Risk Prediction of Arteriovenous Fistula Dysfunction in Hemodialysis Patients Using Routine Clinical Indicators

Abstract

Introduction

Arteriovenous fistula (AVF) is the preferred vascular access for hemodialysis (HD) patients, yet AVF dysfunction remains a prevalent complication in maintenance HD. The risk factors influencing AVF patency are not fully defined. This study aimed to identify key clinical predictors and develop a practical model for predicting AVF dysfunction in HD patients.

Methods

We retrospectively reviewed medical records of HD patients treated between January 1, 2020, and February 28, 2025, at the Hemodialysis Center of the People’s Hospital of Baoan, Shenzhen. Demographic characteristics, history of cardiometabolic disease, and laboratory parameters were evaluated. A Least Absolute Shrinkage and Selection Operator (LASSO) Cox regression model was used to select the most relevant predictors, followed by multivariate Cox proportional hazards regression to construct the final prediction model. Model discrimination was assessed using the concordance index (C-index), and internal validation was performed via bootstrap resampling.

Results

Among 439 patients (median age 53 years; 61.3% male), 46 (10.5%) developed AVF dysfunction over a median follow-up of 2.9 years. LASSO regression identified five variables – total protein, albumin, left ventricular ejection fraction (LVEF), history of hypertension, and history of heart disease – as the most predictive. In the multivariate Cox model, all five variables remained statistically significant: total protein (hazard ratio [HR]: 0.604; 95% confidence interval [CI]: 0.372–0.983), albumin (HR: 0.468; 95% CI: 0.225–0.969), LVEF (HR: 0.627; 95% CI: 0.522–0.753), history of hypertension (HR: 2.234; 95% CI: 1.086–4.598), and history of heart disease (HR: 1.950; 95% CI: 1.024–3.715). The final model yielded a C-index of 0.812 (95% CI: 0.753–0.871), with consistent performance in internal bootstrap validation.

Conclusion

This study identified five routinely available clinical variables as independent predictors of AVF dysfunction in HD patients and developed a nomogram with strong predictive accuracy. This tool may support early risk stratification and guide timely interventions to reduce AVF failure and improve dialysis efficacy.

Introduction

With lifestyle transitions and a rapidly aging population, the global burden of end-stage renal disease (ESRD) continues to rise. Maintenance hemodialysis (MHD) remains the predominant form of renal replacement therapy for patients with ESRD. Among available vascular access options, the autologous arteriovenous fistula (AVF) is the most commonly used and is widely recommended by international guidelines due to its technical simplicity, long functional lifespan, low infection rate, and stable blood flow [1–4].

Successful creation and long-term maintenance of AVF function are essential for ensuring adequate dialysis delivery, optimizing patient prognosis, and enhancing quality of life. However, AVF dysfunction is a common and serious complication, often arising from repeated cannulation, thrombosis, stenosis, decreased blood flow, or infection [3, 5]. Its reported incidence varies widely, ranging from 3.9% to 39% [6]. AVF dysfunction not only compromises dialysis adequacy but also contributes to increased hospitalization, cardiovascular risk, and healthcare costs. Therefore, early identification and prevention of AVF dysfunction represent critical priorities in the care of hemodialysis (HD) patients.

Despite progress in surgical techniques and postoperative care, the incidence of AVF dysfunction remains high, suggesting a complex and multifactorial pathophysiology. Previous studies have identified a range of potential risk factors – including cardiovascular comorbidities, nutritional status, and hemodynamic parameters – that may influence AVF patency [7, 8]. There remains a need for simple, reliable models that can help stratify risk and guide early intervention in everyday dialysis care.

To address this gap, we developed a clinically applicable prediction model for AVF dysfunction using routinely collected demographic, echocardiographic, and biochemical indicators. A nomogram was constructed to facilitate individualized risk estimation. This tool has the potential to support risk-based patient management, enable early identification of high-risk individuals, and improve the overall quality and efficiency of dialysis care.

