Endothelial activation and stress index associated with in-hospital mortality risk in patients with end-stage renal disease: a retrospective analysis based on the MIMIC database and machine learning model development.

Background

End-stage renal disease (ESRD) involves irreversible, severe declines in kidney function, which is fatal without kidney transplantation or dialysis treatment to sustain long-term survival [ 1 , 2 ]. ESRD is classified as Stage 5 in the National Kidney Foundation’s Kidney Disease Outcomes Quality Initiative classification for chronic kidney disease (CKD), defined as a glomerular filtration rate (GFR) less than 15 ml per minute per 1.73 square meters of body surface area, and the necessity for renal replacement therapy [ 1 ]. Numerous risk factors influence the onset and progression of ESRD and CKD, including inflammation, hypertension, elevated fasting blood glucose levels, hyperlipidemia, obesity, proteinuria, and cardiovascular disease (CVD). Additionally, lifestyle factors, such as insufficient physical activity, excessive nutrition, and alcohol consumption, are significant contributors to CKD progression, particularly among individuals aged 65 years and older [ 3 – 5 ]. Decline and loss of kidney function can lead to a series of severe symptoms, including chronic inflammatory states stemming from infection, fluid retention caused by impaired water excretion and hormonal imbalances, anemia resulting from reduced erythropoietin production, hyperthyroidism, hyperphosphatemia, hypocalcemia, and osteoporosis due to disrupted bone metabolism, CVD, and progressive uremia [ 1 , 6 , 7 ]. Taking CVD as an example, most patients with ESRD or those who undergo kidney transplantation may die from CVD complications, such as left ventricular hypertrophy, myocardial infarction, and dyslipidemia. Moreover, the CVD mortality rate among ESRD patients is 20 to 40 times higher than that of the general population [ 1 , 8 , 9 ]. Worse still, even after kidney transplantation or dialysis treatment, surgical interventions significantly diminish the quality of life for ESRD patients, leading to adverse effects, such as pain, itching, fatigue, depression, malnutrition, and medication side effects [ 10 , 11 ]. The persistent rise in ESRD prevalence poses a formidable challenge worldwide, given its significant health and economic burdens [ 12 ]. Further investigation into the incidence and prognosis of ESRD is crucial for developing or refining corresponding medical policies and treatment plans, which are vital not only for the CKD population but for society.

The endothelial activation and stress index (EASIX) is a novel indicator for assessing and reflecting endothelial dysfunction. It is calculated by determining the ratio of three laboratory parameters: platelet count (PLT), creatinine, and lactate dehydrogenase (LDH) [ 13 , 14 ]. Previous articles have demonstrated the crucial role of endothelial dysfunction in the pathogenesis and pathophysiological processes of specific diseases. It can lead to non-relapse mortality in allogeneic stem cell transplantation (alloSCT) and complications, such as sepsis, graft-versus-host disease, sinusoidal obstruction syndrome, and transplant-associated microangiopathy. It is also considered a risk factor for sepsis following alloSCT and a major cause of mortality. EASIX integrates relevant indicators of microvascular and endothelial cell pathology, enabling accurate quantification of endothelial dysfunction and inflammation levels [ 14 , 15 ]. Moreover, elevated levels of EASIX have been demonstrated to be closely linked with the onset and adverse prognosis risks of CVD, stroke, and asthma [ 13 , 16 – 18 ]. Furthermore, studies on tumors and their immune complications, such as metastatic pancreatic cancer and refractory multiple myeloma complicated by cytokine release syndrome, have demonstrated EASIX’s potential for assessing tumor prognosis and stratifying disease risk prediction [ 19 , 20 ].

Extensive research indicates the key role of endothelial dysfunction and inflammation in ESRD and CKD [ 21 – 23 ]. For instance, endothelial damage induced by toxins or inflammatory cytokines, defects in endothelial protective factors, and the transition of renal endothelial cells toward an interstitial phenotype can contribute to CKD and renal fibrosis [ 22 ]. When endothelial dysfunction occurs, abnormalities in renal hemodynamics, systemic hypertension, platelet aggregation, abnormal proliferation of renal cells, and alterations in microvascular morphology and density during acute kidney injury (AKI) may all contribute to CKD progression [ 24 , 25 ]. The activation of the innate immune system induces sterile inflammation, which has become a driving factor in CKD and its high-mortality complication, CVD [ 26 ]. During ESRD progression, the accumulation of uremic toxins, oxidative stress, and dysregulation of immune signaling exacerbate inflammation and endothelial dysfunction, accelerating immune aging characterized by immune dysfunction and chronic low-grade inflammation, thus triggering a vicious cycle [ 27 ]. Given that EASIX reflects endothelial dysfunction and inflammatory states, although its association with prognosis and various diseases has been extensively studied, its predictive value for ESRD warrants further investigation.

In recent years, powerful computational methods in the field of machine learning (ML) have been gradually applied to disease prediction. This interdisciplinary field has opened up new avenues for medical research [ 28 ]. We developed a risk prediction model based on the EASIX score using data from the Medical Information Mart for Intensive Care-IV (MIMIC-IV). Due to the inherent “black box” nature of ML models, their internal data cannot be intuitively visualized. Therefore, we also calculated Shapley Additive exPlanation (SHAP) values to visualize the model and enhance its interpretability, revealing the independent contributions of different features.

Herein, this study sought to investigate the association between EASIX and in-hospital ESRD prognosis within the MIMIC-IV cohort. By establishing ML models and performing SHAP analysis, we aimed to achieve more precise and efficient prediction of in-hospital ESRD prognosis. This approach is expected to create a favorable environment for high-risk ESRD cases, enabling rapid diagnostic assessment and effective therapeutic intervention for high-risk patients at an early stage.

Method

Data source

All data used were sourced from MIMIC-IV 3.1 [ 29 ]. This database contains electronic health record datasets from over 90,000 patients in the intensive care unit (ICU) at Beth Israel Deaconess Medical Center (BIDMC) in Boston, Massachusetts, between 2008 and 2022. The collection of patient information and creation of research resources were reviewed by the BIDMC Institutional Review Board, which waived informed consent and approved the data sharing plan.

Inclusion and exclusion criteria

Based on the availability of analytical data, the following populations were included:

Admitted to the ICU. Having ESRD according to the ICD-9/ICD-10. Aged 18 years or older.

We excluded the following groups:

Patients with missing values for any of the following indicators: creatinine, PLT, or LDH. Patients with ICU admission less than 24 h or more than 30 days.

Data extraction

We extracted patient demographic data (age, sex, and ethnicity), comorbidity data (CKD, anemia, CRRT, hypertension, and ischemic heart disease) based on ICD codes. Laboratory and ICU data were obtained, including anion gap (AG), albumin, serum calcium, creatinine, eosinophil count, blood glucose, hemoglobin, LDH, PLT, serum potassium, red blood cell count, serum sodium, total cholesterol, triglycerides, blood urea nitrogen (BUN), uric acid, and white blood cell count (WBC). We also extracted data on patients’ vital signs (serum glucose, body temperature (in degrees Celsius), and unintentional weight loss exceeding 10 pounds, and medication usage (amlodipine, atorvastatin, warfarin, furosemide, valsartan, and heparin). The sequential organ failure assessment (SOFA), Oxford acute severity of illness score (OASIS), and systemic inflammation response syndrome (SIRS) were also included to assess the severity of ESRD.

EASIX = LDH (U/L)× creatinine (mg/Dl)/PLT (×10 9 /L) [ 17 ]. To avoid the influence of skewed data or extreme values, we performed log2 transformation on EASIX. In addition, we divided patients with different EASIX levels into four equal groups using quartiles.

The study outcome was the risk of 30-day in-hospital mortality (IHM). Follow-up was from admission to 30 days or death.

If a patient had multiple hospital admissions, only the first record was extracted and included. If data on vital signs and comorbidities were repeatedly measured, only the first record within 72 h was included. Data measured outside the follow-up period were considered invalid or missing. Data extraction was performed using pgAdmin software. To avoid selection bias caused by excluding missing samples, we employed the mice package in R software for multiple imputation for data with missing data of less than 25%. Data with missing values exceeding 25% were categorized.

Statistical analysis

Continuous variables were tested for normality using the Shapiro-Wilk test. Normally distributed variables were depicted as means and standard deviations (SDs), and intergroup comparisons were performed via T tests and analysis of variance. Nonnormally distributed variables were delineated as median and interquartile range (IQR), and intergroup comparisons were performed via the Wilcoxon rank sum test or the Kruskal-Wellis rank sum test. Categorical variables were described as sample size or proportion (n%), and intergroup differences were analyzed via the chi-square test or Fisher’s exact test.

