I. Introduction
Chronic kidney disease (CKD), also known as chronic nephropathy, ranks as the 10th leading cause of death globally and continues to represent a major public health burden [1]. In South Korea, approximately 4.6 million individuals are estimated to have CKD, accounting for one in nine adults [2]. Because early symptoms are often minimal, CKD frequently remains undiagnosed until an advanced stage, resulting in a high mortality risk [3]. As diabetes and hypertension are major risk factors, CKD is closely tied to aging and other societal challenges [4]. In 2021, chronic diseases accounted for 7.1% of national healthcare expenditures, underscoring their increasing economic impact [5].
To effectively address CKD-related challenges—such as reducing disease incidence and associated social costs—it is crucial to identify the major factors influencing death among CKD patients. Mortality in CKD is primarily affected by clinical factors such as creatinine levels, hypoalbuminemia, nutritional status, proteinuria, and glomerular filtration rate [6]. These clinical variables are recognized as essential predictors of disease outcomes, particularly in the era of precision medicine and electronic medical record-based healthcare environments [7,8]. Such data provide valuable insights into a patient’s physiological status and disease progression. However, clinical factors are often complex and not readily available in all care settings, limiting the implementation of data-driven decision support systems in primary and secondary healthcare facilities.
To overcome these limitations, it is necessary to focus on general characteristics that are more easily accessible and applicable in routine clinical practice. General characteristics allow for quick patient assessment without extensive testing, thereby enhancing accessibility. Dialysis patients, in particular, depend heavily on healthcare professionals, emphasizing the importance of the clinical decision-making abilities of medical teams. In this context, general characteristics serve as the most immediate and fundamental parameters for clinical decision-making [9]. Notably, CKD clinical guidelines specifically list factors such as age (≥65 years), family history, and other general attributes as risk indicators [10]. General characteristics—including age, comorbidities, and the use of dialysis equipment—have been identified as significant predictors of mortality [6]. This highlights the importance of incorporating general characteristics when evaluating patient prognosis.
Mortality prediction in CKD patients has traditionally relied on probability-based statistical methods, which have long served as the primary analytical approach in this field [11]. More recently, AI models that rely heavily on clinical test results or laboratory data have been developed. Although these methods can yield high accuracy, they often require detailed medical information that may be difficult to obtain or interpret in many healthcare environments [12,13]. In contrast, this study focuses on developing a deep learning model using only general patient characteristics—such as demographics, hospital type, and healthcare utilization patterns—without depending on complex clinical inputs. By applying machine learning-based variable selection before model construction, we sought to improve both model simplicity and transparency. This streamlined approach supports practical application in primary and secondary care settings, where detailed clinical data may be limited. Ultimately, the goal is to develop a simplified yet high-performing prognostic tool that enables timely intervention and improved outcomes in CKD management.
II. Methods
1. Study Design
This study aimed to construct a death prediction model using data from the Korea National Hospital Discharge In-depth Injury Survey (KNHDIS). The target population included all patients discharged from acute care hospitals with ≥100 beds, and the database was created using discharge summaries for this population. Approval for the use of raw data was obtained through the Korea Disease Control and Prevention Agency website, and the study received an exemption from the Institutional Review Board of Kongju National University (No. KNU_IRB_2022-040).
The overall research process is illustrated in Figure 1. First, data preprocessing was performed, followed by a feature selection phase using machine learning to identify key factors influencing death. The least absolute shrinkage and selection operator (LASSO) regression was used for feature selection and compared with logistic regression to evaluate its performance. After selecting relevant features, a deep learning model was developed to predict death, and its predictive accuracy was assessed. Additionally, the deep learning model’s performance was compared with that of random forest and XGBoost algorithms to evaluate relative predictive capability.
