MIMIC-IV-Ext Gender-Expansive Patient Stigmatizing Language Annotations

@article{PhysioNet-mimic-iv-ext-gender-stigma-1.0.0,
  author = {Xue, Liyang and Chayko, Mary and Singh, Vivek},
  title = {{MIMIC-IV-Ext Gender-Expansive Patient Stigmatizing Language Annotations}},
  journal = {{PhysioNet}},
  year = {2026},
  month = oct,
  note = {Version 1.0.0},
  doi = {10.13026/x00d-c603},
  url = {https://doi.org/10.13026/x00d-c603}
}

@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 database supports research on stigmatizing language in clinical documentation concerning gender-expansive patients. It contains expert-reviewed, note-level annotations for 754 deidentified discharge summaries from MIMIC-IV-Note, including 366 notes from gender-expansive patients and 388 comparison notes. The annotations cover credibility and obstinacy, compliance, and descriptors, with an additional flag for misgendering and parallel outcomes that include or exclude misgendering. The project aims to enable studies of documentation disparities, development of natural language processing methods for detecting stigmatizing language, and evaluation of model performance across patient groups. It provides a resource for research on more equitable clinical documentation.

Background

Stigmatizing language in electronic health records can influence clinical decision-making, propagate bias across care encounters, and undermine patient trust [1,2]. Gender-expansive patients, including transgender, nonbinary, gender-queer, agender, and gender-fluid individuals, may be particularly vulnerable to documentation-based stigma [2]. Previous work developed a multistage transfer learning approach for automated detection of stigmatizing language in electronic health records [3]. This resource was created to support reproducibility of the associated study [4] and further research on equitable clinical documentation, automated detection of stigmatizing language, and subgroup-aware evaluation of natural language processing systems. It augments MIMIC-IV-Note [5] with structured note-level annotations while avoiding redistribution of clinical note text.

Methods

Keyword screening and cohort review

The source corpus consisted of 331,794 deidentified discharge summaries from MIMIC-IV-Note [5]. Notes were screened for identity-related keywords to identify candidates for manual review, rather than to assign gender-expansive status automatically. The screening vocabulary covered gender identities, spelling and spacing variants, gender-affirming care, and references to assigned sex or gender. Examples included "transgender," "trans woman," "trans man," "transwoman," "transman," "nonbinary," "non binary," "genderqueer," "gender queer," "genderfluid," "gender fluid," "agender," "gender identity," "FTM," "MTF," "AFAB," "AMAB," "deadname," and "gender reassignment." Broader terms, such as "queer," "intersex," "masculine," and "feminization," were also used to retrieve potentially relevant contexts. A match to one of these terms was not, by itself, considered evidence that the patient was gender-expansive. The full keyword list is provided in Section S1 of the associated publication's Multimedia Appendix 1 [4].

Screening and comparison-note selection yielded 377 candidate gender-expansive notes and 377 candidate comparison notes. All candidates were manually reviewed to determine whether identity-related language referred to the patient rather than a family member, another person, or unrelated or templated content. After review and reclassification, the final corpus contained 366 gender-expansive notes and 388 comparison notes [4]. Notes reclassified into the comparison cohort were retained even when they contained gender-related keywords, avoiding a comparison group in which such vocabulary was artificially absent. The cohort therefore reflects documented identity-related language and does not identify every gender-expansive patient in the source population.

Comparison-note selection

Comparison candidates were randomly sampled from the broader MIMIC source corpus, with the candidate gender-expansive cohort's distributions of race, age, and primary language guiding selection. The aim was to obtain a comparison cohort with approximately similar demographic composition, rather than exact one-to-one matches or identical counts within every joint demographic stratum. The structured gender field was not used for matching because it does not reliably represent affirmed gender identity [4]. Manual review of both candidate groups then established the final cohort assignments, including the reclassification described above.

Annotation categories and examples

Two coders, LX and VKS, completed multiple alignment rounds using notes outside the final analytic dataset to refine the annotation codebook. Each note was reviewed for three nonmutually exclusive stigmatizing language categories adapted from prior work [6]. The definitions and illustrative phrases below follow the codebook in Section S2 of the associated publication's Multimedia Appendix 1 [4]. Examples illustrate the criteria and are not linked to individual records in this release. Classification depended on the surrounding context and whether provider-authored language conveyed judgment, disbelief, blame, or a pejorative characterization, rather than neutral clinical description.