To compare IHM among patients with different EASIX levels, Kaplan-Meier survival curves were used to visualize mortality differences among the four groups using the log-rank test.

The study cohort was randomized into a training cohort and a validation cohort in a 7:3 ratio. The training cohort was applied for feature screening and model building, while the validation cohort was used to verify model performance. In the training cohort, we used the Boruta algorithm and LASSO regression to discern the most core features or confounding factors. In the Boruta algorithm, all real features were copied and shuffled, and the random forest model was used to obtain the Z value of each feature. It identified the importance of features in the dataset by comparing the Z value of each feature with the corresponding “shadow feature”. Features were color-coded as “important” (green), “unimportant” (red), and “pending” (yellow). Features identified as “important” were considered to have passed the screening. LASSO regression used the penalty coefficient Lambda to bring the coefficients of non-contributing features close to 0. After applying the optimal Lambda, variables with coefficients not equal to 0 were considered important features. Finally, the results of the Boruta algorithm and LASSO regression were intersected to obtain the most important features.

Cox proportional hazards regression models were leveraged to explore the association between EASIX levels and IHM in ESRD patients. Three models were constructed during the analysis: Model 1 did not adjust for any covariates; Model 2 adjusted for age, sex, and race; and Model 3 adjusted for age, sex, race, and all screened confounders. Additionally, we calculated the Variance Inflation Factor (VIF) for each variable in the model to rule out the possibility of significant multicollinearity.

In this study, we used restricted cubic splines (RCS) to determine the potential nonlinear relationship using data on the calculated EASIX and age, sex, race, and important characteristics obtained through screening.

To explore the consistency of the analysis results in specific populations, we conducted subgroup analyses based on age, sex, race, warfarin use, hypertension, and SOFA score and performed interaction tests. The results were presented in forest plots.

Finally, combining the EASIX indicators and ML methods, six ML models were established, including logistic regression (LR), random forest (RF), neural network (NN), gradient boosting machine (GBM), support vector machine (SVM), and eXtreme Gradient Boosting (XGboost). We also used receiver operator characteristic (ROC) curves, calibration curves, decision curves, and computational performance metrics to examine and compare the predictive performance, clinical validity, and accuracy of different models and select the optimal one. Furthermore, SHAP analysis was performed on the optimal model to enhance its interpretability.

All data analyses were performed using R version 4.5.1 and R studio. A two-tailed P value < 0.05 was considered statistically significant.

Result

Baseline characteristics

The initial sample size was 2,357 patients with ESRD. First, we excluded 1,196 patients who had missing indicators (LDH, creatinine, and PLT) required to calculate EASIX. Second, 146 patients with ICU hospitalization time of more than 30 days were also excluded, and 18 patients were excluded because their hospitalization time was less than 24 h. The final sample size was 997. The screening process is shown in Fig. 1 .

Fig. 1 Selection of the study population from the MIMIC-IV database

In our overall study population, there were 194 in-hospital deaths, representing approximately 19.5% of the total sample, with males accounting for a larger proportion (61%). The median age and median follow-up duration were 66 years and 9 days, respectively. White individuals (53%) and Black individuals (18%) accounted for a higher proportion of in-hospital deaths (Table 1 ). The median age in the IHM group was 70 years, indicating a higher risk of death among older ESRD patients. This cohort also exhibited greatly higher levels of AG, hemoglobin, LDH, WBC, blood glucose, and SOFA scores. Patients with CKD-related anemia and those who received amlodipine mostly did not experience IHM.

Table 1 All characteristic of the participants (grouped by in-hospital dead) Characteristic Overall Survival Death p -value 2,3 N = 997 1 N = 803 1 N = 194 1 Hospital time (days) 9 [6 to 15] 9 [6 to 15] 9 [4 to 16] 0.087 Age (years) 66 [56 to 75] 65 [55 to 74] 70 [62 to 78] < 0.001*** Sex (%) 0.427 Female 366 (37%) 290 (36%) 76 (39%) Male 631 (63%) 513 (64%) 118 (61%) Race (%) 0.107 Asian 45 (4.5%) 39 (4.9%) 6 (3.1%) Black 225 (23%) 191 (24%) 34 (18%) Hispanic 64 (6.4%) 53 (6.6%) 11 (5.7%) Other 159 (16%) 119 (15%) 40 (21%) White 504 (51%) 401 (50%) 103 (53%) Anemia in CKD (%) 426 (43%) 365 (45%) 61 (31%) < 0.001*** Hypertension (%) 33 (3.3%) 28 (3.5%) 5 (2.6%) 0.525 Ischemic heart disease (%) 507 (51%) 402 (50%) 105 (54%) 0.31 CRRT (%) 181 (18%) 80 (10.0%) 101 (52%) < 0.001*** Amlodipine (%) 119 (12%) 111 (14%) 8 (4.1%) < 0.001*** Atorvastatin (%) 430 (43%) 346 (43%) 84 (43%) 0.958 Furosemide (%) 351 (35%) 296 (37%) 55 (28%) 0.026* Heparin (%) 925 (93%) 751 (94%) 174 (90%) 0.064 Valsartan (%) 22 (2.2%) 22 (2.7%) 0 (0%) 0.013* Warfarin (%) 205 (21%) 177 (22%) 28 (14%) 0.019* AG (mmol/L) 18.0 [15.0 to 21.0] 18.0 [15.0 to 21.0] 19.0 [16.0 to 23.0] 0.005** Albumin (g/dL) 3.20 [2.80 to 3.70] 3.20 [2.90 to 3.70] 3.00 [2.40 to 3.50] < 0.001*** Calcium (mg/dL) 8.60 [8.00 to 9.20] 8.60 [8.00 to 9.10] 8.65 [8.10 to 9.30] 0.172 Creatinine (mg/dL) 5.6 [3.8 to 7.8] 5.9 [3.9 to 8.2] 5.0 [3.3 to 6.5] < 0.001*** Glucose (mg/dL) 123 [95 to 176] 122 [95 to 168] 126 [94 to 190] 0.313 Hemoglobin (g/dL) 9.40 [8.20 to 10.70] 9.30 [8.10 to 10.70] 9.80 [8.70 to 11.00] 0.012* LDH (mmol/L) 260 [197 to 367] 244 [190 to 338] 340 [245 to 610] < 0.001*** PLT (K/µL) 178 [121 to 244] 182 [125 to 244] 159 [100 to 251] 0.013* Potassium (mmol/L) 4.60 [4.10 to 5.20] 4.50 [4.10 to 5.20] 4.65 [4.20 to 5.30] 0.089 Red blood cell (K/µL) 3.11 [2.73 to 3.53] 3.08 [2.71 to 3.52] 3.19 [2.76 to 3.56] 0.17 Sodium (mmol/L) 137.0 [134.0 to 140.0] 137.0 [134.0 to 140.0] 136.0 [133.0 to 139.0] 0.030* Brea nitrogen (mmol/L) 51 [35 to 76] 51 [35 to 77] 51 [36 to 75] 0.766 White blood cell (K/µL) 9.0 [6.6 to 13.2] 8.6 [6.4 to 12.5] 11.4 [7.6 to 17.0] < 0.001*** Serum glucose (mg/dL) 126 [96 to 172] 125 [95 to 167] 129 [100 to 196] 0.101 Unintentional weight loss more than 10 lbs (%) 40 (9.7%) 33 (9.1%) 7 (14%) 0.3 OASIS 33 [26 to 39] 32 [25 to 39] 36 [30 to 43] < 0.001*** SIRS 0.0005 0 17 (1.7%) 15 (1.9%) 2 (1.0%) 1 141 (14%) 128 (16%) 13 (6.7%) 2 342 (34%) 301 (37%) 41 (21%) 3 347 (35%) 271 (34%) 76 (39%) 4 150 (15%) 88 (11%) 62 (32%) SOFA 7.0 [5.0 to 10.0] 7.0 [5.0 to 9.0] 10.0 [8.0 to 13.0] < 0.001*** EASIX 9 [5 to 18] 8 [5 to 16] 12 [6 to 30] < 0.001*** Log2 (EASIX) 3.30 [2.58 to 4.21] 3.23 [2.54 to 4.10] 3.72 [2.86 to 4.96] < 0.001*** EASIX (quartiles) < 0.001*** Q1 250 (25%) 211 (26%) 39 (20%) Q2 249 (25%) 214 (27%) 35 (18%) Q3 249 (25%) 203 (25%) 46 (24%) Q4 249 (25%) 175 (22%) 74 (38%) 1 Median [Q1 to Q3]; n (%) 2 Wilcoxon rank sum test; Pearson’s Chi-squared test; Fisher’s exact test; NA 3 * p < 0.05; ** p < 0.01; *** p < 0.001

In the baseline table stratified by EASIX quartiles (Table 2 ), compared with the group with lower EASIX levels (Q1), the group with high EASIX levels (Q4) exhibited notably higher IHM rates, along with markedly elevated levels of AG, albumin, serum creatinine, LDH, BUN, SOFA scores, and SIRS. Furthermore, individuals with higher EASIX levels were more frequently associated with prior CRRT treatment. Conversely, the Q1 cohort exhibited a considerably higher prevalence of ischemic heart disease and warfarin use than the Q4. Additionally, Q1 patients demonstrated higher levels of PLT, red blood cell count, and WBC (all p < 0.05).

