MaLei at MultiClinSUM: Summarisation of Clinical Documents using Perspective-Aware Iterative Self-Prompting with LLMs

Libo Ren; Lifeng Han; Yee Man Ng

REVIEW 4 major objections 5 minor 19 references

MaLei at MultiClinSUM: Summarisation of Clinical Documents using Perspective-Aware Iterative Self-Prompting with LLMs

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A perspective-aware iterative self-prompting pipeline on GPT-4/4o produced clinical-case summaries with BERTScore F1 of 0.8546 on 3,396 English clinical case reports, indicating strong semantic match with gold references even though lexical

desk verdict An honest shared-task system description whose own internal evidence undercuts its central claim: the refinement loop produced no measurable gain, so read it as a report of one GPT-4o configuration's official scores, not as validation of PA-ISP. read the letter →

arxiv 2509.07622 v1 pith:2PONP7AS submitted 2025-09-09 cs.CL

Libo Ren , Yee Man Ng , Lifeng Han This is my paper

classification cs.CL

keywords ClinicalsummarisationLargelanguagemodelsIterativeself-promptingPerspective-awarepromptingROUGEBERTScorecasereportsPatientcommunication

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 reports a method for automatically summarising clinical case reports with large language models, developed for the multilingual clinical summarisation shared task MultiClinSUM. The authors designed a perspective-aware variant of Iterative Self-Prompting (PA-ISP), in which GPT-4 and GPT-4o generate and then refine their own summarisation prompts guided by five clinical perspectives and by ROUGE-L/BERTScore feedback. On 3,396 English clinical case reports the final system scored 0.8546 BERTScore F1 but only 0.3077 ROUGE-L F1. The authors interpret this gap as evidence that the generated summaries preserve the meaning of the gold-standard references while paraphrasing heavily at the word level, which matters because clinical summaries need semantic fidelity rather than verbatim overlap. The paper therefore positions PA-ISP as a viable method for clinician- and patient-facing summarisation of clinical documents.

What carries the argument

Perspective-aware Iterative Self-Prompting (PA-ISP): a loop in which a meta-prompt—combining chain-of-thought instructions, five clinical perspectives (Patient Presentation, Clinical Presentation, Diagnosis, Treatment/Intervention, Outcome and Follow-up), and metric-based feedback—prompts the model to write a task-specific prompt; GPT-4o then uses that prompt with three few-shot examples to summarise 50 gold training cases, the outputs are scored with ROUGE-L and BERTScore, and the 15 worst-scoring cases are fed back with reflection to revise the prompt. This repeats over five epochs; the first prompt version with no invalid predictions was selected for inference. The machinery carries the c

What would settle it

Generate summaries for the same 3,396 test reports with GPT-4o using a simple one-shot instruction without iterative prompt refinement, and compare BERTScore and ROUGE-L; if scores are within about 0.01 BERTScore F1 of the reported 0.8546, the high semantic fidelity is attributable to the backbone model rather than to PA-ISP.

Watch

Extended reading notes

Core claim

The central claim is that perspective-aware Iterative Self-Prompting on GPT-4/4o generates clinical case summaries that are semantically equivalent to human-written references. With 3,396 English test cases, the official evaluation gave BERTScore precision 0.8784, recall 0.8325, F1 0.8546, while ROUGE-L was 0.4653, 0.2468, 0.3077. The authors argue the gap is a signature of heavy but meaning-preserving paraphrase: 99.35% of instances fall in the 0.8-0.9 BERTScore range, and qualitative review shows summaries cover the key patient, diagnostic, treatment, and outcome information. They further claim this supports deployment of PA-ISP for clinical summarisation to improve communication between p

Load-bearing premise

The effectiveness claim relies on the assumption that the reported semantic quality comes from the perspective-aware iterative self-prompting method itself; the paper does not compare against a plain, non-iterative prompt on the same model.

Editorial extensions

If this is right

If PA-ISP works as reported, clinical reports can be condensed into five-part structured summaries spanning presentation, diagnosis, treatment, and follow-up without losing medical meaning.
The BERTScore/ROUGE split implies clinical-summary evaluation should rely on semantic metrics, since generated text may deliberately paraphrase into more patient-facing language than the gold references use.
The model's tendency to expand abbreviations and add explicit section headers indicates a style mismatch with reference summaries; future prompt revisions that match reference style could raise ROUGE without hurting BERTScore.
The generated observations about missing information, such as absent follow-up details, could help clinicians spot under-documented cases and improve the completeness of clinical reports.
The same perspective-aware loop is directly extensible to the Spanish, French, and Portuguese data in MultiClinSUM and to plain-language patient-facing adaptations.

