Anatomy and Biology: Two Complementary Sides of Breast Cancer Prognostication

Abstract

represent the weighted and transformed average of two separate risk predictions from Cox models derived independently from each data set. The primary prediction end point was 10-year risk of distant recurrence, and the models were compared by using the likelihood ratio test. This statistical test is used to compare the fit of two predictionmodels,onewithallvariablesincludedandanotherwithreduced number of variables. The test is based on the likelihood ratio, which indicates how many times more likely the observed data are in one model than the other. As one might expect, by combining partially independentclinicalandmolecularvariables,theRSPChadsignificantlybetterfit,indicatinggreaterprognosticvalue,thandidRSor a clinical model including tumor size, grade, and age alone. However, the average risk of distant recurrence was not statistically significantly different for the low-, intermediate-, and high-risk cohorts defined by RS and RSPC. For example, the average risk of 10-year distant recurrence for the low-risk cohort defined by RS was 5.9% (95% CI, 4.0% to 7.8%), whereas it was 6.5% (95%CI, 4.5% to 8.5) for the low-risk group defined by RSPC. These results suggest that, although the average predicted risk by risk group is similar, the two models classify individual patients differently and the individual outcomes are more accurately predicted by the RSPC, as indicated by the significantly better fit of that model. Indeed, RSPC classified more patients as having low risk (64% v 54%) and fewer patients as having intermediate risk (18% v 27%) than did RS alone. The largest discordance in risk assignment occurred in small (1 cm), low-grade tumors for which RSPC tendedtoassignlowerriskthanRSandinlarger(3cm),high-grade