{"doi":"10.1093/jamia/ocae312","title":"A dataset and benchmark for hospital course summarization with adapted large language models","abstract":"OBJECTIVE: Brief hospital course (BHC) summaries are clinical documents that summarize a patient's hospital stay. While large language models (LLMs) depict remarkable capabilities in automating real-world tasks, their capabilities for healthcare applications such as synthesizing BHCs from clinical notes have not been shown. We introduce a novel preprocessed dataset, the MIMIC-IV-BHC, encapsulating clinical note and BHC pairs to adapt LLMs for BHC synthesis. Furthermore, we introduce a benchmark of the summarization performance of 2 general-purpose LLMs and 3 healthcare-adapted LLMs. MATERIALS AND METHODS: Using clinical notes as input, we apply prompting-based (using in-context learning) and fine-tuning-based adaptation strategies to 3 open-source LLMs (Clinical-T5-Large, Llama2-13B, and FLAN-UL2) and 2 proprietary LLMs (Generative Pre-trained Transformer [GPT]-3.5 and GPT-4). We evaluate these LLMs across multiple context-length inputs using natural language similarity metrics. We further conduct a clinical study with 5 clinicians, comparing clinician-written and LLM-generated BHCs across 30 samples, focusing on their potential to enhance clinical decision-making through improved summary quality. We compare reader preferences for the original and LLM-generated summary using Wilcoxon signed-rank tests. We further request optional qualitative feedback from clinicians to gain deeper insights into their preferences, and we present the frequency of common themes arising from these comments. RESULTS: The Llama2-13B fine-tuned LLM outperforms other domain-adapted models given quantitative evaluation metrics of Bilingual Evaluation Understudy (BLEU) and Bidirectional Encoder Representations from Transformers (BERT)-Score. GPT-4 with in-context learning shows more robustness to increasing context lengths of clinical note inputs than fine-tuned Llama2-13B. Despite comparable quantitative metrics, the reader study depicts a significant preference for summaries generated by GPT-4 with in-context learning compared to both Llama2-13B fine-tuned summaries and the original summaries (P<.001), highlighting the need for qualitative clinical evaluation. DISCUSSION AND CONCLUSION: We release a foundational clinically relevant dataset, the MIMIC-IV-BHC, and present an open-source benchmark of LLM performance in BHC synthesis from clinical notes. We observe high-quality summarization performance for both in-context proprietary and fine-tuned open-source LLMs using both quantitative metrics and a qualitative clinical reader study. Our research effectively integrates elements from the data assimilation pipeline: our methods use (1) clinical data sources to integrate, (2) data translation, and (3) knowledge creation, while our evaluation strategy paves the way for (4) deployment.","journal":"Journal of the American Medical Informatics Association","year":2024,"id":421297,"datarank":1.0804241508214265,"base_score":3.4339872044851463,"endowment":3.4339872044851463,"self_citation_contribution":0.515098080672772,"citation_network_contribution":0.5653260701486545,"self_endowment_contribution":0.515098080672772,"citer_contribution":0.5653260701486545,"corpus_percentile":81.24854954745881,"corpus_rank":2425,"citation_count":30,"citer_count":29,"citers_with_citation_signal":17,"citers_with_endowment":17,"datacite_reuse_total":0,"is_dataset":true,"is_dataset_confidence":0.9154,"is_data_producer":false,"deposit_databanks":null,"is_oa":true,"file_count":0,"downloads":0,"has_version_chain":false,"published_date":"2024-01-01","fair_score":70.8333,"fair_percentile":91.99021705900336,"algorithm_id":"datarank_citation_only_1hop_v6","ranking_scope":"data_only","authors":[{"id":1159187,"name":"Dave Van Veen","orcid":"0000-0001-9312-1773","position":1,"is_corresponding":false},{"id":802918,"name":"Yamin Arefeen","orcid":"0000-0003-2194-1945","position":2,"is_corresponding":false},{"id":293181,"name":"Jason Hom","orcid":"0000-0002-8404-5796","position":3,"is_corresponding":false},{"id":1213455,"name":"Christian Blüthgen","orcid":"0000-0001-7321-5676","position":4,"is_corresponding":false},{"id":983219,"name":"Eduardo Pontes Reis","orcid":"0000-0001-5110-457X","position":5,"is_corresponding":false},{"id":1047261,"name":"Sergios