A scoping review of models to identify transgender patients in electronic health records

Abstract

Objective

Electronic health records (EHRs) lack a widely adopted standard for recording transgender and gender diverse (TGD) status, complicating research on TGD health. Computational models have been developed to identify TGD individuals in EHRs; however, gaps remain in understanding which components contribute to stronger phenotyping approaches. This scoping review evaluates EHR-based models for identifying TGD individuals, focusing on identifier types, performance, external validation, and ethical reporting to guide best practices.

Materials and Methods

We searched PubMed, CINAHL, Web of Science, and Embase for peer-reviewed articles published before January 2024, following PRISMA-ScR guidelines. Included studies used EHR data to identify TGD individuals, verified TGD status, reported or allowed calculation of positive predictive value (PPV), and listed identifiers. Two authors screened and extracted data. We categorized models by data type and logic (structured, unstructured, and multimodal), summarized PPV distributions, and synthesized author-reported ethical considerations.

Results

Fourteen studies describing 50 models met inclusion criteria. Models using TGD-related diagnostic codes alone (n = 11) or requiring both structured and unstructured data (n = 6) showed the highest mean PPVs (85.3% and 97.1%). Models validated on larger confirmed TGD cohorts reported more stable performance, but external validation was rare. Most studies minimally addressed ethics; only 3 described protective measures or stakeholder engagement.

Discussion

Phenotyping of TGD individuals in EHR data remains heterogeneous in design and ethical transparency. Reported PPVs should be interpreted cautiously, as performance is influenced by study design, sample size, and verification methods.

Conclusions

Our recommendations emphasize the components that strengthen phenotyping approaches—identifier choice, multimodal intersection logic, validation practices, and ethical safeguards—rather than endorsing any single model.

Background and significance

Transgender and gender diverse (TGD) people are those whose gender does not align with the sex assigned to them at birth.¹ Due to stigma and discrimination, TGD people experience numerous health and social inequities, including disproportionately high rates of HIV infection, lifetime suicide events, and severe psychological distress.^2,3 These inequities are more pronounced for TGD people of racial and ethnic minoritized groups, those living with disabilities, and those struggling against economic marginalization.² Successfully addressing these inequities and investigating long-term health outcomes of surgical and medication-based gender-affirming care starts with diverse, rich, systematically collected data.⁴

Prospective data collection, however, is resource-intensive and complicated by diverse care settings, limited numbers of TGD-knowledgeable clinicians and researchers, logistical challenges, inconsistent data collection practices, as well as warranted mistrust of healthcare providers and researchers.^4,5 Mistrust of healthcare providers stems from reported negative experiences, including being misgendered, being asked unnecessary or invasive questions, physical or verbal abuse, and refusal of gender-affirming, emergency, or other care.^2,6,7 Additionally, historic and ongoing ethical missteps in TGD research, including inappropriate categorization, inadequate community engagement, and insufficient protection of participant confidentiality, have eroded confidence in medical and research systems.⁸ A significant concern fueling mistrust is the risk of being outed—the involuntary or non-consensual disclosure of a person’s TGD status or gender history.^4,5 Even voluntary disclosure can carry severe consequences, exposing TGD individuals to misgendering; invasive or unnecessary questioning; violence; denial of gender-affirming, emergency, or other healthcare; loss of employment or housing; and social ostracization.^2,3,5,7,9 For these reasons, TGD health research has primarily relied on cross-sectional or retrospective study designs, frequently leveraging electronic health record (EHR) data.⁴

While using EHR data to identify TGD individuals for health research has potential benefits,^4,10 researchers face substantial technical and ethical challenges. Foremost, the absence of a widely adopted, standard method for codifying TGD status in the EHR—like a 2-step gender verification question⁶—requires that researchers infer gender identity based on administrative (eg, legal sex) or clinical markers (eg, diagnosis codes, procedure codes, or medications such as testosterone, estrogen, puberty blockers, hormone blockers, progesterone, or anti-androgens) not designed for that purpose. These markers, or potential identifiers, are used across the health needs of TGD individuals, as well as cisgender and non-TGD intersex individuals. Inferring gender identity using these non-specific identifiers can result in misrepresentation, erasure, or undue surveillance.^11,12 Data availability often dictates which identifiers researchers can employ; however, no consensus exists on the optimal indicators for accurately identifying TGD status, partly due to variations in institutional data practices, evolving clinical standards, and ethical concerns about inferring gender identity.^6,11,12 This lack of consensus risks unnecessary duplication of research efforts and complicates the ability to benchmark algorithmic performance across studies. Additionally, a growing body of scholarship emphasizes that these methods carry significant ethical implications regarding consent, visibility, misgendering, and data justice that remain underexplored.^11–14

