Objective Clinical prediction models of acute kidney injury （AKI） in patients with ACS after emergency admission was constructed based on the method of Machine Learning （ML）.
Methods Clinical data were collected from January 2018 to December 2020 from patients hospitalized with confirmed ACS in the First Affiliated Hospital of Jiangxi Medical University. （1） The data set was randomly divided into training group （70%） and testing group （30%）, and logistic regression, decision tree, random forest, XGboost, support vector machine （Support Vector Classification, SVC） and LightGBM methods were used in the training group to construct the ACS-AKI prediction model. （2） The best machine learning prediction model variables are screened by SHapley Additive exPlanations （SHAP） and Recursive feature elimination （REF）, a model visualization tool, and compressed into a compact prediction model with reduced features. （3） Finally, the screened prediction models are evaluated in terms of discrimination, calibration and clinical effectiveness using test set data.
Results A total of 499 clinical data of ACS patients were included, of which the incidence of ACS-AKI was 29.5% （147/499）; the LightGBM model showed the best predictive performance in the training set （AUC = 0.78）. Screening of the characteristic variables revealed that the predictors leading to ACS complicated by AKI were admission creatinine, ejection fraction, triglycerides, B-type natriuretic peptide （BNP）, blood urea nitrogen, cystatin C, D-dimer, age, loop diuretics, and extent of myocardial infarction. Finally, we validated the ML model of ACS-AKI using test set data and obtained that the LightGBM model had the best performance in terms of discrimination （AUC: 0.79）, calibration aspects, and clinical benefit.
Conclusion We have constructed and validated a clinical prediction model for concurrent AKI in ACS patients by different ML algorithms and obtained better results. However, this model still needs more external data for validation.
.png)
DownLoad: