DEVELOPING A PREDICTIVE MODEL FOR PROGRESSION TO DIALYSIS IN STAGE 4 SEVERE CHRONIC KIDNEY DISEASE PATIENTS

The global burden of chronic kidney disease (CKD) continues to escalate, with both incidence and prevalence showing sustained upward trends across populations. As such, early detection, prevention, and strategies to slow disease progression have emerged as critical priorities for public health systems worldwide. As CKD advances to end-stage renal disease (ESRD), characterized by a glomerular filtration rate (GFR) of less than 15 ml/min/1.73m², patients typically require renal replacement therapies, such as dialysis or kidney transplantation. These interventions not only place substantial strain on healthcare resources but also significantly affect patients’ quality of life and increase the burden on caregiving and support systems.

This study focuses on individuals diagnosed with stage 4 CKD (15≤GFR<30 ml/min/1.73m²), who received care at a northern tertiary medical center and a central regional teaching hospital in Taiwan. It aims to develop machine learning-based predictive models to identify clinical and demographic risk factors and to assist clinical decision-making and individualized patient management.

A total of 5,508 patients with CKD4 were retrospectively identified from integrated healthcare databases covering the period from January 1, 2016, to December 31, 2023. Predictors: age, gender, BMI, occupation, betel quid chewing, diabetes mellitus, hypertension, cardiovascular disease, heart failure, anemia, proteinuria, total cholesterol, low-density lipoprotein cholesterol, triglycerides, hypertriglyceridemia, hemoglobin, uric acid, serum sodium, serum phosphorus, serum calcium, and serum potassium. The dataset was randomly partitioned into a training (70%) and a testing (30%) subsets using SAS Enterprise Miner (EM) 15.2 (SAS Institute Inc., Cary, NC, USA). Predictive models were developed using five machine learning algorithms—decision tree (DT), random forest (RF), artificial neural network (ANN), logistic regression (LR), and support vector machine (SVM)—within the SAS EM. Model performance was assessed through repeated 10-fold cross-validation and confusion matrix-based metrics. To enhance model interpretability, feature importance was evaluated using SHapley Additive exPlanations (SHAP) values, implemented in Python 3.10 (Python Software Foundation), to identify the most influential predictors of dialysis initiation.

A three-year follow-up was conducted from the time patients were enrolled, at which point their eGFR ranged between 15 and 30 mL/min/1.73 m², 3,323 patients (60.33%) progressed to dialysis, while 2,185 patients (39.67%) remained stable or improved. Among the machine learning models evaluated, the ANN demonstrated the highest predictive performance (AUC=0.757, F1-score=0.762) when trained on the full set of features. SHAP analysis further revealed serum phosphate level, age, heart failure, proteinuria, and sex as the most influential predictors of progression.

Table 1. Average performance scores of the predictive model: Full Model

Figure 1. SHAP value ranking of each variable in predicting CKD stage 4 progression to dialysis

Figure 2. SHAP bee swarm plot of each variable’s impact on the model

This study highlights serum phosphate level, age, heart failure, proteinuria, and sex as the most influential predictors of progression from CKD4 to dialysis. The machine learning-based predictive models developed herein offer a robust, data-driven framework for early identification of high-risk patients, thereby facilitating personalized treatment planning and more efficient allocation of healthcare resource in nephrology practice. These findings underscore the potential of integrative analytics to augment clinical decision-making and support proactive, evidence-informed intervention strategies aimed at delaying dialysis initiation and improving patient outcomes.