Method used to identify adenomyosis and potentially undiagnosed adenomyosis in a large, US electronic health record database

Background The prevalence of adenomyosis is underestimated due to lack of a specific diagnostic code and diagnostic delays given most diagnoses occur at hysterectomy. Objectives To identify women with adenomyosis using indicators derived from natural language processing (NLP) of clinical notes in the Optum Electronic Health Record database (2014-2018), and to estimate the prevalence of potentially undiagnosed adenomyosis. Methods An NLP algorithm identified mentions of adenomyosis in clinical notes that were highly likely to represent a diagnosis. The anchor date was date of first affirmed adenomyosis mention; baseline characteristics were assessed in the 12 months prior to this date. Characteristics common to adenomyosis cases were used to select a suitable pool of women from the underlying population, among whom undiagnosed adenomyosis might exist. A random sample of this pool was selected to form the comparator cohort. Logistic regression was used to compare adenomyosis cases to comparators; the predictive probability (PP) of being an adenomyosis case was assessed. Comparators having a PP >= 0.1 were considered potentially undiagnosed adenomyosis and were used to calculate the prevalence of potentially undiagnosed adenomyosis in the underlying population. Results Among 11 456 347 women aged 18-55 years in the underlying population, 19 503 were adenomyosis cases. Among 332 583 comparators, 22 696 women were potentially undiagnosed adenomyosis cases. The prevalence of adenomyosis and potentially undiagnosed adenomyosis was 1.70 and 19.1 per 1000 women aged 18-55 years, respectively. Conclusions Considering potentially undiagnosed adenomyosis, the prevalence of adenomyosis may be 10x higher than prior estimates based on histologically confirmed adenomyosis cases only.