Reading between the lines

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

A natural next experiment, not reported in the paper, is an ablation that runs GPT-4o with a single fixed prompt and no refinement loop; such a comparison would separate the contribution of iterative self-prompting from the backbone model's intrinsic summarisation ability.
The strong BERTScore with weak ROUGE-L could partly be an artifact of the five-perspective section-header format: if headers and expanded abbreviations are stripped before scoring, lexical overlap with the gold summaries should rise, and this is testable on the released test set.
The precision-recall gap (BERTScore precision 0.878 vs recall 0.833; ROUGE-L precision 0.465 vs recall 0.247) suggests the summariser is conservative, including mostly correct content but omitting some details; a follow-up targeting recall could matter more for clinical completeness than for general readability.
The model's habit of commenting on missing report sections points toward a separate quality-control tool for clinical documentation, a use case the paper mentions only in passing.

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.

Desk Editor's Note

read the letter

The useful part of this paper is the candor. It reports official MultiClinSUM scores on 3,396 English clinical reports, adapts the authors' earlier Iterative Self-Prompting approach with five manually designed clinical perspectives, and then honestly documents where things failed: BERTScore stayed flat around 0.86, ROUGE-L stayed around 0.30 across five refinement epochs, prompt_v2 was chosen only because it fixed one invalid prediction, and the fuller structure-aware and retrieval-based enhancements were never run. The qualitative error analysis is also genuinely informative—longer-than-source summaries, rigid section headers, hallucinated follow-up details, and paraphrases that hurt ROUGE but not BERTScore are all described concretely. That is better than most shared-task writeups.

The load-bearing problem is the one the stress-test note names, and it is confirmed by the paper's own text. Sections 3.3 and 3.4 state that iterative refinement did not improve scores. With no baseline—no zero-shot prompt, no plain few-shot prompt, no fine-tuned encoder-decoder—the official numbers cannot be attributed to the PA-ISP loop. The high BERTScore could be GPT-4o's intrinsic summarization ability. So the concluding claim that PA-ISP "can help summarise lengthy clinical documents" is not supported by the evidence presented. The paper itself almost concedes this, saying more sophisticated strategies beyond self-iterative prompt refinement may be required.

There are also reproducibility gaps: no exact final prompt in the text, no commit hash, and the GitHub link is not populated with code. Those are common in working notes, but they matter here because the only novel component is the prompt, and it is not fully specified.

For a reader tracking MultiClinSUM or prompting strategies for clinical text, the paper is worth a skim for the qualitative findings and the official numbers. For a reader evaluating whether PA-ISP works, the paper should not be cited as evidence for that. It deserves a serious referee only if framed as a system description; as a methods claim it would need a simple baseline and softer conclusions. My own verdict: conditional accept for a workshop, not for a general NLP venue. A quick zero-shot comparison on the 50-sample development batch would have made this a much stronger paper.

Referee Report

4 major / 5 minor

Summary. The paper describes the MaLei team's submission to the MultiClinSUM 2025 shared task on clinical case report summarization. The authors propose Perspective-Aware Iterative Self-Prompting (PA-ISP), in which an LLM generates and iteratively refines task-specific prompts using few-shot examples and metric-based feedback (ROUGE-L and BERTScore). The summarized system is applied to 3,396 English clinical case reports using GPT-4 for prompt generation and GPT-4o for summarization and reflection. The official test-set scores are ROUGE-L F1 0.3077 and BERTScore F1 0.8546. The paper also presents qualitative analysis and error analysis. The central claim is that PA-ISP can help summarize lengthy clinical documents while preserving semantic content.

