Using LLMs for Multilingual Clinical Entity Linking to ICD-10

Ivan Koychev; Svetla Boytcheva; Sylvia Vassileva

REVIEW 4 major objections 4 minor 18 references

Using LLMs for Multilingual Clinical Entity Linking to ICD-10

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A two-stage pipeline—an ICD-10 dictionary for exact matches plus GPT-4.1 one-shot in-context learning—assigns ICD-10 codes to clinical mentions in Spanish and Greek, reaching 0.89 F1 at the category level without any task-specific fine-tuni

desk verdict A clearly described dictionary-plus-GPT-4.1 pipeline with new Spanish and Greek benchmark numbers, but the 'outperforms most supervised models' claim rests on a metric that is never defined and comparisons to systems doing a different task. read the letter →

arxiv 2509.04868 v1 pith:7HL2YOPW submitted 2025-09-05 cs.CL

Sylvia Vassileva , Ivan Koychev , Svetla Boytcheva This is my paper

classification cs.CL

keywords clinicalentitylinkingICD-10codinglargelanguagemodelsin-contextlearningmultilingualNLPdictionary-basedmatchingCodiEspElCardioCC

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

The paper sets out to show that ICD-10 entity linking—assigning the right disease code to a clinical phrase—can be done without any task-specific training data. Its system pairs a language-specific dictionary, which resolves unambiguous terms exactly, with GPT-4.1 in one-shot in-context learning, which codes the leftover mentions using a single example discharge summary. On two benchmark corpora, Greek cardiology letters and Spanish discharge summaries, the pair reaches 0.85 F1 (Greek), 0.89 F1 on Spanish ICD-10 categories, and 0.78 F1 on Spanish subcategories. The authors further claim this outperforms most supervised systems reported for the same benchmarks, a result that would matter because supervised medical coding requires expensive expert-annotated corpora that rarely transfer across languages or hospitals.

What carries the argument

The carrying mechanism is the two-stage cascade itself. Stage one is exact dictionary lookup: mentions are matched against language-specific term-to-code dictionaries (about 88K Spanish terms mapped to up to 4-character ICD-10 codes, and about 11.5K Greek terms mapped to 3-character codes), compiled from official ICD-10 translations and training-set vocabulary; an unambiguous match returns its code immediately. Stage two is one-shot in-context learning: the remaining mentions, marked with asterisks in the discharge summary, are sent to GPT-4.1 with a JSON-output prompt and one LLM-generated example summary in the same language; the model returns a JSON array of term, code, and explanation tr

What would settle it

Take a published supervised CodiEsp system, run it under the paper's exact setting—gold mention spans given, one code per mention, the same code-level F1 metric—and compare; if its score rises above 0.89 when the mention-detection burden is removed, the outperformance claim collapses. A cheaper check: inspect the raw per-mention predictions to confirm every gold mention received exactly one code, because every row of Table 1 reports identical precision, recall, and F1, which only holds when the number of predictions equals the number of gold mentions.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that a fully unsupervised cascade—exact dictionary matching first, GPT-4.1 one-shot in-context learning second—assigns ICD-10 codes to clinical mentions with an F1 of 0.85 on the Greek ElCardioCC corpus and 0.89/0.78 on the Spanish CodiEsp categories and subcategories, and that these numbers place it above most supervised systems trained on the same benchmarks. The dictionary stage provides precision for unambiguous terms; the LLM stage provides recall and context sensitivity for everything else, including paraphrases and out-of-vocabulary concepts. The authors deliberately frame the task as entity linking only: gold mention spans are given, so

Load-bearing premise

The claim of beating most supervised systems assumes those systems were evaluated under the same rules—gold mention spans supplied and the same F1 metric—so their published scores, which mostly come from end-to-end coding with different metrics, are directly comparable to Table 1.

Editorial extensions

If this is right

ICD-10 coding support for a new language reduces to obtaining a term-to-code dictionary in that language: no annotation or fine-tuning is required on the clinical side.
The strong result in the mention-given setting shifts the bottleneck to mention detection, which the paper explicitly leaves as future work; end-to-end systems will inherit NER errors on top of linking errors.
Model generation matters more than prompt effort: GPT-4.1's balanced precision and recall versus GPT-4o's low recall suggests newer versions of general-purpose LLMs, not more elaborate prompting, drive the gains.
On the same benchmarks, the pipeline is competitive with fine-tuned BERT, CRF, and knowledge-graph systems, implying that a deterministic guardrail plus a generalist LLM is a viable substitute for supervised clinical coding models.

