Objectives. Primary care electronic health records provide a rich source of information for inequalities research. However, the reliability and validity of the research derived from these records depends on the completeness and resolution of the codelists (i.e. collection of medical terms/codes) used to identify populations of interest. The aim of this project was to develop comprehensive codelists for identifying people from ethnic minority groups, people with learning disabilities (LD), people with severe mental illness (SMI) and people who are transgender.

Design. We followed a three-stage process to define and extract relevant codelists. First, groups of interest were defined a priori. Next, relevant clinical codes, relating to the groups, were identified by searching Clinical Practice Research Datalink (CPRD) publications, codelist repositories and the CPRD Code Browser. Relevant codelists were extracted and merged according to group, and duplicates were removed. Finally, remaining codes were reviewed by two general practitioners.

Setting. The curated codelists were compared using a representative sample in the United Kingdom. The frequencies of individuals identified using the curated codelists were assessed and compared to widely used alternative codelists.

Participants. Comprehensiveness was assessed in a representative CPRD population of 10,966,759 people.

Results. After removal of duplicates and GP review, codelists were finalized with 325 unique codes for ethnicity, 558 for LD, 499 for SMI, and 38 for transgender. Compared with comparator codelists, an additional 48,017 (76.6%), 52,953 (68.9%) and 508 (36.9%) people with a LD, SMI or transgender code were identified. The proportions identified for ethnicity, meanwhile, were consistent with expectations for the UK (e.g. 6.1% Asian, 2.7% black, 1.4% mixed).

Conclusions. The curated codelists are more sensitive than those widely used in practice and research. Discrepancies between national estimates and primary care records suggest potential record/retention issues. Resolving these requires further investigation and could lead to improved data quality for research.