MIMIC-IV-Ext-EntityCoding: Clinical Named Entity Recognition and Assertion Classification Dataset for ICD Coding

@article{PhysioNet-mimic-iv-ext-entitycoding-1.0.0,
  author = {Douglas, James C and Gan, Yidong and Hachey, Ben and Kummerfeld, Jonathan},
  title = {{MIMIC-IV-Ext-EntityCoding: Clinical Named Entity Recognition and Assertion Classification Dataset for ICD Coding}},
  journal = {{PhysioNet}},
  year = {2026},
  month = sep,
  note = {Version 1.0.0},
  doi = {10.13026/58h9-pm85},
  url = {https://doi.org/10.13026/58h9-pm85}
}

@article{pollard_physionet_2026, title = {{PhysioNet} as a global platform for biomedical research}, volume = {1}, issn = {3005-0693}, url = {https://doi.org/10.1038/s44360-026-00096-z}, doi = {10.1038/s44360-026-00096-z}, abstract = {Twenty-five years after its first public repository was established, the PhysioNet resource for biomedical and clinical data has enabled tens of thousands of research papers. Further efforts are ongoing to advance secure data sharing, community engagement and scalable research infrastructure.}, number = {8}, journal = {Nature Health}, author = {Pollard, Tom and Moody, Benjamin E. and Lehman, Li-wei H. and Gow, Brian J. and Fernandes, Chrystinne and Xie, Chen and Johnson, Alistair and Mark, Roger G. and Heldt, Thomas}, month = aug, year = {2026}, pages = {792--795},}

Abstract

This dataset provides clinical Named Entity Recognition (NER) and Assertion Classification (AC) annotations for 400 discharge summaries from MIMIC-IV-Note (v2.2), designed to support research on computer-assisted ICD coding and related clinical natural language processing tasks. It comprises 85,328 entity annotations across six categories (Normal Finding, Abnormal Finding, Disorder, Procedure, Health Context, and Medication) and 10,922 assertion labels across five categories (Present, Absent, Possible, Hypothetical, and Not associated with patient). In contrast to prior clinical NER resources, the schema was developed with the ICD-10-CM/PCS coding framework in mind, capturing coding-relevant modifiers such as acuity, laterality, and anatomical site within entity spans. The resource supports entity-centric clinical coding pipelines that use annotated entities as compact downstream inputs, and more broadly enables clinical NER and AC research on MIMIC-IV discharge summaries.

Background

Clinical coding, the process of translating free-text patient notes into standardized International Classification of Diseases (ICD) codes, supports billing, reimbursement, hospital planning, and epidemiological research. [1] Manual coding is labor-intensive and prone to delays, contributing to coding backlogs in many health systems. [2] Recent neural approaches to assisted ICD coding typically process entire discharge summaries [3-5], which often contain redundant or peripheral information. This redundancy, sometimes described as "note bloat", inflates computational requirements and can introduce noise during model training. [6,7] Targeted clinical Named Entity Recognition (NER) and Assertion Classification (AC) offer a way to retain the most salient information for clinical coding while discarding text that is unlikely to contribute to a code assignment.

Although clinical NER and AC are well-established tools in clinical NLP, existing clinical NER resources tend to lack the granularity required for accurate ICD coding. The widely used i2b2 2010 shared-task corpus, for example, labels spans with broad "problem", "test", and "treatment" categories. [8] ICD coding, by contrast, requires distinguishing between underlying disorders and their manifestations, and capturing specific modifiers (e.g., acuity, laterality, anatomical site) that influence code assignment. Prior attempts to combine NER and AC for clinical coding [9,10] have demonstrated the value of entity-based pipelines, but neither released ICD-aligned entity annotations.

This dataset was developed as part of recent work on entity-driven clinical coding. [11] It addresses the gap above by providing ICD-aligned entity annotations over MIMIC-IV-Note discharge summaries, capturing within entity spans the modifiers relevant to code assignment, and pairing a subset of those entities with assertion-status labels. While motivated by ICD coding, the schema and annotations are intended to be reusable for clinical NER and AC research more broadly, including use in combination with other clinical NER corpora to train models with broader category coverage.