Reading between the lines

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

Every row in the results table reports identical precision, recall, and F1, which only occurs when the number of returned predictions equals the number of gold mentions; auditing whether every gold mention receives exactly one prediction would directly test whether the headline numbers are per-mention F1 or a coarser aggregate.
A natural extension the authors do not run: apply the same dictionary-plus-LLM cascade to a third language with a small ICD-10 dictionary, such as French or German clinical notes, to see whether the 0.85-0.89 F1 band is a property of the two-stage design or depends on dictionary size and LLM fluency per language.
Because the dictionary baseline alone scores 0.657 (Greek) and 0.546 (Spanish), the LLM is doing most of the coding work; the approach's practical ceiling is therefore set by LLM reliability on rare, context-heavy codes, not by dictionary coverage.

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

Quick take: this paper is a practical pipeline paper, not a methodological breakthrough. The idea — ICD-10 dictionary exact-match first, then GPT-4.1 in-context learning for the leftovers — is sensible, and the authors report strong F1 numbers on CodiEsp (Spanish) and ElCardioCC (Greek). The strengths are real: the pipeline is described well enough to re-implement, the prompts are in the appendix, the limitation section is candid about Azure-only deployment and inference cost, and the no-fine-tuning angle is useful for low-resource clinical NLP.

The soft spot is the evaluation, and it's load-bearing. The paper never defines precision, recall, or F1. Many rows in Table 1 have identical P and R (e.g., 0.657/0.657/0.657, 0.891/0.891/0.891). That identity in standard entity linking only happens if the system returns exactly one prediction per gold mention. For a dictionary-plus-LLM system where GPT-4o alone recalls only 0.462, the combined row should not have P=R unless the authors are scoring accuracy over gold mentions and calling it F1. That isn't necessarily wrong, but it has to be stated.

More importantly, the central claim — "outperforms most of the supervised models for the same benchmark datasets" — is not supported by the tables. The cited supervised results from Section 2 are from end-to-end coding systems that include mention detection and are evaluated under official protocols with MAP or document-level micro/macro F1. Table 1 does not run any supervised baseline under the same gold-mention protocol. So even if the F1 numbers are accurate, they measure a different task than the baselines. The comparison is apples-to-oranges.

I'd also note the GitHub link is promised but not actually present in the arXiv version, and there are no error bars. These are minor by comparison.

Overall: the engineering is plausible and the numbers are probably real in the narrow sense of linking given mentions. But the paper's main conclusion overreaches. It needs a careful evaluation section with a defined metric and a same-task supervised baseline. That is fixable.

Who is this for? People working on multilingual clinical coding, especially low-resource settings, will find the pipeline a useful data point. As submitted, I wouldn't quote the F1 scores in a paper without re-running the evaluation. But I'd send it to peer review: the topic is relevant, the approach is clearly presented, and the evaluation issues are addressable.

Recommendation: major revision with a re-analysis of the metric and a fair baseline.

Referee Report

4 major / 4 minor

Summary. The paper proposes a multistage, LLM-based pipeline for multilingual clinical entity linking to ICD-10. The system first matches mentions against language-specific ICD-10 dictionaries compiled from ICD-10 specifications and from CodiEsp/ElCardioCC training-set terms, then uses GPT-4.1 (and GPT-4o for comparison) in a zero- or one-shot in-context learning setup to assign codes to unmatched mentions. Experiments on the CodiEsp Spanish diagnoses subset and the ElCardioCC Greek dataset report F1 scores of 0.89 (Spanish categories), 0.78 (Spanish subcategories), and 0.85 (Greek categories), with the combined dictionary plus GPT-4.1 one-shot system giving the best results. The central claim is that this unsupervised, no-fine-tuning approach outperforms most supervised models on the same benchmarks.

