Figure 9.

Simulation: distinguishing endometrial, colon, stomach, lung, and breast tumors from normal samples when tumor signal is diluted. The graphs are arranged as in Figure 7. Tumor signal characteristics (gray CDFs) were simulated by mixing 1% tumor signal with 99% randomly picked normal signal. Normal samples are the same as in Figure 7 (black CDFs). A: Diluted tumors were practically indistinguishable from normal samples when relying on m, with an area under the receiver operating characteristic curve (AUC) of 0.54. B–D: By contrast, the capacity for classification persisted over dilutions for the other signal measures, x, y, and z (AUCs of 0.73, 0.75, and 0.63, from left to right). As an example of the use of the convex hull (gray off-diagonal line), C shows an increase in the AUC from 0.75 to 0.79. CDF, cumulative distribution function; FPR, false-positive rate (ie, 1 − specificity); TPR, true-positive rate (ie, sensitivity).