In a narrative review, Beltran et al highlighted 17 papers in which researchers identified TGD individuals from healthcare data in the United States. They concluded that (a) computable phenotypes are valuable when self-reported gender identity is unavailable, (b) self-reported gender identity is the gold standard but is often missing, (c) EHRs offer greater flexibility for verification of TGD status when compared to claims data, (d) diagnosis codes alone can identify most TGD patients when other data are unavailable, and (e) phenotypes combining diagnostic codes with procedure codes, prescriptions, or key-text strings have the highest positive predictive value (PPV).¹⁵ While Beltran et al’s narrative review provided a valuable overview of model components and data sources, several significant gaps remain. First, performance comparisons were presented qualitatively and lacked a granular categorization by identifier categories (ie, differences between TGD-related and non-TGD-related diagnosis codes), evaluation of logic in combining identifiers (ie, requiring both or either), evaluation of performance by sample size, or how these factors are affected by external validation—essential for assessing generalizability. Second, although the authors raised ethical concerns, they were not incorporated into an ethical synthesis. To address these limitations, we conducted a systematic scoping review that focused exclusively on EHR-based models, aligning with the literature’s caution against using claims data to infer gender identity due to epistemic and ethical risks.¹¹ Our approach was designed to enable structured comparison of model architectures and performance, assess reporting of external validation, and examine how ethical considerations are addressed across studies. We hope that this in-depth analysis provides further guidance and support for future researchers interested in TGD health research.

Objective

To evaluate peer-reviewed, EHR-based models for identifying TGD individuals by type of identifier used, number of verified TGD individuals, and external validation; to examine ethical considerations reported by study authors; and to highlight limitations and gaps in the literature to guide future best practices by identifying components that contribute to stronger phenotyping approaches.

Methods

Reporting standards and funding

We used the Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR).¹⁶

Information sources and search strategy

We searched PubMed, Embase, Web of Science, and CINAHL. We last consulted each source on January 19, 2024. Filters used for PubMed and Embase included “Abstract,” “Human,” and “English.” Limits used for Web of Science included “English” and “Document Type = Articles or Review Articles.” Filters for CINAHL included “Peer-Reviewed,” “Abstract,” “English,” “Human,” and “Exclude MEDLINE records.” A reference librarian helped inform the search strategy, shown in Figure 1.

Figure 1.

Concept groupings and Boolean logic used to develop the literature search strategy.

Eligibility criteria

Studies were eligible for inclusion if they (1) identified only TGD patients using data available in EHRs; (2) confirmed TGD identity through a verification process—we limited the allowable verification processes to include patient- or clinician-reported status as well as processes that included manual chart review; (3) explicitly reported PPV or provided sufficient data on true and false positives to allow for its calculation; and (4) published a complete list of identifiers used for TGD identification. We included studies that focused solely on TGD individuals or subpopulations within the TGD umbrella (eg, trans men, trans women, non-binary).

We excluded studies if they (1) used manual chart review to identify TGD individuals; (2) used a previously published cohort (ie, used a cohort of TGD individuals without identifying them independently); (3) did not report PPV or provide enough information to allow for its calculation—even if they reported other performance metrics such as sensitivity or specificity; (4) required extensive manual derivation of PPV due to unclear data presentation—to maintain methodological consistency; or (5) identified non-TGD populations (eg, men who have sex with men). We did not exclude studies based on differing operational definitions of transgender identity.

We limited inclusion to studies that either explicitly reported PPV or provided sufficient data to calculate it. PPV is a commonly reported performance metric¹⁷ and is particularly relevant because it reflects the probability that individuals flagged by a model are verified as TGD by the verification method used in the study (eg, chart review, patient-reported gender). Using PPV provided a consistent metric for comparison, reducing variability introduced by studies reporting different performance measures. Other performance metrics were recorded but not analyzed.

Selection and data collection process

We screened and reviewed all identified studies using Rayyan,¹⁸ a web-based review screening software. We deduplicated articles using ISSN, title, and author. The first 2 authors (R.A.B. and J.U.L.K.) independently evaluated every identified study for both the title/abstract screening and full-text review. We determined inter-rater reliability (IRR)¹⁹ via Cohen’s kappa (κ)²⁰ and interpreted κ via the Landis and Koch guidelines, where “substantial” agreement was deemed adequate.²¹ IRR was calculated following the title/abstract screening and the full-text review. R.A.B. and J.U.L.K. then independently reviewed and extracted relevant study data using a custom REDCap form—an electronic data capture tool hosted at Vanderbilt University^22,23—included in Supplementary Material S1.

