Urinary KIM-1 as an early biomarker for acute kidney injury in acute coronary syndrome: a synthetic data study

Abstract

Acute Kidney Injury (AKI) is a serious complication of Acute Coronary Syndrome (ACS) that increases morbidity and mortality. Traditional markers such as serum creatinine are delayed indicators of renal injury. Kidney Injury Molecule-1 (KIM-1) has been reported as an early biomarker of proximal tubular injury. In the reference ACS cohort used in this work, original patient-level data were not available for secondary analysis, motivating reconstruction of a synthetic dataset to enable additional modeling. A synthetic dataset of 90 ACS patients was reconstructed using published summary statistics from the reference cohort. Demographic, clinical, and biomarker variables were simulated to preserve reported distributions. AKI status was assigned probabilistically based on established clinical risk factors, independent of KIM-1. Urinary KIM-1 was modeled as a continuous variable and generated conditionally to reflect reported separation between AKI and non-AKI groups. Multivariable modeling was performed using ridge logistic regression, Random Forest, and Extreme Gradient Boosting (XGBoost), with performance evaluated using five-fold stratified cross-validation and 1,000-iteration bootstrapping. Urinary KIM-1 levels were higher in simulated AKI patients (4.19 ± 0.50 ng/mL) than in non-AKI patients (2.75 ± 0.43 ng/mL), consistent with the reference cohort summaries. Within the reconstructed dataset, a KIM-1 threshold of 3.38 ng/mL yielded sensitivity of 0.900 and specificity of 0.914. Ridge logistic regression demonstrated high discrimination (AUC = 0.984 ± 0.030), followed by XGBoost (AUC = 0.968 ± 0.024) and Random Forest (AUC = 0.904 ± 0.103). Bootstrap analysis showed stable ridge performance (AUC = 0.983; 95% CI: 0.953–1.000). Under assumptions consistent with the reference cohort summaries, urinary KIM-1 showed strong discriminative performance in multivariable models within a reconstructed synthetic dataset. These results should be interpreted as scenario-based and hypothesis-generating rather than as estimates of real-world diagnostic accuracy. External validation using patient-level data is required before clinical application.