Skip to main content
Figure 1 | Breast Cancer Research

Figure 1

From: Benefits of biomarker selection and clinico-pathological covariate inclusion in breast cancer prognostic models

Figure 1

Performance, model size distribution and variable stability of reduced models for predicting 15-year breast cancer-specific survival. Upper row: The average time-dependent area under the receiver operator characteristic curve (ATD-AUCROC) performances of the full Cox models (FM) and reduced models (RM) derived utilizing 14 of the proteins included in the Oncotype DX assay (left column), the 18-variable full model that incorporates these 14 markers with four additional clinico-pathological variables (middle column) and seven standard clinico-pathological variables (right column) are denoted by circles. The corresponding performances on the training sets are denoted by plus signs. Error bars span ± 1 standard deviation from the average performance of the models. Combining protein plus clinico-pathological variables improved model performance, and variable reduction shown in the reduced models resulted in further improvement. Middle row: The sizes of the 15-year survival reduced Cox models were derived from the expected model size distributions. Bottom row: The variables incorporated in these reduced models were chosen according to their stability (frequency) in the nested cross-validation procedure. Distribution of model sizes and frequency-based stability were derived from the reduced models trained on the outer training set. For example, the average size distribution of the reduced models derived from the protein only variables (left column) is four, and thus the final reduced model includes AURKA, BCL2, CD68 and MYBL2. ER, estrogen receptor; HER, human epidermal growth factor receptor; PR, progesterone receptor.

Back to article page