ABSTRACT Background and Aims Most countries are experiencing an increasing trend of chronic kidney disease (CKD). Preventive strategies, such as early prediction of CKD progression and mortality, are crucial in reducing this disease rate. Although artificial intelligence (AI) models have demonstrated outstanding predictive performance in the prognosis of CKD, no study has been conducted on the 5‐year mortality risk of this disease. This study aims to perform this task to gain a deeper understanding of this technology and optimize preventive and treatment strategies. Methods This retrospective study utilized 1543 CKD hospitalized patients referred to four clinical centers in Tehran City from November 2022 to December 2024. We leveraged AI algorithms to establish prognostic models for the 5‐year mortality risk of CKD. The accuracy and calibration indicators were utilized to determine performance eligibility. The prognostic factors, including demographic characteristics, comorbidities, vital parameters, laboratory findings, and medical treatments, were analyzed using both univariate and multivariate statistics for this purpose. Results Ultimately, 1528 samples were employed for model construction. Our empirical results revealed that Random Forest (RF) (Positive Predictive Value (PPV) of 96.57% and 95% CI of 95.03–98.1, Negative Predictive Value (NPV) of 96.15% and 95% CI of 94.48–98.93, sensitivity of 96.13% and 95% CI of 94.39–98.85, specificity of 96.58% and 95% CI of 94.13–98.54, accuracy of 96.36% and 95% CI of 94.47–99.01, F1 ‐score of 96.35% and 95% CI of 95.06–97.63, Area Under the Receiver Operator Characteristics (AU‐ROC) of 0.941 has more performance efficiency for prognostic purposes. SHAP analysis indicated age, estimated Glomerular Filtration Rate (eGFR), dialysis, mechanical ventilation, Prothrombin Time (PT), and Partial Thromboplastin Time (PTT) as the best predictive factors. Discussion This study revealed that RF could potentially enhance the predictive strength of mortality in patients with CKD and decision‐making in clinical environments.
Nopour et al. (Wed,) studied this question.