Table 2 All characteristics of the participants (grouped by EASIX level) Characteristics Overall Q1 Q2 Q3 Q4 p -value 2,3 N = 997 1 N = 250 1 N = 249 1 N = 249 1 N = 249 1 Hospital dead (%) 194 (19%) 39 (16%) 35 (14%) 46 (18%) 74 (30%) < 0.001*** Hospital time (days) 9 [6 to 15] 10 [6 to 18] 9 [6 to 15] 9 [5 to 14] 9 [6 to 16] 0.049* Age (years) 66 [56 to 75] 68 [57 to 76] 70 [60 to 77] 66 [57 to 73] 63 [52 to 72] < 0.001*** Sex (%) 0.005** Female 366 (37%) 100 (40%) 109 (44%) 83 (33%) 74 (30%) Male 631 (63%) 150 (60%) 140 (56%) 166 (67%) 175 (70%) Race (%) < 0.001*** Asian 45(4.5%) 6 (2.4%) 4 (1.6%) 8 (3.2%) 27 (11%) Black 225(23%) 36 (14%) 51 (20%) 68 (27%) 70 (28%) Hispanic 64 (6.4%) 11 (4.4%) 21 (8.4%) 18 (7.2%) 14 (5.6%) Other 159 (16%) 36 (14%) 43 (17%) 37 (15%) 43 (17%) White 504 (51%) 161 (64%) 130 (52%) 118 (47%) 95 (38%) Anemia in CKD (%) 426 (43%) 104 (42%) 103 (41%) 120 (48%) 99 (40%) 0.235 Hypertension (%) 33 (3.3%) 8 (3.2%) 6 (2.4%) 8 (3.2%) 11 (4.4%) 0.659 Ischemic heart disease (%) 507 (51%) 125 (50%) 138 (55%) 140 (56%) 104 (42%) 0.004** CRRT (%) 181 (18%) 42 (17%) 31 (12%) 44 (18%) 64 (26%) 0.002** Amlodipine (%) 119 (12%) 36 (14%) 31 (12%) 25 (10%) 27 (11%) 0.45 Atorvastatin (%) 430 (43%) 112 (45%) 114 (46%) 125 (50%) 79 (32%) < 0.001*** Furosemide (%) 351 (35%) 100 (40%) 89 (36%) 77 (31%) 85 (34%) 0.197 Heparin (%) 925 (93%) 237 (95%) 231 (93%) 228 (92%) 229 (92%) 0.51 Valsartan (%) 22 (2.2%) 5 (2.0%) 8 (3.2%) 5 (2.0%) 4 (1.6%) 0.642 Warfarin (%) 205 (21%) 60 (24%) 64 (26%) 43 (17%) 38 (15%) 0.008** AG (mmol/L) 18.0 [15.0 to 21.0] 17.0 [14.0 to 19.0] 17.0 [15.0 to 21.0] 19.0 [16.0 to 22.0] 20.0 [17.0 to 24.0] < 0.001*** Albumin (K/µL) 3.20 [2.80 to 3.70] 3.00 [2.50 to 3.40] 3.30 [2.90 to 3.80] 3.30 [2.90 to 3.70] 3.20 [2.80 to 3.60] < 0.001*** Calcium (mg/dL) 8.60 [8.00 to 9.20] 8.60 [8.10 to 9.20] 8.70 [8.20 to 9.10] 8.70 [8.10 to 9.20] 8.40 [7.60 to 9.10] 0.003** Creatinine (mg/dL) 5.6 [3.8 to 7.8] 3.6 [2.4 to 4.6] 5.3 [3.9 to 6.7] 7.0 [5.1 to 9.2] 7.2 [5.3 to 10.6] < 0.001*** Glucose (mg/dL) 123 [95 to 176] 129 [97 to 178] 129 [94 to 178] 117 [92 to 166] 121 [94 to 164] 0.372 Hemoglobin (g/dL) 9.40 [8.20 to 10.70] 9.50 [8.50 to 10.80] 9.30 [8.30 to 10.50] 9.70 [8.40 to 10.70] 9.30 [7.90 to 10.70] 0.508 LDH (mmol/L) 260 [197 to 367] 205 [170 to 254] 232 [186 to 289] 269 [215 to 369] 434 [310 to 792] < 0.001*** PLT (K/µL) 178 [121 to 244] 245 [190 to 306] 187 [137 to 233] 162 [119 to 216] 118 [74 to 176] < 0.001*** Potassium (mmol/L) 4.60 [4.10 to 5.20] 4.40 [4.00 to 4.80] 4.50 [4.10 to 5.00] 4.70 [4.20 to 5.40] 4.70 [4.30 to 5.60] < 0.001*** Red blood cell (K/µL) 3.11 [2.73 to 3.53] 3.18 [2.82 to 3.59] 3.04 [2.73 to 3.47] 3.14 [2.72 to 3.55] 3.08 [2.60 to 3.47] 0.08 Sodium (mmol/L) 137.0 [134.0 to 140.0] 137.0 [134.0 to 140.0] 137.0 [134.0 to 139.0] 137.0 [134.0 to 140.0] 137.0 [134.0 to 139.0] 0.741 BUN (mmol/L) 51 [35 to 76] 37 [26 to 56] 46 [33 to 67] 55 [42 to 77] 67 [45 to 98] < 0.001*** WBC (K/µL) 9.0 [6.6 to 13.2] 10.0 [7.5 to 15.1] 8.9 [6.5 to 13.0] 8.2 [6.1 to 11.6] 8.9 [6.3 to 13.1] < 0.001*** Serum glucose (mg/dL) 126 [96 to 172] 125 [98 to 170] 134 [98 to 177] 122 [94 to 170] 126 [98 to 166] 0.742 Unintentional weight loss more than 10 lbs (%) 40 (9.7%) 14 (13%) 8 (9.5%) 7 (6.2%) 11 (11%) 0.429 OASIS 33 [26 to 39] 32 [25 to 39] 34 [28 to 40] 32 [25 to 38] 32 [26 to 40] 0.126 SIRS 0.004** 0 17 (1.7%) 2 (0.8%) 8 (3.2%) 4 (1.6%) 3 (1.2%) 1 141(14%) 42 (17%) 20 (8.0%) 45 (18%) 34 (14%) 2 342(34%) 74 (30%) 89 (36%) 88 (35%) 91 (37%) 3 347(35%) 103(41%) 86 (35%) 81 (33%) 77 (31%) 4 150(15%) 29 (12%) 46 (18%) 31 (12%) 44 (18%) SOFA 7.0 [5.0 to 10.0] 6.0 [4.0 to 9.0] 7.0 [5.0 to 9.0] 7.0 [5.0 to 9.0] 8.0 [6.0 to 11.0] < 0.001*** EASIX 9 [5 to 18] 3 [2 to 4] 7 [6 to 8] 12 [10 to 15] 31 [22 to 53] < 0.001*** Log2(EASIX) 3.30 [2.58 to 4.21] 2.13 [1.78 to 2.38] 2.93 [2.80 to 3.12] 3.71 [3.50 to 3.99] 4.99 [4.52 to 5.76] < 0.001*** 1 n (%); Median [Q1 to Q3] 2 Pearson’s Chi-squared test; Kruskal-Wallis rank sum test; NA 3 * p < 0.05; ** p < 0.01; *** p < 0.001

Kaplan-Meier survival analysis

Kaplan-Meier curve (Fig. 2 ) illustrated differences in 30-day in-hospital survival rates among ESRD patients at varying exposure levels (log2 [EASIX]). The curve representing the Q4 cohort lay at the lowest point, indicating the highest IHM rate among the four groups. Conversely, the survival curve for the Q1 cohort occupied the highest position, signifying the lowest IHM rate (P of Log rank < 0.0001).