• Credibility and obstinacy: Language that casts doubt on a patient's account or portrays the patient's perspective or intentions as inflexible or difficult. Codebook examples include "claims to see a therapist," "adamant he doesn't smoke," and "insists on being admitted."
• Compliance: Paternalistic or moralizing language that portrays a patient as failing, or being unlikely, to follow medical advice. Examples include "noncompliant" and "struggles with medication and follow-up compliance." Neutral documentation of a declined treatment, adherence barriers, or a clinical plan to encourage adherence was not automatically stigmatizing.
• Descriptors: Negative or judgmental descriptions of a patient's demeanor, behavior, or motives. Examples include "drug-seeking behavior," "concern for secondary gain," and "frequently manipulative." Neutral descriptions of symptoms, diagnoses, or observable behavior were not labeled solely because they described a clinical problem.

Only provider-authored or provider-endorsed language about the patient counted toward stigma labels. A quotation from a patient or family member was not automatically labeled unless the provider framed or endorsed it in a stigmatizing way. A note received a positive subtype label if at least one qualifying instance occurred. Multiple subtype labels could apply to the same note.

Misgendering and annotation reliability

Gender-expansive notes were additionally assessed for misgendering, defined as a pronoun, name, or gendered descriptor inconsistent with affirmed gender identity explicitly established within the note. An illustrative situation is a note documenting a patient's she/her pronouns and then referring to that patient as "he." Medically necessary anatomical or organ-specific descriptions were not labeled as misgendering unless accompanied by an incongruent gendered label or identity-invalidating phrasing. Notes without sufficient evidence of affirmed gender identity were not labeled as misgendering [4]. Misgendering was included in the original descriptor definition. Parallel descriptor and overall stigma labels excluded misgendering while retaining any other qualifying stigmatizing language.

During the alignment rounds, disagreements were discussed by the two coders until consensus was reached, and the shared codebook was refined accordingly. The associated publication reports an overall Cohen's κ of 0.85 and 92.5% agreement achieved through coder alignment [4]. During full annotation, notes were independently reviewed, with approximately 10% double-annotated to assess reliability [4]. The consensus procedure described here applies to the alignment stage and should not be interpreted as joint adjudication of every note in the released corpus.

Data Description

The resource contains one comma-separated value file, mimic_iv_gep_stigmatizing_language_annotations_v1_0_0.csv, with 754 records representing 754 unique MIMIC-IV-Note discharge summaries from 580 patients. Each row represents one clinical note. Some patients contribute more than one note. Clinical note text and direct identifiers such as names, addresses, and contact details are not included. The MIMIC linkage identifiers remain subject to credentialed access requirements [5,7].

Data dictionary

Values below describe the released CSV. Blank fields represent missing source values and have not been imputed. There are 12 missing insurance values and 13 missing marital_status values. All other columns are complete.

