Pith. sign in

REVIEW 5 major objections 5 minor 22 references

Structuring Radiology Reports: Challenging LLMs with Lightweight Models

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A fully finetuned 278M-parameter BERT2BERT model structures radiology reports into a fixed template as well as or better than prompt-adapted LLMs up to 70B parameters, and at a fraction of the compute cost.

desk verdict Useful cost-aware benchmark with a real evaluation-pipeline flaw: the human gold standard may be GPT-4-anchored, so the headline superiority claim is overstated, but the deployment argument holds. read the letter →

arxiv 2506.00200 v2 pith:BBW2FMPV submitted 2025-05-30 cs.CL cs.LG

Johannes Moll , Louisa Fay , Asfandyar Azhar , Sophie Ostmeier , Tim Lueth , Sergios Gatidis , Curtis Langlotz , Jean-Benoit Delbrouck This is my paper
classification cs.CLcs.LG
keywords radiologyreportstructuringlightweightencoder-decodermodelsBERT2BERTlargelanguageLoRAfinetuningGPT-4weakannotationMIMIC-CXRCheXpertPlus
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that converting free-text chest X-ray reports into a standardized template is not a task that requires massive language models. A fully finetuned 278M-parameter BERT2BERT model, trained on roughly 183,000 reports structured by GPT-4, outperforms all LLMs adapted with prefix prompting or in-context learning on a radiologist-reviewed test set. Even when LLMs are finetuned with LoRA, only the Findings section sees modest gains for the largest models, and no LLM beats the lightweight model when Findings and Impression sections are averaged. The authors pair this with a cost analysis showing that the 70B-parameter LLaMA-3 model costs more than 400 times as much in inference time, money, and carbon emissions. If correct, this means that for structured clinical text transformation, small task-specific models are the practical choice for deployment in privacy- and resource-constrained healthcare settings.

What carries the argument

The load-bearing design is BERT2BERT, an encoder-decoder built from two identical pretrained RoBERTa models totaling 278M parameters, fully finetuned on GPT-4-annotated training pairs. The fixed output structure is the RSNA RadReport template RPT144, which defines the section headers (Exam Type, History, Technique, Comparison, Findings, Impression) and the organ-system organization inside Findings. The argument is carried by the human-annotated test set and by radiology-specific evaluation metrics—GREEN and F1-SRR-BERT—that measure clinical and structural correctness rather than just lexical overlap, which is what lets the paper claim clinical parity despite modest lexical scores.

What would settle it

Re-annotate a fresh set of, say, 200 chest X-ray reports by having radiologists write the structured versions from scratch without ever seeing an AI-generated structure, then run the same benchmark; if a LoRA-finetuned LLM (e.g., LLaMA-3-8B) then matches or exceeds the BERT2BERT model on the averaged sections, the paper's central performance claim would be tied to GPT-4-derived ground truth rather than to the models' intrinsic ability.

Watch

Extended reading notes

Core claim

The central claim is that a 278M-parameter BERT2BERT model—two RoBERTa models serving as encoder and decoder, fully finetuned end-to-end on 182,962 GPT-4-generated structured reports—structures free-text chest X-ray reports into the RSNA RPT144 template at or above the level of much larger LLMs. On a 223-report test set reviewed by five radiologists, the lightweight model beats every LLM adapted with prefix prompting or in-context learning, and when results are averaged across the Findings and Impression sections, no adapted LLM, including 70B-parameter LLaMA-3 with LoRA, surpasses it. The few LoRA-finetuned LLMs that beat the lightweight model on the Findings section alone do so by small margins (e.g., BLEU +6.4%, ROUGE-L +4.8%) while using more than 400 times the inference time, cost, and carbon emissions, and the lightweight model is also far cheaper to train.

Load-bearing premise

The human-annotated test set was created with GPT-4's structured output shown to radiologists as a reference, and the lightweight models were trained on GPT-4 labels, so the evaluation may encode GPT-4's formatting preferences in a way that favors the lightweight models.

