Identifying COPD in primary care: targeting patients at the highest risk

Abstract

Aim: To select and internally validate candidate variables for a risk prediction algorithm to detect chronic obstructive pulmonary disease (COPD) in primary care.

Background: COPD is vastly under-diagnosed in primary care and a variety of case finding and screening tools have been proposed to help identify patients with undiagnosed disease. Risk prediction algorithms have been developed for a variety of diseases but the development and use of such models for identification of COPD in the UK is currently limited. We used routine primary care data to identify and internally validate candidate variables for inclusion in a risk prediction model for COPD in primary care.

Method: We performed a literature search to identify potential risk factors associated with COPD. We then extracted data on 17,719 patients with an incident diagnosis of COPD (from 1st January 1990 to 31st March 2006) and 35,944 age-, sex-, and practice-matched controls from the General Practice Research Database and randomised them in a 2:1 ratio to form derivation and validation samples, respectively. The prevalence of a variety of clinical risk factors in the derivation sample (recorded at least 60 days prior to the diagnosis of COPD or equivalent matched time point for controls) was summarised. The unadjusted association between COPD and these risk factors was assessed using fixed effects conditional logistic regression. Candidate predictors were initially scoped from the literature search and then selected for a combined model based on the difference in prevalence between cases and controls, clinical face validity, and the size and statistical significance of their unadjusted odds ratios (ORs≥1.5 with p<0.05). The adjusted ORs were then estimated from a random intercept model. This model was used to formulate a risk prediction algorithm and was tested on the validation sample to assess its accuracy as measured by the area under the receiver operator characteristic curve (AUC or c-statistic) and calibration slope. Analyses were performed using Stata version 10.1.

Results: The mean age in the derivation sample was 69.7 years (SD 11.0) and 51.8% were male.Smoking status, salbutamol prescriptions and dyspnoea were the strongest predictors of COPD. Altogether nine variables were included in the combined risk prediction model including symptoms of wheeze and cough, previous diagnosis of asthma and lower respiratory tract infections (LRTI), and prescriptions of prednisolone and antibiotics for a chest infection. When tested on the validation sample this risk prediction algorithm had an AUC of 0.875 (95% CI 0.87 to 0.88). A cutpoint of 0.3 yielded a sensitivity of 86.6% and specificity of 70.1%. The algorithm performed good estimates against risk group deciles with a calibration slope of 0.98 (95% CI 0.87 to 1.10).

Conclusions: A risk prediction algorithm that includes smoking status, history of asthma and LRTI, symptoms of dyspnoea, wheeze and cough and prescriptions of salbutamol, antibiotics for respiratory infections and prednisolone appears to be highly predictive of incident COPD in primary care. This model, combined with age and sex, will be further developed and externally validated in a large screening trial for COPD in primary care (Birmingham Lung Improvement Studies, TargetCOPD). An externally validated risk prediction algorithm with sufficient sensitivity and specificity could be used to identify patients in primary care who might benefit from spirometry testing to detect COPD.