Gatidis","orcid":"0000-0002-6928-4967","position":6,"is_corresponding":false},{"id":1213456,"name":"N. Clifford","orcid":"0000-0003-2334-3818","position":7,"is_corresponding":false},{"id":1214096,"name":"Joseph Daws","orcid":null,"position":8,"is_corresponding":false},{"id":1214097,"name":"Arash Saber Tehrani","orcid":null,"position":9,"is_corresponding":false},{"id":485995,"name":"Jangwon Kim","orcid":"0000-0003-0228-3502","position":10,"is_corresponding":false},{"id":301558,"name":"Akshay Chaudhari","orcid":"0000-0002-3667-6796","position":11,"is_corresponding":false},{"id":1159189,"name":"Asad Aali","orcid":"0009-0008-2120-5722","position":0,"is_corresponding":true}],"reference_count":43,"raw_metadata":null,"created_at":"2026-07-19T01:57:36.298171Z","pmid":"39786555","pmcid":"PMC11833472","fwci":null,"citation_percentile":null,"influential_citations":0,"oa_status":null,"license":null,"views":0,"total_file_size_bytes":0,"version_count":0,"fair_f":88.8889,"fair_a":81.25,"fair_i":20.0,"fair_r":41.6667,"fair_zscore":1.4401,"fair_rationale":{"fair_score":70.83,"has_llm":true,"taxonomy_version":"fair_taxonomy_v5","dimensions":{"F":{"name":"Findable","score":88.89,"criteria":[{"key":"f_dataset_pid","label":"Persistent identifier for the data","kind":"llm","weight":2.0,"fraction":1.0,"verdict":"yes","evidence":"Aali A, Van Veen D, Arefeen YI, et al. MIMIC-IV-Ext-BHC: Labeled Clinical Notes Dataset for Hospital Course Summarization. PhysioNet. 2024. https://doi.org/10.13026/fh2q-4148.","grounded":true,"rationale":"The reference list entry for the dataset contains a DOI (10.13026/fh2q-4148), a persistent identifier scheme. [majority verdict 'yes' (3/5 passes agreed)]","anchors":["RDA-F1-01D — FAIR Data Maturity Model: 'Data is identified by a persistent identifier' (priorit","RDA-F1-02D — FAIR Data Maturity Model: 'Data is identified by a globally unique identifier'","FsF-F1-02D — F-UJI/FAIRsFAIR: 'Data is assigned a persistent identifier'"],"scored":true,"signal":null},{"key":"f_repository_named","label":"Named repository","kind":"llm","weight":2.0,"fraction":1.0,"verdict":"yes","evidence":"The MIMIC-IV-BHC dataset, published on PhysioNet [23].","grounded":true,"rationale":"PhysioNet is a named data repository that issues accessions and commits to retention.","anchors":["RDA-F4-01M — FAIR Data Maturity Model: metadata is offered so it can be harvested and indexed (","NIH DMS Policy Element 4 (NOT-OD-21-014) — name the repository where data will be archived","NSTC Desirable Characteristics of Data Repositories (2022) — 'Long-Term Sustainability', 'Reten"],"scored":true,"signal":null},{"key":"f_data_availability_statement","label":"Data-availability statement","kind":"llm","weight":2.0,"fraction":1.0,"verdict":"yes","evidence":"To allow researchers to replicate our adaptation and evaluation approach for BHC summarization, we publicly release: 1. The MIMIC-IV-BHC dataset, published on PhysioNet [23].","grounded":true,"rationale":"The data availability statement points to a repository record (PhysioNet, with a DOI accessible via the reference).","anchors":["Colavizza, Hrynaszkiewicz, Staden, Whitaker & McGillivray (2020), 'The citation advantage of li","Springer Nature research data policy — Data Availability Statements: standard statement templat","RDA-F3-01M — metadata clearly and explicitly includes the identifier of the data it describes"],"scored":false,"signal":null},{"key":"f_discovery_metadata","label":"Description of the dataset as an object","kind":"llm","weight":2.0,"fraction":0.5,"verdict":"partial","evidence":"Table 1. a) A sample from MIMIC-IV-BHC, our novel pre-processed dataset extracted from raw MIMIC-IV notes. ... b) Statistics for the MIMIC-IV-BHC subsets, binned across multiple context length ranges for adaptation tasks.","grounded":false,"rationale":"Table 1 provides an itemised inventory of the dataset, showing a sample and statistics. [downgraded to 'partial' — no verifiable quote from the paper] [majority verdict 'partial' (3/5 passes agreed)]","anchors":["RDA-F2-01M — 'Rich metadata is provided to allow discovery' (priority Essential)","FsF-F2-01M — F-UJI: 'Metadata includes descriptive core elements to support data findability'","FsF-R1-01MD — F-UJI: 'Metadata specifies the content of the data'"],"scored":false,"signal":null},{"key":"f_dataset_cited","label":"Dataset formally cited","kind":"llm","weight":1.0,"fraction":1.0,"verdict":"yes","evidence":"Aali A, Van Veen D, Arefeen YI, et al. MIMIC-IV-Ext-BHC: Labeled Clinical Notes Dataset for Hospital Course Summarization. PhysioNet. 