Editorial extensions

If this is right

  • For template-driven restructuring of clinical text, scale is not decisive: a fully finetuned sub-300M model matches or beats LLMs up to 70B when both sections are averaged.
  • Without LoRA finetuning, LLMs underperform badly on this task; prefix prompting and in-context learning leave large gaps on the Findings section, so parameter-efficient finetuning is the only realistic LLM adaptation for structured reporting.
  • The cost gap is large enough to change deployment decisions: the 70B model used more than 400 times the inference time, cost, and CO2 of the lightweight model, and training it emitted about 142 times more CO2.
  • Domain-matched pretraining helps: initializing from biomedical and radiology text (RoBERTa-PM-M3) gives the best lightweight model, though radiology-exclusive pretraining does not always win.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the radiologist-ground-truth test set carries a GPT-4 formatting bias, the lightweight model's margin could shrink against LLMs that never saw those labels; an independent from-scratch annotation would settle this.
  • The pipeline of 'LLM as weak annotator plus small fully finetuned model' is a general recipe: as the annotator model improves, the same recipe should produce better small models for other structured clinical outputs, which this paper does not test.
  • The cost comparison assumes the institution would pay for LLM-scale compute; for organizations that already run an LLM for other purposes, the marginal cost of reuse is lower, which would narrow—but not eliminate—the reported gap.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes lightweight encoder-decoder models (T5 and BERT2BERT, under 300M parameters) for converting free-text chest X-ray radiology reports into a structured template based on RSNA RPT144. Training labels for 182,962 reports from MIMIC-CXR and CheXpert Plus are generated by GPT-4 as a weak annotator, and the models are compared against open-source LLMs (1B–70B) adapted via prefix prompting, in-context learning, and LoRA finetuning. Evaluation uses 223 reports reviewed by five radiologists and a battery of lexical and radiology-specific metrics. The central claim is that the best lightweight BERT2BERT model outperforms all prompt-adapted LLMs when averaged across the Findings and Impression sections, with some LoRA-finetuned LLMs achieving modest gains on Findings at substantially higher computational cost.

Significance. If the result holds, it is practically important: it would show that small, task-specific models can match or exceed prompt-based LLM adaptation for structured radiology reporting at a fraction of the compute, with favorable privacy and deployment characteristics. The study is broad in scope, covering multiple model families, adaptation strategies, and clinically oriented metrics, and it includes a useful cost and environmental-impact analysis. The public release of model checkpoints and the use of multiple seeds for lightweight models are also strengths. However, the evaluation pipeline may favor models trained on GPT-4 labels, and the abstract overstates the prompt-based comparison, so the headline claim needs qualification before the significance can be assessed.