Methods

Data Selection

Four hundred discharge summaries were randomly sampled from MIMIC-IV-Note (version 2.2) [12], stratified by encounter type to provide representation across different clinical scenarios.

Annotation Schema Development

The annotation schema was developed to align with ICD-10-CM/PCS coding requirements. Six entity categories were defined to distinguish disorders from their manifestations, capture procedures and medications, and cover contextual circumstances relevant to ICD-10-CM external-cause codes (V00-Y99) and factors influencing health status codes (Z00-Z99). Five assertion categories were defined to record whether each labeled mention is present, absent, possible, hypothetical, or not associated with the patient.

Entity Categories:

1. Normal Finding: physiologically normal observations or test results that typically do not map to an ICD-10-CM code.
2. Abnormal Finding: pathological or atypical observations, test results, or symptoms.
3. Disorder: underlying pathologies or etiologies giving rise to clinical findings.
4. Procedure: diagnostic or therapeutic interventions, including named devices used during the intervention.
5. Health Context: contextual factors relevant to ICD-10-CM external-cause codes (V00-Y99) and factors influencing health status codes (Z00-Z99), such as socioeconomic status, allergies, code status, injury circumstances, and personal medical history.
6. Medication: mentions of drugs (brand or generic names), preferably together with a relevant action (e.g., started, held, titrated) or route of administration.

Assertion Categories:

1. Present: directly stated as existing for the patient.
2. Absent: explicitly negated (e.g., "no evidence of pneumonia", "denies chest pain").
3. Possible: uncertain or probable (e.g., "probable pneumonia").
4. Hypothetical: speculative or future-oriented (e.g., "if pneumonia were to occur", "considering intubation").
5. Not associated with patient: mention pertains to someone other than the patient (e.g., "family history of diabetes", "mother has osteoporosis").

Entity spans are annotated to include neighboring contextual information that is relevant to coding, such as acuity (e.g., "acute"), laterality (e.g., "left"), and anatomical site (e.g., "lower lobe"). The full set of category definitions, boundary rules, and worked examples is distributed with the dataset in annotation_guidelines.md.

Annotation Process

Primary annotations were performed manually in Label Studio (v1.15) [13] by one of the authors, who holds a medical degree. Entity spans and assertion statuses were assigned in a single pass through each document, following the written guidelines distributed with the dataset (see annotation_guidelines.md).

Quality Control

Although primary annotations were produced by a single annotator, two steps were taken to support annotation quality. First, the written guidelines distributed with the dataset were informed by the ICD-10-CM/PCS coding framework. Second, a second annotator (also an author of the associated paper) independently annotated a random subset of documents containing approximately 1,500 entities using the same written guidelines. Agreement with the primary annotations was measured using exact-span F1 and Cohen's kappa, yielding values of 0.77 and 0.81, respectively. These figures are consistent with inter-annotator agreement reported in earlier clinical NER studies. [14] Disagreements were not reconciled post hoc; the primary annotator's labels are released as the single reference set, and the dataset has not been reviewed by professional clinical coders.

Data Description

The dataset comprises 400 annotated discharge summaries from MIMIC-IV-Note, 85,328 entity annotations, and 10,922 assertion classifications, distributed across the files described below.