Data items

We collected data items at both the study and the model levels. At the study level, we extracted identifying and contextual information, including authors, publication year, study timeframe, country, study purpose, specific models used to identify TGD patients, and any mention of ethical considerations such as privacy concerns, consent, data governance, and involvement of the TGD community. At the model level, we extracted the type of data used (eg, ICD-9/10 codes, gender markers, RxNorm codes, clinical notes), the specific identifiers applied (eg, 302*, F64*, “transgender,” etc), the logical conditions under which identifiers were combined (eg, at least one ICD-10 F64* code and “transgender” in clinical notes), the verification method used to confirm TGD status (eg, manual chart review, provider-reported gender—a structured data field only editable by clinicians²⁴—or patient-reported gender), the number of confirmed TGD patients (ie, the number of TGD individuals that were identified by the model and confirmed to be TGD through a verification process), and any reported performance metrics.

Study-reported ethical considerations

We characterized ethical issues as reported by study authors, rather than evaluating studies against a formal framework. We aimed to capture how authors acknowledged ethical considerations at the time of writing, not to critique their adequacy retrospectively. To highlight gaps and guide future work, we present broader implications for ethically grounded TGD research in the Discussion section.

Synthesis of results

We extracted the identifiers and the logical conditions under which identifiers were combined from each model included in the review. We categorized models into 5 groups based on the identifier(s) used: (1) TGD-specific diagnosis codes alone (TGD-ICD), (2) other structured data excluding TGD-ICD codes (Other Coded), (3) a combination of TGD-ICD codes and other structured data (TGD-ICD + Other Coded), (4) unstructured clinical text (Text-based), and (5) multimodal approaches that integrated both structured and unstructured data (Multimodal). For clarity, structured data in the EHR refer to standardized fields such as diagnosis codes, medication orders, and demographic variables. In contrast, unstructured data encompass free-text clinical notes, pathology reports, patient messages, and other narrative documentation that typically require natural language processing or manual review for analysis.^17,25  Table 1 showcases specific identifiers within each category.

Table 1.

Identifier categories and example identifiers used in TGD identification models.

Identifier category	Data type	Description	Example identifiers and logical conditions
TGD-ICD	Structured	ICD-CM codes indicative of TGD identity	ICD-9-CM: 302.50, 302.51, 302.52, 302.53, 302.6, 302.85
			ICD-10-CM: F64.0, F64.1, F64.2, F64.8, F64.9
Other Coded	Structured	Any structured data except TGD-ICDs	ICD-9-CM: 277.6, Other deficiencies of circulating enzymes; ICD-10-CM: E34.9, Unspecified endocrine disorder
			Gender, Legal Sex, and/or Assigned Sex at Birth: Male, Female, Non-binary, etc.
			Medications: RxNorm Codes
			Procedures: ICD-PCS, CPT Codes
TGD-ICD + Other Coded	Structured	Any combination of TGD-ICDs and Other Coded	ICD-9-CM 302.5 AND (CPT 54520, OR 54521 OR 54522 OR 54530)26
Text-based	Unstructured	Keywords and phrases found in unstructured clinical text	eg, “transgender,” “genderqueer,” etc.
Multimodal	Structured and/or Unstructured	Requires both structured (TGD-ICD, Other Coded, TGD-ICD + Other Coded) AND unstructured (text-based) data	TGD-ICD AND 1 or more keywords: (ICD-9-CM 302.51 OR 302.52 OR 302.53 OR 302.6 OR 302.85) AND (“gender dysphoria” OR “genderqueer” OR “MTF” OR “FTM” OR “transgender” OR “transsexual”)27
		Requires either structured (TGD-ICD, Other Coded, TGD-ICD + Other Coded) OR unstructured (text-based) data	TGD-ICD OR Gender = Transgender OR 1 or more keywords: (ICD-10-CM F64, F64.0, F64.1, F64.2, F64.8) OR (Gender = Transgender) OR (“transgender,” “transsexual,” “transvestite,” “gender identity,” “gender dysphoria,” “gender reassignment”)28

We synthesized findings across 4 domains: study-level characteristics, model-level characteristics and performance, as well as author-reported limitations and ethical considerations. At the study level, we summarized (a) general information (eg, first author, publication year, country, study timeframe, identifier categories used, number of models, and reported performance metrics), (b) data sources and study design, and (c) methods for verifying TGD status. At the model level, we documented the data categories used, specific identifiers applied, and trends in performance across the 5 data categories defined in Table 1. Finally, we summarized limitations and ethical considerations as reported by study authors. Because our objective was to map and characterize existing TGD identification models rather than evaluate methodological rigor, we did not conduct a formal quality assessment or risk of bias appraisal.

