Development and validation of an inflammatory index-based interpretable machine learning model for mortality risk stratification in hemodialysis patients

Abstract

Background

Hemodialysis patients with end-stage renal disease have high all-cause mortality, with chronic low-grade inflammation as a key prognostic factor. Existing mortality prediction models lack both accuracy and clinical interpretability, and no studies have systematically integrated multiple inflammatory indices (neutrophil-to-lymphocyte ratio, monocyte-to-lymphocyte ratio, etc.) into interpretable tools for this population.

Methods

A single-center retrospective cohort study included 512 hemodialysis patients (Jan 2021–Oct 2024) from The Central Hospital of Wuhan, split into training (70%), validation (15%), and test (15%) sets per TRIPOD guidelines. Fifteen baseline clinical variables and five inflammatory indices were collected. Missing data were imputed, data normalized, and oversampling used to address imbalance. Twelve models (9 traditional machine learning, 1 neural network, 2 ensembles) were built, optimized via tenfold cross validation, and interpreted with SHapley Additive exPlanations.

Results

At follow-up (Oct 30, 2024), 212 (41.4%) patients died. Non-survivors differed significantly from survivors in myocardial infarction (16.0% vs. 2.7%, p < 0.001), neutrophil-to-lymphocyte ratio (median: 4.1 vs. 3.6, p = 0.012), dialysis vintage (42.5 vs. 77.5 months, p < 0.001), and age (62.2 ± 12.3 vs. 58.1 ± 13.3 years, p < 0.001). The Stacking model performed best (AUC = 0.983, accuracy = 0.922), outperforming logistic regression (AUC = 0.703). Body mass index, myocardial infarction history, and neutrophil-to-lymphocyte ratio were top predictors.

Conclusions

The interpretable stacking model enables accurate mortality risk stratification for hemodialysis patients. Future multi-center validation and multi-modal data integration will enhance its generalizability for clinical application.

Introduction

End-stage renal disease (ESRD) represents the terminal phase of chronic kidney disease (CKD), distinguished by a glomerular filtration rate (GFR) falling below 15 ml per minute. This condition profoundly diminishes patients' quality of life and markedly elevates the risk of premature mortality [1]. The management of ESRD primarily depends on renal replacement therapy, encompassing both dialysis and kidney transplantation. Dialysis is further categorized into hemodialysis and peritoneal dialysis, each modality presenting distinct benefits and limitations. Hemodialysis is a prevalent therapeutic option for individuals with ESRD, significantly contributing to life sustenance and enhancement of quality of life. Empirical studies have demonstrated that hemodialysis effectively facilitates the removal of toxins and excess fluids from the body, thereby ameliorating patients' health conditions to a measurable degree [2]. However, hemodialysis is also accompanied by a series of complications and challenges. Chronic low-grade inflammation in hemodialysis patients has been substantiated by numerous studies. This condition is not only a prevalent complication of hemodialysis but is also significantly linked to a range of adverse clinical outcomes. Research indicates that this inflammatory state in hemodialysis patients is associated with several factors, including uremic toxins, oxidative stress, and immune system dysfunction [3, 4]. The buildup of uremic toxins in dialysis patients is a key cause of microinflammation. Research indicates that medium-molecular-weight uremic toxins, like β2-microglobulin, are strongly linked to cardiovascular diseases [5]. Furthermore, the accumulation of uremic toxins contributes to endothelial cell dysfunction, thereby intensifying the inflammatory response. While dialysis, particularly high-flux dialysis and hemofiltration techniques, facilitates the removal of certain uremic toxins, some toxins remain challenging to eliminate. This persistence may partly account for the ongoing state of microinflammation [6, 7]. Oxidative stress constitutes a significant contributor to microinflammation observed in patients undergoing dialysis. The dialysis procedure markedly elevates oxidative stress levels, which in turn leads to an augmented presence of inflammatory markers, including C-reactive protein (CRP) and interleukin-6 (IL-6). This oxidative stress not only directly triggers inflammatory responses but also intensifies the inflammatory condition by damaging cellular membranes and DNA[8–10]—a factor that further highlights the importance of investigating how inflammatory status impacts the prognosis of hemodialysis patients.

Alataş et al. demonstrated that antioxidant capacity and anthropometric markers directly reflect inflammatory load in hemodialysis patients, [11] strengthening the rationale for incorporating inflammatory indices into our predictive model. In addition, Arslan linked cardiometabolic risk, body composition, and inflammation—findings that parallel our top SHAP predictors (BMI, myocardial infarction history, NLR).[12] These studies underscore the multifaceted nature of inflammation and its interplay with nutritional status and cardiovascular health in hemodialysis patients, further supporting the need to systematically integrate inflammatory indices into mortality prediction models.

