Fig. 2: Algorithm performance on held-out datasets.

In the left figure, shown in black is the performance on the Hold-out Data (582 slides from BLC3001, BLC2002 and TCGA); in red the performance on the subset with closest population to the one in the deployment setting (BLC3001 subset with 420 slides from the 582 slides) and in yellow the performance on an independent dataset with slides from multiple tumor tissues (i.e., PAN-Tumor with 361 slides). Performances are summarized in the legend by area under the curve (AUC). The sensitivity, specificity and estimated molecular testing reduction rate given the algorithm performance and dataset FGFR+ prevalence values are shown in the table.