Fig. 3. Pan-GI cancer detection accuracy using informative plasma DMRs identified from gitBS.

(A) Plasma samples of each GI cancer were randomly partioned into a training set (70%) and test set (30%) 10 times. Training sets of all GI cancers were pooled for training a pan-GI cancer prediction model. Representative ROC curve and AUC scores for the combined test sets are shown. (B) Use of informative plasma DMRs from A to predict pan-GI cancer in tissue samples.