The influence of inflammatory status on the prognosis of patients undergoing hemodialysis constitutes a significant area of research. Numerous studies have demonstrated a strong correlation between inflammatory markers and both the survival rates and cardiovascular events in dialysis patients. For instance, one study identified a significant association between elevated levels of high-sensitivity C-reactive protein (hs-CRP) and decreased survival rates among dialysis patients. Furthermore, in individuals who are current or former smokers, tumor necrosis factor-α (TNF-α) levels below a certain threshold have been linked to improved survival outcomes [13]. A separate study identified the Systemic Inflammatory Response Index (SIRI) as an effective prognostic indicator for diabetic patients receiving maintenance hemodialysis [14]. Monitoring and managing inflammation is vital for improving the prognosis of hemodialysis patients. Early identification and intervention can enhance survival rates and reduce cardiovascular events, highlighting the need for attention to inflammatory markers in clinical practice. Inflammatory markers, including the neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), and systemic immune-inflammation index (SII), are predominantly utilized in cardiovascular research, yet their application in the context of hemodialysis remains comparatively limited [14].

In contrast, integrating multiple inflammatory indices offers distinct advantages: it captures the multidimensional nature of systemic inflammation (e.g., neutrophil-mediated acute inflammation and monocyte-mediated chronic inflammation), mitigates the impact of measurement variability, and enhances the robustness of prognostic assessments. For instance, Ye et al. developed an inflammatory composite score incorporating the neutrophil-to-lymphocyte ratio (NLR), monocyte-to-lymphocyte ratio (MLR), and platelet-to-lymphocyte ratio (PLR) in maintenance hemodialysis (MHD) patients [15]. Their analysis revealed that the high-score group (score = 3) was independently associated with an increased risk of all-cause mortality (hazard ratio [HR] = 4.562, 95% confidence interval [CI] = 1.342–15.504, P = 0.015). Notably, this composite model exhibited significantly superior prognostic performance compared to any single inflammatory marker, effectively avoiding the limitation of single indices (e.g., C-reactive protein) that are prone to interference from acute stress. Further supporting this, Li et al. conducted a study incorporating the systemic immune-inflammation index (SII) and demonstrated that the multi-index combination (SII + NLR + MLR) yielded a mean increase of 0.13 in the area under the receiver operating characteristic curve (AUC) for predicting mortality in MHD patients (AUC = 0.78 for the combination vs. 0.65 for NLR alone and 0.63 for MLR alone) [16]. The core advantage of this integrated approach lies in its ability to simultaneously capture innate immune activation and immune cell imbalance—key pathological features of persistent inflammation in end-stage kidney disease (ESKD) that single inflammatory indices often fail to fully encompass.

In recent years, the utilization of machine learning techniques in the management of hemodialysis patients has advanced considerably, particularly in the prediction and management of complications associated with dialysis. By analyzing clinical data from patients, machine learning models have demonstrated efficacy in forecasting acute adverse events during hemodialysis, such as life-threatening arrhythmias and refractory intradialytic hypotension. This capability facilitates timely clinical interventions and the optimal allocation of healthcare resources [17]. Furthermore, machine learning is employed to predict post-dialysis fatigue (PDF). These models are capable of identifying key factors associated with PDF and offer more intuitive explanations through the SHapley Additive exPlanations (SHAP) method. This approach aids clinicians in gaining a deeper understanding of the decision-making process [18].

Consequently, this research utilizes clinical data from patients alongside validation indices to develop an interpretable machine learning model specifically tailored for hemodialysis patients. The objective is to construct a predictive model for mortality risk in this patient population. This model is intended to assist clinicians in evaluating the impact of patients' micro-inflammatory status on their prognosis and to facilitate the formulation of personalized treatment plans informed by the model's decision-making framework. The findings are presented as follows.

Methods

Study population

This study was a single-center retrospective cohort study, which included 512 patients who underwent hemodialysis at The Central Hospital of Wuhan from January 1, 2021 to October 2024. The data sources were all from the electronic medical record system. In this study, patients were randomly allocated into a training set (70%), a validation set (15%), and a test set (15%). The model's development and validation adhered to the guidelines outlined in the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) statement. [19] This study was approved by the Ethics Committee of The Central Hospital of Wuhan (No.WHZXKYL2024-115). Given that it is a retrospective study, the requirement for informed consent was waived. The study strictly adhered to the Helsinki Declaration.

Inclusion criteria

Subjects included in this study were required to meet all the following conditions: (1) aged 18 years or older; (2) receiving dialysis treatment no fewer than 2 times per week; and (3) each dialysis session lasting no less than 3 h.

Exclusion criteria

Subjects were excluded from this study if they met any of the following conditions: (1) receiving dialysis via a temporary catheter; (2) the proportion of missing clinical data exceeding 30%; and (3) having left the research institution with no access to data regarding their.

Data collection

The baseline data for this study were obtained from the electronic medical record system of our hospital and the hospital’s dedicated hemodialysis management database. A total of 15 fundamental patient information items were collected, including demographic characteristics (gender, age, education level), etiological diagnosis of renal disease, comorbidities (hypertension, diabetes, secondary hyperparathyroidism, hyperphosphatemia, heart failure, cerebral infarction, myocardial infarction(MI)), dialysis-related indicators (dialysis frequency, dialysis vintage, vascular access type(VAT)), and body mass index(BMI).