Significance. If the central claim were fully supported, the paper would provide a practical, training-free approach to clinical summarization with high semantic fidelity, a useful contribution for shared-task venues. Strengths include the use of official shared-task evaluation on a large test set, a candid qualitative analysis that identifies concrete failure modes, and a public GitHub repository for reproducibility. However, the reported quantitative results are not sufficient to establish the effectiveness of the PA-ISP method: the paper provides no non-ISP baseline, and its own development-phase observations indicate that iterative refinement did not improve scores. The significance of the work therefore hinges on whether the authors can supply controlled evidence isolating the contribution of the iterative self-prompting loop.

major comments (4)

[§3.3, §3.4] The paper's own development results show no quantitative improvement from iterative refinement: after five epochs, BERTScore remained around 0.86 and ROUGE-L around 0.30, and prompt_v2 was adopted only because it fixed one invalid prediction. This directly undermines the attribution of the final test-set scores to the ISP loop. Please report development-set scores for prompt_v1 versus prompt_v2, and where possible a comparison against a non-ISP prompt, so that the effect of the iterative component is quantified.
[§4.2] No baseline system is reported. The official evaluation table contains only the MaLei submission, and there is no comparison with zero-shot prompting, plain few-shot prompting, or a fine-tuned encoder-decoder model. Without such a baseline, the high BERTScore (0.8546 F1) cannot be attributed to PA-ISP rather than to GPT-4o's intrinsic summarization capability. Add at least a zero-shot or plain few-shot baseline on the 50-sample development set, and if feasible on a random subsample of the test set, with matched decoding settings.
[§3.5 vs. §3.6] The description of the final system is internally inconsistent. Section 3.5 describes a Similar Case Retrieval-based Few-shot Augmentation used 'to enhance test-time generation,' but Section 3.6 states that this augmentation was not performed due to time constraints. The final system configuration must be stated unambiguously. In addition, the post-processing in Section 3.6 (regenerating summaries up to five times and replacing longer outputs with the original full text) is not quantified or ablated; its effect on the reported aggregate scores should be reported.
[§4.2.1, Figure 4] The interpretation of the score distribution is speculative. The text says the left-skewed ROUGE-L distribution 'suggests that ISP may have inaccurately generalized the structure of the examples,' but a score distribution alone cannot identify the cause. This claim should be softened or supported with side-by-side examples or a targeted analysis.

minor comments (5)

[Throughout] The metric name is written inconsistently: 'BERTScore' and 'BERT-score' are both used. Please standardize to 'BERTScore'.
[Figures 3 and 4] The grouped bar chart and histogram are difficult to read in the provided version. Please ensure high-resolution figures with labeled axes and readable legends.
[§4.2.1] The sentence 'BERTScore is overall more than twice as high as ROUGE-L' is imprecise: the F1 values are 0.8546 and 0.3077, so a ratio-based wording would be clearer.
[§3.4] The phrase 'native instructions for the five clinical perspectives' is unclear; presumably 'initial instructions' or 'original instructions' is intended.
[References] Reference [3] is cited for ISP, but a fuller description of the ISP method's design choices and relationship to prior prompt-optimization work would help contextualize the contribution.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: official test-set scores are independent of the prompt-selection procedure; self-citations are motivational, not load-bearing.

full rationale

The reported test-set scores (ROUGE-L F1 0.3077, BERTScore F1 0.8546 on 3,396 cases) come from the official MultiClinSUM evaluation. The prompt used for inference (prompt_v2) was selected from a 50-sample gold-standard subset (§3.3), and the retrieval/structure-augmentation stages were not applied to the test set (§3.4, §3.5). Therefore the evaluation is not fitted to the test labels, and no 'prediction' reduces by construction to a training fit. The self-citation of [3] (the authors' PerAnsSumm paper) is used only to motivate the ISP approach and to indicate prior experience; it is not the evidence for the present quantitative claim. The paper's own admission that iterative refinement did not improve ROUGE-L (§3.4) and the absence of a non-ISP baseline are evidential weaknesses regarding attribution of the scores to PA-ISP, but they are not circularity: they concern whether the method is necessary, not whether the derivation is definitionally forced. No equation or construction equates an input with an output. Hence no circular step.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities. Its free parameters are hand-chosen hyperparameters of the prompting pipeline, listed above. The central assumptions are that LLM self-prompting works as advertised, that ROUGE-L/BERTScore capture quality, and that the designed perspective structure is appropriate; none of these are tested against alternative configurations.

free parameters (5)

Number of prompt refinement epochs = 5
Chosen by hand; the paper reports that five epochs were run and that scores plateaued, but no grid search or stopping criterion is given.
Number of few-shot examples in meta-prompt = 3
Three representative examples were selected for initial prompt generation; no systematic search over example count.
Number of low-scoring summaries used for refinement = 15
For each iteration, the 15 summaries with lowest ROUGE-L were used as feedback; chosen without reported sensitivity analysis.
Regeneration threshold and tries = regenerate if summary > input; up to 5 tries; replace with original if still longer
A post-hoc rule to handle overlong outputs; the threshold and try count are arbitrary hand choices.
Selection of prompt_v2 = Prompt version 2
Selected because it eliminated invalid predictions, not because evaluation scores improved; this is a manual, outcome-based selection on the validation subset.

