Derivation And Evaluation Of Machine Learning Approaches To Predict Acute Kidney Injury After Acute Myocardial Infarction

Objective: To establish prediction models of acute kidney injury(AKI)after acute myocardial infarction(AMI)based on machine learning(ML)approaches and evaluate their performance.Methods: The clinical data of patients with acute myocardial infarction from 2008 to 2019 in the Medical Information Mart for Intensive Care database(MIMIC-Ⅳ v1.0)and from 2013 to 2018 in Zhongda Hospital affiliated to Southeast University are extracted respectively according to the inclusion and exclusion criteria.Data including demographics,medical history,medications,procedures,vital signs and laboratory tests are defined as features and AKI determined according to the Kidney Disease: Improving Global Outcomes(KDIGO)criteria is defined as the label.Binary classification models are derived based on ML algorithms provided by scikit-learn in Python,namely logistic regression,support vector machines,stochastic gradient descent,nearest neighbors,naive Bayes,decision trees,random forests,adaptive boosting,gradient boosting and neural network.The performance of each model is evaluated through confusion matrix as well as accuracy score,precision score,recall score,F1 score and receiver operating characteristic(ROC)curve.Results: A total of 5865 patients in MIMIC-Ⅳ database and 2024 patients in Zhongda Hospital affiliated to Southeast University with AMI are included in this study,among which 3946(67.3%)and 649(32.1%)had AKI,while 1919(32.7%)and 1375(67.9%)did not.The gradient boosting model and the random forest model show the best performance,with areas under the ROC curve(AUC)of 0.78,0.78(based on MIMIC-Ⅳ database)and 0.75,0.74(based on Zhongda Hospital),respectively.Features such as invasive ventilation,vital signs,creatinine levels and use of diuretics or norepinephrine are most important in the models.Conclusion: ML models represented by gradient boosting and random forest can well predict the risk of AKI after AMI.