Fig. 2 KM survival curve. KM curves show the survival rates at 30 days for each quartile. Q1 (0.598–2.580), Q2 (2.580–3.305), Q3 (3.305–4.212), and Q4 (4.212–9.901)

Feature selection

In the training cohort, feature selection was implemented via LASSO regression and the Boruta algorithm. The intersection of the results from both methods was selected as the final modeling variables.

Figure 3 displays the LASSO regression results. In Fig. 3 B, the vertical dashed line on the right (lambda.min) indicated the lambda value of the penalty coefficient at the point of minimum mean squared error, which would be used for selection. The parallel standard error line to its left (lambda.1se) represented the lambda value after adding one unit of standard error. The optimal value we obtained was lambda.min = 0.004781334. The features selected via LASSO regression included log2(EASIX), CKD-related anemia, hypertension, CRRT, atorvastatin use, amlodipine use, furosemide use, heparin use, valsartan use, warfarin use, AG, serum calcium, creatinine, eosinophil count, LDH, lymphocyte count, monocyte count, pH, PLT, serum potassium, red blood cell count, serum sodium, WBC, non-invasive systolic blood pressure, unintentional weight loss ≥ 10 pounds, OASIS, SIRS, and SOFA scores.

Fig. 3 Featured selection based on the Boruta algorithm. ( A ) Plots for the LASSO regression coefficients over different values of the penalty parameter. The vertical dashed lines correspond to the lambda.min and lambda.1se from the cross-validation. ( B ) Coss-validation plot for the penalty term. The dashed lines represent the lambda.min and lambda.1se

Figure 4 displays the results of the Boruta analysis. Blue blocks represented shadow features, green blocks denoted features considered important and suitable for screening, red blocks indicated unimportant features, and yellow blocks represented features of uncertain importance that did not appear in this study. The features collected via the Boruta algorithm included log2(EASIX), BUN, unintentional weight loss > 10 pounds, non-invasive systolic blood pressure, AG, SIRS, PLT, creatinine, WBC, LDH, CRRT, and SOFA (Fig. 4 ).

Fig. 4 Featured selection based on the Boruta algorithm. The horizontal axis is each variable, and the vertical axis is the Z value of each variable. The box plot shows the Z value of each variable during model calculation. The green boxes represent important variables, and the red boxes represent unimportant variables

Finally, the results from both approaches were intersected to obtain four key features: log2(EASIX), CRRT, AG, creatinine, LDH, PLT, WBC, non-invasive systolic blood pressure, unintentional weight loss > 10 pounds, SIRS, and SOFA. To avoid the impact of multicollinearity on the model and the redundant selection of components of the SOFA score, we excluded creatinine, LDH, PLT, and SOFA, with log2(EASIX), CRRT, AG, WBC, non-invasive systolic blood pressure, unintentional weight loss > 10 pounds, and SIRS retained as the final variable.

Association analysis

In Cox regression analyses (Table 3 ), when EASIX was included as a continuous variable (log2(EASIX)), an increase in EASIX levels was substantially positively associated with IHM among ESRD patients across all three models. The corresponding log2(EASIX) hazard ratios (HR) were 1.237 (95% CI: 1.136–1.348), 1.332 (95% CI: 1.216–1.459), and 1.286 (95% CI: 1.138–1.453), respectively. When EASIX was analyzed as a categorical variable segmented by quartiles (IQR), patients with high EASIX levels (Q4) had a higher IHM risk than those with low EASIX levels (Q1). This result held true after adjusting for confounding factors in the models. The Q4 data showed HR = 2.226 (95% CI, 1.509–3.283), HR = 2.908 (95% CI, 1.939–4.361), and HR = 2.506 (95% CI, 1.543–4.070), respectively. Trend tests also demonstrated consistent patterns.

Table 3 Association between EASIX Levels and in-hospital mortality risk in patients with ESRD Characteristic Model 1 Model 2 Model 3 HR 95% CI p -value HR 95% CI p -value HR 95% CI p -value Log2(EASIX) 1.237 1.136, 1.348 < 0.001 1.332 1.216, 1.459 < 0.001 1.286 1.138, 1.453 < 0.001 EASIX (quartiles) Q1 — — — — — — Q2 1.148 0.725, 1.818 0.6 1.123 0.707, 1.785 0.6 1.063 0.626, 1.804 0.8 Q3 1.752 1.137, 2.701 0.011 1.865 1.210, 2.875 0.005 1.525 0.931, 2.499 0.094 Q4 2.226 1.509, 3.283 < 0.001 2.908 1.939, 4.361 < 0.001 2.506 1.543, 4.070 < 0.001 Test for trend 1.88 1.422, 2.485 < 0.001 2.292 1.713, 3.066 < 0.001 2.008 1.422, 2.835 < 0.001 Model 1 was unadjusted Model 2 was adjusted for age, sex and race Model 3 was adjusted for age, sex, race, CRRT, AG, unintentional weight loss > 10 pounds, non-invasive systolic blood pressure, WBC and SIRS. Q1 (0.598–2.580), Q2 (2.580–3.305), Q3 (3.305–4.212), and Q4 (4.212–9.901) Abbreviations: CI: Confidence Interval, HR: Hazard Ratio

Potential non-linear relationship

RCS analysis (Fig. 5 ) revealed that in the unadjusted, partially adjusted, and fully adjusted models, when log2(EASIX) > 3.302, for an increase in EASIX levels, the IHM risk among ESRD patients increased in an approximately linear fashion. This trend was consistent across all three models, with non-linear P-values for log2(EASIX) of Model 1: P = 0.396, Model 2: P = 0.127, and Model 3: P = 0.192. VIF values were all below 10, indicating no significant multicollinearity, and the model is relatively robust.

Fig. 5 RCS analysis of mortality. ( A ) Model 1 was unadjusted. ( B ) Model 2 was adjusted for age, sex, and race. ( C ) Model 3 was adjusted for age, sex, race, log2(EASIX), CRRT, AG, WBC, non-invasive systolic blood pressure, unintentional weight loss > 10 pounds, and SIRS

Subgroup analysis

Subgroup analyses based on several population characteristics (age, sex, ethnicity, hypertension, warfarin use, SOFA score) (Fig. 6 ) showed that in patients aged ≥ 60 years, male patients, white and black patients, patients without hypertension, patients not using warfarin, and patients with a SOFA score ≥ 7, the positive link between elevated EASIX levels and increased IHM risk in ESRD patients remained significant and demonstrated high consistency across all subgroups (all HR > 1, all P < 0.001).

Fig. 6 Results from subgroup analysis. The outcome was the hospital death

Additionally, we observed significant interactions between EASIX and both sex (P for sex = 0.005) and race (P for race = 0.044).

Establishment and validation of the prediction model

We established five ML models: RF, NN, GBM, SVM, XGboost, and traditional LR. For each model, we performed 10-fold cross-validation with 100 iterations using random search to select the optimal hyperparameter combination, which was then applied for modeling. The optimal hyperparameters for each model are listed in Table S1 .

The traditional LR model was visualized as a nomogram (Fig. 7 ), which added up the scores of each variable to obtain total points. Finally, the probability corresponding to total points was the probability of IHM in ESRD patients.

Fig. 7 Predictive nomogram of mortality, in which the total points correspond to a death probability at the bottom, summing each value of the variable

These models were then validated. ROC curves (Fig. 8 A) showed that in the training set, the area under the curve (AUC) for the LR model and the five ML models were: LR = 0.82, GBM = 0.904, NN = 0.821, RF = 0.913, SVM = 0.861, and XGboost = 0.802. In the validation set, the AUCs were: LR = 0.82, GBM = 0.811, NN = 0.832, RF = 0.82, SVM = 0.776, and XGboost = 0.764. The calibration curve analysis and goodness-of-fit tests were also performed for the predicted and observed mortality risks. The calibration curves for each model are shown in Fig. 8 B. In both the training and validation sets, the calibration curves for GBM fitted well with the reference lines, and their predicted and observed values were highly consistent, indicating excellent predictive performance. DCA curves (Fig. 8 C) showed that each model demonstrated a significant positive net clinical benefit compared to no treatment and all treatments, demonstrating robust clinical effectiveness. Finally, the performance metrics of the ML models were calculated (Table S2 ). In the training set, XGboost showed the highest sensitivity (0.881), followed by SVM (0.863). In the validation set, XGboost (0.879) and RF (0.824) showed higher sensitivity. In the training set, GBM showed the highest sensitivity (0.906), followed by RF (0.827); GBM (0.701) also exhibited high specificity in the training set. The XGboost and GBM models demonstrated favorable performance in terms of their F1 scores.

