I commend the attempt of McCowan and colleagues1 at developing a prediction rule for an important condition. If validated it will be a welcome addition to the armoury of GPs trying to rule out breast cancer confidently in women presenting with breast symptoms.
Perhaps inadvertently, the article provides an interesting case study on some of the difficulties in developing diagnostic instruments for low prevalence target disorders. Generating an adequate sample size for the target disorder (breast cancer) is difficult when, per thousand women, there are so few cases seen in general practice.
To circumvent this problem, the authors used data from a specialist-clinic setting to extract the relevant variables for their logistic regression model. In such a setting, it is not just the prevalence of breast cancer that is likely to be much higher than in general practice. Patients attending specialist clinics have already undergone some diagnostic filtering by GPs. The effect of filtering is that only those patients that are considered to have potentially serious conditions get referred on to the clinics. Thus not only is the prevalence of disease increased in the clinic setting, but also the severity of disease is increased. This may affect the significance of some of the explanatory variables in the model.
Nonetheless, using data from specialist clinics is understandable given the difficulty of deriving data from an unselected general practice population. Furthermore, it is not necessarily a concern if the resulting prediction rule proves valid in the intended population, that is, general practice. Unfortunately, in this instance, validating the tool proved difficult owing to the sample size in the validation cohort being underpowered.
Currently, the prediction tool represents a step in the right direction but needs much larger validation studies in unselected general practice populations before its uptake in general practice may be recommended.
- © British Journal of General Practice, January 2011