Methods

Study Design and Population

This retrospective cohort study consecutively screened patients with ESRD who underwent MHD at the Hemodialysis Center of the Department of Nephrology, People’s Hospital of Baoan Shenzhen, from January 1, 2020, to February 28, 2025. Two well-trained physicians independently reviewed the medical records, extracted the relevant data and then cross-checked the datasets for accuracy. Only data that passed this verification were included in the analysis. The study included adult patients who were receiving MHD three times per week via an autogenous AVF. Exclusion criteria were as follows: (1) presence of severe comorbidities, including malignancies, mental illness, or cognitive impairment; (2) history of organ transplantation; (3) severe infections or coagulation disorders; (4) pregnancy or lactation; (5) dysfunction of other vital organs (e.g., liver or lung); and (6) incomplete clinical data. Finally, a total of 439 adult patients were included in the analysis, of whom only one had a history of prior AVF use (shown in Fig. 1). The study was approved by the Ethics Committee of the People’s Hospital of Baoan Shenzhen.

Fig. 1.

Flowchart of participant selection and analysis strategy. ESRD, end-stage renal disease; MHD, maintenance hemodialysis; LASSO, Least Absolute Shrinkage and Selection Operator; ROC, receiver operating characteristic; C-index, concordance index.

Potential Predictive Variables

The selection of candidate predictors was informed by prior clinical knowledge and pathophysiological considerations, focusing on variables that are routinely obtainable in standard HD care [6, 9, 10]. All variables were collected within 1 week prior to the creation of the AVF, including demographic characteristics, history of cardiometabolic disease, and laboratory parameters.

Demographic information included age, sex, and body mass index (BMI). History of cardiometabolic disease was determined based on the presence or absence of diabetes mellitus, hypertension, heart disease, and cerebrovascular disease. The diagnostic criteria for these comorbidities are detailed in online supplementary Table 1 (for all online suppl. material, see https://doi.org/10.1159/000550128) [11].

Laboratory parameters included a broad range of hematological, biochemical, and cardiovascular indicators relevant to dialysis outcomes. These comprised white blood cell count, hemoglobin concentration, neutrophil percentage, thrombin time, activated partial thromboplastin time, plasma fibrinogen, prothrombin time, and the international normalized ratio. Renal and metabolic parameters included serum creatinine, bicarbonate, blood urea nitrogen, uric acid, β2-microglobulin, last estimated glomerular filtration rate prior to AVF creation which was calculated by the Modification of Diet in Renal Disease (MDRD) 4-variable formula, cystatin C, total protein, albumin, and the albumin-to-globulin ratio. Nutritional and lipid-related indicators, such as low-density lipoprotein cholesterol, triglycerides, apolipoprotein AⅠ, apolipoprotein B, atherosclerosis index, parathyroid hormone, 25-hydroxyvitamin D, unsaturated iron-binding capacity, transferrin, and serum iron were also included. In addition, left ventricular ejection fraction (LVEF), measured via echocardiography, was incorporated as a marker of cardiac function, given its potential impact on vascular flow dynamics. Together, these variables were selected to capture a comprehensive clinical profile of patients prior to AVF creation, and to reflect factors that are potentially associated with AVF patency and dysfunction risk in the context of maintenance HD.

Definition of Outcomes

AVF dysfunction was defined as one or more of the following: AVF blood flow <200 mL/min, failure to meet dialysis prescription requirements, difficulty in cannulation, or reduced dialysis adequacy. Reduced dialysis adequacy is defined as a failure to maintain a single-pool Kt/V of at least 1.2.

Fistula duration was calculated as the number of days from AVF creation to the first percutaneous transluminal angioplasty procedure in patients with dysfunction. For patients without AVF dysfunction, follow-up time was calculated from AVF creation to the earliest occurrence of death, kidney transplantation, loss to follow-up, or study endpoint (February 28, 2025), whichever came first.

