Research Background：End-stage renal disease (ESRD) represents the terminal stage of chronic kidney disease (CKD) progression. By 2020, the global ESRD population exceeded 12 million, with China accounting for nearly 30%, the highest in the world. Renal replacement therapy (RRT), including hemodialysis (HD), peritoneal dialysis (PD), and kidney transplantation, is the primary treatment for ESRD. Over 3.5 million patients worldwide receive maintenance dialysis, 90% of whom undergo HD. According to the Chinese National Renal Data System (CNRDS), the total number of dialysis patients in China approached 1 million in 2022, with maintenance hemodialysis (MHD) patients reaching 840,000-a 3.5-fold increase from 2012. Addressing the rapid growth in dialysis demand by improving medical quality and promoting social reintegration has become a global healthcare priority.Dialysis adequacy is a critical survival indicator for MHD patients. Currently, clinical practice relies on the Urea Reduction Ratio (URR) and Kt/V to assess adequacy. However, the 2002 HEMO study demonstrated that high-flux dialysis based on Kt/V did not improve survival rates. Consequently, existing metrics are criticized for failing to reflect the clearance of middle-molecule toxins and lacking a direct correlation with clinical outcomes, quality of life, and long-term prognosis. Identifying which indicators and computational models best assess dialysis adequacy remains an unresolved challenge in nephrology.Recently, advancements in Artificial Intelligence (AI) have offered new research avenues for assessing dialysis adequacy. Since 2005, studies using Artificial Neural Networks (ANN) and various machine learning (ML) models-including Decision Trees (DT), Random Forest (RF), Support Vector Machines (SVM), and Extreme Gradient Boosting (XGBoost)-have demonstrated superior predictive performance (AUROC up to 0.874) compared to traditional linear regression and formulas (e.g., Smye, Daugirdas). Despite this progress, existing studies often lack comprehensive variables such as dietary nutrition, neuropsychiatric status, and physical function. Furthermore, most current models are "diagnostic" in nature-identifying differences between stable "normal" and "diseased" states-making them suitable for diagnosis but insufficient for early intervention.Therefore, this retrospective study leverages real-world data (RWD) from Huashan Hospital to identify early warning factors for inadequate dialysis. The investigators aim to construct an "Ultra-early AI Warning Model" and an "AI Diagnostic Model" to form an Intelligent Decision-Making System for Hemodialysis, providing precise clinical intervention recommendations.