File Structure

• entity_annotations.csv: entity spans and NER labels. A comma-delimited file with the following columns:
  • entity_id (int): a unique identifier for an entity, used to map entities between NER and AC labels.
  • note_id (str): the note_id of the discharge summary in MIMIC-IV-Note.
  • start (int): the character start index of the entity span (inclusive), defined over the processed text in mimic-iv_notes_subset.csv.
  • end (int): the character end index of the entity span (exclusive), defined over the processed text in mimic-iv_notes_subset.csv.
  • entity_label (str): the assigned NER label for the span (see "Label Naming Conventions" below for permitted values).
• assertion_annotations.csv: entity spans and assertion-status labels for the subset of entities that received an assertion annotation. A comma-delimited file with the following columns:
  • entity_id (int): a unique identifier for an entity, used to map entities between NER and AC labels.
  • note_id (str): the note_id of the discharge summary in MIMIC-IV-Note.
  • start (int): the character start index of the entity span (inclusive), defined over the processed text in mimic-iv_notes_subset.csv.
  • end (int): the character end index of the entity span (exclusive), defined over the processed text in mimic-iv_notes_subset.csv.
  • assertion_label (str): the assigned AC label for the span (see "Label Naming Conventions" below for permitted values).
• assertion_sentences.csv: the sentence containing each assertion-annotated entity, with sentence-local offsets, supplied for the convenience of users training assertion classifiers without needing to recompute sentence boundaries. A comma-delimited file with the following columns:
  • entity_id (int): a unique identifier for an entity, used to map entities between NER and AC labels.
  • note_id (str): the note_id of the discharge summary in MIMIC-IV-Note.
  • text (str): the sentence in which the entity occurs.
  • entity_text (str): the surface text of the entity within text.
  • entity_start (int): the character start index of the entity span within text (inclusive).
  • entity_end (int): the character end index of the entity span within text (exclusive).
  • assertion_label (str): the assigned AC label for the span.
• mimic-iv_notes_subset.csv: the subset of 400 discharge summaries from MIMIC-IV-Note used for annotation, after the lightweight preprocessing described under "Document Preprocessing" below. A comma-delimited file with the following columns:
  • note_id (str): the note_id of the discharge summary in MIMIC-IV-Note.
  • text (str): the processed text of the note.
• entity_ids.txt: a plain-text list of every entity_id that appears in entity_annotations.csv, one identifier per line. Provided as a convenience for filtering or sharding entities without reparsing the annotation CSV.
• assertion_ids.txt: a plain-text list of the entity_id values that carry an assertion label, one per line. Equivalent to the set of entity_id values appearing in assertion_annotations.csv and assertion_sentences.csv.
• annotation_guidelines.md: the annotation guidelines used to label each note, including category definitions, boundary rules, and worked examples.
• README.md: a quick-reference user guide for the files in the download.

Label Naming Conventions

For readability, entity and assertion category names are written in Title Case throughout this description (e.g., "Abnormal Finding", "Not associated with patient"). In the released data files, the corresponding label strings use lowercase snake_case (e.g., abnormal_finding, not_associated_with_patient). The mapping is one-to-one. The exact set of permitted values in each field is:

entity_label (in entity_annotations.csv): normal_finding, abnormal_finding, disorder, procedure, health_context, medication

assertion_label (in assertion_annotations.csv and assertion_sentences.csv): present, absent, possible, hypothetical, not_associated_with_patient

Entity Statistics

Assertion Statistics

Document Preprocessing

The 400 sampled discharge summaries from MIMIC-IV-Note [12] underwent lightweight cleaning prior to annotation. This included tidying redundant whitespace, removing hard-wrapped formatting, and segmenting each document into newline-delimited sentences. The pipeline was implemented using the medspaCy extension (v1.2.0) [15] for spaCy (v3.7.6) [16], including its sentence-splitting component. The preprocessed notes are released in mimic-iv_notes_subset.csv. Aside from these formatting changes, the textual content of each note was left unchanged.

Character Offsets

Character offsets in entity_annotations.csv and assertion_annotations.csv are defined with respect to the processed text in mimic-iv_notes_subset.csv, not the raw MIMIC-IV-Note content. These note-level offsets are the authoritative alignment for the NER and AC annotations. Sentence-level offsets in assertion_sentences.csv (entity_start and entity_end) are defined with respect to the sentence in the text column of that file, and text[entity_start:entity_end] == entity_text within each row. The entity_id column links every assertion-annotated entity back to the note-level offsets.