Fig. 8 ( A ) ROC curve of machine learning algorithms for comparing performance; ( B ) Calibration curve of machine learning algorithms; ( C ) Decision curve of machine learning algorithms. RF: random forest, NN: neural network, GBM: gradient boosting machine, SVM: support vector machine, XGboost: eXtreme Gradient Boosting

After comprehensively considering the AUC values of ROC curves, the fit of the calibration curves, DCA curves, and various performance metrics, we concluded that the GBM model demonstrated strong predictive capability and balanced overall performance without obvious flaws, making it the most suitable ML model for predicting IHM risk in ESRD patients based on EASIX levels.

SHAP analysis

Figure S1 shows the bee warm plot (Figure S1 A) and bar chart (Figure S1 B) from the SHAP analysis. Figure S1 A demonstrates that the WBC, non-invasive systolic blood pressure, log2(EASIX), SIRS, and AG all positively contributed to the predictive power of the GBM model in estimating the mortality risk. Figure S1 B shows the ranking of each feature in predicting mortality outcomes, with WBC having the highest predictive value, followed by non-invasive systolic blood pressure, log2(EASIX), SIRS, and AG. Figure S1 C shows the dependency analysis in SHAP. Figure S1 C and Figure S1 D show waterfall plot and force plot of the SHAP analysis, with yellow graphs representing features that increase mortality prediction and purple graphs representing features that decrease mortality prediction. We randomly selected a case that died in the hospital (Figure S1 C) and a case that did not die in the hospital (Figure S1 D) from the research cohort. Group D showed the specific contribution values and proportions of the four features to the prediction results in predicting the patient’s death. The cumulative contribution of these four features was − 1.36, greater than the baseline value of -1.82; therefore, the model predicts that the patient will die. The cumulative contribution of these features in Group E was − 2.03, less than the baseline value of -1.82; hence, the model predicts that the patient will survive. The prediction results were consistent with the actual observations. These visualizations and images help enhance the interpretability of the GBM model, providing deeper insight into how the model makes predictions and the contribution of each feature to the prediction results.

Discussion

This study, a retrospective analysis based on the MIMIC-IV database, aimed to investigate the association between EASIX and IHM risk in ESRD patients. Our results established a significant positive link, and this association persisted after adjustment for covariates and was consistent across subgroups, including male patients, patients over 60 years of age, and white and black patients, confirming the validity of our findings. This is the first study to clarify the link between EASIX and IHM risk in ESRD patients. To ensure the rigor and scientific validity of feature selection, we used two authoritative screening methods: the Boruta algorithm and LASSO regression. EASIX was identified as an important feature in both analyses, consistent with our analysis results in Cox regression. Based on this observation, we believe that EASIX has the potential to become a reliable marker for predicting IHM in ESRD patients. Finally, we incorporated seven selected features (log2(EASIX), CRRT, AG, WBC, non-invasive systolic blood pressure, unintentional weight loss > 10 pounds, and SIRS) into LR and five other ML models. Validation and comparison revealed that GBM was the optimal model for predicting IHM risk in ESRD patients, demonstrating excellent performance. SHAP analysis of the GBM model revealed that the EASIX level made a significant contribution to predicting mortality risk within the model. This further validated our conclusion that the EASIX level is a crucial factor in predicting the IHM risk among ESRD patients.

Regarding the EASIX indicator, numerous studies have previously been conducted, but they have primarily focused on stem cell transplantation or its complications, endothelial diseases, immune disorders, and neoplastic diseases [ 30 , 31 ]. For instance, EASIX is a key prognostic indicator for the risk of sepsis following alloSCT [ 15 ]. Coagulation dysfunction may promote hemorrhagic lesions during traumatic brain injury, and its mortality risk can be predicted by an increase in EASIX scores [ 32 ]. A German study indicated that hyperbilirubinemia frequently occurred after alloSCT due to endothelial complications (whose risk is closely associated with elevated EASIX levels), often portending poor prognosis, with the risk of hyperbilirubinemia showing a strong positive correlation with EASIX [ 33 ]. The modified EASIX score also enabled early identification of primary refractory classical Hodgkin lymphoma [ 34 ]. Additionally, a multicenter cohort study found that EASIX served as a prognostic indicator for lymphoma-associated hemophagocytic lymphohistiocytosis, being closely associated with early mortality in this disease [ 35 ]. In summary, as a novel marker for assessing and reflecting endothelial dysfunction, EASIX correlates with multiple diseases and adverse outcomes.

Based on the research foundation of EASIX in immune diseases, endothelial diseases, and inflammation-related mechanisms, this study addresses the prognostic association between EASIX and CKD, a current research gap, demonstrating that elevated EASIX levels are closely linked to poor outcomes in ESRD. ESRD progression undoubtedly involves complex pathophysiological mechanisms, and EASIX’s response to inflammatory states and endothelial dysfunction may serve as a key anchor point in its relationship with CKD. A study on diabetic nephropathy indicates that oxidative stress and inflammation are key factors leading to endothelial dysfunction and diabetic nephropathy. Elevated levels of tumor necrosis factor-α and interleukin-6 promote glomerulosclerosis and tubulointerstitial fibrosis, leading to worsening renal function. Concurrently, inflammation induces oxidative stress, resulting in excessive production of reactive oxygen species and reactive nitrogen species. This further triggers cellular damage and dysfunction, activates immune cells, and leads to the massive release of inflammatory cytokines, thereby exacerbating the disease [ 16 ]. Renal fibrosis is a common clinical feature in CKD and a key factor in its progression to ESRD. Inflammation and myofibroblast activation constitute its initial processes. Alterations in the inflammatory microenvironment, resulting from cellular damage and abnormal chemokine secretion, trigger fibrosis and the transformation of vascular endothelial cells, pericytes, and fibroblasts into myofibroblasts. The continuous accumulation of extracellular matrix accelerates renal interstitial hypoxia and loss of nephrons [ 36 , 37 ]. A study on the immunological mechanisms of kidney disease indicates that innate immune cells, such as granulocytes and macrophages, can induce AKI by secreting inflammatory mediators, like cytokines and enzymes, as well as releasing extracellular traps. Macrophages and dendritic cells can further drive disease progression toward CKD through pro-fibrotic mechanisms [ 38 ]. Regarding the predictive interpretation of EASIX for ESRD, endothelial dysfunction assessed through inflammation is closely associated with mortality and CVDs in ESRD patients. For instance, inflammation exhibits a significant amplifying effect on the impact of ADMA on the severity of atherosclerosis in ESRD patients, all of which may contribute to poor prognosis [ 39 ]. In summary, in ESRD patients with complex and heterogeneous pathological conditions, endothelial injury and stress responses may arise from multiple competing pathophysiological mechanisms. EASIX can serve as a convergent pathway that quantifies various pathogenic factors, such as inflammation, ischemia, and malignancy, which share common endothelial injury processes, thereby demonstrating its predictive ability distinct from other single factors.

In subgroup analyses, we identified a significant interaction between the exposure indicator EASIX and sex subgroups ( P = 0.005). Specifically, higher EASIX levels were strongly linked with poor ESRD outcomes in the male cohort ( P < 0.001). We propose several potential explanations. Despite higher CKD prevalence in women, more male patients undergo renal replacement therapy due to the natural decline in GFR. Additionally, the protective effect of female estrogen and the damaging effect of male testosterone, combined with more unhealthy lifestyle habits (smoking, alcohol consumption, high-fat diet) among male patients, result in faster renal decline, poorer CKD/ESRD outcomes, and greater quality-of-life deterioration in males [ 40 ].

The GBM model was selected due to its optimal clinical applicability. Although its discriminative ability (AUC) was comparable to that of RF, GBM demonstrated the best calibration performance, with its calibration curve closest to the diagonal, indicating more accurate and reliable predicted probabilities. In contrast, RF showed evident signs of overfitting on the validation set, which is a critical flaw in clinical prediction models. Other models (e.g., XGboost, SVM) exhibited significant shortcomings in key metrics, such as sensitivity, specificity, or calibration. Therefore, GBM was identified as the preferred model based on its excellent calibration and overall balanced performance. Additionally, the optimal model we selected—GBM—also exhibits significant advantages over traditional LR models. Traditional LR models have considerable limitations, such as only incorporating normally distributed variables and being restricted to capturing linear relationships [ 41 ]. In contrast, the GBM model, as a ML approach, can identify nonlinear relationships while delivering superior predictive performance and accuracy. During the validation process of ML models, we observed pronounced overfitting in the XGboost model. This may stem from the relatively small study cohort ( n = 997) combined with XGboost’s strong learning capability, which is prone to overfitting [ 42 ].

