BackgroundAssociation rule mining (ARM) is a structured data mining method used to discover numerically important and often unexpected associations between variables in a dataset without being hypothesis driven. Although widely applied in business, ARM has been little used in public health. Smoking is the largest avoidable cause of death and health inequalities in the UK, and varenicline is one of the most effective treatments to help smokers quit. We used ARM methods to identify the comorbidity characteristics of patients who typically do not receive varenicline for smoking cessation in UK primary care to identify numerically important groups in which varenicline is not being used. MethodsFrom the Health Improvement Network primary care database we obtained data on demographic and lifestyle characteristics, medical diagnoses based on Quality and Outcomes Framework disease indicators, and prescription of varenicline in all current smokers aged at least 16 years in 2011. ARM identifies rules of the form A→B, where A and B are disjoined sets of items such that B is likely to occur when A occurs. We therefore identified rules for sets of characteristics (A) of patients who did and did not receive a prescription for varenicline during 2011 (B). All (A) sets were required to contain a minimum of 500 patients, and we ordered rules according to their confidence, which is the probability of B given A. Findings were compared with traditional risk factor analysis of predictors of receiving a varenicline prescription by logistic regression. Ethics approval was granted from the EPIC Scientific Review Committee. Findings477 620 smokers were included in the analysis, of whom 19 316 (4%) were prescribed varenicline. Patients in the top ten rules for receiving varenicline were all heavy smokers, without obesity, excessive alcohol intake, hypertension, anxiety, depression, or psychotic disorders. Rates of receiving varenicline in these groups were at least 13·3%, which is much higher than in the general smoking population (4%). Patients in the initial top ten rules for not receiving varenicline were lighter smokers without comorbidity; however, of smokers with at least one comorbidity, those not prescribed varenicline were typically older (aged >60 years), non-heavy smokers without chronic obstructive pulmonary disease, but with a psychotic disorder or dementia; in these groups, fewer than 0·7% of patients were prescribed varenicline. Logistic regression also identified these characteristics, but did not provide the context of numerical importance. InterpretationARM analysis shows that varenicline is systematically underused in people with severe mental health problems, a group in whom smoking prevalence is extremely high and has changed very little over recent years. Although use of varenicline is subject to caution in these groups, the risk of continued smoking is likely to far exceed any risk arising from varenicline use. ARM has thus identified an important area of under-access of one of the most effective smoking cessation treatments. FundingThe authors are members of the UK Centre for Tobacco and Alcohol Studies, and the study is funded by the centre.
Read full abstract