Among these, dialysis vintage was specifically defined as the time interval from the patient’s first session of maintenance hemodialysis (the initiation of regular hemodialysis treatment as per clinical guidelines) to either the study’s follow-up endpoint (October 30, 2024) or the patient’s date of death if it occurred before this endpoint.

The selection of dialysis vintage as a key dialysis-related indicator was based on two practical considerations: first, due to the retrospective nature of this study, complete documentation of the initial CKD diagnosis date (required for calculating CKD duration) and standardized records of the ESRD confirmation timeline (required for consistent ESRD duration calculation) were unavailable for 38.7% of patients (n = 198). This was primarily because a subset of patients was transferred from external primary care clinics or other hospitals, where historical diagnostic data were not fully linked to our hospital’s electronic medical record system. Second, dialysis vintage is a clinically validated surrogate for long-term renal disease burden: it directly reflects the cumulative exposure of patients to hemodialysis therapy and is closely associated with prognostic factors (e.g., dialysis-related inflammatory activation, vascular access complication risk, nutritional adaptation) that overlap with the potential impacts of CKD/ESRD duration.

In addition, five inflammatory indices were calculated utilizing biochemical test data from the first month of the patient’s hemodialysis enrollment in our hospital, including neutrophil-to-lymphocyte ratio, monocyte-to-lymphocyte ratio, systemic inflammatory response index, systemic immune-inflammation index, and aggregate index of systemic inflammation. The relevant calculation formulas are as follows:

NLR = neutrophil count (Neu)/lymphocyte count (Lym), calculated as neutrophil count divided by lymphocyte count. NLR is a well-validated marker for mortality in hemodialysis patients, with Neuen et al.[20] demonstrating its utility in predicting cardiovascular and all-cause mortality, and Li et al.[21] further linking elevated NLR to increased cardiovascular mortality risk in this population.

MLR = monocyte count (Mon)/lymphocyte count (Lym), defined as monocyte count divided by lymphocyte count. MLR captures monocyte-mediated chronic inflammation and lymphocyte dysfunction, two critical pathological processes in hemodialysis patients. Chen et al.[22] validated its association with all-cause mortality in a multicenter study involving 2387 hemodialysis patients, supporting its inclusion in our analysis.

SIRI = (neutrophil count (Neu) × monocyte count (Mon))/lymphocyte count (Lym), computed as (neutrophil count × monocyte count)/lymphocyte count. Han et al. [14] confirmed that SIRI predicts major adverse cardiovascular events (MACE) in maintenance hemodialysis patients with pre-existing cardiovascular disease. Given that MACE is a leading cause of death in hemodialysis populations, SIRI aligns with our study’s focus on mortality risk stratification.

SII = (neutrophil count (Neu) × platelet count (Plt))/lymphocyte count (Lym), calculated as (neutrophil count × platelet count)/lymphocyte count. SII integrates inflammatory and coagulation markers—key inter-related pathological factors in hemodialysis patients. Zhu et al. [23] validated its value in predicting cardiovascular events in this population, justifying its inclusion as a comprehensive inflammatory metric.

AISI = (neutrophil count (Neu) × platelet count (Plt) × monocyte count (Mon))/lymphocyte count (Lym), defined as (neutrophil count × platelet count × monocyte count)/lymphocyte count. AISI expands on SII by incorporating monocyte counts, enhancing the quantification of multi-dimensional inflammatory responses. While no study has specifically validated AISI in hemodialysis patients, its design adheres to the evidence-based framework of SII [23] and MLR [22]—two indices already validated in hemodialysis populations—supporting its use for comprehensive inflammatory burden assessment.

Outcome measures

The outcome measure of this study was patient mortality, with the follow-up endpoint being October 30, 2024.

Data processing

The original data set included a wide range of clinical variables. For missing values, imputation was conducted using the Mice package to address this issue. Following this step, baseline indicators were retained, and inflammatory biomarkers were calculated individually for each patient. Given the significant variability in feature values across the data set, data normalization was implemented to standardize the data. To prevent overfitting of the model, feature engineering was adopted for feature selection; meanwhile, oversampling techniques were applied to correct the imbalance in the data set. Model optimization was carried out using a tenfold cross-validation strategy coupled with a grid search method to identify the optimal parameter combinations. To guarantee the predictive efficacy of the model, feature selection, normalization, and standardization procedures were implemented solely within the confines of the training data set. All continuous features (e.g., BMI, NLR, and dialysis vintage) were standardized using Z-score normalization (mean = 0, standard deviation = 1) during preprocessing, ensuring consistent scaling across features. Categorical features (e.g., myocardial infarction history and vascular access type) were one-hot encoded, ensuring they were treated consistently with continuous features in the SVM model (which is sensitive to feature scale).

