Case-finding for common mental disorders of anxiety and depression in primary care: an external validation of routinely collected data

Background: The robustness of epidemiological research using routinely collected primary care electronic data to support policy and practice for common mental disorders (CMD) anxiety and depression would be greatly enhanced by appropriate validation of diagnostic codes and algorithms for data extraction. We aimed to create a robust research platform for CMD using population-based, routinely collected primary care electronic data. Methods: We developed a set of Read code lists (diagnosis, symptoms, treatments) for the identification of anxiety and depression in the General Practice Database (GPD) within the Secure Anonymised Information Linkage Databank at Swansea University, and assessed 12 algorithms for Read codes to define cases according to various criteria. Annual incidence rates were calculated per 1000 person years at risk (PYAR) to assess recording practice for these CMD between January 1st 2000 and December 31st 2009. We anonymously linked the 2799 MHI-5 Caerphilly Health and Social Needs Survey (CHSNS) respondents aged 18 to 74 years to their routinely collected GP data in SAIL. We estimated the sensitivity, specificity and positive predictive value of the various algorithms using the MHI-5 as the gold standard. Results: The incidence of combined depression/anxiety diagnoses remained stable over the ten-year period in a population of over 500,000 but symptoms increased from 6.5 to 20.7 per 1000 PYAR. A 'historical' GP diagnosis for depression/anxiety currently treated plus a current diagnosis (treated or untreated) resulted in a specificity of 0.96, sensitivity 0.29 and PPV 0.76. Adding current symptom codes improved sensitivity (0.32) with a marginal effect on specificity (0.95) and PPV (0.74). Conclusions: We have developed an algorithm with a high specificity and PPV of detecting cases of anxiety and depression from routine GP data that incorporates symptom codes to reflect GP coding behaviour. We have demonstrated that using diagnosis and current treatment alone to identify cases for depression and anxiety using routinely collected primary care data will miss a number of true cases given changes in GP recording behaviour. The Read code lists plus the developed algorithms will be applicable to other routinely collected primary care datasets, creating a platform for future e-cohort research into these conditions.