In summary, this study first illustrates a significant correlation between EASIX levels and the IHM risk in ESRD patients, establishing a ML predictive model and interpreting it through SHAP analysis. As a novel biomarker reflecting inflammation and endothelial dysfunction, the EASIX metric, along with our ML model, assists in early identifying and monitoring ESRD patients at high risk of mortality. This enables timely intervention, treatment, and risk management, thereby mitigating clinical workloads and reducing adverse patient outcomes. Furthermore, EASIX retains substantial potential for predicting additional severe diseases. Our study heightens awareness of EASIX while providing theoretical and data support for subsequent in-depth investigations.

However, our study also has certain limitations that warrant mention. First, given its retrospective nature, our analysis cannot determine a causal relationship between EASIX and the IHM risk in ESRD patients; further prospective cohort studies are warranted to clarify this. Additionally, at the start of the study, we excluded patients with missing LDH data. Since LDH is often measured in patients with more critical conditions or specific pathological conditions, this may introduce selection bias where data availability is linked to outcomes, which warrants attention. Second, the MIMIC-IV database originates from the United States, so the generalizability of our conclusions to research cohorts in Asia and Africa requires further confirmation. Third, as this analysis was conducted at a single center, external validation through prospective multicenter research is warranted. Finally, due to limitations in the database data types, certain variables potentially relevant to our research topic could not be included in the analysis, such as genetic characteristics, radiomics, and dietary patterns. Future studies could explore multi-data, multi-modal approaches to address these gaps.It should be noted that EASIX was originally developed for transplant-associated microangiopathy (TMA), with its components—platelets, creatinine, and LDH—intended to reflect endothelial damage and complement activation. However, in critically ill ESRD patients admitted to the ICU, these parameters are often independently affected by multiple non-endothelial processes, including uremia, dialysis-associated hemolysis, effects of continuous renal replacement therapy (CRRT), sepsis, disseminated intravascular coagulation (DIC), and cardiorenal syndrome.

Conclusion

Elevated EASIX levels show a significant positive correlation with increased IHM risk among ESRD patients, with consistent trends observed across multiple subgroups. ML prediction models incorporating this metric demonstrate robust predictive performance. These findings may aid in the early identification, intervention, and management of high-risk ESRD patients, though further validation in prospective cohorts is warranted.

Supplementary Information

Below is the link to the electronic supplementary material.

Supplementary Material 1

Supplementary Material 2

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Acknowledgements

Not applicable.

Author contributions

Liandong Chen: Conceptualization, Methodology, Formal analysis, Writing - Original Draft, Writing - Review & Editing, Software, Writing - Review & Editing, Investigation, Data Curation; Bin Fu: Supervision, Project administration, Resources, Funding acquisition. All authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

This study was supported by Youth Fund Project of Humanities and Social Sciences Research of Ministry of Education in 2024 (Grant number 24YJCZH448), The Second Batch of Provincial-Level Teaching Reform Project for Postgraduates in Zhejiang Province During the 14th Five-Year Plan Period (Grant number JGCG2024254), Traditional Chinese Medicine Science and Technology Plan of Zhejiang Province (Grant number 2026ZF42).

Data availability

The datasets analysed during the current study are available from the MIMIC-IV database 3.1.

Declarations

Ethics approval and consent to participate

The MlMIC-IV database was approved by the Massachusetts Institute of Technology (Cambridge, MA) and Beth Israel Deaconess Medical Center (Boston, MA), and consent was obtained for the collection of original data. All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. Therefore, ethical approval and the need for informed consent were waived for this study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Abbreviations

Anion gap

Acute kidney injury

Allogeneic stem cell transplantation

Beth Israel Deaconess Medical Center

Blood urea nitrogen

Chronic kidney disease

Cardiovascular disease

The endothelial activation and stress index

End-stage renal disease

Gradient boosting machine

Glomerular filtration rate

Intensive care unit

In-hospital mortality

Interquartile range

Lactate dehydrogenase

Continuous renal replacement therapy

Logistic regression

Medical information mart for intensive care-IV

Machine learning

Neural network

Oxford acute severity of illness scor

Parameters: platelet count

Restricted cubic splines

Random forest

Receiver operator characteristic

Standard deviations

Shapley additive explanation

Systematic inflammation response syndrome

Sequential organ failure assessment

Support vector machine

White blood cell count

eXtreme gradient boosting

References

1. Abbasi MA, Chertow GM, Hall YN. End-stage renal disease. BMJ Clin Evid. 2010;2010. PMC3217820 21418665

2. Wouk N End-Stage Renal Disease: Medical Management Am Fam Physician 2021 104 5 493 9 34783494 Wouk N. End-Stage Renal Disease: Medical Management. Am Fam Physician. 2021;104(5):493–9. 34783494

3. Iseki K Factors influencing the development of end-stage renal disease Clin Exp Nephrol 2005 9 1 5 14 10.1007/s10157-005-0341-3 15830267 Iseki K. Factors influencing the development of end-stage renal disease. Clin Exp Nephrol. 2005;9(1):5–14. 10.1007/s10157-005-0341-3. 15830267 10.1007/s10157-005-0341-3

4. Ekart R Hojs R Obese and diabetic patients with end-stage renal disease: Peritoneal dialysis or hemodialysis? Eur J Intern Med 2016 32 1 6 10.1016/j.ejim.2016.03.016 27067614 Ekart R, Hojs R. Obese and diabetic patients with end-stage renal disease: Peritoneal dialysis or hemodialysis? Eur J Intern Med. 2016;32:1–6. 10.1016/j.ejim.2016.03.016. 27067614 10.1016/j.ejim.2016.03.016

5. GBD 2023 Kidney Failure with Replacement Therapy Collaborators Global, regional, and national prevalence of kidney failure with replacement therapy and associated aetiologies, 1990–2023: a systematic analysis for the Global Burden of Disease Study 2023 Lancet Glob Health 2025 13 8 e1378 95 10.1016/s2214-109x(25)00198-6 40712611 GBD 2023 Kidney Failure with Replacement Therapy Collaborators. Global, regional, and national prevalence of kidney failure with replacement therapy and associated aetiologies, 1990–2023: a systematic analysis for the Global Burden of Disease Study 2023. Lancet Glob Health. 2025;13(8):e1378–95. 10.1016/s2214-109x(25)00198-6. 40712611 10.1016/S2214-109X(25)00198-6

6. Cobo G Lindholm B Stenvinkel P Chronic inflammation in end-stage renal disease and dialysis Nephrol Dial Transpl 2018 33 suppl3 iii35 40 10.1093/ndt/gfy175 PMC6168801 30281126 Cobo G, Lindholm B, Stenvinkel P. Chronic inflammation in end-stage renal disease and dialysis. Nephrol Dial Transpl. 2018;33(suppl3):iii35–40. 10.1093/ndt/gfy175. 10.1093/ndt/gfy175 PMC6168801 30281126

7. Peacock M Calcium metabolism in health and disease Clin J Am Soc Nephrol 2010 5 Suppl 1 S23 30 10.2215/cjn.05910809 20089499 Peacock M. Calcium metabolism in health and disease. Clin J Am Soc Nephrol. 2010;5(Suppl 1):S23–30. 10.2215/cjn.05910809. 20089499 10.2215/CJN.05910809

8. Collins AJ Li S Ma JZ Herzog C Cardiovascular disease in end-stage renal disease patients Am J Kidney Dis 2001 38 4 Suppl 1 S26 9 10.1053/ajkd.2001.27392 11576917 Collins AJ, Li S, Ma JZ, Herzog C. Cardiovascular disease in end-stage renal disease patients. Am J Kidney Dis. 2001;38(4 Suppl 1):S26–9. 10.1053/ajkd.2001.27392. 11576917 10.1053/ajkd.2001.27392

9. Matsushita K Ballew SH Wang AY Kalyesubula R Schaeffner E Agarwal R Epidemiology and risk of cardiovascular disease in populations with chronic kidney disease Nat Rev Nephrol 2022 18 11 696 707 10.1038/s41581-022-00616-6 36104509 Matsushita K, Ballew SH, Wang AY, Kalyesubula R, Schaeffner E, Agarwal R. Epidemiology and risk of cardiovascular disease in populations with chronic kidney disease. Nat Rev Nephrol. 2022;18(11):696–707. 10.1038/s41581-022-00616-6. 36104509 10.1038/s41581-022-00616-6

10. Valderrábano F Jofre R López-Gómez JM Quality of life in end-stage renal disease patients Am J Kidney Dis 2001 38 3 443 64 10.1053/ajkd.2001.26824 11532675 Valderrábano F, Jofre R, López-Gómez JM. Quality of life in end-stage renal disease patients. Am J Kidney Dis. 2001;38(3):443–64. 10.1053/ajkd.2001.26824. 11532675 10.1053/ajkd.2001.26824

