Published in:
Open Access 01-12-2017 | Research article
Development of an algorithm for determining smoking status and behaviour over the life course from UK electronic primary care records
Published in: BMC Medical Informatics and Decision Making | Issue 1/2017
Login to get accessAbstract
Background
Patients’ smoking status is routinely collected by General Practitioners (GP) in UK primary health care. There is an abundance of Read codes pertaining to smoking, including those relating to smoking cessation therapy, prescription, and administration codes, in addition to the more regularly employed smoking status codes. Large databases of primary care data are increasingly used for epidemiological analysis; smoking status is an important covariate in many such analyses. However, the variable definition is rarely documented in the literature.
Methods
The Secure Anonymised Information Linkage (SAIL) databank is a repository for a national collection of person-based anonymised health and socio-economic administrative data in Wales, UK. An exploration of GP smoking status data from the SAIL databank was carried out to explore the range of codes available and how they could be used in the identification of different categories of smokers, ex-smokers and never smokers. An algorithm was developed which addresses inconsistencies and changes in smoking status recording across the life course and compared with recorded smoking status as recorded in the Welsh Health Survey (WHS), 2013 and 2014 at individual level. However, the WHS could not be regarded as a “gold standard” for validation.
Results
There were 6836 individuals in the linked dataset. Missing data were more common in GP records (6%) than in WHS (1.1%). Our algorithm assigns ex-smoker status to 34% of never-smokers, and detects 30% more smokers than are declared in the WHS data. When distinguishing between current smokers and non-smokers, the similarity between the WHS and GP data using the nearest date of comparison was κ = 0.78. When temporal conflicts had been accounted for, the similarity was κ = 0.64, showing the importance of addressing conflicts.
Conclusions
We present an algorithm for the identification of a patient’s smoking status using GP self-reported data. We have included sufficient details to allow others to replicate this work, thus increasing the standards of documentation within this research area and assessment of smoking status in routine data.