabel-stratified split (66% train+validation; 33% test), preserving label prevalence. Performance was evaluated with accuracy, F1, Cohen's $\kappa$, AUC, PPV, NPV, and recall. On 24 organ-finding combinations with $>$15 positives, HSMP-BERT achieved mean F1 0.83$\pm$0.08 and $\kappa$ 0.65$\pm$0.17, outperforming the SMP zero-shot baseline (F1 0.49$\pm$0.07, $\kappa$ 0.06$\pm$0.07) and standard fine-tuning (F1 0.30$\pm$0.27, $\kappa$ 0.27$\pm$0.34; paired t-test $p < 10^{-7}$). Hierarchical inference cuts runtime 5.1$\times$ vs. traditional inference. Applied to all reports, it revealed associations among ileal wall thickening, stenosis, and pre-stenotic dilatation, plus age- and sex-specific trends in inflammatory findings. HSMP-BERT offers a scalable solution for structured extraction in radiology, enabling population-level analysis of Crohn's disease and demonstrating AI's potential in low-resource settings.

Figures

Figures reproduced from arXiv: 2509.04519 by Dan Turner, Gili Focht, Hadas Ben-Atya, Liam Hazan, Moti Freiman, Naama Gavrielov, Ruth Cytter-Kuint, Talar Hagopian, Zvi Badash.

Figure 1
Figure 1. Figure 1: Data selection flowchart. The leftmost arm of the diagram shows how radiology reports [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Hierarchical SMP-BERT training framework for structured data ex [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the Hierarchical SMP-BERT inference framework for structured data extrac [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (a) Distribution of organ-specific findings and (b) filtered organ-finding combinations [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Distribution of organ involvement and pathology prevalence stratified by sex and age [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 32 canonical work pages

  1. [1]

    Rethinking radiology reports: A survey of referring physicians’ perspectives,

    P. Reschke, C. Booz, L. D. Gruenewald, V. Koch, E. H¨ ohne, A. G¨ okduman, K. Eichler, J. Schl¨ uchtermann, T. J. Vogl, and J. Gotta, “Rethinking radiology reports: A survey of referring physicians’ perspectives,” European Journal of Radiology, vol. 187, 6 2025

  2. [2]

    Structured reporting in radiology enables epidemiological analysis through data mining: urolithiasis as a use case,

    T. Jorg, M. C. Halfmann, N. R¨ olz, R. Mager, D. P. dos Santos, C. D¨ uber, P. Mildenberger, and L. M¨ uller, “Structured reporting in radiology enables epidemiological analysis through data mining: urolithiasis as a use case,” Abdominal Radiology, vol. 48, pp. 3520–3529, 11 2023

  3. [3]

    Practical guide to natural language processing for radiology,

    A. Mozayan, A. R. Fabbri, M. Maneevese, I. Tocino, and S. Chheang, “Practical guide to natural language processing for radiology,” Radiographics, vol. 41, pp. 1446–1453, 9 2021

  4. [4]

    Transformer-based structuring of free-text radiology report databases,

    S. Nowak, D. Biesner, Y. C. Layer, M. Theis, H. Schneider, W. Block, B. Wulff, U. I. Atten- berger, R. Sifa, and A. M. Sprinkart, “Transformer-based structuring of free-text radiology report databases,” European Radiology, vol. 33, pp. 4228–4236, 6 2023

  5. [5]

    Transformation of free-text radiology reports into structured data,

    M. M. Graf, K. K. Bressem, and L. C. Adams, “Transformation of free-text radiology reports into structured data,” Radiologie, vol. 65, pp. 249–256, 4 2025

  6. [6]

    Automatic structuring of radiology reports with on-premise open-source large language models,

    P. Wo´ znicki, C. Laqua, I. Fiku, A. Hekalo, D. Truhn, S. Engelhardt, J. Kather, S. Foersch, T. A. D’Antonoli, D. P. dos Santos, B. Baeßler, and F. C. Laqua, “Automatic structuring of radiology reports with on-premise open-source large language models,” European Radiology, vol. 35, pp. 2018–2029, 4 2025

