Development of a natural language-processing application for LGBTQ+ status in mental health records

Lesbian, gay, bisexual, transgender, queer and related community (LGBTQ+) individuals have significantly increased risk for mental health problems. However, research on inequalities in LGBTQ+ mental healthcare is limited because LGBTQ+ status is usually only contained in unstructured, free-text sections of electronic health records.

This study investigated whether natural language processing (NLP), specifically the large language model, Bi-directional Encoder Representations from Transformers (BERT), can identify LGBTQ+ status from this unstructured text in mental health records.

Using electronic health records from a large mental healthcare provider in south London, UK, relevant search terms were identified and a random sample of 10 000 strings extracted. Each string contained 100 characters either side of a search term. A BERT model was trained to classify LGBTQ+ status.

Among 10 000 annotations, 14% (1449) confirmed LGBTQ+ status while 86% (8551) did not. These other categories included LGBTQ+ negative status, irrelevant annotations and unclear cases. The final BERT model, tested on 2000 annotations, achieved a precision of 0.95 (95% CI 0.93-0.98), a recall of 0.93 (95% CI 0.91-0.96) and an F1 score of 0.94 (95% CI 0.92-0.97).

LGBTQ+ status can be determined using this NLP application with a high success rate. The NLP application produced through this work has opened up mental health records to a variety of research questions involving LGBTQ+ status, and should be explored further. Additional work should aim to extend what has been done here by developing an application that can distinguish between different LGBTQ+ groups to examine inequalities between these groups.