11. O’Connor NR Corcoran AM End-stage renal disease: symptom management and advance care planning Am Fam Physician 2012 85 7 705 10 22534347 O’Connor NR, Corcoran AM. End-stage renal disease: symptom management and advance care planning. Am Fam Physician. 2012;85(7):705–10. 22534347

12. Nahas ME Bello A Chronic kidney disease: the global challenge Lancet 2005 365 9456 331 40 10.1016/s0140-6736(05)17789-7 15664230 Nahas ME, Bello A. Chronic kidney disease: the global challenge. Lancet. 2005;365(9456):331–40. 10.1016/s0140-6736(05)17789-7. 15664230 10.1016/S0140-6736(05)17789-7

13. Zhao E Chen Y Xie H Wang R Gao Y Dong B Association between the Endothelial Activation and Stress Index (EASIX) and all-cause and cardiovascular mortality in patients with diabetes and prediabetes Eur J Med Res 2025 30 1 730 10.1186/s40001-025-03009-0 40790498 PMC12337403 Zhao E, Chen Y, Xie H, Wang R, Gao Y, Dong B, et al. Association between the Endothelial Activation and Stress Index (EASIX) and all-cause and cardiovascular mortality in patients with diabetes and prediabetes. Eur J Med Res. 2025;30(1):730. 10.1186/s40001-025-03009-0. 40790498 10.1186/s40001-025-03009-0 PMC12337403

14. Luft T Benner A Terzer T Jodele S Dandoy CE Storb R EASIX and mortality after allogeneic stem cell transplantation Bone Marrow Transpl 2020 55 3 553 61 10.1038/s41409-019-0703-1 PMC8082940 31558788 Luft T, Benner A, Terzer T, Jodele S, Dandoy CE, Storb R, et al. EASIX and mortality after allogeneic stem cell transplantation. Bone Marrow Transpl. 2020;55(3):553–61. 10.1038/s41409-019-0703-1. 10.1038/s41409-019-0703-1 PMC8082940 31558788

15. Korell F Schreck N Müller-Tidow C Dreger P Luft T Pre-transplant EASIX and sepsis after allogeneic stem cell transplantation Intensive Care Med 2022 48 6 753 5 10.1007/s00134-022-06676-3 35429256 PMC9012907 Korell F, Schreck N, Müller-Tidow C, Dreger P, Luft T. Pre-transplant EASIX and sepsis after allogeneic stem cell transplantation. Intensive Care Med. 2022;48(6):753–5. 10.1007/s00134-022-06676-3. 35429256 10.1007/s00134-022-06676-3 PMC9012907

16. Chen S Guo L Ma X Wang S Wu J Wu L Association between Endothelial Activation and Stress Index and mortality in diabetic nephropathy ICU patients: A retrospective cohort study PLoS ONE 2025 20 8 e0329233 10.1371/journal.pone.0329233 40811428 PMC12352680 Chen S, Guo L, Ma X, Wang S, Wu J, Wu L, et al. Association between Endothelial Activation and Stress Index and mortality in diabetic nephropathy ICU patients: A retrospective cohort study. PLoS ONE. 2025;20(8):e0329233. 10.1371/journal.pone.0329233. 40811428 10.1371/journal.pone.0329233 PMC12352680

17. He Y Li Y Xiaojin Y Wu D Jiang W Xie X Endothelial activation and stress index for prediction of mortality in asthma Front Med (Lausanne) 2025 12 1622944 10.3389/fmed.2025.1622944 40703256 PMC12283289 He Y, Li Y, Xiaojin Y, Wu D, Jiang W, Xie X. Endothelial activation and stress index for prediction of mortality in asthma. Front Med (Lausanne). 2025;12:1622944. 10.3389/fmed.2025.1622944. 40703256 10.3389/fmed.2025.1622944 PMC12283289

18. Huang Y Li Z Wang J Wang D Yin X Endothelial activation and stress index is a reliable predictor for the prevalence and mortality outcomes of stroke Sci Rep 2025 15 1 23285 10.1038/s41598-025-06595-5 40603416 PMC12223256 Huang Y, Li Z, Wang J, Wang D, Yin X. Endothelial activation and stress index is a reliable predictor for the prevalence and mortality outcomes of stroke. Sci Rep. 2025;15(1):23285. 10.1038/s41598-025-06595-5. 40603416 10.1038/s41598-025-06595-5 PMC12223256

19. Cavdar VC, Gokmen Y, Aric M, Altunkaya T, Erdem CG, Gulturk I, et al. The prognostic value of the EASIX score in patients with metastatic pancreatic cancer. Diagnostics (Basel). 2025;15(14). 10.3390/diagnostics15141740. 10.3390/diagnostics15141740 PMC12293124 40722492

20. Frenking JH, Zhou X, Wagner V, Hielscher T, Kauer J, Mai EK, et al. EASIX-guided risk stratification for complications and outcome after CAR T-cell therapy with ide-cel in relapsed/refractory multiple myeloma. J Immunother Cancer. 2024;12(10). 10.1136/jitc-2024-009220. 10.1136/jitc-2024-009220 PMC11459298 39379098

21. Carollo C, Sorce A, Cirafici E, Ciuppa ME, Mulè G, Caimi G. Silent inflammation, loud consequences: decoding NLR across renal, cardiovascular and metabolic disorders. Int J Mol Sci. 2025;26(17). 10.3390/ijms26178256. 10.3390/ijms26178256 PMC12428752 40943179

22. Jourde-Chiche N Fakhouri F Dou L Bellien J Burtey S Frimat M Endothelium structure and function in kidney health and disease Nat Rev Nephrol 2019 15 2 87 108 10.1038/s41581-018-0098-z 30607032 Jourde-Chiche N, Fakhouri F, Dou L, Bellien J, Burtey S, Frimat M, et al. Endothelium structure and function in kidney health and disease. Nat Rev Nephrol. 2019;15(2):87–108. 10.1038/s41581-018-0098-z. 30607032 10.1038/s41581-018-0098-z

23. Huang P Mai Y Zhao J Yi Y Wen Y Association of systemic immune-inflammation index and systemic inflammation response index with chronic kidney disease: observational study of 40,937 adults Inflamm Res 2024 73 4 655 67 10.1007/s00011-024-01861-0 38489048 Huang P, Mai Y, Zhao J, Yi Y, Wen Y. Association of systemic immune-inflammation index and systemic inflammation response index with chronic kidney disease: observational study of 40,937 adults. Inflamm Res. 2024;73(4):655–67. 10.1007/s00011-024-01861-0. 38489048 10.1007/s00011-024-01861-0

24. Basile DP The endothelial cell in ischemic acute kidney injury: implications for acute and chronic function Kidney Int 2007 72 2 151 6 10.1038/sj.ki.5002312 17495858 Basile DP. The endothelial cell in ischemic acute kidney injury: implications for acute and chronic function. Kidney Int. 2007;72(2):151–6. 10.1038/sj.ki.5002312. 17495858 10.1038/sj.ki.5002312

25. Vercellotti GM Tolins JP Endothelial activation and the kidney: vasomediator modulation and antioxidant strategies Am J Kidney Dis 1993 21 3 331 43 10.1016/s0272-6386(12)80757-8 8447314 Vercellotti GM, Tolins JP. Endothelial activation and the kidney: vasomediator modulation and antioxidant strategies. Am J Kidney Dis. 1993;21(3):331–43. 10.1016/s0272-6386(12)80757-8. 8447314 10.1016/s0272-6386(12)80757-8

26. Speer T Dimmeler S Schunk SJ Fliser D Ridker PM Targeting innate immunity-driven inflammation in CKD and cardiovascular disease Nat Rev Nephrol 2022 18 12 762 78 10.1038/s41581-022-00621-9 36064794 Speer T, Dimmeler S, Schunk SJ, Fliser D, Ridker PM. Targeting innate immunity-driven inflammation in CKD and cardiovascular disease. Nat Rev Nephrol. 2022;18(12):762–78. 10.1038/s41581-022-00621-9. 36064794 10.1038/s41581-022-00621-9

27. Song Z Tsou S Martin F Kayumov M Xiao Y Zhou H Kidney Disease as a Driver of Immunosenescence: Mechanisms and Potential Interventions J Am Soc Nephrol 2025 10.1681/asn.0000000896 40938671 PMC12861586 Song Z, Tsou S, Martin F, Kayumov M, Xiao Y, Zhou H, et al. Kidney Disease as a Driver of Immunosenescence: Mechanisms and Potential Interventions. J Am Soc Nephrol. 2025. 10.1681/asn.0000000896. 40938671 10.1681/ASN.0000000896 PMC12861586