Results

Study selection

Our search identified 287 articles across all sources. After removing 134 duplicates, we excluded 120 studies based on the title and abstract screening. We retrieved the full text for all 33 remaining articles and excluded 19 articles for the following reasons: 6 did not publish a complete list of identifiers, 7 did not include a performance metric or required extensive manual PPV derivation, 2 used a previously identified cohort, 2 were not TGD-specific, and 2 did not validate the TGD status of identified individuals. We extracted data from the remaining 14 studies. Figure 2 illustrates this process. Cohen’s κ was 0.98 for the title and abstract screening (153 independently screened study titles/abstracts) and 0.74 for the full-text review (33 independently reviewed studies), indicating “almost perfect” and “substantial” agreement,²¹ respectively.

Figure 2.

PRISMA-ScR flow diagram depicting the identification, screening, eligibility, and inclusion of studies in the scoping review.

Study-level characteristics

Table 2 summarizes key study characteristics, including the first author, year of publication, country where the study was conducted, data years, identifiers used, number of models, and reported performance metrics.

Table 2.

Study-level characteristics.

Author (Year)	Country	Study years	Identifier category	No. of models	Performance metrics	TGD verification
Alpert et al. (2021)29	United States	2014-2019	TGD-ICD, Other Coded	3	TP, FP	Manual Chart Review
Beach et al. (2023)30	United States	2019-2021	TGD-ICD, Other Coded, Text-based	3	PPV	Manual Chart Review
Chyten-Brennan et al. (2021)31	United States	1997-2017	TGD-ICD, Other Coded, Text-based, Multimodal	5	PPV	Manual Chart Review
Ehrenfeld et al. (2019)27	United States	2001-2016	Multimodal	1	FPRa	Manual Chart Review
Foer et al. (2019)28	United States	2015-2019	Multimodal	1	Sens, Spec, TP, FP, TN, FN	Manual Chart Review
Guo et al. (2020)32	United States	2012-2019	TGD-ICD + Other Coded, Multimodal	2	PPV, Sens, Spec, NPV, F1	Manual Chart Review
Hua et al. (2023)33	United States	2017-2022	Text-based, Multimodal	10	PPV, Sens, Spec, NPV, Accuracy, F1, AUROC, AUPRC	Manual Chart Review, Provider-reported Gender
Nik-Ahd et al. (2023)26	United States	2000-2021	TGD-ICD, TGD-ICD + Other Coded	10	PPV, Sens, Spec	Manual Chart Review
Quinn et al. (2017)10	United States	2006-2014	TGD-ICD + Other Coded, Text-based, Multimodal	3	PPV	Rule-based and Manual Chart Review
Rich et al. (2020)24	Canada	1996-2013	TGD-ICD + Other Coded	4	PPV, Sens, Spec	Provider-reported Gender
Roblin et al. (2016)34	United States	2006-2014	TGD-ICD + Other Coded, Text-based, Multimodal	4	PPV	Rule-based and Manual Chart Review
Streed et al. (2023)35	United States	2008-2021	TGD-ICD + Other Coded	1	PPV, Sens, Spec, NPV, AUROC, AUPR, TP, FP, TN, FN	Patient-reported Gender
Wolfe et al. (2021)36	United States	2006-2018	TGD-ICD, Other Coded	2	PPV, TP, FP	Manual Chart Review
Xie et al. (2020)37	United States	2015-2018	Text-based	1	PPV, Sens, Spec, NPV, F	Manual Chart Review

^aEhrenfeld et al. (2019) reported a “false-positive rate” as a performance metric; however, one study author confirmed that this value is a false discovery rate (FP/[TP+FP]). False-positive rates require the calculation of the number of true negatives (FP/[FP+TN]), which was not possible according to their study design.²⁷

Abbreviations: AUPR, area under the precision-recall curve; AUROC, area under the receiver operating characteristic curve; F, f-score; F1, F1-score; FN, false negatives; FP, false positives; FPR, false-positive rate; NPV, negative predictive value; PPV, positive predictive value; Sens, sensitivity; Spec, specificity; TP, true positives; TN, true negatives.

Data sources and study design

Most studies (n = 13) were conducted in the United States,^10,26–37 with one conducted in Canada.²⁴ Data sources varied widely and included a cancer registry²⁹; academic medical centers from the Northeast,^28,31,33 Southeast,^27,32,34 Midwest,³⁰ and West Coast³⁷; the Department of Veterans Affairs (VA)^26,36; a community health center³⁵; and multi-site data repositories.^10,24 Two studies incorporated HIV registries, enabling provider-reported gender markers to serve as an identifier or as a reference standard for verification.^24,31 Most studies (n = 12) followed a retrospective diagnostic validation study design,^{10,26,28–37} to varying degrees, while others employed a retrospective observational study design (n = 2).^24,27 All included studies focused broadly on identifying TGD populations. One study (Nik-Ahd et al²⁶) also reported models developed specifically for trans women, alongside models for broadly defined TGD cohorts.

