Fig. 2

Machine learning approach M1. 50 most promising DMCs (hybridization and capture approach) combined with CA19-9 values for distinguishing PBC from pancreatitis and controls. A: PCA based on the 50 most informative DMCs combined with CA19-9 values for the conditions control (blue), pancreatitis (black), and PBC (red). Variances explained: PC1 = 56.75%, PC2 = 9.05%. B: ROC curve (AUC = 0.85) of PBC predicition scores for the identification cohort C2. The red dot indicates the determined optimal threshold value for the PBC prediction score that maximizes sensitivity and specificity with a defined minimum sensitivity of 90%. C: Boxplot of PBC prediction scores from the identification cohort C2 with the optimized classification threshold of 0.15 (gray line). D: ROC curve (AUC = 0.88) of PBC prediction scores for the validation cohort C3 including IMPNs. The red dot indicates the threshold value for classifying PBCs and high grade IPMNs with a minimum sensitivity of 90%. E: Boxplot of the PBC prediction scores from the validation cohort C3 including low and high grade IPMNs and the pre-determined PBC classification threshold of 0.15 (gray line). F: Kaplan-Meier curve for the survival of PBC patients from the validation cohort C3. Follow-up of 44 months after diagnosis. Separation of PBC group (n = 10) into two subgroups by the pre-determined PBC classification threshold of 0.15