2024. https://doi.org/10.13026/fh2q-4148.","grounded":true,"rationale":"The dataset appears as a reference-list entry in the bibliography. [majority verdict 'yes' (4/5 passes agreed)]","anchors":["FORCE11 Joint Declaration of Data Citation Principles (2014) — data should be cited as a first-","RDA-F3-01M — metadata clearly and explicitly includes the identifier of the data it describes","FsF-F3-01M — F-UJI: 'Metadata includes the identifier of the data it describes'"],"scored":true,"signal":null}]},"A":{"name":"Accessible","score":81.25,"criteria":[{"key":"a_data_openly_accessible","label":"Access route free of preconditions","kind":"llm","weight":2.0,"fraction":1.0,"verdict":"yes","evidence":"To allow researchers to replicate our adaptation and evaluation approach for BHC summarization, we publicly release: 1. The MIMIC-IV-BHC dataset, published on PhysioNet [23].","grounded":true,"rationale":"The text states the data are publicly released on PhysioNet with no precondition mentioned.","anchors":["RDA-A1.1-01D — 'Data is accessible through a free access protocol'","FsF-A1-01M — F-UJI: 'Metadata contains access level and access conditions of the data'","NSTC Desirable Characteristics of Data Repositories (2022) — 'Free and Easy Access'"],"scored":true,"signal":null},{"key":"a_access_conditions_stated","label":"Access level labelled","kind":"llm","weight":1.0,"fraction":1.0,"verdict":"yes","evidence":"To allow researchers to replicate our adaptation and evaluation approach for BHC summarization, we publicly release: 1. The MIMIC-IV-BHC dataset, published on PhysioNet [23].","grounded":true,"rationale":"The paper uses the phrase 'publicly release', which is a natural-language equivalent of 'open access', thus labeling the access level.","anchors":["FsF-A1-01M — F-UJI: 'Metadata contains access level and access conditions of the data'","RDA-A1-01M — metadata contains information to enable the user to get access to the data","COAR Controlled Vocabularies — Access Rights v1.0 (open / embargoed / restricted / metadata-onl"],"scored":false,"signal":null},{"key":"a_controlled_access_for_sensitive","label":"Gatekeeper for sensitive data","kind":"llm","weight":0.5,"fraction":0.0,"verdict":"no","evidence":null,"grounded":false,"rationale":"The dataset is de-identified and publicly released; no gatekeeper is named for this study's own data. [majority verdict 'no' (4/5 passes agreed)]","anchors":["NIH Genomic Data Sharing Policy (NOT-OD-14-124) — controlled-access via a Data Access Committee","RDA-A1.2-01D — 'Data is accessible through an access protocol that supports authentication and ","NIH DMS Policy Element 5 (NOT-OD-21-014) — Access, Distribution, or Reuse Considerations (conse"],"scored":false,"signal":null},{"key":"a_timeline_retention","label":"Availability timing & retention","kind":"llm","weight":0.5,"fraction":0.5,"verdict":"partial","evidence":"To allow researchers to replicate our adaptation and evaluation approach for BHC summarization, we publicly release: 1. The MIMIC-IV-BHC dataset, published on PhysioNet [23].","grounded":true,"rationale":"The paper states the data are available now ('publicly release') but does not commit to how long they will persist. [majority verdict 'partial' (4/5 passes agreed)]","anchors":["NIH DMS Plan Element 4 (NOT-OD-21-014) — Data Preservation, Access, and Associated Timelines","NSTC Desirable Characteristics (2022), Organizational Infrastructure: 'Retention Policy'","RDA-A2-01M — 'Metadata is guaranteed to remain available after data is no longer available'"],"scored":false,"signal":null}]},"I":{"name":"Interoperable","score":20.0,"criteria":[{"key":"i_open_nonproprietary_format","label":"Open file format","kind":"llm","weight":1.0,"fraction":0.0,"verdict":"no","evidence":null,"grounded":false,"rationale":"No file format is named for the released dataset.","anchors":["FsF-R1.3-02D — F-UJI: 'Data is available in a file format recommended by the target research co","RDA-R1.3-02D — data is expressed in a machine-understandable community standard","RDA-I1-01D — data uses a knowledge representation expressed in a standardised format"],"scored":true,"signal":null},{"key":"i_community_standard_vocabulary","label":"Community standard / vocabulary","kind":"llm","weight":1.0,"fraction":0.0,"verdict":"no","evidence":null,"grounded":false,"rationale":"No community standard for data or metadata is named; the paper only describes its own structure.","anchors":["RDA-R1.3-01M — 'Metadata complies with a community standard' (priority Essential)","RDA-R1.3-01D — 'Data complies with a community standard'","RDA-I2-01M — '(Meta)data use vocabularies that follow FAIR principles'"],"scored":false,"signal":null},{"key":"i_qualified_references","label":"Identifiers for the resources the data depend on","kind":"llm","weight":0.5,"fraction":1.0,"verdict":"yes","evidence":"MIMIC-IV-Note: Deidentified free-text clinical notes. PhysioNet. 