28. Lin S Lu W Wang T Wang Y Leng X Chi L Predictive model of acute kidney injury in critically ill patients with acute pancreatitis: a machine learning approach using the MIMIC-IV database Ren Fail 2024 46 1 2303395 10.1080/0886022x.2024.2303395 38264967 PMC10810629 Lin S, Lu W, Wang T, Wang Y, Leng X, Chi L, et al. Predictive model of acute kidney injury in critically ill patients with acute pancreatitis: a machine learning approach using the MIMIC-IV database. Ren Fail. 2024;46(1):2303395. 10.1080/0886022x.2024.2303395. 38264967 10.1080/0886022X.2024.2303395 PMC10810629

29. Johnson A, Bulgarelli L, Pollard T, Gow B, Moody B, Horng S et al. MIMIC-IV. In: 3.1 edn. Physio Net. 2024.

30. Pennisi M Sanchez-Escamilla M Flynn JR Shouval R Alarcon Tomas A Silverberg ML Modified EASIX predicts severe cytokine release syndrome and neurotoxicity after chimeric antigen receptor T cells Blood Adv 2021 5 17 3397 406 10.1182/bloodadvances.2020003885 34432870 PMC8525234 Pennisi M, Sanchez-Escamilla M, Flynn JR, Shouval R, Alarcon Tomas A, Silverberg ML, et al. Modified EASIX predicts severe cytokine release syndrome and neurotoxicity after chimeric antigen receptor T cells. Blood Adv. 2021;5(17):3397–406. 10.1182/bloodadvances.2020003885. 34432870 10.1182/bloodadvances.2020003885 PMC8525234

31. Kordelas L Terzer T Gooley T Davis C Sandmaier BM Sorror M EASIX-1year and late mortality after allogeneic stem cell transplantation Blood Adv 2023 7 18 5374 81 10.1182/bloodadvances.2022008617 37477588 PMC10509665 Kordelas L, Terzer T, Gooley T, Davis C, Sandmaier BM, Sorror M, et al. EASIX-1year and late mortality after allogeneic stem cell transplantation. Blood Adv. 2023;7(18):5374–81. 10.1182/bloodadvances.2022008617. 37477588 10.1182/bloodadvances.2022008617 PMC10509665

32. Băetu AE, Mirea LE, Cobilinschi C, Grințescu IC, Grințescu IM. Platelet contribution and endothelial activation and stress index-potential mortality predictors in traumatic brain injury. Int J Mol Sci. 2024;25(14). 10.3390/ijms25147763. 10.3390/ijms25147763 PMC11276696 39063005

33. Dai H Penack O Radujkovic A Schult D Majer-Lauterbach J Blau IW Early bilirubinemia after allogeneic stem cell transplantation-an endothelial complication Bone Marrow Transpl 2021 56 7 1573 83 10.1038/s41409-020-01186-6 PMC8263345 33517355 Dai H, Penack O, Radujkovic A, Schult D, Majer-Lauterbach J, Blau IW, et al. Early bilirubinemia after allogeneic stem cell transplantation-an endothelial complication. Bone Marrow Transpl. 2021;56(7):1573–83. 10.1038/s41409-020-01186-6. 10.1038/s41409-020-01186-6 PMC8263345 33517355

34. Hunutlu F, Öztop H, Gürsoy V, Ersal T, Elgün E, Yavuz Ş, et al. Modified endothelial activation and stress index: a new predictor for survival outcomes in classical hodgkin lymphoma treated with doxorubicin-bleomycin-vinblastine-dacarbazine-based therapy. Diagnostics (Basel). 2025;15(2). 10.3390/diagnostics15020185. 10.3390/diagnostics15020185 PMC11764177 39857069

35. Liu Y Ren Y Gao L Ahmed S Lu X Li B Endothelial Activation and Stress Index Serves as a Predictor for Survival in Lymphoma-Associated Hemophagocytic Lymphohistiocytosis: A Retrospective Multicenter Cohort Study of Jiangsu Cooperative Lymphoma Group Am J Hematol 2025 10.1002/ajh.70063 40919852 Liu Y, Ren Y, Gao L, Ahmed S, Lu X, Li B, et al. Endothelial Activation and Stress Index Serves as a Predictor for Survival in Lymphoma-Associated Hemophagocytic Lymphohistiocytosis: A Retrospective Multicenter Cohort Study of Jiangsu Cooperative Lymphoma Group. Am J Hematol. 2025. 10.1002/ajh.70063. 40919852 10.1002/ajh.70063

36. Venkatachalam MA Weinberg JM Kriz W Bidani AK Failed Tubule Recovery AKI-CKD Transition, and Kidney Disease Progression J Am Soc Nephrol 2015 26 8 1765 76 10.1681/asn.2015010006 25810494 PMC4520181 Venkatachalam MA, Weinberg JM, Kriz W, Bidani AK, Failed Tubule Recovery AKI-CKD. Transition, and Kidney Disease Progression. J Am Soc Nephrol. 2015;26(8):1765–76. 10.1681/asn.2015010006. 25810494 10.1681/ASN.2015010006 PMC4520181

37. Yuan Q Tang B Zhang C Signaling pathways of chronic kidney diseases, implications for therapeutics Signal Transduct Target Ther 2022 7 1 182 10.1038/s41392-022-01036-5 35680856 PMC9184651 Yuan Q, Tang B, Zhang C. Signaling pathways of chronic kidney diseases, implications for therapeutics. Signal Transduct Target Ther. 2022;7(1):182. 10.1038/s41392-022-01036-5. 35680856 10.1038/s41392-022-01036-5 PMC9184651

38. Meng XM Wang L Nikolic-Paterson DJ Lan HY Innate immune cells in acute and chronic kidney disease Nat Rev Nephrol 2025 21 7 464 82 10.1038/s41581-025-00958-x 40263532 Meng XM, Wang L, Nikolic-Paterson DJ, Lan HY. Innate immune cells in acute and chronic kidney disease. Nat Rev Nephrol. 2025;21(7):464–82. 10.1038/s41581-025-00958-x. 40263532 10.1038/s41581-025-00958-x

39. Tripepi G Mattace Raso F Sijbrands E Seck MS Maas R Boger R Inflammation and asymmetric dimethylarginine for predicting death and cardiovascular events in ESRD patients Clin J Am Soc Nephrol 2011 6 7 1714 21 10.2215/cjn.11291210 21642364 Tripepi G, Mattace Raso F, Sijbrands E, Seck MS, Maas R, Boger R, et al. Inflammation and asymmetric dimethylarginine for predicting death and cardiovascular events in ESRD patients. Clin J Am Soc Nephrol. 2011;6(7):1714–21. 10.2215/cjn.11291210. 21642364 10.2215/CJN.11291210

40. Carrero JJ Hecking M Chesnaye NC Jager KJ Sex and gender disparities in the epidemiology and outcomes of chronic kidney disease Nat Rev Nephrol 2018 14 3 151 64 10.1038/nrneph.2017.181 29355169 Carrero JJ, Hecking M, Chesnaye NC, Jager KJ. Sex and gender disparities in the epidemiology and outcomes of chronic kidney disease. Nat Rev Nephrol. 2018;14(3):151–64. 10.1038/nrneph.2017.181. 29355169 10.1038/nrneph.2017.181

41. Balachandran VP Gonen M Smith JJ DeMatteo RP Nomograms in oncology: more than meets the eye Lancet Oncol 2015 16 4 e173 80 10.1016/s1470-2045(14)71116-7 25846097 PMC4465353 Balachandran VP, Gonen M, Smith JJ, DeMatteo RP. Nomograms in oncology: more than meets the eye. Lancet Oncol. 2015;16(4):e173–80. 10.1016/s1470-2045(14)71116-7. 25846097 10.1016/S1470-2045(14)71116-7 PMC4465353

42. Xie W Li Y Meng X Zhao M Machine learning prediction models and nomogram to predict the risk of in-hospital death for severe DKA: A clinical study based on MIMIC-IV, eICU databases, and a college hospital ICU Int J Med Inf 2023 174 105049 10.1016/j.ijmedinf.2023.105049 37001474 Xie W, Li Y, Meng X, Zhao M. Machine learning prediction models and nomogram to predict the risk of in-hospital death for severe DKA: A clinical study based on MIMIC-IV, eICU databases, and a college hospital ICU. Int J Med Inf. 2023;174:105049. 10.1016/j.ijmedinf.2023.105049. 10.1016/j.ijmedinf.2023.105049 37001474