Modeling Approaches for Estimating Breast Cancer Probability Using Conventional Risk Factors and Continuous Assessment of Mammographic Parenchymal Density

Abstract

Introduction. Currently available breast cancer (BC) risk models demonstrate limited prognostic value.

Aim. To develop multiparametric BC probability models incorporating both conventional risk factors and a quantitative measure of mammographic density, and to assess their accuracy.

Materials and Methods. The study included 82 women with histologically confirmed BC and 269 women without BC. In all participants, conventional BC risk factors were recorded, and mammographic parenchymal density along with its mean annual change was assessed. Classification models were developed using binary logistic regression, linear discriminant analysis, hierarchical cluster analysis, and decision tree analysis.

Results. Correlation analysis revealed weak correlations between BC risk factors and the presence of disease (r² = 0.060–0.367) in most cases. The only exception was a moderate correlation (r² = 0.609) between a history of breast biopsy and BC probability. When mammographic density change was not accounted for, the binary logistic regression model yielded the best performance: for the entire sample, sensitivity was 67.50 % and specificity was 94.42 %. For patients with ACR A–B breast density, these values were 75.86 % and 92.42 %, respectively; for patients with ACR C–D density, 80.56 % and 89.26 %, respectively. When mammographic density change was incorporated, the sensitivity and specificity of the binary logistic regression model increased to 81.82 % and 97.62 %, respectively, for the entire sample, and reached 100.00 % and 100.00 % for both ACR A–B and ACR C–D subgroups.

Conclusion. Combining mammographic density and its mean annual change with conventional BC risk factors within a binary logistic regression model provides high predictive accuracy (88.40–100.00 %).