2023. https://www.physionet.org/content/mimic-iv-note/2.2/.","grounded":true,"rationale":"The paper references the MIMIC-IV-Note dataset with a URL, which is an identifier for an external resource. [majority verdict 'yes' (3/5 passes agreed)]","anchors":["RDA-I3-01M — '(meta)data include references to other (meta)data'","RDA-I3-03M — 'metadata includes qualified references to other metadata'","FsF-I3-01M — F-UJI: 'Metadata includes links between the data and its related entities'"],"scored":false,"signal":null}]},"R":{"name":"Reusable","score":41.67,"criteria":[{"key":"r_reuse_license","label":"Reuse licence","kind":"llm","weight":2.0,"fraction":0.0,"verdict":"no","evidence":null,"grounded":false,"rationale":"No license or terms of use are named for the dataset.","anchors":["RDA-R1.1-01M — 'Metadata includes information about the licence under which the data can be reu","RDA-R1.1-02M — 'Metadata refers to a standard reuse licence'","RDA-R1.1-03M — 'Metadata refers to a machine-understandable reuse licence'"],"scored":true,"signal":null},{"key":"r_provenance_methods","label":"Provenance of the data","kind":"llm","weight":1.0,"fraction":1.0,"verdict":"yes","evidence":"Using regular expressions and natural language toolkit (NLTK) in Python, we apply: whitespace removal, uniform spacing, section identification, header standardization, extraneous symbol removal, and tokenization.","grounded":true,"rationale":"The paper names the specific tools (NLTK, Python) used to produce the dataset.","anchors":["RDA-R1.2-01M — 'Metadata includes provenance information according to community- specific standa","FsF-R1.2-01M — F-UJI: 'Metadata includes provenance information about data creation or generati","W3C PROV-O (W3C Recommendation, 2013) — the entity/activity/agent model of provenance"],"scored":false,"signal":null},{"key":"r_documentation_codebook","label":"Documentation / codebook","kind":"llm","weight":1.0,"fraction":0.0,"verdict":"no","evidence":"Table 1. a) A sample from MIMIC-IV-BHC, our novel pre-processed dataset extracted from raw MIMIC-IV notes. ... b) Statistics for the MIMIC-IV-BHC subsets, binned across multiple context length ranges for adaptation tasks.","grounded":false,"rationale":"The variable and file definitions are provided inside the article (Table 1), not as a separate documentation object shipped with the data. [downgraded to 'no' — no verifiable quote from the paper] [majority verdict 'no' (3/5 passes agreed)]","anchors":["RDA-R1-01M — '(Meta)data are richly described with a plurality of accurate and relevant attribu","FsF-R1-01MD — F-UJI: 'Metadata specifies the content of the data'","NIH DMS Policy Element 3 (NOT-OD-21-014) — Standards (documentation and metadata to accompany t"],"scored":false,"signal":null},{"key":"r_versioning","label":"Snapshot identified","kind":"llm","weight":0.5,"fraction":0.0,"verdict":"no","evidence":null,"grounded":false,"rationale":"Neither a version string nor a specific date is given for the dataset snapshot; the reference year (2024) is insufficient to pin a snapshot. [majority verdict 'no' (4/5 passes agreed)]","anchors":["DataCite Metadata Schema 4.6 — the 'Version' property","RDA-R1.2-01M — provenance information (which version was used is provenance)","NSTC Desirable Characteristics of Data Repositories (2022) — 'Provenance', 'Retention Policy'"],"scored":true,"signal":null},{"key":"x_code_availability","label":"Analysis code available","kind":"llm","weight":1.0,"fraction":1.0,"verdict":"yes","evidence":"The underlying code for data pre-processing, LLM adaptation, and evaluation: github.com/StanfordMIMI/clin-bhc-summ.","grounded":true,"rationale":"A machine-resolvable code repository URL is provided.","anchors":["NIH DMS Policy Element 2 (NOT-OD-21-014) — 'Related Tools, Software and/or Code'","FAIR4RS Principles v1.0 (Chue Hong et al., 2022; RDA/FORCE11/ReSA) — FAIR Principles for Resear","FORCE11 Software Citation Principles (Smith, Katz & Niemeyer, 2016, PeerJ CS 2:e86)"],"scored":true,"signal":null},{"key":"x_funding_attribution","label":"Funder and award number","kind":"llm","weight":0.5,"fraction":1.0,"verdict":"yes","evidence":"A.C. receives research support from NIH grants R01 HL167974, R01 HL169345, R01 AR077604, R01 EB002524, R01 AR079431, and P41 EB027060;","grounded":true,"rationale":"Multiple award numbers are listed for the funding supporting the work.","anchors":["DataCite Metadata Schema 4.6 — 'FundingReference' property (funderName, funderIdentifier, award","Crossref Funder Registry — canonical funder identifiers for funding metadata","RDA-F2-01M — rich metadata provided to allow discovery (funding is part of the descriptive reco"],"scored":true,"signal":null}]}},"actions":[{"key":"r_reuse_license","dimension":"R","label":"Reuse licence","action":"Attach a standard, machine-readable open licence to the deposit — CC0 or CC BY, which is what Horizon Europe and most funders expect — and print the licence identifier in the paper. 'Free to use' is not a licence: it grants nothing a reuser's institution can rely on.","anchors":["yes","partial","no"],"verdict":"no","current":0.0,"evidence":null,"why":"No license or terms of use are named for the dataset.","gain":16.67,"priority":"essential","scored":true},{"key":"i_open_nonproprietary_format","dimension":"I","label":"Open file format","action":"Release the data in an open, community-standard format (CSV/TSV, JSON, HDF5, NetCDF, FASTQ, VCF, NIfTI…) instead of — or alongside — any proprietary or instrument-native format, and name the format in the paper. A dataset that needs a €2,000 licence to open is not reusable.","anchors":["yes","partial","no"],"verdict":"no","current":0.0,"evidence":null,"why":"No file format is named for the released dataset.","gain":8.33,"priority":"important","scored":true},{"key":"r_versioning","dimension":"R","label":"Snapshot identified","action":"Version the deposit and cite the exact version analysed (a version-specific DOI, or an accession with its version suffix). A reader reproducing your work against 'the current release' is reproducing it against a different dataset.","anchors":["yes","partial","no"],"verdict":"no","current":0.0,"evidence":null,"why":"Neither a version string nor a specific date is given for the dataset snapshot; the reference year (2024) is insufficient to pin a snapshot. [majority verdict 'no' (4/5 passes agreed)]","gain":4.17,"priority":"useful","scored":true},{"key":"f_discovery_metadata","dimension":"F","label":"Description of the dataset as an object","action":"Add a 'Data Records' section: itemise every file in the deposit and every variable or sample it holds, with counts and units. Describe the dataset as an object in its own right, not as a by-product of the findings — this is what makes it discoverable to someone who is not looking for your paper.","anchors":["yes","partial","no"],"verdict":"partial","current":0.5,"evidence":"Table 1. a) A sample from MIMIC-IV-BHC, our novel pre-processed dataset extracted from raw MIMIC-IV notes. ... b) Statistics for the MIMIC-IV-BHC subsets, binned across multiple context length ranges for adaptation tasks.","why":"Table 1 provides an itemised inventory of the dataset, showing a sample and statistics. [downgraded to 'partial' — no verifiable quote from the paper] [majority verdict 'partial' (3/5 passes agreed)]","gain":0.0,"priority":"essential","scored":false},{"key":"i_community_standard_vocabulary","dimension":"I","label":"Community standard / vocabulary","action":"Adopt and NAME your domain's data standard — the minimum-information checklist, metadata schema, or ontology your community uses (MIAME/MINSEQE, ISA-Tab, BIDS, an OBO ontology, HL7 FHIR/OMOP) — and say which one you followed. A reporting checklist standardises your paper; it does nothing for your data. In clinical / human-subjects, describe the data with OMOP CDM, CDISC SDTM or HL7 FHIR.","anchors":["yes","partial","no"],"verdict":"no","current":0.0,"evidence":null,"why":"No community standard for data or metadata is named; the paper only describes its own structure.","gain":0.0,"priority":"important","scored":false},{"key":"r_documentation_codebook","dimension":"R","label":"Documentation / codebook","action":"Ship a README and a data dictionary IN the deposit — every file, every variable, its units, its allowed values, its missing-value codes. It is the cheapest single thing that makes a dataset usable by someone who was not in the lab, and a table buried in the article does not travel with the data.","anchors":["yes","partial","no"],"verdict":"no","current":0.0,"evidence":"Table 1. a) A sample from MIMIC-IV-BHC, our novel pre-processed dataset extracted from raw MIMIC-IV notes. ... b) Statistics for the MIMIC-IV-BHC subsets, binned across multiple context length ranges for adaptation tasks.","why":"The variable and file definitions are provided inside the article (Table 1), not as a separate documentation object shipped with the data. [downgraded to 'no' — no verifiable quote from the paper] [majority verdict 'no' (3/5 passes agreed)]","gain":0.0,"priority":"important","scored":false},{"key":"a_controlled_access_for_sensitive","dimension":"A","label":"Gatekeeper for sensitive data","action":"Route sensitive data through an institutional gatekeeper — deposit in a controlled- access repository (dbGaP, EGA) with a Data Access Committee and a published DUA — rather than through the corresponding author's inbox. An author-gated dataset dies with the author's email address, and 'on reasonable request' has been shown repeatedly not to yield data. For sensitive/human clinical / human-subjects data, use a controlled-access repository such as dbGaP or EGA.","anchors":["yes","partial","no"],"verdict":"no","current":0.0,"evidence":null,"why":"The dataset is de-identified and publicly released; no gatekeeper is named for this study's own data. [majority verdict 'no' (4/5 passes agreed)]","gain":0.0,"priority":"useful","scored":false},{"key":"a_timeline_retention","dimension":"A","label":"Availability timing & retention","action":"State when the data become available AND how long they will be retained — cite the repository's preservation policy. NIH DMS Element 4 asks for both; most papers give neither.","anchors":["yes","partial","no"],"verdict":"partial","current":0.5,"evidence":"To allow researchers to replicate our adaptation and evaluation approach for BHC summarization, we publicly release: 1. The MIMIC-IV-BHC dataset, published on PhysioNet [23].","why":"The paper states the data are available now ('publicly release') but does not commit to how long they will persist. [majority verdict 'partial' (4/5 passes agreed)]","gain":0.0,"priority":"useful","scored":false}],"suggestions":["Attach a standard, machine-readable open licence to the deposit — CC0 or CC BY, which is what Horizon Europe and most funders expect — and print the licence identifier in the paper. 'Free to use' is not a licence: it grants nothing a reuser's institution can rely on.","Release the data in an open, community-standard format (CSV/TSV, JSON, HDF5, NetCDF, FASTQ, VCF, NIfTI…) instead of — or alongside — any proprietary or instrument-native format, and name the format in the paper. A dataset that needs a €2,000 licence to open is not reusable.","Version the deposit and cite the exact version analysed (a version-specific DOI, or an accession with its version suffix). A reader reproducing your work against 'the current release' is reproducing it against a different dataset.","Add a 'Data Records' section: itemise every file in the deposit and every variable or sample it holds, with counts and units. Describe the dataset as an object in its own right, not as a by-product of the findings — this is what makes it discoverable to someone who is not looking for your paper.","Adopt and NAME your domain's data standard — the minimum-information checklist, metadata schema, or ontology your community uses (MIAME/MINSEQE, ISA-Tab, BIDS, an OBO ontology, HL7 FHIR/OMOP) — and say which one you followed. A reporting checklist standardises your paper; it does nothing for your data. In clinical / human-subjects, describe the data with OMOP CDM, CDISC SDTM or HL7 FHIR."],"model":"deepseek/deepseek-v4-flash","agent_version":"fair_agent_v8","fulltext_source":"unpaywall_pdf"},"fair_model":"deepseek/deepseek-v4-flash","fair_agent_version":"fair_agent_v8","fair_fulltext_source":"unpaywall_pdf","fair_has_llm":true,"fair_computed_at":"2026-07-20T11:36:40.705193Z","clinical_trials":[],"software_tools":[],"db_accessions":[],"linked_datasets":[],"topics":[]}