The study included patients diagnosed with CKD (N18.X: N18.1, N18.2, N18.3, N18.4, N18.5, N18.9). A total of 12,680 patients were included. Data from 2016 to 2021 were divided into a training set (2016–2020; 10,498 patients) and a test set (2021; 2,182 patients) (Figure 2). To ensure the model’s temporal validity and real-world applicability, data from 2016–2020 were used for model training and validation, while the 2021 dataset—the most recent available—was reserved for testing. This approach was designed to simulate forward-looking prediction performance in clinical practice. The dependent variable was death, and independent variables were categorized into demographics and patient characteristics, medical institution characteristics, and healthcare utilization characteristics (Supplement A) [14].
2. Feature Selection
LASSO regression, a machine learning-based variable selection technique, was applied for feature extraction. LASSO regularization (L1 norm) was used to shrink regression coefficients, effectively eliminating less relevant explanatory variables [15]. This approach was chosen for its ability to perform variable selection and regularization simultaneously, thus promoting model simplicity and interpretability. Although alternative methods such as recursive feature elimination or mutual information-based selection could have been used, LASSO was selected for its practicality and wide adoption in clinical predictive modeling. For comparison, a traditional logistic regression model was also applied. Logistic regression is a well-established method for binary classification; however, it is prone to multicollinearity when handling high-dimensional data, which can lead to biased coefficient estimates and reduced predictive stability [16].
3. Deep Learning
A deep learning model was trained to predict death in CKD patients using TensorFlow. The model architecture consisted of an input layer, two hidden layers, and an output layer. Dropout regularization (50%) was applied to prevent overfitting. The rectified linear unit (ReLU) activation function was used in the hidden layers, while a sigmoid activation function was employed in the output layer. The Adam optimizer was used for efficient gradient-based parameter updates. Performance comparisons were made against random forest and XGBoost models. Hyperparameter tuning was conducted using grid search, with the following configurations: for random forest (n_estimators = 10, max_depth = 6, min_samples_leaf = 8, and min_samples_split = 8) and for XGBoost (n_estimators = 1,000, max_depth = 5, and min_ child_weight = 3) [17,18]. Model performance was evaluated using receiver operating characteristic (ROC) curves, accuracy, and root mean squared error (RMSE). Comparative analyses among models were conducted to validate predictive performance for mortality in CKD patients.
III. Results
1. Patient Characteristics
The general characteristics of the study participants are summarized in Table 1 (Supplement B). Among the participants, men accounted for 56.2%, and the mean age was 63.3 years. Regarding insurance coverage, 81.0% were covered by national health insurance, while 18.4% were medical benefits recipients. The majority of CKD patients were classified as stage 5 CKD (74.9%), followed by unspecified CKD (14.5%). With respect to hospital location, 31.5% were located in metropolitan areas, while 24.5% were in Seoul. In terms of hospital size, institutions with 500–999 beds were the most common (52.5%), whereas hospitals with 300–499 beds were the least common (10.7%). Outpatient admissions accounted for 66.9% of the total, the average length of stay (LOS) was 12.4 days, and the mean adjusted Charlson Comorbidity Index (CCI) was approximately 0.9.
2. Feature Selection Results
Variable selection using LASSO regression identified eight key factors influencing death in CKD patients (Table 2). The most influential factor was LOS (coefficient = 0.023), followed by emergency admissions (0.016), age (0.013), CCI (0.008), and regional area (0.003).
For comparison, traditional logistic regression analysis was also performed, and the results indicated that five factors were significantly associated with death (Table 3). Older age was linked to a higher risk of death (odds ratio [OR] = 1.05; 95% confidence interval [CI], 1.04–1.06). Smaller hospitals (100–299 beds) had a higher death rate than larger hospitals (≥1,000 beds) (OR = 1.98; 95% CI, 1.32–2.97). Emergency admissions were associated with a higher death rate than outpatient admissions (OR = 0.30; 95% CI, 0.23–0.38). Longer LOS (OR = 1.02; 95% CI, 1.01–1.01) and higher CCI scores (OR = 1.32; 95% CI, 1.20–1.45) were also significant predictors of mortality.
The variable selection methods were evaluated using the ROC curve and RMSE values for both LASSO and logistic regression models (Table 4). For LASSO, the area under the ROC curve (AUC) was 82.06%, whereas for logistic regression, it was 79.95%. The RMSE for logistic regression was 0.1930, indicating minimal difference between the two methods. Based on these findings, this study adopted LASSO—a machine learning methodology—for variable selection in model development.