TGD verification methods

Streed et al was the only study to use patient-reported gender³⁵; considered the gold standard. Other studies relied on manual chart review,^{26–32,36,37} provider-reported gender,²⁴ or employed a rule-based system that incorporated diagnostic codes, keywords, and manual chart review.^10,34 One study combined provider-reported gender and manual chart review.³³ Rule-based verification systems incorporated diagnostic codes, keywords, and manual chart review to supplement unavailable data. In Quinn et al, the authors relied on the presence of multiple ICD-9 codes (either 2 different ICD-9 codes or the same ICD-9 code appearing at 2 different dates) when identifying TGD individuals using only ICD-9 codes. In contrast, if a keyword were present, the chart would be reviewed and reconciled by 2 independent reviewers.¹⁰

Model-level characteristics

Across the 14 included studies, we identified 50 models—46 unique and 4 external validations of previously published models. Among these, 28 models relied solely on structured data: 11 used TGD-ICD codes only^{24,26,29–31,36}; 6 used Other Coded data only^29–31,36; and 11 used a combination of TGD-ICD codes and Other Coded data.^{10,24,26,32,34,35} Thirteen models incorporated unstructured text: 7 used regular expression matching,^{10,30,31,33,34,37} 2 used deep learning,³³ and 4 used statistical or machine learning approaches.³³ Additionally, 9 models relied on multimodal approaches.^{10,27,28,31–34} We analyzed model performance by identifier category (Figure 3) and by the number of confirmed TGD patients (Figure 4). We did not assess performance by verification method because most studies (n = 10) relied on manual chart review. All study- and model-level details are available in Supplementary Material S2.

Figure 3.

Box-and-whisker plot showing model performance across six identifier categories.

Figure 4.

Model performance by number of confirmed TGD patients.

Model performance by identifier category

Figure 3 presents a box-and-whisker plot illustrating the distribution of PPV across different identifier categories. Given the small number of models within several categories, we chose not to conduct formal statistical testing and instead summarized the distributions of model performance descriptively. Among the 11 models using TGD-ICD codes alone, the mean PPV was 89.8% (excluding one outlier, PPV of 40.4%). Including the outlier, the mean PPV was 85.3% with a standard deviation of 17.2% and a coefficient of variation of 20.1%. The 6 models that relied solely on Other Coded Data had a mean PPV of 27.8%, a standard deviation of 28.3%, and a coefficient of variation of 102%, suggesting a high degree of variability in performance. The 11 models using TGD-ICD codes with Other Coded Data had an average PPV of 73.8%, a standard deviation of 28.9%, and a coefficient of variance of 39.1%. The 13 text-based models had an average PPV of 79.4%, a standard deviation of 28.1%, and a coefficient of variance of 35.4%. The 6 multimodal approaches that required both structured and unstructured data had an average PPV of 97.1%, a standard deviation of 2.83%, and a coefficient of variance of 2.9%. Finally, the 3 models that employed a multimodal approach, requiring either structured or unstructured data, had an average PPV of 58.7%, a standard deviation of 46.7%, and a coefficient of variance of 79.6%. Table 3 showcases these results.

Table 3.

Model performance by identifier category.

Identifier category	Mean PPV	PPV Standard deviation	PPV Coefficient of variance
TGD-ICD only	85.3%	17.2%	0.201
Other Coded data only	27.8%	28.3%	1.02
TGD-ICD + Other Coded data	73.8%	28.9%	0.391
Text-based data only	79.4%	28.1%	0.354
Structured AND Unstructured data	97.1%	2.83%	0.029
Structured OR Unstructured data	58.7%	46.7%	0.796

Model performance by number of confirmed TGD patients

Figure 4 is a box-and-whisker plot showing the distribution of PPV across models, grouped by the number of confirmed TGD individuals in quartiles. We chose quartiles to avoid arbitrary binning and ensure a balanced representation of models across the range of confirmed TGD individuals. Quartile 1 (Q1) included models that confirmed between 2 and 15 TGD individuals, quartile 2 (Q2) confirmed between 16 and 51 TGD individuals, quartile 3 (Q3) confirmed between 52 and 1397 TGD individuals, and quartile 4 (Q4) confirmed between 1398 and 5616 TGD individuals. As expected, Q1 exhibited the most considerable PPV variability, while Q4 had the least. Despite differences in sample size, models across all quartiles had median PPVs between 75% and 100%, with Q4 approaching a median PPV of 100%. More than half of the models confirmed fewer than 100 TGD individuals.

