Fig. 6: Establishment, validation and evaluation of the bile lipid classifier for discriminating between benign control and CCA samples.

a The establishment workflow of the CCA classifier. Diagnostic modelling was conducted using data from Cohort 1 and Cohort 2. Optimal combinations of bile lipid markers were identified through the LASSO binary logistic regression model. These diagnostic models were then validated in another cohort. The best diagnostic model was selected on the basis of cross-validation results across the two cohorts. Additionally, the modelling and validation process was performed with sex stratification to account for potential sex-specific differences in lipid profiles. b–d Diagnostic performance of the serum markers for CCA. The figure highlights the optimal cutoff points for each marker in this study cohort, along with the sensitivity (SEN), specificity (SPE), and area under the curve (AUC) values, including the 95% confidence interval (CI) for the AUC. e-l Diagnostic performance was assessed by modelling with the positive ion mode datasets of Cohort 1 (e, f) and Cohort 2 (g, h) and the negative ion mode datasets of Cohort 1 (i, j) and Cohort 2 (k, l). The other cohort served as a validation set. This assessment was conducted both with and without sex stratification. Waterfall plots for predicting the probability of CCA using sex-neutral (m), male (n), and female (o) optimal diagnostic models; the optimal cutoff points, SEN, SPE and AUC are labelled. Each stream represents the predicted probability of disease according to the classifier, and the colours indicate the gold standard of diagnosis: red represents CCA, and blue represents benign cases. p, q SEN, SPE and AUC for the BileScreen, BileMet and the bile lipid classifier (BileLipid).