3. Deep Learning
1) Deep learning model performance
The training and validation performance of the deep learning-based predictive model is illustrated in Figure 3. The loss value decreased as the number of training iterations increased, demonstrating the stability of the model. The final model achieved a training accuracy of 97.04% and a validation accuracy of 96.84%.
The predictive model for mortality in CKD patients was developed using deep learning techniques. To assess its effectiveness, the model was compared with random forest and XGBoost algorithms (Table 5, Figure 4). The AUC of the deep learning model was 0.83, with an accuracy of 96.83%. In comparison, random forest achieved an AUC of 0.79, and XGBoost achieved 0.71. These results confirm that the deep learning model outperformed traditional machine learning techniques in predicting mortality among CKD patients.
2) Final model evaluation
Using the 2021 test dataset, the final model was validated and achieved a loss value of 0.1207 and an accuracy of 96.84% (Table 6). These findings demonstrate that deep learning-based prediction models can serve as effective tools for identifying high-risk CKD patients and supporting clinical decision-making to improve patient outcomes.
IV. Discussion
This study aimed to identify the factors influencing death in CKD patients and to predict prognosis based on general characteristics using data from the KNHDIS, collected by the Korea Disease Control and Prevention Agency from 2016 to 2021. The key factors influencing death included prolonged hospitalization, emergency admissions, older age, higher CCI scores, specific regions (other area and Gyeonggi Province), hospitals with 100–299 beds, and health insurance coverage. The results of this study raise several important points for discussion.
First, emergency hospitalizations and prolonged hospital stays were strongly associated with a higher risk of death in CKD patients. This association can be attributed to the progressive nature of CKD, wherein many patients remain unaware of their deteriorating renal function until an acute event necessitates emergency hospitalization and urgent dialysis [19]. Emergency dialysis is frequently associated with severe complications, including infections and electrolyte imbalances, which significantly increase the risk of death within 3 to 12 months [15]. Previous studies have reported that CKD patients requiring emergency hospitalization have higher one-year mortality rates than those receiving scheduled treatment [3].
Moreover, patients requiring emergency care typically experience longer hospital stays, often exceeding twice the duration of those undergoing planned dialysis sessions [20]. Additionally, patients who die from CKD tend to have hospitalization durations more than four times longer than those who survive [6]. These findings suggest that prolonged hospital stays not only reflect disease severity but also exacerbate poor prognosis.
Reducing hospital length of stay is therefore a critical factor in decreasing death risk among CKD patients. To mitigate the negative effects of emergency admissions and prolonged inpatient stays, regular nephrology consultations and structured monitoring programs are essential. Ensuring timely initiation of renal replacement therapies—including hemodialysis, peritoneal dialysis, or kidney transplantation—can prevent acute deterioration, reduce the need for emergency interventions, and ultimately improve long-term survival in CKD patients.
Second, age was a significant predictor of death in CKD patients. According to the Korea Centers for Disease Control and Prevention, CKD prevalence increases with age, with individuals aged ≥60 years accounting for nearly 30% of all cases [21]. The incidence of CKD rises by 1.22 times per year of age, and the risk of death increases by 1.05 times [3,4]. Given that clinical guidelines identify age ≥65 years as a risk factor for CKD [11], tailored interventions and refined clinical strategies for older patients are essential. Furthermore, because awareness of chronic diseases is generally low [22], some older adults may already have CKD without realizing it. Considering that the mean age of subjects in this study was 63.3 years, it is important to recognize a broader age range (including those over 65 years) as a major factor influencing mortality in CKD. As the global population of older adults continues to rise, the development of age-stratified and evidence-based clinical guidelines will be increasingly important for effective CKD management.