Statistical analysis

Data were collected with Microsoft Excel and processed in Python 3.13. Categorical variables are shown as counts (percentages) and analyzed with chi-square or Fisher's exact tests. Continuous variables are presented as mean ± standard deviation and analyzed with an independent samples t test if normally distributed, or the Mann–Whitney U test otherwise. Missing data were under 5%; multiple imputation was used for numerical variables and mode imputation for categorical ones, assuming data were missing at random.

Model development

The data set comprising 512 patients was partitioned into training and validation subsets using a 7:1.5:1.5 ratio.

Total number of models: the cumulative total of the aforementioned three categories amounts to 12 models. Model quantity statistics (categorized by type): (1) traditional machine learning models: this category comprises nine models, specifically logistic regression (LR), elastic net, decision tree (DT), random forest (RF), gradient boosting machine (GBM), XGBoost, LightGBM, support vector machine (SVM), and Naive Bayes; (2) neural network model: this category includes 1 model, specifically the artificial neural network (ANN); and (3) ensemble learning strategies: this category encompasses two models, namely, weighted voting and stacking.

Model interpretation

To address the complexity of machine learning algorithms, SHAP (SHapley Additive exPlanations) values were utilized in this study to elucidate the models. Through the visualization of individual feature contributions, we sought to improve clinicians' understanding of how these features influence outcomes Table 1. SHAP values quantify the predictive contribution of each feature to the model’s output (mortality risk prediction) by decomposing the model’s decision-making process. They do not establish direct biological causality between features and outcomes. Due to technical limitations in computing stable SHAP values for the heterogeneous Stacking ensemble, we conducted interpretability analysis using the top-performing base model with reliable SHAP support (SVM, AUC = 0.981).

Table 1.

Comparison of baseline characteristics between the two groups of patients

Features	Death group(n = 212)	Survival group(n = 300)	p value
Gender (%)			0.127
Male	150(70.8)	192(64.0)
Female	62(29.2)	108(36.0)
Etiological diagnosis (%)			0.067
Chronic glomerulonephritis	40(18.9)	52(17.13)
Diabetic nephropathy	55(25.9)	49(16.13)
Hypertensive renal damage	97(45.8)	169(56.4)
Polycystic kidney	8(3.8)	8(2.7)
Henoch–Schoenlein purpura nephritis	0	2(0.7)
Hereditary nephropathy	0	2(0.7)
Nephrotic Syndrome	0	2(0.7)
Others	12(5.7)	16(5.3)
Education Level (%)			< 0.001
Illiterate	44(20.8)	30(10.0)
Primary education	38(17.9)	52(17.3)
Junior high school	40(18.9)	96(32.0)
High school	38(17.9)	54(18.0)
Three year college	8(3.8)	22(7.3)
Undergraduate and above	44(20.8)	46(15.3)
Vascular access type (%)			< 0.001
Tunneled cuffed catheter	64(30.2)	44(14.7)
Arteriovenous fistula	148(69.8)	256(85.3)
Hypertension (%)			0.196
No	28(13.2)	28(9.3)
Yes	184(86.2)	272(90.7)
Diabetes (%)			0.043
No	134(63.2)	216(72.0)
Yes	78(36.8)	84(28.0)
Secondary hyperparathyroidism (%)			0.016
No	166(78.3)	206(68.7)
Yes	46(21.7)	94(31.3)
Hyperphosphatemia (%)			< 0.001
No	178(84.0)	210(70.0)
Yes	34(16.0)	90(30.0)
Heart failure (%)			0.296
No	136(64.2)	206(68.7)
Yes	76(35.8)	84(31.3)
Gout (%)			0.49
No	202(95.3)	290(96.7)
Yes	10(4.7)	10(3.3)
Cerebral infarction (%)			0.817
No	172(81.1)	246(82.0)
Yes	40(18.9)	54(18.0)
Myocardial infarction (%)			< 0.001
No	178(84.0)	292(97.3)
Yes	34(16.0)	8(2.7)
NLR (median [IQR])	4.1 [3.0, 5.9]	3.6 [2.7, 5.5]	0.012
MLR (median [IQR])	0.4 [0.3, 0.6]	0.4 [0.3, 0.5]	0.109
SIRI (median [IQR])	1.9 [1.3, 3.1]	1.6 [1.1, 2.4]	0.017
SII (median [IQR])	803.9 [492.1, 1147.5]	684.3 [455.9, 1084.9]	0.08
AISI (median [IQR])	360.7 [221.3, 632.7]	304.1 [184.6, 515.7]	0.068
Dialysis frequency (median [IQR])	3.0 [3.0, 3.0]	3.0 [2.0, 3.0]	0.056
Dialysis vintage (months)(median [IQR])	42.5 [29.0, 71.0]	77.5 [54.0, 104.0]	< 0.001
BMI (kg/m2)(median [IQR])	22.1 [20.7, 23.5]	22.3 [20.0, 24.5]	0.233
Age (years)	62.2 ± 12.3	58.1 ± 13.3	< 0.001

Results

Patient characteristics

A total of 614 patients were recruited for this study (Fig. 1), of whom 512 were ultimately included. The study population comprised 300 survivors and 212 non-survivors. The two patient groups differed significantly in Education Level, Vascular Access Type, Diabetes, Secondary Hyperparathyroidism, Hyperphosphatemia, NLR, SIRI, Dialysis Vintage, and Age (p < 0.05).Of the 512 included patients, 212 (41.4%) died during follow-up. The time from baseline predictor assessment to death ranged from 3.2 to 46.8 months (mean: 19.7 ± 11.5 months), confirming that outcomes occurred after baseline measurements rather than concurrently.

