Figure 3

Model I and II results. ROC curves and Kaplan-Meier (KM) curves produced with Model I and II (data from Desmedt et al. [20], Miller et al. [14], and Sotiriou et al. [15] was used for training, and the data from Wang et al. [13] and Pawitan et al. [16] for testing). Solid lines show the mode of the curves? distributions; dotted lines show the 95% Bayesian confidence intervals. (A) ROC curves for Model I. Red and green lines represent Model I results after training on multiple datasets with informative and noninformative priors, receptively. Blue lines represent Model I after training on a single dataset. The area under the curves (AUC) are 0.76 (0.71; 0.80), 0.57 (0.55; 0.58), and 0.51 (0.44; 0.60), respectively (numbers in parentheses show the Bayesian confidence intervals). The predictive performance of Model I trained with multiple datasets and informative priors was found to be significantly better than when using noninformative priors and with only a single dataset for training (Bayesian p-value < 0.0005 in both cases). (B) KM curves [44] for patients predicted by Model I (at 80% specificity) indicate the probability that patients develop distant metastases before (red) and after (cyan) 5 years. (C) Survival times predicted by Model II. The KM curves show survival of patients belonging to the predicted percentiles 0-33, 34-66, 67-100 (red, green, and red curves, respectively). The difference between the three groups is significant (Bayesian P-value < 0.003).