Dynamic machine learning prediction of persistent AKI after cardiac surgery with modifiable perioperative risk factors

Figures

• 

  Figure 1. Patient flow diagram MAP, mean arterial pressure; CVP, central venous pressure.

• 

  Figure 2. ROC and PR curves of the machine learning models (all performance metrics were evaluated on the test set comprising 595 patients. Model 1 includes only preoperative variables; model 2 incorporates preoperative, intraoperative, and immediate postoperative variables [within 1 h of intensive care unit admission]; and model 3 includes preoperative, intraoperative, and postoperative variables up to 48 h after surgery) (A) ROC curves for the three ensemble models and the Thakar score. The AUROC values with 95% CIs are as follows: model 1 = 0.808 (95% CI: 0.745–0.869); model 2 = 0.837 (95% CI: 0.778–0.897); model 3 = 0.878 (95% CI: 0.834–0.922); and Thakar Score = 0.725 (95% CI: 0.663–0.802). (B) PR curves for the same models. The AUPRC values with 95% CIs are as follows: model 1 = 0.237 (95% CI = 0.155–0.375); model 2 = 0.330 (95% CI: 0.202–0.478); model 3 = 0.361 (95% CI: 0.227–0.508); and Thakar Score = 0.190 (95% CI: 0.122–0.286). ROC, receiver operating characteristic; PR, precision-recall; ENS, ensemble; AUROC, area under the ROC curve; CI, confidence interval; AUPRC, area under the PR curve.

• 

  Figure 3. SHAP summary plot of the top 10 important features (A) Model 1 includes only preoperative variables. (B) Model 2 incorporates preoperative, intraoperative, and immediate postoperative variables (within 1 h of intensive care unit admission). (C) Model 3 includes preoperative, intraoperative, and postoperative variables up to 48 h after surgery. EuroSCORE II, European System for Cardiac Operative Risk Evaluation II; eGFR: estimated glomerular filtration rate; Preop, preoperative; Hb, hemoglobin; METs: metabolic equivalents; ALT, alanine transaminase; Postop, postoperative; pRBC, packed red blood cell; Intraop, intraoperative; AUC, area under the curve; PP, perfusion pressure; CPB, cardiopulmonary bypass.

• 

  Figure 4. SHAP dependence plots for key predictors in model 3 The SHAP dependence plots show how feature values affect individual predictions. Each point represents a patient, with the feature value on the x axis and the SHAP value on the y axis indicating the contribution to predicted risk. eGFR, estimated glomerular filtration rate; Postop, postoperative; pRBC, packed red blood cell; Intraop, intraoperative; Hb, hemoglobin; Preop, preoperative; EuroSCORE II, European System for Cardiac Operative Risk Evaluation II; AUC, area under the curve; PP, perfusion pressure; CPB, cardiopulmonary bypass.