Seeing the unseen: how can we best identify transgender women within the Veterans Affairs healthcare system's electronic medical record?

One challenge in transgender research is reliably identifying patients through electronic medical records data, as there is no universal transgender International Classification of Diseases (ICD) code, but rather multiple ICD codes that can be used. To explore the sensitivity and specificity of 5 commonly used ICD codes to identify transgender patients overall and transgender women specifically (assigned male sex at birth) by using data from the Veterans Affairs (VA), the largest integrated health system in the United States. Patients aged ≥18 years were identified via ICD-9 codes 302.5 and 302.6 (Ninth Revision) and ICD-10 codes F64.0, F64.8, and F64.9 (Tenth Revision) using VA health records from 2000 to 2021 and stratified by bilateral orchiectomy status. Detailed chart review was performed on 32 randomly selected patients for each code (half with and half without orchiectomy) to confirm transgender status and to perform descriptive analyses. For each ICD code, rates of confirmed transgender status ranged from 88% to 100% for those with and without an orchiectomy, with the majority being transgender women (consistent with most veterans being assigned male sex at birth). Most transgender women (66%-100%) were undergoing estrogen gender-affirming therapy. The majority of provider-driven entries of transgender status took place from 2011 to 2020, with 75% of entries made from 2011 to 2020, consistent with increased recognition and societal acceptance of this population. False negatives were detected at a rate of 15%. Based upon these 5 ICD codes alone, we estimate that the VA has records for 9,449 to 10,738 transgender individuals. All 5 codes are very sensitive in identifying transgender patients, and the combination of these codes with orchiectomy is extremely sensitive in identifying transgender women, specifically. Major strengths of the study are the use of universal ICD codes and a large patient sample size that spans health records nationally and across multiple decades, potentially making our data more generalizable. The main limitation of this study is that subanalyses were performed on a limited number of patients, which prevented us from capturing all false positives and thus from calculating specificity for each code. Similarly, our true negatives were derived from a small, random subset of the population; as such, our calculation for specificity is an estimate. This study highlights a novel method to identify transgender women and paves the wa