assumptions (4)

domain assumption Large language models can generate and iteratively refine task-specific prompts that improve summarization performance.
Carried over from prior PerAnsSumm work (reference [3]); not re-established here with a baseline or controlled experiment.
domain assumption ROUGE-L and BERTScore are suitable metrics to guide prompt refinement and to judge summary quality.
Used as optimization targets during prompt selection and as evaluation metrics; the paper itself discusses ROUGE-L's limitations, so the metric choice is a recognized but untested assumption.
ad hoc to paper The five clinical perspectives (Patient Presentation, Clinical Presentation, Diagnosis, Treatment, Outcome) are the correct structure for clinical summaries.
Manually designed by the authors; no evidence is provided that this structure matches gold-standard summaries, and the paper notes it may have hurt ROUGE-L.
domain assumption Gold-standard summaries from the shared task are valid references for measuring summarization quality.
Standard practice in shared tasks; the paper relies on the official references and evaluation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MaLei at MultiClinSUM: Summarisation of Clinical Documents using Perspective-Aware Iterative Self-Prompting with LLMs." pith.science (2026). https://pith.science/paper/2PONP7AS

@misc{pith2026250907622,
  author       = {Pith},
  title        = {Pith review of: MaLei at MultiClinSUM: Summarisation of Clinical Documents using Perspective-Aware Iterative Self-Prompting with LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2PONP7AS}},
  note         = {Machine review of arXiv:2509.07622}
}

read the original abstract

Efficient communication between patients and clinicians plays an important role in shared decision-making. However, clinical reports are often lengthy and filled with clinical jargon, making it difficult for domain experts to identify important aspects in the document efficiently. This paper presents the methodology we applied in the MultiClinSUM shared task for summarising clinical case documents. We used an Iterative Self-Prompting technique on large language models (LLMs) by asking LLMs to generate task-specific prompts and refine them via example-based few-shot learning. Furthermore, we used lexical and embedding space metrics, ROUGE and BERT-score, to guide the model fine-tuning with epochs. Our submission using perspective-aware ISP on GPT-4 and GPT-4o achieved ROUGE scores (46.53, 24.68, 30.77) and BERTscores (87.84, 83.25, 85.46) for (P, R, F1) from the official evaluation on 3,396 clinical case reports from various specialties extracted from open journals. The high BERTscore indicates that the model produced semantically equivalent output summaries compared to the references, even though the overlap at the exact lexicon level is lower, as reflected in the lower ROUGE scores. This work sheds some light on how perspective-aware ISP (PA-ISP) can be deployed for clinical report summarisation and support better communication between patients and clinicians.

Figures

Figures reproduced from arXiv: 2509.07622 by the authors.

**Figure 1.** Perspective-aware Iterative Self-Prompting (PA-ISP) Illustration Diagrams from the MaLei team. 2. Clinical Presentation: key symptoms and signs. 3. Diagnosis: relevant investigations, tests, conclusions. 4. Treatment/Intervention: medications, surgeries, therapies. 5. Outcome and Follow-up: results of treatment, current status. The metrics we used are ROUGE-L (for lexical overlap) and BERTScore (for embedding semant… view at source ↗

**Figure 2.** Data Partitioning and Usage in the MultiClinSUM Pipeline in our pipeline. We completed summary generation for 3,396 English full-text cases. Other languages (Spanish, Portuguese, and French) will be explored in future work. 4.2. Submission outcome of MaLei from the shared task 4.2.1. Quantitative Results For the MultiClinSUM2025 shared task that we attended [18], the results of 3,396 submitted English test summaries… view at source ↗

**Figure 3.** Evaluation Metric Comparison: BERTScore vs ROUGE-L [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 5.** Figure 5: Example of an original report next to the generated summary using the ISP technique, in which the generated summary covers key aspects in the original report. Generally, the summaries tend to cover key features in the clinical report and contain logical paraphrasing, …

**Figure 6.** Figure 6: Example of output during prompting procedure in which the generated summary is much longer and detailed than the reference summary. Furthermore, we found that our prompt design resulted in the LLM including section headers with every generated summary (see [PITH_FULL_…