  7. [7]

    Clinical tractor: A framework for automatic natural language understanding of clinical practice guidelines,

    D. R. Schlegel, K. Gordon, C. Gaudioso, and M. Peleg, “Clinical tractor: A framework for automatic natural language understanding of clinical practice guidelines,” AMIA Annual Sym- posium Proceedings, vol. 2019, pp. 784–793, 3 2020

  8. [8]

    A scoping review of large language model based approaches for information extraction from radiology reports,

    D. Reichenpfader, H. M¨ uller, and K. Denecke, “A scoping review of large language model based approaches for information extraction from radiology reports,” npj Digital Medicine 2024 7:1 , vol. 7, pp. 1–12, 8 2024

  9. [9]

    Performance of an open-source large language model in extracting information from free-text radiology reports,

    B. L. Guellec, A. Lef` evre, C. Geay, L. Shorten, C. Bruge, L. Hacein-Bey, P. Amouyel, J. P. Pruvo, G. Kuchcinski, and A. Hamroun, “Performance of an open-source large language model in extracting information from free-text radiology reports,” Radiology: Artificial Intelligence , vol. 6, 7 2024. 13

  10. [10]

    Improving large language model applications in biomedicine with retrieval-augmented generation: a systematic review, meta-analysis, and clinical development guidelines,

    S. Liu, A. B. McCoy, and A. Wright, “Improving large language model applications in biomedicine with retrieval-augmented generation: a systematic review, meta-analysis, and clinical development guidelines,” Journal of the American Medical Informatics Association , vol. 32, pp. 605–615, 4 2025

  11. [11]

    Radiorag: Online retrieval-augmented generation for radiology question answering,

    S. T. Arasteh, M. Lotfinia, K. Bressem, R. Siepmann, L. Adams, D. Ferber, C. Kuhl, J. N. Kather, S. Nebelung, and D. Truhn, “Radiorag: Online retrieval-augmented generation for radiology question answering,” Radiology. Artificial intelligence, vol. 7, 7 2025

  12. [12]

    Clinical, cultural, computational, and regulatory considerations to deploy ai in radiology: Perspectives of rsna and miccai experts,

    M. G. Linguraru, S. Bakas, M. Aboian, P. D. Chang, A. E. Flanders, J. Kalpathy-Cramer, F. C. Kitamura, M. P. Lungren, J. Mongan, L. M. Prevedello, R. M. Summers, C. C. Wu, M. Adewole, and C. E. Kahn, “Clinical, cultural, computational, and regulatory considerations to deploy ai in radiology: Perspectives of rsna and miccai experts,” Radiology: Artificial ...

  13. [13]

    Challenges in using large language models for simplifying radiology reports,

    Y. Artsi, V. Sorin, and E. Klang, “Challenges in using large language models for simplifying radiology reports,” Journal of the American College of Radiology , vol. 22, p. 405, 4 2025

  14. [14]

    Implementing large language models in healthcare while balancing control, collaboration, costs and security,

    F. Dennst¨ adt, J. Hastings, P. M. Putora, M. Schmerder, and N. Cihoric, “Implementing large language models in healthcare while balancing control, collaboration, costs and security,” npj Digital Medicine, vol. 8, pp. 1–4, 12 2025

  15. [15]

    Current applications and challenges in large language models for patient care: a systematic review,

    F. Busch, L. Hoffmann, C. Rueger, E. H. van Dijk, R. Kader, E. Ortiz-Prado, M. R. Makowski, L. Saba, M. Hadamitzky, J. N. Kather, D. Truhn, R. Cuocolo, L. C. Adams, and K. K. Bressem, “Current applications and challenges in large language models for patient care: a systematic review,” Communications Medicine 2025 5:1 , vol. 5, pp. 1–13, 1 2025

  16. [16]