Fig. 1.

Flowchart of patient enrollment and data set partition

Characteristics of variables

To visualize pairwise associations, a feature correlation bubble heatmap (Fig. 2) was generated, where color intensity (ranging from deep blue to deep red) denoted the direction of Pearson correlations (blue = negative, red = positive), and bubble size represented the magnitude of correlation (larger bubbles = stronger |r|). Strong positive correlations were observed among inflammatory indices (e.g., NLR, MLR, and SII) and between metabolic conditions (HTN and DM) and cardiovascular comorbidities (HF, CI, and MI). In contrast, etiological diagnosis and BMI showed weaker or inverse (blue-toned) correlations with most metabolic and inflammatory parameters.

Fig. 2.

Feature correlation bubble heatmap (20 variables, Pearson)

The results of machine learning

A total of 12 models were constructed and their performance was assessed using metrics, such as area under the curve, accuracy, precision, recall, and F1 score. Based on these evaluations, two ensemble models were subsequently developed utilizing distinct methodologies, namely, weighted voting and stacking. The performance evaluations of these models are detailed in Table 2. According to the data presented, the SVM emerges as the most effective individual model, attaining an AUC of 0.981 (Fig. 3), the Stacking ensemble model surpasses this performance, achieving an AUC of 0.983, and exhibits outstanding results across various other evaluative metrics. To ensure internal validity, we implemented strict data set partitioning (training: 70%, validation: 15%, test: 15%) with stratification by mortality status (to preserve event rate across sets) and tenfold cross-validation on the training set for hyperparameter optimization. The Stacking model achieved optimal performance on the test set (AUC = 0.983, accuracy = 0.922) as reported in the main results, and tenfold cross-validation on the training set confirmed consistent predictive stability (mean performance metrics aligned closely with the test set). This consistency—reflected in minimal fluctuations across cross-validation folds and alignment with the independent test set—indicates the model captures reproducible risk signals (e.g., the consistent prognostic value of BMI, NLR, and myocardial infarction) rather than center-specific noise, laying a stable groundwork for external validation.

Table 2.

Performance metrics results of different models

Models	AUC	Accuracy	Precision	Recall	F1 Score
Logistic regression (LR)	0.703	0.636	0.545	0.750	0.632
Elastic Net	0.697	0.675	0.585	0.750	0.658
Decision tree (DT)	0.822	0.753	0.710	0.688	0.698
Random forest (RF)	0.930	0.896	0.833	0.933	0.882
Gradient boosting (GBM)	0.971	0.876	0.853	0.936	0.891
XGBoost	0.953	0.885	0.823	0.938	0.882
LightGBM	0.956	0.942	0.887	0.978	0.909
SVM	0.981	0.883	0.811	0.938	0.870
Naive Bayes	0.633	0.610	0.531	0.531	0.531
Neural network (ANN)	0.896	0.909	0.838	0.969	0.899
Weighted voting	0.926	0.921	0.875	0.914	0.899
Stacking	0.983	0.922	0.882	0.938	0.909

Fig. 3.

ROC–AUC curves of 12 prediction models

To effectively illustrate the predictive performance of the various models, confusion matrices were employed to display their classification outcomes (Fig. 4). In general, ensemble strategies, such as stacking and weighted voting, along with advanced tree-based models, such as XGBoost and LightGBM, exhibited a lower rate of misclassification in comparison with traditional models, such as logistic regression.

Fig. 4.

Confusion matrices of all models

Visualization of feature importance

To ensure an objective interpretation of the selected variables, we utilized SHAP to clarify the contribution of these variables to the mortality risk among hemodialysis patients within the model. Fig. 5 presents the risk factors that contribute most significantly to the model, as identified by the mean absolute SHAP values. To identify the primary predictors for the SVM model, we visualized the top 15 features ranked by their importance (see Fig. 5).BMI was identified as the most influential feature, with an importance score of 0.070, closely followed by myocardial infarction at 0.069, and the NLR at 0.066. It is noteworthy that inflammatory and anthropometric indices, such as BMI, myocardial infarction, and NLR, along with vascular access type (VAT, 0.065), demonstrated greater importance than demographic variables (e.g., gender, 0.061) and metabolic variables [e.g., diabetes mellitus (DM), 0.060]. This highlights the significant role of systemic inflammation and clinical phenotypes in enhancing the predictive performance of the SVM model.