**Figure 7.** Figure 7: Example of an instance in the test set where the generated summary is longer than and very similar to the original report. such a patient would require close monitoring and treatment adjustments based on laboratory and clinical responses.” [PITH_FULL_IMAGE:figures/ful…

**Figure 8.** Figure 8: An example of starting prompt A. MaLei Team Online Resources The sources for the MaLei Team at MultiClinSUM shared task 2025 will be available via • GitHub https://github.com/Libo-Ren/MultiClinSum, • Our earlier PerAnSumm page from Manchester Bees https://github.com/pa…

**Figure 9.** Figure 9: Instruction for Prompt Refinement [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

19 extracted references · 17 canonical work pages

[1]

A. M. Stiggelbout, A. H. Pieterse, J. C. De Haes, Shared decision making: concepts, evidence, and practice, Patient education and counseling 98 (2015) 1172–1179

work page 2015
[2]

Stiggelbout, I

A. Stiggelbout, I. Griffioen, J. Brands, M. Melles, J. Rietjens, M. Kunneman, M. Van Der Kolk, C. Van Eijck, D. Snelders, Metro mapping: development of an innovative methodology to co- design care paths to support shared decision making in oncology, BMJ evidence-based medicine 28 (2023) 291–294

work page 2023
[3]

Romero, L

P. Romero, L. Ren, L. Han, G. Nenadic, The manchester bees at peranssumm 2025: Iterative self-prompting with claude and o1 for perspective-aware healthcare answer summarisation, in: Proceedings of the Second Workshop on Patient-Oriented Language Processing (CL4Health), 2025, pp. 340–348

work page 2025
[4]

Naumann, A

T. Naumann, A. Ben Abacha, S. Bethard, K. Roberts, D. Bitterman (Eds.), Proceedings of the 6th Clinical Natural Language Processing Workshop, Association for Computational Linguistics, Mexico City, Mexico, 2024. URL: https://aclanthology.org/2024.clinicalnlp-1.0/

work page 2024
[5]

Collier, P

N. Collier, P. Ruch, A. Nazarenko (Eds.), Proceedings of the International Joint Workshop on Natural Language Processing in Biomedicine and its Applications (NLPBA/BioNLP), COLING, Geneva, Switzerland, 2004. URL: https://aclanthology.org/W04-1200/

work page 2004
[6]

Demner-Fushman, S

D. Demner-Fushman, S. Ananiadou, M. Miwa, K. Roberts, J. Tsujii (Eds.), Proceedings of the 23rd Workshop on Biomedical Natural Language Processing, Association for Computational Linguistics, Bangkok, Thailand, 2024. URL: https://aclanthology.org/2024.bionlp-1.0/

work page 2024
[7]

L. Han, S. Gladkoff, G. Erofeev, I. Sorokina, B. Galiano, G. Nenadic, Neural machine translation of clinical text: an empirical investigation into multilingual pre-trained language models and transfer-learning, Frontiers in Digital Health 6 (2024) 1211564

work page 2024
[8]

Z. Ling, Z. Li, P. Romero, L. Han, G. Nenadic, Malei at the plaba track of trec 2024: Roberta for term replacement–llama3. 1 and gpt-4o for complete abstract adaptation, arXiv preprint arXiv:2411.07381 (2024)

arXiv 2024

Show all 19 references

[9]

Z. Li, S. Belkadi, N. Micheletti, L. Han, M. Shardlow, G. Nenadic, Investigating large language models and control mechanisms to improve text readability of biomedical abstracts, in: 2024 IEEE 12th International Conference on Healthcare Informatics (ICHI), IEEE, 2024, pp. 265–274

2024
[10]

Belkadi, L

S. Belkadi, L. Han, Y. Wu, G. Nenadic, Exploring the value of pre-trained language models for clinical named entity recognition, in: 2023 IEEE International Conference on Big Data (BigData), IEEE, 2023, pp. 3660–3669

2023
[11]

H. Tu, L. Han, G. Nenadic, Extraction of medication and temporal relation from clinical text using neural language models, in: 2023 IEEE International Conference on Big Data (BigData), IEEE, 2023, pp. 2735–2744

2023
[12]

Y. Cui, L. Han, G. Nenadic, Medtem2. 0: Prompt-based temporal classification of treatment events from discharge summaries, in: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop), 2023, pp. 160–183

2023
[13]

Romero, L

P. Romero, L. Han, G. Nenadic, Medication extraction and entity linking using stacked and voted ensembles on llms, in: Proceedings of the Second Workshop on Patient-Oriented Language Processing (CL4Health), 2025, pp. 303–315