Significance. If the reported scores are valid and the comparison to prior work is fair, the paper would make a useful practical contribution: it would show that multilingual ICD-10 entity linking can be performed without task-specific fine-tuning, using only a dictionary plus an API-based LLM. The pipeline is clearly described, the prompt is included, and the code is promised to be available. However, the significance is currently tempered by three issues: (1) the evaluation metric is never defined and Table 1 contains rows where precision equals recall, suggesting an unconventional scoring protocol; (2) no supervised baseline is run under the same gold-mention entity-linking protocol, so the headline 'outperforms most supervised models' is not supported by the evidence presented; and (3) the claimed 'fully unsupervised' status is contradicted by the use of training-set terms in the dictionaries.

major comments (4)

[§5, Table 1] The paper never defines how precision, recall, and F1 are computed for the entity-linking task. This is not a minor omission: many rows in Table 1 report identical values for P, R, and F1 (e.g., Dict 0.657/0.657/0.657, Dict+GPT-4.1 1-shot 0.891/0.891/0.891 on Spanish categories). In standard entity linking, P=R holds only if the number of system-produced predictions equals the number of gold mentions. The table does not state whether the system is required to output exactly one code per gold mention, whether missing predictions are counted as false negatives without corresponding false positives, or how extra predictions are penalized. If the score is effectively accuracy over gold mentions, that is a legitimate metric but it must be stated as such, and it is not directly comparable to precision/recall/F1 reported by prior work. The F1 values are the basis for every claim in the paper, s
[§1 and §2] The claim that the system 'outperforms most of the supervised models for the same benchmark datasets' is not supported by the evidence in Table 1. Section 2 cites prior results that use different protocols and metrics: MAP for López-García et al. and Costa et al., macro-F1 for Boyle et al. and Puts et al., micro-F1 for Li et al., and several systems perform end-to-end coding with mention detection rather than gold-mention entity linking. No supervised baseline is run under the same gold-mention protocol as Table 1. Without a same-protocol comparison, the stated superiority over supervised models is an inference, not a result. The authors should either add same-protocol supervised baselines (even a simple fine-tuned encoder model) or restrict the claim to 'competitive with reported results' and explicitly note the protocol differences.
[§3 and §6] The paper repeatedly calls the approach 'fully unsupervised' (Section 1) and says it 'does not require specific training data' (Section 6). However, Section 3 states that the dictionaries are compiled using 'CodeEsp train set terms' and 'ElCardioCC train set terms'. This means labeled training data is used to build the dictionary, albeit not for gradient-based training. The claim as written is misleading and should be qualified: the approach requires a dictionary derived from ICD-10 translations and, optionally, from training-set terms; if the train-set terms are removed, the reported scores may change, and the paper should report the sensitivity to this choice.
[§5, Experiments] The experiments use temperature 0.5 for the LLM, yet no variance or repeated-run statistics are reported. With stochastic decoding, the reported F1 differences (e.g., 0.8906 vs. 0.8902 on Spanish categories) are within the range of run-to-run variability. In addition, prompt design, temperature, and the one-shot example appear to be selected using the test benchmarks themselves. This should be acknowledged as a limitation, and ideally the authors should report multiple runs or use a development set for prompt selection. This is not the central flaw, but it affects the reliability of the numerical comparisons in Table 1.

minor comments (4)

[Table 1] Formatting issues: the row 'GPT-4o 1-shot' has values '0.5966 0.597 0.597' with inconsistent decimal places, and several rows have misaligned columns. Please clean up the table so each column is unambiguous.
[Appendix A] The one-shot example is said to be 'generated using GPT-4o' but no details are given about how it was selected, whether it is the same for all test instances, and whether it came from the training set. Please provide the example or a summary of its provenance.
[Footnote 4] The GitHub link for code and prompts is mentioned but the URL is not included in the footnote. Either provide the link or remove the reference until the code is released.
[Table 2] The Spanish phrase 'onefroma' appears to be a typo (likely 'one forma' or a misspelling). Please check the excerpt and the expected/predicted codes.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: dictionary/LLM pipeline is evaluated on external benchmarks with no self-citation or fitted-input reduction.