Fig. 5.

Feature importance ranking (SVM and SHAP)

To understand the SVM's predictive mechanism, we used SHAP and created a SHAP dot plot (Fig. 6).The plot's horizontal axis shows SHAP values, indicating a feature's contribution to the model output (SHAP > 0 favors the "positive" class, SHAP < 0 favors the "negative" class). The vertical axis lists features, and dot colors represent feature values (red for high, blue for low).

Fig. 6.

SHAP dot plot for key predictive features (SVM)

Inflammatory-related indices demonstrated a significant impact on the predictive model. NLR displayed distinct clustering patterns, with high NLR values (represented by red dots) being associated with positive SHAP values, thereby driving predictions towards the target class. Conversely, low NLR values (represented by blue dots) were linked to negative SHAP values. Similarly, myocardial infarction, a condition closely associated with systemic inflammation, and the MLR exhibited clear directional effects. Elevated values of these inflammation-related features (red dots) tended to result in positive SHAP values, favoring target-class predictions, whereas lower values (blue dots) were associated with negative SHAP values. These findings highlight the critical role of inflammatory indices in influencing the predictive behavior of the SVM, thereby affirming their clinical relevance to the modeled outcome.

To analyze individual predictions of the SVM model, we used SHAP force plots (Fig. 7). In the first sample (top panel), dialysis vintage (DV, SHAP = −1.507) and BMI, SHAP = −1.083) had significant negative impacts, while gender (SHAP = 0.688) had a positive effect. In the second sample (bottom panel), monocyte-to-lymphocyte ratio (MLR, SHAP = 1.166) and cerebral infarction (CI, SHAP = 2.104) made strong positive contributions, emphasizing the importance of inflammatory pathways in the model's prediction.

Fig. 7.

SHAP force plots for individual mortality predictions (SVM)

Discussion

In this study, we utilized a total of ten individual models alongside two ensemble models. The findings demonstrated that, among the individual models, the SVM exhibited superior performance. Furthermore, both ensemble models showed outstanding efficacy across multiple metrics, suggesting potential practical applications in clinical settings. For visualization and interpretive purposes, we employed the SHAP method to quantify the predictive contribution of each feature to the model’s mortality risk output. The feature importance ranking plot (based on mean absolute SHAP values) identified BMI, myocardial infarction history, NLR, vascular access type, and gender as the top five features most strongly associated with the model’s predictive performance. These results align with extensive clinical evidence linking these factors to hemodialysis patient prognosis, supporting the biological plausibility of their predictive roles—though it is important to emphasize that SHAP-derived feature importance reflects associations within the model rather than establishing direct biological causality.

The SVM model has been widely applied in the field of kidney diseases, playing an important role in the early detection, diagnosis, and personalized treatment of diseases. In recent years, with the rapid development of machine learning technology, the SVM model has received extensive attention and application in the research and clinical practice of kidney diseases. The SVM model excels in early kidney disease detection and diagnosis. Research indicates that combining SVM with feature selection methods significantly enhances diagnostic accuracy for CKD. Utilizing a filter subset evaluator and best-first search engine, the SVM model achieves a 98.5% accuracy rate in CKD diagnosis [24].

In this study, the ensemble models (Weighted Voting and Stacking) demonstrated significantly superior performance in predicting mortality risk compared to single models: stacking ranked first among all models with an AUC of 0.983, and together with Weighted Voting, they achieved the highest accuracy of 0.922 and the highest precision of 0.882 while maintaining a high recall rate of 0.938 and a stable F1 score of 0.909. This performance not only far surpassed that of traditional linear models (such as logistic regression with an AUC of 0.703 and elastic net with an AUC of 0.697) and shallow tree models (Decision Tree with an AUC of 0.822), but also outperformed most advanced tree-based models (such as LightGBM with an AUC of 0.956 and XGBoost with an AUC of 0.953), and even slightly exceeded the optimal single model SVM (with an AUC of 0.981). Its core advantage lies in effectively reducing prediction bias and variance by integrating the complementary capabilities of different single models (such as the capture of linear relationships by linear models and the fitting of nonlinear interactions by tree-based models), [25, 26] achieving balanced optimization of multiple evaluation metrics, and providing a more reliable tool for clinically accurately identifying high-risk patients, avoiding missed diagnoses and excessive interventions.

The influence of dialysis access, NLR, cardiovascular diseases, and BMI on the mortality rates of hemodialysis patients constitutes a multifaceted subject necessitating examination from diverse perspectives. Notably, cardiovascular disease represents a primary cause of mortality among hemodialysis patients. Inflammatory markers, particularly the NLR, have demonstrated a significant correlation with the prediction of cardiovascular events and overall mortality [22, 23].

