Development And Validation Of Risk Prediction Model For Contrast-Induced Acute Kidney Injury:A Multicenter Cohort Study

The 21st century has seen a global pandemic of cardiovascular disease.The rapid growth of percutaneous coronary interventions in China has led to an explosion in the usage of contrast media and the consequent increase in the incidence of contrast-induced acute kidney injury(CI＿AKI),which in turn has increased the medical burden on society.There is a lack of clinically effective treatment for CI＿AKI,and early prevention of the disease is essential around risk factors and safe doses of contrast agents.However,there is a lack of such studies in China.In view of this,this study collected and used data from a multi-centre cohort in China to conduct a study on the incidence of CI＿AKI,preoperative and perioperative risk prediction models and safe contrast dose levels.ObjectivesCI＿AKI risk prediction is a major clinical problem of multidisciplinary interest in cardiovascular and nephrology,and an important public health topic around which clinical epidemiological studies of incidence and early prediction are of crucial theoretical and practical value.Our study firstly analyzes the incidence of CI＿AKI and provide a basis for the development of CI＿AKI prevention strategies and clinical practice guidelines.Then,this study validates the performance,risk stratification generalization and clinical accessibility of international published preoperative and perioperative CI＿AKI risk prediction models,and to provide clues for the development of risk models suitable for the Chinese population.Furthermore,this study constructs a preoperative risk prediction model for CI＿AKI with good predictive performance and applicability to the Chinese population,and to provide ideas for the prevention of CI＿AKI.Finally,we investigate the dose-effect relationship in ratio of contrast volume-to-glomerular filtration(CV/GFR)and to provide a reference for the safe use of contrast dose.Methods1.Incidence rates and their confidence intervals were estimated.Atotal of 4271 patients who met the inclusion and exclusion criteria for percutaneous coronary intervention or Coronary angiography(PCI/CAG)in 12 central hospitals across China were included in the multicentre cohort.Boostrap self-sampling was used to obtain crude incidence rates,standardised incidence rates and corresponding 95%confidence intervals(CIs)for CI＿AKI under the four prevailing clinical definitions.The four CI＿AKI definitions were CI＿AKIA,CI＿AKIB,CI＿AKIC,CI＿AKID.2.External validation of existing models.A total of 18,868 papers were searched and screened,and 10 preoperative and 16 perioperative CI＿AKI prediction models were included for external validation analysis.In this study,four dimensions of model performance evaluation metrics,including discrimination,calibration,reclassification,and decision curve analysis,were used to externally validate and compare these 26 prediction models.3.Preoperative risk model building.Firstly,our study used the random forest method to multi-fill the missing data and applied time-ordered Holdout cross-validation to partition the training and external validation sets.The prediction model algorithms first used Stepwise-Logistic,LASSO(Least absolute shrinkage and selection operator)Logistic and Randomforest-Logistic,Extreme gradient boosting(XGBoost)with Shapley additive explanations(SHAP)value for feature screening of the training set,followed by the construction of subsequent machine learning prediction models using integrated algorithms such as Bayesian networks,support vector machines,decision trees,random forests,XGBoost,SuperLearner,etc.The optimal hyperparameters of the corresponding algorithms were obtained based on empirical methods and automatic grid search.The CI＿AKI model strategy utilized the area under the receiver operating characteristic(AUC)single value priority strategy and hyperparameter orthogonal strategy,and the model prediction effectiveness was evaluated using four indexes of discrimination,calibration degree,reclassification degree and decision curve analysis.4.CV/GFR dose effect analysis.The dose-effect relationships of 10 types of CV/GFR equations in CI＿AKI were systematically evaluated,and restricted cubic spline was applied to describe the dose-response relationships.The Odds ratio(OR)of CV/GFR cut points