Statistical Analysis

Continuous variables were described as mean ± standard deviation or median with interquartile range, depending on data distribution. Categorical variables were presented as counts and percentages. Univariate analyses were assessed using Student’s t tests for normally distributed variables and Wilcoxon rank-sum tests for non-normally distributed variables. Variables with more than 30% missing values (e.g., low-density lipoprotein cholesterol, triglycerides, apolipoprotein AⅠ and B, atherosclerosis index, 25-hydroxyvitamin D, unsaturated iron-binding capacity, transferrin, and serum iron) were excluded to ensure statistical reliability, as high missingness may bias parameter estimates and distort true associations, and a random forest algorithm implemented via the R package missForest was applied to impute variables with acceptable levels of missing data [12].

To reduce model overfitting and identify the most informative predictors, Least Absolute Shrinkage and Selection Operator (LASSO) Cox regression was applied to the 27 candidate variables [13]. LASSO applies L1 regularization to shrink less informative coefficients toward zero and performs variable selection simultaneously. The optimal penalty parameter (λ) was determined through 10-fold cross-validation using the “one-standard-error” rule to balance parsimony and performance. This procedure was implemented using the glmnet package in R [14].

Variables selected by the LASSO procedure were then included in a multivariate Cox proportional hazards regression model. The Cox proportional hazards regression model assumes that candidate variables significantly influence AVF dysfunction risk and that their hazard ratios (HRs) remain proportional over the follow-up period. The purpose of this step was to identify variables with the highest predictive value for AVF dysfunction risk, expressed as HRs with 95% confidence intervals (CIs). The proportional hazards assumption was assessed using the Kolmogorov-type supremum test.

Based on the final model, a nomogram was developed using the rms package in R [15] to provide individualized risk prediction for AVF dysfunction. Model discrimination was evaluated using the concordance index (C-index), and calibration was assessed by comparing predicted and observed probabilities through calibration curves. Internal validation was performed using 1,000 bootstrap resamples to assess the model’s performance stability [16, 17]. Time-dependent receiver operating characteristic curves were used to evaluate model discrimination at 1-, 2-, and 3-year time points. Decision curve analysis was conducted using the dcurves package [18] to evaluate the clinical utility of the model by quantifying net benefit across a range of threshold probabilities.

All analyses were conducted using R software version 4.2.1. Statistical significance was defined as a two-sided p value <0.05.

Results

Participant Characteristics

A total of 439 participants were included in this study. The median age was 53 years, and 61.3% of participants were male. The median BMI was 22.4 kg/m². Among participants, the respective prevalence rates of hypertension, diabetes mellitus, heart disease, and cerebrovascular disease were 51.9%, 29.6%, 18.5%, and 4.6%. Over a median follow-up duration of 2.9 years, 46 (10.5%) participants developed AVF dysfunction (Table 1).

Table 1.