NLR serves as a straightforward and readily accessible marker of inflammation and demonstrates a correlation with the Erythropoietin Resistance Index (ERI), thereby suggesting its potential utility in the management of anemia [27]. NLR, an indicator of inflammatory processes, demonstrates a significant correlation with other biochemical markers, including C-reactive protein.[20] In patients undergoing hemodialysis, an elevated NLR is correlated with increased levels of cardiovascular risk indicators, including pulse pressure (PP), left ventricular mass index (LVMI), and intima–media thickness (IMT) [21]. Beyond the NLR, additional inflammation-related biomarkers, including white blood cell count (WBC) and 8-hydroxy-2-deoxyguanosine (8-OHdG), have been demonstrated to correlate with mortality risk among hemodialysis patients [28, 29]. These studies further underscore the significant role of inflammation in influencing mortality risk among dialysis patients and highlight the potential value of utilizing multiple biomarkers in combination. NLR is a crucial inflammatory marker for predicting mortality risk in hemodialysis patients. Combined with other biomarkers, it enhances risk assessment, aiding clinicians in developing better treatment strategies. This research highlights inflammation's role in dialysis and suggests new research avenues.

Furthermore, nutritional status and BMI are critical determinants influencing the prognosis of patients undergoing hemodialysis. Research indicates that the concomitant presence of a low BMI and an elevated waist-to-hip ratio substantially heightens the risk of mortality from all causes, including cardiovascular diseases [30]. In a study conducted on Korean patients undergoing hemodialysis, it was observed that the mortality risk was significantly elevated in the low BMI cohort compared to the normal BMI cohort, whereas the mortality risk was reduced in the overweight and obese cohorts [31]. In addition, changes in BMI are also related to mortality, especially in patients with a decrease in BMI during the initial stage of dialysis, where mortality increases significantly, and Arslan’s study demonstrated same trend [12, 32]. Patients with low BMI often have low levels of serum albumin and prealbumin. These indicators are associated with malnutrition and inflammatory status, thereby increasing the mortality rate [33]. Although serum albumin was not measured, BMI aligns with literature showing nutritional markers correlate with inflammatory load [11].

This study observed a close interaction between inflammation and nutritional status through a correlation bubble heatmap: in the heatmap, there was a clear correlation trend between inflammatory indices (such as NLR and SII) and nutrition-related indices (such as BMI and albumin), and this correlation further affected patient mortality risk stratification. This mechanism can be confirmed by the study of Alataş et al.[11]—their team found in HD patients that total dietary antioxidant capacity (reflecting nutritional quality) was significantly negatively correlated with inflammatory load (such as CRP and IL-6), and that nutritional intervention could modulate inflammation levels. Combining the results of this study, it is known that the “inflammation–nutrition association” in HD patients is not a simple accompanying relationship, but there may be a bidirectional regulatory pathway. This provides a theoretical basis for clinically improving inflammation levels by optimizing nutritional status and also strengthens the clinical significance of incorporating both inflammatory and nutritional indices into the prediction model in this study. In summary, a low BMI in hemodialysis patients is linked to negative factors, such as malnutrition and inflammation, highlighting the importance of maintaining proper weight and nutrition for better outcomes.

Our findings on inflammatory indices in HD patients are linked to the oxidative stress–antioxidant capacity–inflammation axis. Hemodialysis increases oxidative stress through immune cell activation and mitochondrial dysfunction, leading to inflammation and reduced antioxidant defenses. Dietary antioxidants can break this cycle by neutralizing ROS and reducing inflammation [34–36]. Although we did not measure oxidative stress directly, this mechanism supports the biological plausibility of our observed links between inflammatory indices and cardiometabolic outcomes.

The choice and management of dialysis access significantly influence patient survival outcomes. The incidence of catheter-related infections is correlated with factors, such as low hemoglobin levels, the presence of diabetes, and the duration of catheter use, each of which can elevate the risk of patient mortality [37]. A meta-analysis has indicated that the catheter infection rate among hemodialysis patients can reach up to 11.7%,Patients undergoing hemodialysis exhibit an elevated risk of developing catheter-related infections if they present with diabetes, reduced hemoglobin levels, advanced age, prolonged catheterization duration, diminished serum albumin levels, utilization of a femoral vein catheter, or an increased number of catheters [38]. Our measured variable (vascular access type) aligns with literature linking tunneled cuffed catheters to infection risk [37, 38], a key driver of mortality in hemodialysis patients.

Following arteriovenous fistula surgery, there is a marked increase in the diameter of the vein and the blood flow within the brachial artery over a 12-week postoperative period. This augmentation in blood flow results in an elevated cardiac output, as the heart accommodates the increased volume of returning blood. Such an increase in cardiac output represents an adaptive cardiac response to the augmented blood volume load. This heightened cardiac load is not only evident in the alteration of cardiac output but may also manifest as an enlargement in the end-diastolic diameter of the left ventricle. By adjusting the rotational speed of a continuous-flow left ventricular assist device (LVAD) to synchronize with the cardiac cycle, it is possible to modulate the end-diastolic volume (EDV) of the left ventricle, thereby influencing the cardiac load [39, 40].