full rationale

The paper's derivation chain is empirical: a dictionary exact-match stage plus GPT-4.1 in-context generation, evaluated on the external CodiEsp and ElCardioCC benchmarks. No step defines or forces its target result by construction. The dictionaries are compiled from ICD-10 translations in Spanish/Greek and from the train splits of CodiEsp and ElCardioCC; using train-split terms to build a dictionary is supervised signal, not circularity, because the reported results are on held-out test data and the LLM predictions are independently generated. No parameter is fitted to the test labels and then renamed as a prediction; prompt and temperature choices are ordinary model-selection decisions, not fitted inputs called predictions. The paper contains no self-citations, so there is no load-bearing self-citation chain and no imported uniqueness theorem. The main weakness—comparing the headline F1 to prior supervised baselines that perform end-to-end coding with mention detection and different metrics—is a validity and comparability concern, not circularity: the reported numbers do not reduce by construction to the paper's own inputs. The identical P/R/F1 entries in Table 1 suggest an accuracy-style evaluation over gold mentions, but that is a metric-definition issue, not a self-referential derivation. Accordingly, no circular steps are present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The system introduces no new scientific entities. Its load-bearing inputs are domain assumptions about benchmark labels, mention availability, dictionary quality, and LLM output parseability, plus a small set of hand-chosen generation hyperparameters.

free parameters (3)

LLM temperature = 0.5
Chosen by hand for LLM generation; affects output determinism and recall. Section 5.
max_tokens = 6000
Set to accommodate discharge summary lengths; may limit the number of codes generated in one call. Section 5.
one-shot in-context example = one GPT-4o-generated discharge summary per language
The single in-context example is selected by the authors and can materially change LLM outputs. Section 4, Appendix A.

assumptions (4)

domain assumption The gold ICD-10 annotations in CodiEsp and ElCardioCC are correct and follow the official encoding guidelines.
Evaluation uses benchmark labels as ground truth; any noise in these labels transfers to reported F1. Section 3.
domain assumption Entity mentions are already given; the system is not responsible for detecting spans.
Section 4 states mention detection is future work, so reported scores cover only linking. Section 5, Conclusion.
domain assumption The compiled Spanish and Greek dictionaries are accurate and contain all unambiguous clinical terms expected in the test data.
Dictionary is built from ICD-10 specifications plus train-set terms, Sections 3 and 4; no validation of dictionary coverage or errors is reported.
domain assumption GPT-4.1 responses are parsed reliably; failed parses do not systematically bias precision or recall.
Evaluation assumes the JSON from the model covers every mention; Section 5 discusses code retrieval issues but no parse failure statistics are given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Using LLMs for Multilingual Clinical Entity Linking to ICD-10." pith.science (2026). https://pith.science/paper/7HL2YOPW

@misc{pith2026250904868,
  author       = {Pith},
  title        = {Pith review of: Using LLMs for Multilingual Clinical Entity Linking to ICD-10},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7HL2YOPW}},
  note         = {Machine review of arXiv:2509.04868}
}

read the original abstract

The linking of clinical entities is a crucial part of extracting structured information from clinical texts. It is the process of assigning a code from a medical ontology or classification to a phrase in the text. The International Classification of Diseases - 10th revision (ICD-10) is an international standard for classifying diseases for statistical and insurance purposes. Automatically assigning the correct ICD-10 code to terms in discharge summaries will simplify the work of healthcare professionals and ensure consistent coding in hospitals. Our paper proposes an approach for linking clinical terms to ICD-10 codes in different languages using Large Language Models (LLMs). The approach consists of a multistage pipeline that uses clinical dictionaries to match unambiguous terms in the text and then applies in-context learning with GPT-4.1 to predict the ICD-10 code for the terms that do not match the dictionary. Our system shows promising results in predicting ICD-10 codes on different benchmark datasets in Spanish - 0.89 F1 for categories and 0.78 F1 on subcategories on CodiEsp, and Greek - 0.85 F1 on ElCardioCC.

Figures

Figures reproduced from arXiv: 2509.04868 by the authors.

