Pre-procedural machine learning model including prior SGLT2 inhibitor therapy improves early risk stratification for contrast-associated kidney injury in acute coronary syndrome
C Azcondo San Segundo, J M Andreu Cayuelas, F Buendia Santiago, A Meseguer Hernandez, I Gonzalez Fernandez, J Merino Romero, M I Banos Maturano, L Risco Eres, M B Villamarin Heredia, E Y Mite Salazar, B De Berardinis, M R Marmol Lozano, J A Castillo MorenoAbstract
Introduction
Contrast-associated acute kidney injury (CA-AKI) is a frequent complication in Acute Coronary Syndrome (ACS) with established prognostic impact. Early pre-procedural risk stratification is crucial, particularly with the emergence of nephroprotective therapies such as SGLT2 inhibitors (SGLT2i). The widely used Mehran score shows only moderate discrimination and relies on post-procedural variables, limiting its applicability before contrast exposure. Machine learning models have demonstrated improved predictive performance for CA-AKI; however, to our knowledge, none have specifically incorporated contemporary nephroprotective therapies such as SGLT2 inhibitors into pre-procedural risk prediction models. Integrating these therapies into non-linear pre-procedural models may enable more accurate early risk assessment.
Objectives
To evaluate whether a pre-procedural Machine Learning Random Forest (RF) model, including prior SGLT2i therapy, improves CA-AKI risk prediction compared with the standard post-procedural Mehran score.
Methods
Prospective observational study including 884 ACS patients undergoing coronary angiography. The cohort was temporally split into derivation (n=538) and validation (n=346) samples. A Random Forest model was trained using pre-procedural variables: age, sex, diabetes, Killip class, admission hemoglobin, baseline eGFR (CKD-EPI), and prior SGLT2i use. Performance was compared with the Mehran score using C-statistics and categorical Net Reclassification Improvement (NRI).
Results
CA-AKI incidence was similar in derivation and validation cohorts (7.4% vs 7.2%, p=0.740). In the validation cohort, the RF model showed higher discrimination than the Mehran score (AUC 0.755 vs 0.641), although the difference did not reach statistical significance (DeLong p>0.05), likely due to the low event rate. Reclassification analysis suggested meaningful clinical utility, with a total NRI of 19.8%, mainly driven by correct down-classification of non-events (non-event NRI 44.8%). Using its optimal threshold, the RF model identified 78.4% of patients as low risk prior to contrast exposure, with a CA-AKI rate of 2.2% (NPV 97.8%), compared with 31.0% using Mehran (NPV 96.2%).
Conclusions
A simple Random Forest model based exclusively on pre-procedural clinical variables, including SGLT2i therapy, improves early CA-AKI risk stratification compared with the standard Mehran score. This approach enables identification of low-risk patients before contrast exposure, potentially supporting tailored nephroprotective strategies.