major comments (5)
  1. [Section 3.1 / Limitation 6 / Table 7] The central comparison may be confounded by the origin of the gold standard. The training labels for the lightweight and LoRA-finetuned models were generated by GPT-4 (Section 3.1), and Limitation 6 states that GPT-4 "was used for data annotation and provided as a reference for radiologist." If radiologists created or corrected the 223 human-reviewed test reports with GPT-4 output in view, the ground truth inherits GPT-4 formatting and content preferences; models trained on exactly those GPT-4 labels are then evaluated against a standard they were explicitly optimized to mimic, while prompt-based LLMs did not see those labels. Excluding GPT-4 from the final tables removes direct self-comparison but leaves the confound intact. The authors should either clarify the radiologist annotation protocol in detail, or re-annotate a subset without GPT-4 reference and report the performance margins on that subset. This is load-bearing for the claim that the lightweight model outperforms all prompt-adapted LLMs.
  2. [Abstract / Section 4.3 / Table 7] The abstract's headline claim is stronger than the results. It states that the best lightweight model "outperforms all LLMs adapted using prompt-based techniques on a human-annotated test set," but Table 7 shows that in the Impression section LLaMA-3-70B with Prefix+ICL achieves GREEN 89.2 and F1-SRR-BERT 88.3, compared with BERT2BERT's 77.8 and 85.8, and the F1-RadGraph score is 69.0 vs. 62.1. Section 4.3 correctly limits the claim to "when averaged across both sections," but this qualification is missing from the abstract, the contributions list, and parts of the Discussion. The abstract should be rewritten to state the section-specific exceptions and the averaging convention.
  3. [Section 4.3 / Table 7] No confidence intervals or statistical tests are reported for any LLM condition. Only lightweight models were trained with multiple seeds (three, with error bars in Figure 3); each LLM adaptation appears to be a single run evaluated on 223 human-reviewed reports. Several margins that support the "no LLM surpasses" claim are small (e.g., Findings F1-SRR-BERT: BERT2BERT 72.3 vs. LLaMA-3-8B LoRA 74.0 and LLaMA-3-3B LoRA 75.4; Impression F1-SRR-BERT: BERT2BERT 85.8 vs. LLaMA-3-8B LoRA 83.7). Without bootstrap intervals or rerun variability, the reader cannot tell whether the ordering is stable, and the "modest gains" characterization is not quantified against sampling noise. Please add confidence intervals or a paired significance test for the main comparisons.
  4. [Section 4.3 / Figure 5 / Limitation 7] The largest model is evaluated under an unequal training budget. LLaMA-3-70B with LoRA was trained for one epoch instead of five (Figure 5 caption, Limitation 7), and it is precisely this under-trained configuration that allows the statement "when averaged across both sections, no LLM surpasses the lightweight model" to hold; in the Impression section, the 70B Prefix+ICL model already surpasses BERT2BERT (see Major 2). The claim should be framed as "under the authors' fixed adaptation resources," not as a scale-independent property, and the authors should report what happens when the 70B LoRA model is trained for five epochs or argue explicitly why one epoch is a fair resource-equated comparison.
  5. [Section 3.2 / Section 5] The scoring of template adherence is described inconsistently. Section 3.2 says metrics are computed for identified organ systems and averaged across identified systems, while Section 5 says the evaluation "assigned zero points to missing or inconsistently labeled headers." These are different protocols: the former ignores missing headers, the latter penalizes them. Which one was implemented, and how did zero-point scoring affect the reported margins? Because the prompt-based LLMs were less likely to reproduce exact header names (Table 8 shows 35 different organ-system names for LLaMA-3-70B), this detail directly affects the comparison between lightweight/LoRA models and prompting baselines.
minor comments (5)
  1. [Section 4.4 / Table 8] The text reports 223 human-reviewed test reports, but Table 8's caption says 233 test samples; please reconcile the numbers.
  2. [Section 3.2 / Tables 4–7] The metric is referred to as "F1-SRRG-Bert", "F1-SRR-BERT", and "F1-Score" in different places; use a single name throughout.
  3. [Section 5] The statement that the lightweight model "operated at less than 25% of their inference cost" is not supported by Table 1, where the lightweight model's single-sample inference cost is 28.7% of LLaMA-3-3B's; please specify the reference model and batch scenario.
  4. [Section 5] In the qualitative analysis paragraph, "lighweight" is a typo for "lightweight".
  5. [Tables 5–7] The tables use "Radgraph" while the text and Section 3.2 use "F1-RadGraph"; standardize the capitalization.

Circularity Check

1 steps flagged · score 4.0 of 10

Human gold standard inherits GPT-4 formatting: models trained on GPT-4 labels are evaluated against a test set annotated with GPT-4 as reference, partially forcing the lightweight model's margin over prompt-based LLMs.

  1. fitted input called prediction [Section 3.1 and Limitations, sixth; abstract claim]
    "However, since it was used for data annotation and provided as a reference for radiologist, its results may be biased in its favor. To account for this, we excluded GPT-4 from the discussion to avoid misleading comparisons."

    The lightweight model and LoRA-finetuned LLMs were trained on GPT-4-generated structured labels (Section 3.1: 'we employed GPT-4 as a weak annotator'), and the same GPT-4 outputs were 'provided as a reference for radiologist' when creating the human-annotated test set. The test set therefore inherits GPT-4's formatting and content preferences. Evaluating the lightweight model on that test set and comparing it with prompt-based LLMs—which never saw the GPT-4 labels—partly measures GPT-4 mimicry rather than independent clinical quality. Excluding GPT-4 from the comparison removes direct self-comparison but not the indirect advantage for every model trained on GPT-4 labels, so the abstract's margin is partially built into the evaluation protocol.