    Medical ethics of large language models in medicine,

    J. C. L. Ong, S. Y.-H. Chang, W. William, A. J. Butte, N. H. Shah, L. S. T. Chew, N. Liu, F. Doshi-Velez, W. Lu, J. Savulescu, and D. S. W. Ting, “Medical ethics of large language models in medicine,” NEJM AI , vol. 1, p. 14, 6 2024

  17. [17]

    Cybersecurity threats and mitigation strategies for large language models in health care,

    T. A. D’Antonoli, A. S. Tejani, B. Khosravi, C. Bluethgen, F. Busch, K. K. Bressem, L. C. Adams, M. Moassefi, S. Faghani, and J. W. Gichoya, “Cybersecurity threats and mitigation strategies for large language models in health care,” Radiology. Artificial intelligence, vol. 7, 7 2025

  18. [18]

    Lchaim - investigating long context reasoning in hebrew,

    E. Malul, O. Perets, Z. Mor, Y. Kassel, and E. Sulem, “Lchaim - investigating long context reasoning in hebrew,” in Findings of the Association for Computational Linguistics: ACL 2025 (W. Che, J. Nabende, E. Shutova, and M. T. Pilehvar, eds.), pp. 7928–7939, Association for Computational Linguistics, 7 2025

  19. [19]

    Veef-multi-llm: Effective vocabulary expansion and parameter efficient finetuning towards multilingual large language models,

    J. Sha, M. Zhu, C. Feng, and Y. Shang, “Veef-multi-llm: Effective vocabulary expansion and parameter efficient finetuning towards multilingual large language models,” in Proceedings of the 31st International Conference on Computational Linguistics (O. Rambow, L. Wanner, M. Apidianaki, H. Al-Khalifa, B. D. Eugenio, and S. Schockaert, eds.), pp. 7963–7981, ...

  20. [20]

    Implementing large language models in health care: Clinician-focused review with interactive guideline,

    H. Y. Li, J. F. Fu, and A. Python, “Implementing large language models in health care: Clinician-focused review with interactive guideline,” Journal of medical Internet research , vol. 27, p. e71916, 7 2025. 14

  21. [21]

    Performance of large language models in numerical versus semantic medical knowledge: Cross-sectional benchmarking study on evidence-based questions and answers,

    E. Avnat, M. Levy, D. Herstain, E. Yanko, D. B. Joya, M. T. Katz, D. Eshel, S. Laros, Y. Dagan, S. Barami, J. Mermelstein, S. Ovadia, N. Shomron, V. Shalev, and R. E. E. Abdul- nour, “Performance of large language models in numerical versus semantic medical knowledge: Cross-sectional benchmarking study on evidence-based questions and answers,” Journal of ...

  22. [22]

    D. H. Bruining, E. M. Zimmermann, E. V. Loftus Jr, W. J. Sandborn, C. G. Sauer, S. A. Strong, and S. of Abdominal Radiology Crohn’s Disease-Focused Panel, “Consensus recom- mendations for evaluation, interpretation, and utilization of computed tomography and mag- netic resonance enterography in patients with small bowel crohn’s disease,”Radiology, vol. 28...

  23. [23]

    Promptbert: Improving bert sentence embeddings with prompts,

    T. Jiang, J. Jiao, S. Huang, Z. Zhang, D. Wang, F. Zhuang, F. Wei, H. Huang, D. Deng, and Q. Zhang, “Promptbert: Improving bert sentence embeddings with prompts,” in Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (Y. Goldberg, Z. Kozareva, and Y. Zhang, eds.), pp. 8826–8837, Association for Computational Linguistics...