Externally validated models

Four of the 50 models we extracted were external validations of previously published approaches. Quinn et al externally validated 3 models initially developed by Roblin et al across the entire Kaiser Permanente System; Roblin et al had initially developed these models with data from Kaiser Permanente Georgia. Hua et al externally validated 2 models initially developed by Guo et al. We only included one as an external validation; however, because Guo et al only tested one of the models in the validation set in their original publication. The original models and their external validations are shown in Table 4. As expected, PPV performance decreased across all externally validated models. Notably, the well-performing multimodal approaches originally published by Roblin and Guo—externally validated by Quinn and Hua, respectively—showed only minor reductions in performance, with PPV decreasing by just 2% and 6.1%, respectively, despite substantial increases in the number of confirmed TGD individuals by factors of 36 and 315, respectively. These models, however, applied strict definitions of TGD identity (requiring both a diagnostic code and a keyword), making them highly specific but lower in sensitivity.³³

Table 4.

Original and externally validated models and their performance.

Study	Identifier category	Identifiers	Condition	No. of Confirmed TGD individuals	PPV (CI)
Roblin et al. (2016)34	TGD-ICD + Other Coded data only	TGD-ICD-9-CM, ICD-9-CM V-codes + KPGA codes	≥1 TGD-ICD-9-CM Code(s) OR ≥1 Institution-specific codes	14	56% (35-75)
Quinn et al. (2017)10,a				612	54%
Roblin et al. (2016)34	Text-based	Keywords (text-based)	≥1 Keyword(s)	62	45% (37-54)
Quinn et al. (2017)10,a				1876	26%
Roblin et al. (2016)34	Multimodal—Structured AND Unstructured data	TGD-ICD-9-CM, ICD-9-CM V-codes + KPGA codes, Keywords	(≥1 TGD-ICD-9-CM Code(s) OR ≥1 Institution-specific codes) AND ≥1 Keyword(s)	109	100% (96-100)
Quinn et al. (2017)10,a				3968	98%
Guo et al. (2020)32	Multimodal—Structured AND Unstructured data	TGD ICD-9-CM, TGD ICD-10-CM, Gender, Keywords	Gender = Transgender OR (≥1 Diagnostic Code(s) AND ≥1 Keyword(s))	5	100%
Hua et al. (2023)33,a				1575	93.9%

Abbreviation: KPGA, Kaiser Permanente Georgia.

^aExternal validation.

Study-reported limitations

Author-disclosed limitations varied across studies; see Supplementary Material S2 for study-level details. The most frequently cited was incomplete or inaccurate provider documentation, reported in 11 studies (79%), followed by concerns about generalizability due to single-site data or institutional differences, reported by 9 studies (64%). Likely false negatives were noted in 4 studies (29%), as were concerns of missing identifier categories. Patient-related biases were also described, such as models only capturing TGD individuals who disclose their identity (5 studies, 36%), seek gender-affirming care (4 studies, 29%), access gender affirming services at specific institutions (2 studies, 14%), and have access to healthcare (3 studies, 21%). A small sample size was reported in 2 studies (14%), and the need to update algorithms as terminology evolves was identified in 3 studies (21%).

Three studies (21%) reported difficulty identifying non-binary or gender non-conforming individuals due to diagnosis-based or binary classification systems that may misclassify or omit people who reject binary gender categories or present a binary identity to access care. Gender fluidity was acknowledged in 1 study (7%), as a limitation since recorded gender may not reflect identity over time.³⁵ Technical constraints of machine and deep learning approaches were noted in 1 study (7%), primarily related to potential lost data signal from methodologic choices (eg, restricting context windows to only parts of a note or limited word corpora in NLP-based approaches).³³ Some patterns reflected model type where structured-data approaches more often cited documentation and coding issues, while text-based approaches emphasized context, negation, and terminology drift.

Study-reported ethical considerations

Seven studies made high-level recommendations related to ethical data practices and systemic change. Commonly cited recommendations included standardized, routine, and self-reported gender identity data collection within the EHR,^{27,29–31,35,37} ideally in close consultation with the TGD community to ensure non-stigmatizing and patient-centered processes.^29,30 Beach et al emphasized the importance of explicit consent, transparency about data use, and prioritizing patient autonomy.³⁰ Some studies called for education and training for providers on TGD health^26,27,29 as well as the incorporation of nondiscrimination and grievance processes to address discrimination faced by TGD individuals.²⁹ Streed et al called for the use of organ inventories to avoid gender-based assumptions when addressing TGD-related health needs.³⁵

Four studies directly addressed ethical considerations of TGD phenotyping in EHR data. Chyten-Brennan et al and Streed et al described steps to mitigate privacy and security risk by limiting data access for chart review, de-identifying data after chart review, using de-identified data from the start of the study, and masking small counts in the results.^31,35 Chyten-Brennan et al was the only study to explicitly consider the risk-benefit tradeoff of performing phenotyping work with sensitive gender data.³¹ Hua et al respected patient autonomy by excluding individuals who had selected “chose not to disclose” in structured sex or gender fields.³³ Quinn et al uniquely involved TGD community members in study design and interpretation, reflecting a participatory approach to ethical engagement.¹⁰

Five studies, however, did not mention or explicitly consider ethical dimensions at all.^{24,28,32,34,36} Some studies occupied a more ambiguous space where they implemented practices with ethical implications but did not engage in explicit discussion of the ethical rationale or consequences of those choices.^10,33 A table of author-disclosed ethical considerations can be found in Supplementary Material S2.