full rationale

The paper contains no mathematical derivation chain, so most circularity patterns do not apply. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from prior work, and no ansatz is smuggled in via citation. The main comparison is an empirical benchmark with held-out human review, and the results are supported by multiple metrics, including external lexical metrics (BLEU, ROUGE-L, BERTScore, F1-RadGraph) as well as author-proposed metrics (GREEN, F1-SRR-BERT). The author-proposed metrics are self-citations, but they are not the sole load-bearing evidence; the qualitative analysis and external metrics corroborate the central trends. The one substantive circularity concern is the evaluation pipeline: GPT-4 was used both to generate training labels and as a reference for radiologists annotating the test set, as acknowledged in Limitation 6. This means the 'human-annotated test set' is not fully independent of the training signal for models trained on GPT-4 labels, which partially advantages the lightweight model over prompt-based LLMs that did not see those labels. Because the radiologists performed the final annotation and the paper is transparent about the GPT-4 reference, this is a partial evaluation-protocol circularity rather than a complete reduction of the central claim. Score 4 reflects this partial dependency while recognizing that the lightweight model's advantage is not purely tautological.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central benchmark rests on a chain of choices: GPT-4 weak labels, a small human test set, self-developed clinical metrics, LoRA hyperparameters, and a scoring rule that rewards exact template adherence. None of these are fitted constants in a derivation, but they are load-bearing assumptions about what 'correct structuring' means and whether the LLM comparison is fair.

free parameters (5)
  • LoRA rank r = 8
    Chosen by hand for all LLM finetuning; controls how much of each LLM is adapted and affects the fairness of the comparison (Section 3.4).
  • LoRA alpha = 8
    Set equal to rank; scaling factor for LoRA updates (Appendix A.4).
  • LLM finetuning epochs = 5 (1 for LLaMA-3-70B)
    Training budget is a free choice; the 70B model's single epoch likely understates its potential and shapes the cost comparison (Section 4.3, Table 1 caption).
  • ICL example selection = 1-shot and 2-shot, manually chosen
    Manually selected to represent the data distribution; authors flag selection bias (Limitations, fourth).
  • Max sequence lengths = input 370, output 120-286 tokens
    Determines how much of a report is processed and may truncate long reports (Appendix A.3).
assumptions (5)
  • domain assumption GPT-4-generated structured labels are accurate enough to serve as training supervision.
    Training uses 182,962 reports labeled by GPT-4 with no systematic validation beyond a human-reviewed test set (Section 3.1).
  • domain assumption The human-annotated test set of 223 reports is unbiased and representative for both datasets.
    All headline conclusions rest on this small set; no confidence intervals are reported for LLM runs (Section 3.1, Table 7).
  • domain assumption GREEN and F1-SRR-BERT scores reflect clinical quality rather than format mimicry.
    These are the primary metrics; both were introduced by the same research group and may encode template preferences (Sections 3.2, 4).
  • ad hoc to paper LoRA finetuning with rank 8 for up to 5 epochs is a fair representative of LLM adaptation.
    The comparison of lightweight full finetuning vs LLM LoRA depends on this. The authors acknowledge full finetuning was not extensively explored and 70B got one epoch (Limitations, seventh).
  • ad hoc to paper Scoring gives zero points for missing or inconsistent organ headers.
    Section 3.2 and Section 5: Findings are parsed by exact organ headers; missing or inconsistently labeled headers receive zero for that system, which favors models trained on the exact template.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Structuring Radiology Reports: Challenging LLMs with Lightweight Models." pith.science (2026). https://pith.science/paper/BBW2FMPV

@misc{pith2026250600200,
  author       = {Pith},
  title        = {Pith review of: Structuring Radiology Reports: Challenging LLMs with Lightweight Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BBW2FMPV}},
  note         = {Machine review of arXiv:2506.00200}
}
read the original abstract