Usage Notes

Intended Uses

This dataset is intended as a research resource for clinical natural language processing on MIMIC-IV discharge summaries, focused initially on entity-driven ICD coding [11] but also useful for clinical NER and AC research more broadly. Representative use cases include:

• Training and evaluation of clinical NER models aligned with ICD-10-CM/PCS coding conventions, including recognition of coding-relevant modifiers (e.g., acuity, laterality, anatomical site) within entity spans.
• Training and evaluation of assertion classifiers that distinguish whether each mention of a clinical concept is present, absent, possible, hypothetical, or associated with someone other than the patient.
• Construction of "entity-only" representations of discharge summaries to reduce input length and training cost for downstream coding models. [11]
• Combining with other clinical NER/AC corpora (e.g., the i2b2 2010 challenge corpus [8]) to train clinical NER and AC models, where the granular ICD-aligned categories in this resource can supplement the broader categories used in earlier corpora.
• Interpretability and evidence-extraction research, by aligning model predictions or attention-based attributions with clinician-annotated entity spans.
• Information extraction, cohort identification, and downstream clinical NLP applications (e.g., problem-list construction, phenotyping, note search) over MIMIC-IV discharge summaries.
• More generally, as a benchmark resource for clinical NER and AC research over discharge summaries, including model comparison, ablation studies, and teaching.

Linking

• Entities can be linked to their source document via the note_id column.
• NER and AC labels can be collated for a given entity via the entity_id column, which is shared across entity_annotations.csv, assertion_annotations.csv, and assertion_sentences.csv.
• The note_id column follows the MIMIC-IV-Note convention and can be used to link records to other MIMIC-IV modules.
• Access to the original unprocessed discharge summaries requires credentialed access to MIMIC-IV-Note [12].

Known Limitations

Users should be aware of the following limitations when working with this dataset:

• Single primary annotator. All released labels originate from a single annotator with medical training. A second annotator labeled approximately 1,500 entities for quality assessment (exact-span F1 = 0.77, Cohen's kappa = 0.81), but the dataset has not been reviewed by professional clinical coders, and inter-annotator disagreements were not reconciled.
• Institution-specific source data. MIMIC-IV-Note contains discharge summaries from a single U.S. academic medical center (Beth Israel Deaconess Medical Center) [12], so models trained on this resource may not generalize to other healthcare settings, patient populations, or note types.
• Assertion subset is intentionally skewed. Of the 85,328 entity annotations, 10,922 carry an assertion label. Within that subset, rarer categories (Hypothetical, Possible, Not associated with patient) were prioritized during annotation in order to expose downstream classifiers to more diverse patterns. As a consequence, the label distribution within the assertion subset does not reflect natural prevalence in clinical text. Users for whom matching natural prevalence is important may wish to combine these labels with additional sources (see [11]).
• English-only content. All annotations are over English-language clinical text.
• Offsets defined over processed text. Character offsets in entity_annotations.csv and assertion_annotations.csv are defined with respect to mimic-iv_notes_subset.csv rather than the raw MIMIC-IV-Note text.

Release Notes

This is the initial release (1.0.0) of the dataset, accompanying the associated paper. [11]

Ethics

This dataset is a direct extension of MIMIC-IV-Note [12] and is derived entirely from the deidentified discharge summaries in that source. It adds annotation labels (entity spans, assertion statuses, and the sentence-level excerpts surrounding annotated entities) over the existing deidentified text without altering its content. No new protected health information has been introduced: the released artifacts consist only of category labels, character offsets, entity texts drawn from the already deidentified source, and excerpts of sentences from that source.

Because the dataset is derived from MIMIC-IV-Note, it inherits and is distributed under the same access policy, license, and data use agreement as MIMIC-IV-Note [12]. Users must hold valid PhysioNet credentialing and accept the PhysioNet Credentialed Health Data Use Agreement for MIMIC-IV-Note before accessing this resource.

Users should also be aware that models trained on single-institution data may not generalize to all patient populations.

Acknowledgements

This work was supported in part by the Australian Research Council through a Discovery Early Career Researcher Award. We thank the MIT Laboratory for Computational Physiology, PhysioNet, and Beth Israel Deaconess Medical Center for making the MIMIC-IV and MIMIC-IV-Note resources available to the research community, and the anonymous reviewers who provided constructive feedback on our associated paper.

Conflicts of Interest

The authors have no conflicts of interest to declare.