**Figure 2.** Example prompt for ICD-10 coding in Greek. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 16 canonical work pages

[1]

Joseph S Boyle, Antanas Kascenas, Pat Lok, Maria Liakata, and Alison Q O'Neil. 2023. Automated clinical coding using off-the-shelf large language models. arXiv preprint arXiv:2310.06552

arXiv 2023
[2]

Joao Costa, In \^e s Lopes, Andr \'e V Carreiro, David Ribeiro, and Carlos Soares. 2020. https://ceur-ws.org/Vol-2696/paper_187.pdf Fraunhofer aicos at clef ehealth 2020 task 1: Clinical code extraction from textual data using fine-tuned bert models. In CLEF (Working Notes)

work page 2020
[3]

Dimitriadis, V

D. Dimitriadis, V. Patsiou, E. Stoikopoulou, A. Toumpas, A. Kipouros, D. Papadopoulos, A. Bekiaridou, K. Barmpagiannos, A. Vasilopoulou, A. Barmpagiannos, A. Samaras, G. Giannakoulas, and G. Tsoumakas. 2025. Overview of ElCardioCC Task on Clinical Coding in Cardiology at BioASQ 2025 . In CLEF 2025 Working Notes

work page 2025
[4]

Nuria Garc \' a-Santa, Kendrick Cetina, L Cappellato, C Eickhoff, N Ferro, and A Nev \'e ol. 2020. https://ceur-ws.org/Vol-2696/paper_111.pdf Fle at clef ehealth 2020: Text mining and semantic knowledge for automated clinical encoding. In CLEF (Working Notes)

work page 2020
[5]

Sch \" a fer Henning and Friedrich C. 2020. https://ceur-ws.org/Vol-2696/paper_212.pdf Multilingual ICD -10 Code Assignment with Transformer Architectures using MIMIC - III Discharge Summaries . Conference and Labs of the Evaluation Forum

work page 2020
[6]

Katarina Karlin and Diana Amin. 2025. https://www.diva-portal.org/smash/get/diva2:1959279/FULLTEXT01.pdf From prompt to icd-10: Evaluating prompt-based decoder large language models for icd-10 coding using a swedish clinical dataset

work page 2025
[7]

Rumeng Li, Xun Wang, and Hong Yu. 2024. Exploring llm multi-agents for icd coding. arXiv preprint arXiv:2406.15363

arXiv 2024
[8]

Guillermo L \'o pez-Garc \' a, Jos \'e M Jerez, and Francisco J Veredas. 2020. https://ceur-ws.org/Vol-2696/paper_101.pdf Icb-uma at clef e-health 2020 task 1: Automatic icd-10 coding in spanish with bert . In CLEF (Working Notes)

work page 2020

Show all 18 references

[9]

Ola Maatouk. 2025. https://www.diva-portal.org/smash/get/diva2:1933618/FULLTEXT02.pdf Leveraging llms for icd coding and uncertainty estimation: Can the model's awareness of the hierarchical structureof icd-10 codes impact its prediction performance?

2025
[10]

Zulfat Miftahutdinov and Elena Tutubalina. 2018. Deep learning for icd coding: Looking for medical concepts in clinical documents in english and in french. In Experimental IR Meets Multilinguality, Multimodality, and Interaction: 9th International Conference of the CLEF Associ...

2018
[11]

Antonio Miranda-Escalada, Aitor Gonzalez-Agirre, Jordi Armengol-Estapé, and Martin Krallinger. 2020. Overview of automatic clinical coding: annotations, guidelines, and solutions for non-english clinical cases at codiesp track of CLEF eHealth 2020. In Working Notes of Conferen...

2020
[12]

Akram Mustafa, Usman Naseem, and Mostafa Rahimi Azghadi. 2025. Large language models vs human for classifying clinical documents. International Journal of Medical Informatics, page 105800

2025
[13]

Rudransh Pathak, Gabriel Vald, Yusuf Sermet, and Ibrahim Demir. 2024. Utilizing large language models to predict icd-10 diagnosis codes from patient medical records. In 2024 IEEE MIT Undergraduate Research Technology Conference (URTC), pages 1--5. IEEE

2024
[14]

Sander Puts, Catharina ML Zegers, Andre Dekker, and I \ n igo Bermejo. 2025. Developing an icd-10 coding assistant: Pilot study using roberta and gpt-4 for term extraction and description-based code selection. JMIR Formative Research, 9:e60095