Radiology reports are critical for clinical decision-making but often lack a standardized format, limiting both human interpretability and machine learning (ML) applications. While large language models (LLMs) have shown strong capabilities in reformatting clinical text, their high computational requirements, lack of transparency, and data privacy concerns hinder practical deployment. To address these challenges, we explore lightweight encoder-decoder models (<300M parameters)-specifically T5 and BERT2BERT-for structuring radiology reports from the MIMIC-CXR and CheXpert Plus datasets. We benchmark these models against eight open-source LLMs (1B-70B), adapted using prefix prompting, in-context learning (ICL), and low-rank adaptation (LoRA) finetuning. Our best-performing lightweight model outperforms all LLMs adapted using prompt-based techniques on a human-annotated test set. While some LoRA-finetuned LLMs achieve modest gains over the lightweight model on the Findings section (BLEU 6.4%, ROUGE-L 4.8%, BERTScore 3.6%, F1-RadGraph 1.1%, GREEN 3.6%, and F1-SRR-BERT 4.3%), these improvements come at the cost of substantially greater computational resources. For example, LLaMA-3-70B incurred more than 400 times the inference time, cost, and carbon emissions compared to the lightweight model. These results underscore the potential of lightweight, task-specific models as sustainable and privacy-preserving solutions for structuring clinical text in resource-constrained healthcare settings.

Figures

Figures reproduced from arXiv: 2506.00200 by the authors.

Figure 1
Figure 1. Overview of our study and qualitative comparison. An unstructured radiology report is structured using [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Left: Dataset generation from free-form radiology reports to structured radiology reports using GPT-4 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Performance comparison of lightweight models, initialized from pretrained models of increasing domain [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison of LLM Adaptation Methods and the best performing lightweight model (BERT2BERT [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Model performance of LLaMA-3 models of increasing size. (Left/Right) The figure shows the GREEN scores and BERTScores for adaptation using Prefix+ICL and LoRA finetuning, respectively. The re￾sult for the LLaMA-3-70B model with LoRA finetuning is indicated with a dashe…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 13 canonical work pages

  1. [1]

    Provide the relevant information as specified for each section

    Section Headers: Each section should start with the section header followed by a colon. Provide the relevant information as specified for each section

    work page

  2. [2]

    arXiv e-prints, pages arXiv–2405

    Chexpert plus: Hundreds of thousands of aligned radiology texts, images and patients. arXiv e-prints, pages arXiv–2405. Dong Chen, Shuo Zhang, Yueting Zhuang, Siliang Tang, Qidong Liu, Hua Wang, and Mingliang Xu. 2024a. Improving large models with small models: Lower costs and better performance. arXiv preprint arXiv:2406.15471. Qi Chen, Yutong Xie, Biao ...

    arXiv 2023

  3. [3]

    Journal of Machine Learning Research, 25(70):1–53

    Scaling instruction-finetuned language models. Journal of Machine Learning Research, 25(70):1–53. Raul Salles de Padua and Imran Qureshi. 2024. Leverag- ing summary of radiology reports with transformers. Artificial Intelligence in Health, 1(4):85–96. Jean-Benoit Delbrouck, Pierre Chambon, Christian Bluethgen, Emily Tsai, Omar Almusa, and Curtis P Langlot...

    arXiv 2024

  4. [4]

    Do not add or extrapolate information beyond what is found in the original report

    Content Restrictions: Strictly include only the content that is relevant to the structured sections provided. Do not add or extrapolate information beyond what is found in the original report. If the original report doesn’t contain the information necessary to generate a section, write the section header and then leave the section empty. Do not make up an...

    work page

  5. [5]

    Journal of Cancer Research and Clinical Oncology, 150(10):1–12

    Proof-of-concept study of a small language model chatbot for breast cancer decision support–a transparent, source-controlled, explainable and data- secure approach. Journal of Cancer Research and Clinical Oncology, 150(10):1–12. Suchin Gururangan, Ana Marasovi ´c, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A Smith. 2020. Don’t stop pre...

    arXiv 2020

  6. [6]

    Ensure all findings are numbered

    Impression: Summarize the key findings with a numbered list from the most to the least clinically relevant. Ensure all findings are numbered. The radiology report to improve is the following: \{report\} A.2 Overview of model checkpoints and pre-training data Table 2: Pretrained T5 models used for initialization along with details of their pretraining corp...

    work page 2020

  7. [8]