Discussion

Summary of findings

This scoping review revealed substantial heterogeneity in methods used to identify TGD individuals in EHR data. TGD-specific ICD codes consistently outperformed other structured data, while multimodal models performed best when structured and unstructured data were combined using intersection logic (eg, TGD-ICD Codes and keywords). Union-based combinations (eg, TGD-ICD Codes or keywords) showed weaker performance. Model performance was generally less reliable in studies with small validation samples. External validation was uncommon but reinforced the relative strength of TGD-ICD codes and intersected multimodal models over text-based approaches. Authors frequently noted limitations related to generalizability and data completeness. However, few studies addressed ethical considerations such as consent, privacy, or community engagement.

Comparison with existing literature

Similar to Beltran et al, we found that models using diagnosis codes offered relatively strong performance; however, our review clarifies that this strength applies only to TGD-specific ICD codes, as general diagnostic codes performed substantially worse. Both reviews noted improved performance when structured and unstructured data were combined, but we expand on this by showing that gains were primarily limited to models using intersection logic, while union-based combinations were less effective.¹⁵ We also observed that models validated on larger cohorts of confirmed TGD individuals tended to report more stable performance, a factor not examined in Beltran. In contrast to Beltran et al, we restricted our analysis to EHR-based models, motivated by concerns that claims data cannot be independently verified through chart review or self-reported gender and may carry heightened ethical risks when used to infer TGD status.¹¹

Ethical considerations in TGD identification models

Phenotyping TGD individuals using EHR data has traditionally been framed as a technical or methodological challenge; however, recent scholarship has emphasized that such practices are not ethically neutral. As awareness of gender diversity and sociopolitical threats to TGD communities grows, so too does recognition that decisions about how to identify TGD individuals carry ethical consequences. Cato et al explicitly apply the principles of the Belmont Report—respect for persons, beneficence, and justice—to phenotyping practices. They argue that EHR-based inferences about sensitive attributes like gender identity risk violating patient autonomy if individuals are unaware that such characteristics could be inferred from their data. They also warn that such inferences, if disclosed or acted upon without patient knowledge, could damage trust and create inequities in care. To mitigate these harms, Cato et al recommend transparency about phenotyping practices, ongoing consent mechanisms, and safeguards to protect against provider bias and misuse.¹²

Alpert et al further critique current TGD phenotyping approaches as reductive and politically fraught. They argue that many algorithms reinforce binary gender norms, erasing gender-diverse realities and failing to address structural risks. Alpert et al encourage researchers to disaggregate relevant clinical indicators (eg, people with cervixes as evidenced by procedure codes) rather than infer gender identity, and to critically evaluate the potential for surveillance or exclusion in how phenotypes are defined and applied. These concerns align with broader principles of data justice, emphasizing that classification practices can reflect and reinforce social hierarchies and must be designed with consideration for their political and ethical context.¹¹

Adams et al offer complementary guidance for researchers and institutional review boards involved in TGD health research. They stress the importance of community consultation, reflexivity, and risk-benefit analysis when designing studies that could expose individuals to unintended visibility or harm. Even well-intentioned projects can have negative consequences if researchers fail to account for the lived experiences and vulnerabilities of TGD participants. Adams et al urge researchers to consider how their work may shape future research norms, data governance, and participant trust—particularly in populations that have experienced medical marginalization.¹³

While recent scholarship emphasizes the ethical stakes of TGD phenotyping, few studies in our review meaningfully engaged these concerns. Some raised issues aligned with frameworks from Cato, Alpert, and Adams—such as autonomy, transparency, and community engagement—but did not incorporate these principles into their methods. A smaller group applied select ethical practices, but did not always discuss their rationale. Some studies did neither. This limited engagement underscores the need for future research to adopt more explicit, rigorous, and community-informed ethical approaches.

Implications for research and practice

Standardization and reproducibility

We found wide variations in how studies reported model logic, performance metrics, and verification methods, which limited reproducibility and cross-study comparison. Future studies should adopt consistent reporting standards, including clear documentation of data sources, model logic, validation methods, and cohort sizes (ie, number of records evaluated, number of potential individuals identified, and the number confirmed through verification). Standardization would also support external validation—an area where few studies to date have contributed, but which is crucial for ensuring generalizability across institutions and populations.