Third, hospital size was identified as a crucial factor influencing mortality among CKD patients. This study found that patients admitted to hospitals with 100–299 beds had a higher mortality risk compared to those in larger hospitals. This hospital category includes facilities referred to as “comprehensive hospitals” under South Korean medical law. CKD patients undergoing renal replacement therapy such as hemodialysis receive care in artificial kidney units, and their prognosis is influenced by the clinical competence of the medical staff in these units [23]. Typically, nurses in artificial kidney units reach optimal proficiency after approximately 3 years of work experience [24]. However, as of 2018, comprehensive hospitals had the lowest proportion of nurses with more than 2 years of experience [25]. Because hospitals with fewer than 300 beds play a central role in managing basic diagnoses and providing chronic rehabilitation care, targeted strategies are needed to enhance clinical competency and improve CKD patient outcomes within these institutions.
Fourth, hospital location—particularly facilities located in “other areas” and Gyeonggi Province—was also a significant determinant of mortality among CKD patients. The term “other area” refers to regions excluding Seoul, Gyeonggi Province, and other metropolitan cities, including Gangwon, Chungbuk, Chungnam, Jeonbuk, Jeonnam, Gyeongbuk, Gyeongnam, and Jeju Provinces. These areas generally have fewer general hospitals and nephrologists, resulting in reduced healthcare accessibility and a higher risk of death [26,28]. The disparity in healthcare infrastructure between metropolitan and non-metropolitan areas may contribute to delays in CKD diagnosis and treatment, thereby worsening patient outcomes.
The Gyeonggi region, meanwhile, exhibited a lower standardized death rate but had the highest absolute number of deaths, likely reflecting its large population of older adults and high healthcare demand [27,28]. As of 2020, Gyeonggi Province had 521 nephrologists and the second-highest number of healthcare facilities nationwide, including tertiary hospitals, general hospitals, clinics, and other institutions, totaling 21,151 facilities, second only to Seoul [27]. This robust medical infrastructure may attract CKD patients seeking specialized care, which could partially explain the observed regional differences in death rates.
Fifth, sex emerged as a factor influencing death in CKD patients, with men exhibiting a significantly higher risk of death than women. It is well established that men have a higher overall mortality risk and a slower rate of decline in kidney function than women [29]. Additionally, smoking is frequently used as a proxy variable for men and is recognized as a contributor to renal fibrosis—the ultimate pathological manifestation of CKD [30]—making it a notable risk factor for CKD-related mortality.
Finally, the strength of this study lies in its integration of machine learning and deep learning methodologies to improve mortality prediction in CKD patients. Unlike traditional statistical models that depend on hypothesis testing and assumptions about data distribution, this study employed accuracy-based evaluation metrics to enhance predictive performance. Moreover, performing variable selection before implementing deep learning helped address one of the common limitations of deep learning models—namely, the difficulty in interpreting the relative importance of specific factors.
This study has several limitations. First, the model was developed using data from the Korean healthcare system, which may limit its generalizability to other populations. External validation using internationally diverse datasets is needed to ensure broader applicability. Second, although our two-step modeling approach—LASSO feature selection followed by multilayer perceptron (MLP) classification—was designed to improve interpretability, we did not compare it with alternative architectures such as TabNet or Transformers. Future research should evaluate more complex models and examine their relative advantages depending on data structure and clinical context. Third, although the model achieved satisfactory predictive performance, it was not optimized for class imbalance. Additional performance metrics, including precision, recall, and AUC-PR, could further clarify model behavior. Furthermore, explainability tools such as SHAP or LIME were not utilized in this study, but their application could enhance clinicians’ understanding of individual risk factors and increase trust in model-driven decisions. Given the model’s simplicity and reliance on general characteristics, it holds strong potential for early triage and decision support, particularly in resource-limited healthcare environments. Interestingly, the model demonstrated a slightly higher training loss than validation loss, likely due to the use of dropout regularization and early stopping, which helped mitigate overfitting during training.
Despite these limitations, this study demonstrates that general patient characteristics can serve as valuable prognostic indicators, enabling the development of a high-performance predictive model for CKD mortality. Future studies should focus on refining predictive algorithms, incorporating additional clinical variables, and validating these findings across broader and more diverse populations to further enhance clinical applicability and support evidence-based decision-making in CKD management.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation
Supplement1