References

1. Dong H, Falis M, Whiteley W, Alex B, Matterson J, Ji S, et al. Automated clinical coding: what, why, and where we are? NPJ Digit Med. 2022;5(1):159. doi:10.1038/s41746-022-00705-7.
2. Campbell S, Giadresco K. Computer-assisted clinical coding: a narrative review of the literature on its benefits, limitations, implementation and impact on clinical coding professionals. Health Inf Manag J. 2020;49(1):5-18. doi:10.1177/1833358319851305.
3. Mullenbach J, Wiegreffe S, Duke J, Sun J, Eisenstein J. Explainable prediction of medical codes from clinical text. In: Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). New Orleans, Louisiana: Association for Computational Linguistics; 2018. p. 1101-11. doi:10.18653/v1/N18-1100.
4. Huang CW, Tsai SC, Chen YN. PLM-ICD: automatic ICD coding with pretrained language models. In: Proceedings of the 4th Clinical Natural Language Processing Workshop. 2022. p. 10-20. doi:10.18653/v1/2022.clinicalnlp-1.2.
5. Edin J, Junge A, Havtorn JD, Borgholt L, Maistro M, Ruotsalo T, et al. Automated medical coding on MIMIC-III and MIMIC-IV: a critical review and replicability study. In: Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '23). 2023. p. 2572-82. doi:10.1145/3539618.3591918.
6. Searle T, Ibrahim Z, Teo J, Dobson R. Estimating redundancy in clinical text. J Biomed Inform. 2021;124:103938. doi:10.1016/j.jbi.2021.103938.
7. Liu J, Capurro D, Nguyen A, Verspoor K. "Note bloat" impacts deep learning-based NLP models for clinical prediction tasks. J Biomed Inform. 2022;133:104149. doi:10.1016/j.jbi.2022.104149.
8. Uzuner Ö, South BR, Shen S, DuVall SL. 2010 i2b2/VA challenge on concepts, assertions, and relations in clinical text. J Am Med Inform Assoc. 2011;18(5):552-6. doi:10.1136/amiajnl-2011-000203.
9. Nath N, Lee SH, Lee I. Application of specialized word embeddings and named entity and attribute recognition to the problem of unsupervised automated clinical coding. Comput Biol Med. 2023;165:107422. doi:10.1016/j.compbiomed.2023.107422.
10. DeYoung J, Shing HC, Kong L, Winestock C, Shivade C. Entity anchored ICD coding [Internet]. Amazon Science; 2022 [cited 2026 Apr 27]. Available from: https://www.amazon.science/publications/entity-anchored-icd-coding.
11. Douglas JC, Gan Y, Hachey B, Kummerfeld JK. Less is more: explainable and efficient ICD code prediction with clinical entities. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Vienna, Austria: Association for Computational Linguistics; 2025. p. 30835-47. doi:10.18653/v1/2025.acl-long.1489.
12. Johnson A, Pollard T, Horng S, Celi L A, Mark R. MIMIC-IV-Note: Deidentified free-text clinical notes (version 2.2). PhysioNet. 2023. RRID:SCR_007345. Available from: https://doi.org/10.13026/1n74-ne17.
13. Tkachenko M, Malyuk M, Holmanyuk A, Liubimov N. Label Studio: data labeling software [Internet]. 2020-2024 [cited 2026 Apr 27]. Available from: https://github.com/HumanSignal/label-studio.
14. Deleger L, Li Q, Lingren T, Kaiser M, Molnar K, Stoutenborough L, et al. Building gold standard corpora for medical natural language processing tasks. AMIA Annu Symp Proc. 2012;2012:144-53.
15. Eyre H, Chapman AB, Peterson KS, Shi J, Alba PR, Jones MM, et al. Launching into clinical space with medspaCy: a new clinical text processing toolkit in Python. AMIA Annu Symp Proc. 2021;2021:438-47.
16. Honnibal M, Montani I, Van Landeghem S, Boyd A. spaCy: industrial-strength natural language processing in Python [Internet]. Zenodo; 2020 [cited 2026 Apr 27]. Available from: https://doi.org/10.5281/zenodo.1212303.