    PloS one, 19(8):e0307741

    Gpt-4 as an x data annotator: Unraveling its performance on a stance classification task. PloS one, 19(8):e0307741. NCBI. 1996. PubMed. NCBI. 2000. PubMed Central (pmc). OpenAI. 2022. Gpt-3.5. https://openai.com/. Sophie Ostmeier, Justin Xu, Zhihong Chen, Maya Varma, Louis Blankemeier, Christian Bluethgen, Arne Edward Michalson, Michael Moseley, Curtis La...

    arXiv 1996

  8. [9]

    arXiv preprint arXiv:2306.00176

    Automated annotation with generative ai re- quires validation. arXiv preprint arXiv:2306.00176. Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu. 2002. Bleu: a method for automatic evalu- ation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computa- tional Linguistics, pages 311–318. Branislav Pecher, ...

    arXiv 2002

Show all 22 references
  1. [10]

    Transactions of the Associ- ation for Computational Linguistics, 8:264–280

    Leveraging pre-trained checkpoints for se- quence generation tasks. Transactions of the Associ- ation for Computational Linguistics, 8:264–280. Jaromir Savelka, Kevin D Ashley, Morgan A Gray, Hannes Westermann, and Huihui Xu. 2023. Can gpt- 4 support analysis of textual data i...

    2023 arXiv

  2. [11]

    arXiv preprint arXiv:2305.01146

    Radadapt: Radiology report summarization via lightweight domain adaptation of large language models. arXiv preprint arXiv:2305.01146. An Yan, Julian McAuley, Xing Lu, Jiang Du, Eric Y Chang, Amilcare Gentili, and Chun-Nan Hsu. 2022. Radbert: adapting transformer-based language...

    2022 arXiv

  3. [12]

    Turbo 1106 preview

    Evaluating progress in automatic chest x-ray radiology report generation. Patterns, 4(9). Kuo Zhang, Xiangbin Meng, Xiangyu Yan, Jiaming Ji, Jingqian Liu, Hua Xu, Heng Zhang, Da Liu, Jingjia Wang, Xuliang Wang, et al. 2025. Revolutionizing health care: The transformative impac...

    2025 arXiv

  4. [14]

    Identifiers: Remove sentences where identifiers have been replaced with consecutive underscores (’\_\_\_’)

  5. [15]

    Do not reference previous studies or historical data

    Findings and Impression Sections: Focus solely on the current examination results. Do not reference previous studies or historical data

  6. [17]

    Exam Type: Provide the specific type of examination conducted

  7. [18]

    History: Provide a brief clinical history and state the clinical question or suspicion that prompted the imaging

  8. [19]

    Technique: Describe the examination technique and any specific protocols used

  9. [20]

    Comparison: Note any prior imaging studies reviewed for comparison with the current exam

  10. [21]

    Start with the organ system name followed by a colon, then list observations

    Findings: Describe all positive observations and any relevant negative observations for each organ or organ system under distinct headers. Start with the organ system name followed by a colon, then list observations. Here is the corresponding template: Organ 1: - Observation 1...

  11. [2019]

    Scientific data, 6(1):317

    Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317. Alistair EW Johnson, Tom J Pollard, Lu Shen, Li-wei H Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody, Peter Szolovits, Leo Anthony Celi, and...

    2016 arXiv

  12. [2020]

    PhysioNet

    Mimic-iv. PhysioNet. Available online at: https://physionet. org/content/mimiciv/1.0/(accessed August 23, 2021), pages 49–55. Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih- ying Deng, Roger G Mark, and Steven Horng

    2021

  13. [2023]

    In Proceedings of the 2nd Workshop on Pattern-based Approaches to NLP in the Age of Deep Learning, pages 103–113

    A hybrid of rule-based and transformer-based approaches for relation extraction in biodiversity lit- erature. In Proceedings of the 2nd Workshop on Pattern-based Approaches to NLP in the Age of Deep Learning, pages 103–113. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, A...

    2024

  14. [2024]

    Artificial Intelligence in Medicine , 154:102924

    Reshaping free-text radiology notes into struc- tured reports with generative question answering transformers. Artificial Intelligence in Medicine , 154:102924. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav...

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.