Machine Learning-Based Prediction and Feature Attribution Analysis of Contrast-Associated Acute Kidney Injury in Patients with Acute Myocardial Infarction

TL;DR

This study uses machine learning to better predict acute kidney injury in heart attack patients and identifies key risk factors using explainable AI.

Contribution

The study introduces interpretable machine learning models for predicting contrast-associated acute kidney injury with higher accuracy than traditional methods.

Findings

An ensemble machine learning model achieved the highest AUC of 0.721 for predicting CA-AKI.

The model showed a high negative predictive value (0.942), effectively identifying low-risk patients.

Inflammatory markers like NLR and baseline renal indices were identified as key predictors.

Abstract

Background and Objectives: Contrast-associated acute kidney injury (CA-AKI) is a frequent and clinically significant complication in patients with acute myocardial infarction (AMI) undergoing coronary angiography. Early and accurate risk stratification remains challenging with conventional models that rely on linear assumptions and limited variable integration. This study aimed to evaluate and compare the predictive performance of multiple machine learning (ML) algorithms with traditional logistic regression and the Mehran risk score for CA-AKI prediction and to explore key determinants of risk using explainable artificial intelligence methods. Materials and Methods: This retrospective, single-center study included 1741 patients with AMI who underwent coronary angiography. CA-AKI was defined according to KDIGO criteria. Multiple ML models, including gradient boosting machine (GBM), random forest (RF), XGBoost, support vector machine, elastic net, and standard logistic regression were developed using routinely available clinical and laboratory variables. A weighted ensemble model combining the best-performing algorithms was constructed. Model discrimination was assessed using area under the receiver operating characteristic curve (AUC), along with sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Model interpretability was evaluated using feature importance and SHapley Additive exPlanations (SHAP). Results: CA-AKI occurred in 356 patients (20.4%). In multivariable logistic regression, lower left ventricular ejection fraction, higher contrast volume, lower sodium, lower hemoglobin, and higher neutrophil-to-lymphocyte ratio (NLR) were independently associated with CA-AKI. Among ML approaches, the weighted ensemble model demonstrated the highest discriminative performance (AUC 0.721), outperforming logistic regression and the Mehran risk score (AUC 0.608). Importantly, the ensemble model achieved a consistently high NPV (0.942), enabling reliable identification of low-risk patients. Explainability analyses revealed that inflammatory markers, particularly NLR, along with sodium, uric acid, baseline renal indices, and contrast burden, were the most influential predictors across models. Conclusions: In patients with AMI undergoing coronary angiography, interpretable ML models, especially ensemble and gradient boosting-based approaches, provide superior risk stratification for CA-AKI compared with conventional methods. The high negative predictive value highlights their clinical utility in safely identifying low-risk patients and supporting individualized, risk-adapted preventive strategies.

Figures