729 - Automated Identification of Transgender and Gender Non-conforming patients from Electronic Health Record Data

Transgender youth suffer a disproportionate amount of discrimination, and significant healthcare disparities continue to exist. Due to lack of standardized methodologies for patient identification in electronic health records (EHR), it is difficult to accurately disaggregate transgender persons for population-based studies to ascertain how these disparities affect patients. Automated identification of patients could help to implement additional healthcare interventions, improve patient safety, and reduce healthcare disparities. We compared informatics-based identification methodologies including the Gender, Sex, and Sexual Orientation (GSSO) ontology, Natural Language Processing (NLP), and International Classification of Diseases (ICD) codes to identify the best-performing method for automated patient identification of transgender and gender non-conforming patients from EHR.


Objective:

The purpose is to evaluate automated identification of transgender or gender non-conforming individuals from EHR data.


Design/Methods:

The Cincinnati Children’s Hospital pediatric transgender clinic was started in 2013 and has seen more than 2,000 patients since opening. Patients with any hospital visit during the study period were included. We extracted data on patients who may be transgender or gender-nonconforming based one of three methods: evaluating the GSSO ontology or NLP applied to their clinical notes, a hand-curated database of patient visits, and a gold standard of an ICD9 or ICD10 diagnosis (F64.*) with a visit to the clinic. Chart review was performed to confirm accuracy on all patients and reviewers were blinded to the identification method. Data are described with descriptive statistics.


Results:

Compared to methods relying on only ICD codes or hand-curated methods, the NLP and GSSO ontology demonstrated excellent accuracy at identification, including those who may not have had a formal diagnosis listed in their chart. Patients identified only through the text-based methods should be manually screened for inclusion in population-based studies. These methods could provide an automated, scalable tool for researchers aiming to design population-based studies.


Conclusion(s):

Compared to methods relying on only ICD codes or hand-curated methods, the NLP and GSSO ontology demonstrated excellent accuracy at identification, including those who may not have had a formal diagnosis listed in their chart. Patients identified only through the text-based methods should be manually screened for inclusion. These methods could provide an automated, scalable tool for researchers aiming to design population-based studies.