Development of a prognostication model for breast cancer in British Columbia

Detecting breast cancer in its early stages improves patient outcomes, as the disease is often more treatable at early stages. In British Columbia, BC Cancer Breast Screening provides screening mammography as the primary means of early detection. As with any screening test, false positive results do occur. However, false positive results from screening mammography have been shown to be associated with the risk of a future breast cancer. Moreover, the mammographic features that indicate a mammogram is abnormal have been shown to stratify the increase in risk. Models to predict breast cancer risk currently have a modest ability to distinguish between who will be diagnosed and who will not, with concordances in the 0.55 to 0.71 range. Using risk factor, screening, and diagnosis data from BC Cancer Breast Screening, we build several models that incorporate mammographic features with the aim of creating a risk prediction model that provides a better risk prediction than current models. We fit a Cox proportional hazards model, a time varying Cox model, an accelerated failure time model and a Poisson regression model. We found that mammographic features are associated with an increased risk of breast cancer and increased model concordance by about 0.006. However, while the internal calibration was satisfactory, as it was close to 1, the discrimination of the models was unsatisfactory, with concordances around 0.63. Compared to several models in the literature, our model performance was similar, with concordances near 0.63 and internal calibration close to 1 for both ours and the published models examined. We conclude that breast cancer risk prediction is not substantially improved by the incorporation of mammographic features in a categorical format.Science, Irving K. Barber Faculty of (Okanagan)Computer Science, Mathematics, Physics and Statistics, Department of (Okanagan)Graduat