was calculated by a multifactorial Stepwise-Logistic method,and the corresponding sensitivity analysis was performed using Propensity Score(PS)and Directed Acyclic Graph(DAG).Results1.Incidence of CI＿AKI.The crude incidence rates and 95%CIs for the four definitions of CI＿AKI in PCI/CAG patients were,from highest to lowest,CI＿AKIA 7.02%(6.47%to 7.61%),CI＿AKIC 6.06%(5.55%to 6.61%),CI＿AKID 3.44%(3.06%to 3.86%),CI＿AKIB 1.66%(1.40%-1.96%).The standardised incidence rates and 95%CIs were,in descending order,CI＿AKIA 6.91%(6.36%to 7.49%),CI＿AKIC 5.97%(5.46%to 6.51%),CI＿AKID 3.28%(2.91%to 3.68%)and CI＿AKIB 1.57%(2.91%to 3.68%).There was a trend towards a higher incidence of all four types of CI＿AKI in elderly patients,with the highest incidence in the age group above 80 years.CI＿AKIA and CI＿AKIC are more likely to occur in the female population.2.Results of validation of foreign prediction models in the Chinese population.Among the 26 international prediction models,the optimal model for CI＿AKI was inconsistent across definitions.The AUC distributions in the preoperative prediction models ranged from 0.512 to 0.638 for AUCA,0.621 to 0.800 for AUCB,0.497 to 0.632 for AUCc,and 0.610 to 0.734 for AUCD;the AUC distributions in the perioperative prediction models ranged from 0.505 to 0.611 for AUCA and 0.551 to 0.782 for AUCB,Among the different CI＿AKI definitions,Tsai,Brown 2015,Ando ACEF,and Brown 2008 models were in the top 4 for preoperative risk prediction performance,and Abe,Marenzi,MehranA,and Uyarel models were in the top 4 for perioperative risk prediction performance.The prediction performance of Abe,Marenzi,MehranA and Uyarel models was in the top 4.3.Construction and validation results of preoperative prediction models based on various theories and algorithms.In the validation set defined by four CI＿AKI,StepwiseLogistic,LASSO-Logistic,RF-Logistic,XGBoost-XGBoost,XGBoost-NBN,XGBoostSVM,XGBoost-C50,XGBoost-RF and XGBoost-SuperLearner models took AUC values ranging from 0.696 to 0.856,0.754 to 0.913,0.678 to 0.868,0.719 to 0.901,0.690 to 0.852,0.657 to 0.883,and given the degree of discrimination,calibration,reclassification,clinical generalization ability,and number of independent variables,the LASSO-Logistic model performed the best under the four definitions and the findings were consistent.4.Results of a single indicator analysis for optimal prediction of safe dose of contrast.Of the 10 CV/GFR formulas,the CV/CKD-EPI formula performed best in terms of discrimination,calibration,reclassification,and simplicity of the equation.CV/CKD-EPI≥1.78 as the cut-point threshold resulted in an AUC of 0.736 and significantly higher reclassification than the other formulas.Multifactor-corrected OR=2.66(95%CI:1.504.72),DAG-corrected OR=2.64(95%CI:1.48-4.70),and propensity score nearest neighbor matching-corrected OR=1.89(95%CI:1.76-5.08)were consistent with the findings of the main analysis and sensitivity analysis,both indicating that when CV/CKD-EPI>1.78,the risk of developing CI＿AKI increased sharply.Conclusions1.The crude and standardised incidence rates of CI＿AKI differed significantly between the four definitions.The incidence of CI＿AKI under the above definitions tended to be higher in elderly patients,and CI＿AKIA and CI＿AKIC were more likely to be higher in the female population.2.The validation of foreign prediction models in the Chinese population was poor.The performance of 26 international risk prediction models based on prospective multicentre cohort validation in the Chinese population was poor,especially for the earlier prevention and clinical significance of preoperative models,and there is an urgent need to develop suitable preoperative risk prediction models for CI＿AKI in the Chinese population.3.The best preoperative predictive risk model was constructed by LASSO-Logistic.This statistical learning model has better discrimination in the CI＿AKI preoperative risk prediction model than traditional Stepwise-Logistic and machine learning XGBoost-ML models,and the corresponding calibration degree,reclassification degree and clinical generalization performance are stronger,and the model is worth promoting in Chinese clinical practice after visualization.4.The CV/CKD-EPI was the best among the 10 CV/GFR formulas.CV/CKD-EPI≥1.78 can be used as a reliable and concise predictor of the safe dose of CI＿AKI contrast agent.