In summary, based on a single-center cohort of 512 hemodialysis patients and an interpretable stacking model (AUC = 0.983), this study clearly identifies the core interactive mechanisms driving mortality risk in this population: inflammatory indices represented by NLR and SIRI amplify systemic inflammatory burden through the synergy of neutrophil-mediated acute inflammation and monocyte-driven chronic inflammation. This inflammatory state forms a bidirectional amplification loop with myocardial infarction (incidence: 16.0% in the non-survivor group vs. 2.7% in the survivor group)—where local myocardial injury exacerbates systemic inflammation and vice versa—while being implicitly modulated by BMI (BMI < 22 kg/m² amplifies the pro-inflammatory effects of NLR). Furthermore, these inflammatory processes interact with dialysis-related factors (shorter dialysis vintage: 42.5 months vs. longer vintage: 77.5 months; higher proportion of tunneled cuffed catheters: 30.2% vs. 14.7%) to establish a vicious cycle of "inadequate dialysis → inflammatory activation → organ damage". Collectively, these multidimensional interactions (inflammation–cardiovascular–nutrition dialysis) not only serve as key drivers of mortality risk but also explain why the ensemble model outperforms traditional single models. This mechanistic insight highlights that clinical prognosis management for hemodialysis patients must move beyond single-dimensional interventions and instead simultaneously address inflammation control (e.g., monitoring NLR and SIRI), cardiovascular protection (e.g., preventing myocardial infarction), nutritional maintenance (e.g., avoiding low BMI), and dialysis access optimization (e.g., prioritizing arteriovenous fistula). The interpretable model developed in this study thus provides a reliable tool for such multidimensional risk stratification and targeted clinical interventions.

This study has several limitations that should be noted. First, specific nutritional markers (e.g., serum albumin and dietary antioxidant capacity)—which Alataş et al. linked to inflammatory load in hemodialysis (HD) patients [11]—were not included. This is due to the retrospective design: our electronic health record system only routinely collects BMI as a nutritional proxy, and detailed dietary intake or serial albumin data were not systematically recorded. This gap may have hindered full elucidation of the nutrition–inflammation bidirectional relationship and validation of dietary antioxidants’ modulatory role.

Second, detailed body composition measures (e.g., visceral fat area and skeletal muscle mass)—identified by Arslan as independent mortality predictors in HD cohorts [12]—were unavailable. BMI, while accessible, cannot distinguish between fat and muscle mass (with divergent effects on inflammation and cardiometabolic risk), and limited access to DXA/BIA data may have introduced residual confounding from heterogeneous body composition effects.

Third, the single-center, retrospective design poses three key constraints: ① it limits the model’s external validity (as local clinical practices, patient demographics, and comorbidity profiles may differ across institutions); ② it precludes causal inference on the inflammation–nutrition–cardiometabolic pathway; and ③ the absence of Kt/V (a core indicator of dialysis adequacy) and external multi-center validation raises concerns. Kt/V directly reflects uremic toxin clearance efficiency, and its omission stems from incomplete records of urea clearance rate and dialysis duration in early enrolled patients, as well as heterogeneous dialysis modes that complicate uniform interpretation. Without external validation, we cannot rule out the possibility that the model’s high performance (AUC = 0.983) partly stems from overfitting to center-specific patterns rather than universal mortality risk drivers in HD patients.

Conclusion

In the present study, to develop a predictive model for the mortality risk of hemodialysis patients with end-stage renal disease, we conducted a single-center retrospective cohort study enrolling 512 hemodialysis patients, divided them into training (70%), validation (15%), and test (15%) sets, and constructed 12 models (9 traditional machine learning models, 1 artificial neural network, and 2 ensemble models: weighted voting and stacking). We used SHAP to interpret feature contributions and comprehensively evaluated model performance via metrics, such as AUC, accuracy, and precision. The Stacking ensemble model showed the best performance (AUC = 0.983), outperforming single models and traditional algorithms, with BMI, myocardial infarction (MI), and NLR identified as the top predictive features. Reasonable use of this interpretable model can assist in high-precision mortality risk stratification of hemodialysis patients, support clinicians’ decision-making, save time, and reduce misjudgment. In the future, we will conduct multi-center external validation to enhance the model’s generalizability, integrate multi-modal data (e.g., longitudinal dialysis records) with advanced algorithms, and optimize solutions for data imbalance to further improve the model’s comprehensiveness and applicability.

Author contributions

Each author made substantial contributions to this work: all participated in the conception or design of the study, collection, processing, or interpretation of data; took part in drafting the manuscript or critically revising it for substantive intellectual content; provided final approval of the published version; and agree to be accountable for the accuracy and integrity of all parts of the work.

Funding

This study was not supported by any grants.

Data availability

The datasets during and/or analysed during the current study available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of The Central Hospital of Wuhan (approval number WHZXKYL2024-115). Given the retrospective nature of the study, the committee waived the requirement for informed consent.

Competing interests

The authors declare no competing interests.