Fig. 3

Predictive performance of ML models in the internal validation cohort and external cohorts. a Comparison of AUCs among the Mehran score, LR, LASSO + LR, SVM, RF, GBDT and XGBT models in the internal validation cohort. b Calibration curve of Mehran score (ECE = 0.046), LR (ECE = 0.037), LASSO + LR (ECE = 0.046), SVM (ECE = 0.055), RF (ECE = 0.068), GBDT (ECE = 0.061) and XGBT (ECE = 0.054) models in the internal validation cohort. A smaller value of the expected calibration error (ECE) represents better calibration. c Decision curve analysis (DCA) of the Mehran score, LR, LASSO + LR, SVM, RF, GBDT and XGBT models in the internal validation cohort. The risk threshold represents the cutoff above which a patient may develop CIAKI. Net benefit is a weighted measure between true and false positives depending on the threshold. The None line and ALL line intersect at the point of risk threshold = 0.181, which also represents the internal CIAKI incidence. DCA of models above the NONE and ALL lines is considered clinically useful. d Comparison of AUCs among the Mehran score, LR, LASSO + LR, SVM, RF, GBDT and XGBT models in the external validation cohort. e Calibration curve of Mehran score (ECE = 0.048), LR (ECE = 0.111), LASSO + LR (ECE = 0.114), SVM (ECE = 0.116), RF (ECE = 0.125), GBDT (ECE = 0.116) and XGBT (ECE = 0.143) models in the external validation cohort. f DCA of the Mehran score, LR, LASSO + LR, SVM, RF, GBDT and XGBT models in the external validation cohort