1. note_id: MIMIC-IV-Note identifier for the discharge summary. Stored as a string in the form subject_id-DS-note_sequence. Values are unique within this file and can be used to retrieve the source note [5].
2. subject_id: MIMIC identifier for the patient. Integer-valued identifier used to link patient-level source records [7]. It can repeat because the file is note-level.
3. insurance: Insurance category recorded in the source data. Observed values are Private, Medicare, Medicaid, and Other. A blank indicates missing information.
4. language: Recorded primary language. Observed values are English, Kabuverdianu, Russian, American Sign Language, Chinese, Spanish, Portuguese, Thai, Other, Haitian, and Arabic. Other is a source category, not a missing value.
5. marital_status: Recorded marital status. Values are MARRIED, DIVORCED, SINGLE, and WIDOWED. A blank indicates missing information.
6. race: Original race/ethnicity category from the source data, retained without regrouping. Observed values are WHITE, BLACK/CARIBBEAN ISLAND, NATIVE HAWAIIAN OR OTHER PACIFIC ISLANDER, WHITE - OTHER EUROPEAN, HISPANIC/LATINO - PUERTO RICAN, BLACK/AFRICAN AMERICAN, BLACK/CAPE VERDEAN, WHITE - RUSSIAN, HISPANIC/LATINO - DOMINICAN, OTHER, UNKNOWN, ASIAN - CHINESE, PATIENT DECLINED TO ANSWER, ASIAN, SOUTH AMERICAN, ASIAN - SOUTH EAST ASIAN, WHITE - EASTERN EUROPEAN, HISPANIC/LATINO - GUATEMALAN, UNABLE TO OBTAIN, HISPANIC/LATINO - MEXICAN, HISPANIC/LATINO - COLUMBIAN, HISPANIC/LATINO - SALVADORAN, HISPANIC/LATINO - CUBAN, HISPANIC OR LATINO, WHITE - BRAZILIAN, PORTUGUESE, and AMERICAN INDIAN/ALASKA NATIVE. Unknown, declined, and unobtainable values remain explicitly recorded as source categories.
7. gender: Structured source gender field. F denotes female and M denotes male in that field. It should not be interpreted as affirmed gender identity [8].
8. anchor_age: Age in years in the patient's MIMIC anchor year, not necessarily age at the date of the note. Observed values range from 18 to 91. MIMIC assigns 91 to patients older than 89 in the anchor year, so 91 is a deidentification value rather than an exact age [7].
9. race_grouped: Broad race/ethnicity grouping derived after trimming whitespace and converting race to uppercase. Values beginning with WHITE, BLACK, ASIAN, or HISPANIC map to WHITE, BLACK/AFRICAN AMERICAN, ASIAN, or HISPANIC/LATINO, respectively. All remaining source values, including unknown or declined categories, map to OTHER.
10. language_grouped: Derived language category. English indicates that language is English. Non-English indicates any other recorded language category.
11. age_group: Ordinal group derived from pooled anchor_age cut points of 29, 46, and 62 years. 0 = 18-29, 1 = 30-46, 2 = 47-62, and 3 = 63-91 in the released age values. The last group includes the MIMIC top-coded value of 91.
12. Credibility and Obstinacy: Binary subtype annotation. 1 indicates at least one instance meeting the credibility and obstinacy criteria in Methods. 0 indicates no annotated instance.
13. Compliance: Binary subtype annotation. 1 indicates at least one instance meeting the compliance criteria. 0 indicates no annotated instance.
14. Descriptors: Binary descriptor annotation including misgendering. 1 indicates at least one qualifying negative descriptor or annotated instance of misgendering. 0 indicates neither.
15. GEP: Manually reviewed cohort membership. 1 denotes a note concerning a gender-expansive patient. 0 denotes a comparison-cohort note. This is not a stigmatizing language outcome or a comprehensive gender identity variable.
16. Misgendering: Binary misgendering annotation. 1 indicates an annotated instance meeting the Methods definition. 0 indicates no annotated instance. Only gender-expansive notes were assessed for this outcome, so zeros in comparison notes should not be interpreted as a comprehensive assessment of gender-related errors.
17. label: Overall binary stigma outcome including misgendering. 1 indicates that at least one of Credibility and Obstinacy, Compliance, or Descriptors equals 1. 0 indicates that all three equal 0.
18. Descriptors_exclude_misgendering: Parallel descriptor annotation excluding misgendering. 1 indicates at least one qualifying nonmisgendering descriptor. 0 indicates none. A note with both misgendering and another qualifying descriptor remains 1.
19. label_exclude_misgendering: Overall binary stigma outcome excluding misgendering. 1 indicates that at least one of Credibility and Obstinacy, Compliance, or Descriptors_exclude_misgendering equals 1. 0 indicates that all three equal 0.

Annotation distribution

Counts were computed from this release. Percentages use the number of notes in each cohort as the denominator. Categories overlap and their counts should not be summed to obtain the number of stigmatizing notes.

Misgendering was assessed only in gender-expansive notes. The zero comparison count represents the released coding, not evidence that every comparison note was assessed for misgendering.

Dataset updates

This release includes updates to selected demographic variables compared with the dataset used for the associated publication [4]. The underlying notes and stigmatizing-language annotations remain unchanged. Analyses using the updated variables may yield results that differ from those reported in the publication.

Usage Notes

Use of this resource requires credentialed access to MIMIC-IV and MIMIC-IV-Note and continued compliance with the applicable PhysioNet data use agreement [5,7]. The file does not contain clinical note text. Credentialed users can retrieve corresponding source notes locally by joining on note_id and, where appropriate, subject_id. The annotations and source records must not be redistributed to individuals who lack the required access.

Intended research uses include:

• Describing the prevalence and overlap of stigmatizing language categories and examining differences across the documented patient cohorts and demographic groups.
• Developing and evaluating note-level natural language processing models for overall stigma detection or multilabel subtype classification after linking to the source notes.
• Comparing analyses that include and exclude misgendering to distinguish identity-related language errors from other stigmatizing descriptors.
• Evaluating subgroup error rates and fairness-aware detection approaches as potential documentation auditing tools with human review.

The associated study found a higher prevalence of stigmatizing language in gender-expansive notes than in comparison notes, including when misgendering was excluded from the outcome [4].