2025
[15]

Jurica Seva, Madeleine Kittner, Roland Roller, and Ulf Leser. 2017. https://ceur-ws.org/Vol-1866/paper_70.pdf Multi-lingual icd-10 coding using a hybrid rule-based and supervised classification approach at clef ehealth 2017. In CLEF (Working Notes)

2017
[16]

Hsiao S., and Tzong-Han Tsai Richard

Chen Sheng-Wei, Lai Po-Ting, Tsai Yi-Lin, Kuan-Chieh Chung Jay, S. Hsiao S., and Tzong-Han Tsai Richard. 2014. https://citeseerx.ist.psu.edu/document?repid=rep1&type=pdf&doi=0b5662e8ed5af98c2cad9217ee6eb4a14fd63659 Ncu IISR System for NTCIR -11 MedNLP -2 Task . NTCIR Conferenc...

2014
[17]

Ali Soroush, Benjamin S Glicksberg, Eyal Zimlichman, Yiftach Barash, Robert Freeman, Alexander W Charney, Girish N Nadkarni, and Eyal Klang. 2023. Assessing gpt-3.5 and gpt-4 in generating international classification of diseases billing codes. medRxiv, pages 2023--07

2023
[18]

Pierre Zweigenbaum and Thomas Lavergne. 2016. https://doi.org/10.18653/v1/w16-6113 Hybrid methods for ICD -10 coding of death certificates . In Proceedings of the Seventh International Workshop on Health Text Mining and Information Analysis , pages 96--105. Association for Com...

2016 doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Joseph S Boyle, Antanas Kascenas, Pat Lok, Maria Liakata, and Alison Q O'Neil. 2023. Automated clinical coding using off-the-shelf large language models. arXiv preprint arXiv:2310.06552

arXiv 2023

[2] [2]

Joao Costa, In \^e s Lopes, Andr \'e V Carreiro, David Ribeiro, and Carlos Soares. 2020. https://ceur-ws.org/Vol-2696/paper_187.pdf Fraunhofer aicos at clef ehealth 2020 task 1: Clinical code extraction from textual data using fine-tuned bert models. In CLEF (Working Notes)

work page 2020

[3] [3]

Dimitriadis, V

D. Dimitriadis, V. Patsiou, E. Stoikopoulou, A. Toumpas, A. Kipouros, D. Papadopoulos, A. Bekiaridou, K. Barmpagiannos, A. Vasilopoulou, A. Barmpagiannos, A. Samaras, G. Giannakoulas, and G. Tsoumakas. 2025. Overview of ElCardioCC Task on Clinical Coding in Cardiology at BioASQ 2025 . In CLEF 2025 Working Notes

work page 2025

[4] [4]

Nuria Garc \' a-Santa, Kendrick Cetina, L Cappellato, C Eickhoff, N Ferro, and A Nev \'e ol. 2020. https://ceur-ws.org/Vol-2696/paper_111.pdf Fle at clef ehealth 2020: Text mining and semantic knowledge for automated clinical encoding. In CLEF (Working Notes)

work page 2020

[5] [5]

Sch \" a fer Henning and Friedrich C. 2020. https://ceur-ws.org/Vol-2696/paper_212.pdf Multilingual ICD -10 Code Assignment with Transformer Architectures using MIMIC - III Discharge Summaries . Conference and Labs of the Evaluation Forum

work page 2020

[6] [6]

Katarina Karlin and Diana Amin. 2025. https://www.diva-portal.org/smash/get/diva2:1959279/FULLTEXT01.pdf From prompt to icd-10: Evaluating prompt-based decoder large language models for icd-10 coding using a swedish clinical dataset

work page 2025

[7] [7]

Rumeng Li, Xun Wang, and Hong Yu. 2024. Exploring llm multi-agents for icd coding. arXiv preprint arXiv:2406.15363

arXiv 2024

[8] [8]

Guillermo L \'o pez-Garc \' a, Jos \'e M Jerez, and Francisco J Veredas. 2020. https://ceur-ws.org/Vol-2696/paper_101.pdf Icb-uma at clef e-health 2020 task 1: Automatic icd-10 coding in spanish with bert . In CLEF (Working Notes)

work page 2020