Phenotyping practice

Researchers should prioritize direct markers of TGD identity, such as TGD-specific ICD codes or self-reported gender, instead of relying on inference from loosely associated indicators (eg, medication orders or procedure codes, alone). Multimodal models also demonstrated strong performance when structured and unstructured data were combined using intersection logic. When inference is unavoidable (eg, in claims data without confirmatory fields), logic should be disaggregated to reflect relevant clinical features without assigning a categorical gender identity.¹¹ Finally, studies with small confirmed TGD cohorts should report PPV cautiously, as small denominators may inflate apparent accuracy and limit confidence in model validity.

Ethical and participatory design

Ethical considerations should be integrated throughout the research process, not appended retrospectively. Where feasible and safe, researchers should engage TGD individuals and communities in model development, validation, and interpretation, reflecting broader calls for community guidance in shaping the use of sensitive health data.¹⁴ Such engagement may include input on classification logic, privacy risks, or potential downstream uses of phenotyping. Studies should also report steps taken to protect participants—such as de-identification, masking of small cell counts, and explicit discussion of potential harms and limitations.

Guidance for future research

This review is intended to guide best practices by emphasizing how components of phenotyping approaches—rather than isolated PPVs—shape validity and ethical responsibility. As phenotyping practices evolve, careful attention to transparency, rigor, and equity will be essential for developing computational tools that meaningfully serve gender-diverse populations.

Strengths and limitations of review

Our scoping review offers a comprehensive synthesis of electronic phenotyping models developed to identify TGD individuals in EHR data. To our knowledge, this is the first study to systematically evaluate the performance and architecture of these models while integrating author-reported ethical considerations. We applied a rigorous, PRISMA-ScR-informed methodology, dual-reviewed all identified studies, and dual-extracted data at both the study and model levels. We also developed a classification framework that distinguishes between uses of structured, unstructured, and multimodal data. Our findings provide clear guidance for future model development, validation, and ethical reporting practices in this rapidly evolving area of research.

This work has several limitations, however. First, our review was limited to studies using EHR data and excluded those based solely on claims or survey data. While this decision was intentional, reflecting our interest in models with verifiable ground truth, it may have excluded studies that used alternative, community-based, or insurance-derived datasets. Second, the included studies were predominantly conducted in the United States, with one study in Canada, which limits the generalizability of existing models to gender-diverse populations beyond North America. Third, the classification of models required some interpretation based on inconsistently described study methods. Fourth, our ethical synthesis is necessarily limited to what authors reported. Fifth, based on our inclusion criteria, we would not have captured studies published in languages other than English or non-peer-reviewed grey literature. Sixth, while we collected and evaluated our search terms based on guidance from clinicians familiar with TGD care, keywords from Guo et al,³² and the Vanderbilt University library staff, we may not have included all relevant search terms and missed pertinent studies in the literature. Seventh, we only examined studies that emphasized TGD individuals, not any study involving TGD individuals. Eighth, while PPV provided a consistent basis for comparing models, differences in how studies verified TGD status may influence reported values. Finally, because definitions of transness evolve, comparability across studies and the long-term validity of findings may be affected.

We also note that the ICD coding systems used to identify TGD individuals have evolved and will likely continue to do so. As such, the performance of models using historical ICD codes may not directly translate to contemporary EHR systems or future implementations. Ongoing validation and adaptation will be essential to ensuring the continued relevance and reliability of TGD phenotyping approaches.

Conclusions

Electronic phenotyping of TGD individuals in EHR data remains a valuable but ethically complex area of research. This review highlights variability in model design, validation practices, and ethical reflection. Our recommendations emphasize the components that strengthen TGD phenotyping approaches—such as identifier choice, multimodal logic, sample size, and ethical safeguards—rather than endorsing any single model. As this field advances, future work should prioritize transparent reporting, external validation, and approaches that center the safety, agency, and lived experiences of TGD populations.

Author contributions

Robert A. Becker (Conceptualization [true], Formal analysis [true], Writing—original draft [true], Writing—review & editing [true]), Jhansi U.L. Kolli (Formal analysis [true], Investigation [true], Writing—original draft [true], Writing—review & editing [true]), and Colin G. Walsh (Resources [true], Supervision [true], Writing—review & editing [true])

Supplementary material

Supplementary material is available at Journal of the American Medical Informatics Association online.

Funding

This research was supported by the National Institutes of Health [Grant Number T15007450-20].

Conflicts of interest

The authors declare no conflicts of interest.

Data availability

The data underlying this article are available in the article and in its online supplementary material.