The gender-expansive cohort was identified through explicit identity-related language in clinical documentation and should be interpreted as a keyword-identified documentation cohort rather than a representative sample of all gender-expansive patients. The data originate from a single health system, and several demographic subgroups are small [4]. Labels are note-level indicators, not counts or locations of stigmatizing phrases. A zero label does not establish that a note is free of every possible form of bias. Because patients can contribute multiple notes, users should account for patient-level dependence in analyses and consider patient-disjoint evaluation splits when developing new models. This release contains the full corpus and does not prescribe a training or testing split.

Users should consider these limitations when conducting subgroup analyses or evaluating generalizability. The resource is intended for research rather than clinical decision-making, automatic inference of a patient's gender identity, or unreviewed assessment of individual clinicians. Analysis and model code associated with the study are available in the accompanying GitHub repository [9].

Release Notes

Version 1.0.0

Initial release of the complete note-level annotation dataset as a single CSV file.

Ethics

This study used deidentified electronic health record data from MIMIC-IV and MIMIC-IV-Note. The study was reviewed and approved by the Rutgers University Institutional Review Board (study ID Pro2025000546) with an exempt determination for secondary analysis of deidentified data. The study involved no direct interaction with human participants, and the institutional review board granted a waiver of informed consent because the research involved minimal risk and used existing deidentified data [4].

The original MIMIC-IV data collection was approved by the institutional review boards of Beth Israel Deaconess Medical Center and the Massachusetts Institute of Technology, with a waiver of informed consent [4,7].

Access to the source data requires completion of the applicable human subjects research training and execution of a data use agreement through PhysioNet [5,7]. This annotation resource does not redistribute clinical note text and is intended to remain subject to the relevant MIMIC-IV and MIMIC-IV-Note access requirements.

Acknowledgements

The authors received no financial support for this work.

Conflicts of Interest

The authors have no conflicts of interest to declare.

References

1. Goddu AP, O'Conor KJ, Lanzkron S, Saheed MO, Saha S, Peek ME, et al. Do words matter? Stigmatizing language and the transmission of bias in the medical record. J Gen Intern Med. 2018;33(5):685-691. doi:10.1007/s11606-017-4289-2. Available from: https://doi.org/10.1007/s11606-017-4289-2
2. Alpert AB, Mehringer JE, Orta SJ, Redwood E, Hernandez T, Rivers L, et al. Experiences of transgender people reviewing their electronic health records, a qualitative study. J Gen Intern Med. 2023;38(4):970-977. doi:10.1007/s11606-022-07671-6. Available from: https://doi.org/10.1007/s11606-022-07671-6
3. Xue L, Rahman AMM, Senteio CR, Singh VK. Automated detection of stigmatizing language in electronic health records (EHRs) using a multi-stage transfer learning approach. J Am Med Inform Assoc. 2026;33(2):283-294. doi:10.1093/jamia/ocaf193. Available from: https://doi.org/10.1093/jamia/ocaf193
4. Xue L, Chayko M, Singh VK. Stigmatizing language in gender-expansive patient records: corpus development, disparity analysis, and natural language processing–based detection study. J Med Internet Res. 2026;28:e91089. doi:10.2196/91089. Available from: https://doi.org/10.2196/91089
5. Johnson A, Pollard T, Horng S, Celi LA, Mark R. MIMIC-IV-Note: Deidentified free-text clinical notes (version 2.2). PhysioNet. 2023. RRID:SCR_007345. doi:10.13026/1n74-ne17. Available from: https://doi.org/10.13026/1n74-ne17
6. Harrigian K, Zirikly A, Chee B, Ahmad A, Links A, Saha S, et al. Characterization of stigmatizing language in medical records. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). Toronto, Canada: Association for Computational Linguistics; 2023. p. 312-329. doi:10.18653/v1/2023.acl-short.28. Available from: https://doi.org/10.18653/v1/2023.acl-short.28
7. Johnson AEW, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 2023;10(1):1. doi:10.1038/s41597-022-01899-x. Available from: https://doi.org/10.1038/s41597-022-01899-x
8. Kronk CA, Everhart AR, Ashley F, Thompson HM, Schall TE, Goetz TG, et al. Transgender data collection in the electronic health record: current concepts and issues. J Am Med Inform Assoc. 2022;29(2):271-284. doi:10.1093/jamia/ocab136. Available from: https://doi.org/10.1093/jamia/ocab136
9. Xue L. EHR GEP Study: analysis and model code for stigmatizing language in gender-expansive patient records. GitHub. 2026. Available from: https://github.com/liyangxue/EHR_GEP_Study (accessed August 28, 2026).