2025
[14]

L. Ren, S. Belkadi, L. Han, W. Del-Pinto, G. Nenadic, Synthetic4health: Generating annotated synthetic clinical letters, Frontiers in Digital Health 7 (2025) 1497130

2025
[15]

L. Ren, S. Belkadi, L. Han, W. Del-Pinto, G. Nenadic, Beyond reconstruction: generating privacy- preserving clinical letters, in: Proceedings of the Sixth Workshop on Privacy in Natural Language Processing, 2025, pp. 60–74

2025
[16]

A. Paul, D. Shaji, L. Han, W. Del-Pinto, G. Nenadic, Deidclinic: A multi-layered framework for de-identification of clinical free-text data, arXiv preprint arXiv:2410.01648 (2024)

2024 arXiv
[17]

Agarwal, M

S. Agarwal, M. S. Akhtar, S. Yadav, Overview of the PerAnsSumm 2025 shared task on perspective- aware healthcare answer summarization, in: S. Ananiadou, D. Demner-Fushman, D. Gupta, P. Thompson (Eds.), Proceedings of the Second Workshop on Patient-Oriented Language Processing ...

2025
[18]

Rodríguez-Ortega, E

M. Rodríguez-Ortega, E. Rodríguez-Lopez, S. Lima-López, C. Escolano, M. Melero, L. Pratesi, L. Vigil-Gimenez, L. Fernandez, E. Farré-Maduell, M. Krallinger, Overview of MultiClinSum task at BioASQ 2025: evaluation of clinical case summarization strategies for multiple language...

2025
[455]

doi:10.18653/v1/2025.cl4health-1

URL: https://aclanthology.org/2025.cl4health-1.41/. doi:10.18653/v1/2025.cl4health-1. 41

2025 doi

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

A. M. Stiggelbout, A. H. Pieterse, J. C. De Haes, Shared decision making: concepts, evidence, and practice, Patient education and counseling 98 (2015) 1172–1179

work page 2015

[2] [2]

Stiggelbout, I

A. Stiggelbout, I. Griffioen, J. Brands, M. Melles, J. Rietjens, M. Kunneman, M. Van Der Kolk, C. Van Eijck, D. Snelders, Metro mapping: development of an innovative methodology to co- design care paths to support shared decision making in oncology, BMJ evidence-based medicine 28 (2023) 291–294

work page 2023

[3] [3]

Romero, L

P. Romero, L. Ren, L. Han, G. Nenadic, The manchester bees at peranssumm 2025: Iterative self-prompting with claude and o1 for perspective-aware healthcare answer summarisation, in: Proceedings of the Second Workshop on Patient-Oriented Language Processing (CL4Health), 2025, pp. 340–348

work page 2025

[4] [4]

Naumann, A

T. Naumann, A. Ben Abacha, S. Bethard, K. Roberts, D. Bitterman (Eds.), Proceedings of the 6th Clinical Natural Language Processing Workshop, Association for Computational Linguistics, Mexico City, Mexico, 2024. URL: https://aclanthology.org/2024.clinicalnlp-1.0/

work page 2024

[5] [5]

Collier, P

N. Collier, P. Ruch, A. Nazarenko (Eds.), Proceedings of the International Joint Workshop on Natural Language Processing in Biomedicine and its Applications (NLPBA/BioNLP), COLING, Geneva, Switzerland, 2004. URL: https://aclanthology.org/W04-1200/

work page 2004

[6] [6]

Demner-Fushman, S

D. Demner-Fushman, S. Ananiadou, M. Miwa, K. Roberts, J. Tsujii (Eds.), Proceedings of the 23rd Workshop on Biomedical Natural Language Processing, Association for Computational Linguistics, Bangkok, Thailand, 2024. URL: https://aclanthology.org/2024.bionlp-1.0/

work page 2024

[7] [7]

L. Han, S. Gladkoff, G. Erofeev, I. Sorokina, B. Galiano, G. Nenadic, Neural machine translation of clinical text: an empirical investigation into multilingual pre-trained language models and transfer-learning, Frontiers in Digital Health 6 (2024) 1211564

work page 2024

[8] [8]

Z. Ling, Z. Li, P. Romero, L. Han, G. Nenadic, Malei at the plaba track of trec 2024: Roberta for term replacement–llama3. 1 and gpt-4o for complete abstract adaptation, arXiv preprint arXiv:2411.07381 (2024)

arXiv 2024