  24. [24]

    Zhuang, J

    Y. Zhuang, J. Zhang, X. Li, C. Liu, Y. Yu, W. Dong, and K. He, “Autonomous interna- tional classification of diseases coding using pretrained language models and advanced prompt learning techniques: Evaluation of an automated analysis system using medical text,” JMIR Medical Informatics, vol. 13, 2025

  25. [25]

    Reichenpfader, J

    D. Reichenpfader, J. Knupp, S. U. von D¨ aniken, R. Gaio, F. Dennst¨ adt, G. M. Cereghetti, A. Sander, H. Hiltbrunner, K. Nairz, and K. Denecke, “Enhancing bidirectional encoder rep- resentations from transformers (bert) with frame semantics to extract clinically relevant infor- mation from german mammography reports: Algorithm development and validation,...

  26. [26]

    Leveraging prompt-learning for structured information extraction from crohn’s disease radiology reports in a low-resource language,

    L. Hazan, N. Gavrielov, R. Reichart, T. Hagopian, M.-L. Greer, R. Cytter-Kuint, G. Focht, D. Turner, and M. Freiman, “Leveraging prompt-learning for structured information extraction from crohn’s disease radiology reports in a low-resource language,” in Proceedings of the 6th Clinical Natural Language Processing Workshop (T. Naumann, A. B. Abacha, S. Beth...

  27. [27]

    Development and valida- tion of novel algorithms to identify patients with inflammatory bowel diseases in israel: an epi-iirn group study,

    M. Y. Friedman, M. Leventer-Roberts, J. Rosenblum, N. Zigman, I. Goren, V. Mourad, N. Led- erman, N. Cohen, E. Matz, D. Z. Dushnitzky, N. Borovsky, M. B. Hoshen, G. Focht, M. Avit- zour, Y. Shachar, Y. Chowers, R. Eliakim, S. Ben-Horin, S. Odes, D. Schwartz, I. Dotan, E. Israeli, Z. Levi, E. I. Benchimol, R. D. Balicer, and D. Turner, “Development and val...

  28. [28]

    Hero: Roberta and longformer hebrew language models,

    V. Shalumov and H. Haskey, “Hero: Roberta and longformer hebrew language models,” arXiv preprint arXiv:2304.11077, 2023

  29. [29]

    Don’t stop pretraining: Adapt language models to domains and tasks,

    S. Gururangan, A. Marasovi´ c, S. Swayamdipta, K. Lo, I. Beltagy, D. Downey, and N. A. Smith, “Don’t stop pretraining: Adapt language models to domains and tasks,” in Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (D. Jurafsky, J. Chai, N. Schluter, and J. Tetreault, eds.), pp. 8342–8360, Association for Computatio...

  30. [30]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, Volume 1 (Long and Short Papers) (J. Burstein, C. Doran, and T. Solorio, e...

  31. [31]

    Biobert: A pre-trained biomedical language representation model for biomedical text mining,

    J. Lee, W. Yoon, S. Kim, D. Kim, S. Kim, C. H. So, and J. Kang, “Biobert: A pre-trained biomedical language representation model for biomedical text mining,” Bioinformatics, vol. 36, pp. 1234–1240, 2 2020

  32. [32]

    Clinicalbert: Modeling clinical notes and predict- ing hospital readmission,

    K. Huang, J. Altosaar, and R. Ranganath, “Clinicalbert: Modeling clinical notes and predict- ing hospital readmission,” arXiv preprint arXiv:1904.05342 , 2019

  33. [33]

    A survey of gpt-3 family large language models including chatgpt and gpt-4,

    K. S. Kalyan, “A survey of gpt-3 family large language models including chatgpt and gpt-4,” Natural Language Processing Journal, vol. 6, p. 100048, 3 2024

  34. [34]

    Biogpt: A generative transformer-based framework for personalized genomic medicine and rare disease diagnosis,

    G. Al-Kateb, E. Cengiz, and M. G¨ ok, “Biogpt: A generative transformer-based framework for personalized genomic medicine and rare disease diagnosis,” Mesopotamian Journal of Artificial Intelligence in Healthcare, vol. 2025, pp. 154–164, 7 2025. 16 Supporting information Table S1: Label prevalence for each of the selected labels (with at least 15 positive...