Risk analysis and warning model of post-noncardiac surgery acute kidney injury: A retrospective observational study

Postoperative acute kidney injury (PO-AKI) following noncardiac surgery remains a major clinical challenge, for which effective early warning models are still lacking. This study aimed to identify risk factors for noncardiac PO-AKI and to develop and validate a practical early warning model based on readily available clinical parameters across multiple surgical subspecialties. In this study, large-scale data from noncardiac surgical patients were extracted from the electronic medical record system of Xuanwu Hospital to analyze risk factors. Subsequently, an early warning model for PO-AKI was developed by integrating logistic regression (LR) and deep learning (DL) approaches. Internal and external validations were performed to assess model performance. A total of 19,152 patients were included in the analysis. Eleven independent risk variables were identified using LR, including age, surgical department, prolonged operation duration, contrast agent exposure, elevated serum creatinine, proteinuria, hemoglobin reduction, elevated fasting blood glucose, serum potassium abnormalities, neutrophil-to-lymphocyte ratio, and D-dimer levels. A corresponding risk score was subsequently constructed. The receiver operating characteristic analysis demonstrated areas under the curve (AUCs) of 0.7679 for the LR model and 0.7578 for the derived risk score. In the validation cohort, the AUCs were 0.7856 and 0.7122, respectively. The DL model further improved predictive performance, achieving AUCs of 0.97 and 0.93 in the validation and test sets, respectively. Precision–recall analysis showed area under the precision–recall curve values of 0.96 and 0.74, with corresponding F 1 scores of 0.945 and 0.897. Big data analysis revealed an overall incidence of PO-AKI after noncardiac surgery of 3.69%. A robust early warning model incorporating 11 clinically practical parameters was developed. The integration of LR and DL approaches improved predictive performance while maintaining interpretability.