Demographic and clinical characteristics of maintenance HD patients at baseline

Characteristics	Overall	Incident AVF dysfunction		p value
		yes	no
Participants, n	439	46	393	​
Demographics
Gender, N (%)
Men	269 (61.28)	33 (71.74)	236 (60.05)	0.168
Women	170 (38.72)	13 (28.26)	157 (39.95)	​
Age, years	53.00 [44.00, 62.00]	57.50 [46.25, 69.50]	53.00 [43.00, 61.00]	0.05
BMI, kg/m2	22.41 [20.36, 24.63]	24.01 [20.73, 25.67]	22.24 [20.35, 24.44]	0.039
Diabetes, N (%)
Yes	130 (29.61)	18 (39.13)	112 (28.50)	0.186
No	309 (70.39)	28 (60.87)	281 (71.50)	​
Hypertension, N (%)
Yes	228 (51.94)	35 (76.09)	193 (49.11)	0.001
No	211 (48.06)	11 (23.91)	200 (50.89)	​
Heart disease, N (%)
Yes	81 (18.45)	16 (34.78)	65 (16.54)	0.005
No	358 (81.55)	30 (65.22)	328 (83.46)	​
Cerebrovascular disease, N (%)
Yes	20 (4.56)	6 (13.04)	14 (3.56)	0.011
No	419 (95.44)	40 (86.96)	379 (96.44)	​
Fistula duration, day	1,073.00 [516.50, 1,933.50]	293.00 [132.50, 546.00]	1,216.00 [647.00, 2,062.00]	<0.001
Dialysis vintage, month	31.00 [15.00, 56.50]	16.50 [6.25, 37.25]	33.00 [16.00, 59.00]	<0.001
Clinical laboratory data
White blood cell count, ×109/L	6.80 [5.50, 8.21]	7.14 [5.99, 8.47]	6.74 [5.45, 8.17]	0.468
Hemoglobin concentration, g/L	96.28 (16.91)	93.41 (18.18)	96.62 (16.75)	0.224
Percentage of neutrophils	0.68 (0.12)	0.71 (0.09)	0.67 (0.12)	0.064
Thrombin time, s	17.30 [16.50, 18.20]	17.45 [16.50, 18.30]	17.30 [16.50, 18.20]	0.86
Activate partial thromboplastin time, s	28.57 [26.17, 30.90]	29.70 [26.60, 30.90]	28.40 [26.00, 30.90]	0.519
Plasma fibrinogen, g/L	3.82 [2.99, 4.70]	3.93 [3.28, 4.60]	3.80 [2.95, 4.76]	0.472
Plasma prothrombin time, s	12.10 [11.40, 12.90]	12.10 [11.53, 12.60]	12.10 [11.40, 12.90]	0.937
International normalized ratio	1.04 [0.99, 1.11]	1.04 [1.00, 1.10]	1.04 [0.99, 1.11]	0.841
Creatinine, μmol/L	869.58 [625.48, 1,130.82]	988.60 [783.28, 1,151.60]	850.65 [596.18, 1,125.65]	0.016
Bicarbonate, mmol/L	21.52 [19.00, 24.53]	20.73 [18.29, 22.79]	21.80 [19.03, 24.74]	0.075
Urea nitrogen, mmol/L	19.11 [13.93, 24.60]	21.40 [15.20, 25.29]	19.00 [13.70, 24.10]	0.156
Uric acid, μmol/L	391.66 (138.68)	413.29 (100.65)	389.13 (142.37)	0.264
β2 microglobulin, mg/L	21.81 [17.24, 25.47]	23.08 [17.89, 24.85]	21.72 [17.09, 25.61]	0.898
Glomerular filtration rate, mL/min/1.73 m2	5.39 [3.94, 7.98]	5.47 [4.34, 7.54]	5.39 [3.89, 8.14]	0.961
Cystatin, mg/L	6.45 [4.85, 7.87]	6.94 [5.57, 7.53]	6.43 [4.78, 7.89]	0.229
White globulin ratio	1.16 [1.00, 1.32]	1.15 [1.01, 1.27]	1.17 [1.00, 1.32]	0.374
Total protein, g/L	69.80 (7.88)	64.75 (8.21)	70.40 (7.63)	<0.001
Albumin, g/L	37.71 (5.45)	34.46 (6.11)	38.09 (5.24)	<0.001
Low-density lipoprotein cholesterol, mmol/L	2.77 [2.13, 3.59]	3.30 [2.62, 4.77]	2.56 [2.07, 3.30]	<0.001
Triglyceride, mmol/L	1.11 [0.78, 1.74]	1.10 [0.80, 1.49]	1.14 [0.76, 1.90]	0.396
Apolipoprotein AⅠ, g/L	1.07 [0.89, 1.25]	1.06 [0.86, 1.21]	1.08 [0.90, 1.28]	0.604
Apolipoprotein B, g/L	0.87 (0.29)	0.88 (0.28)	0.86 (0.29)	0.668
Atherosclerosis index, mmol/L	2.80 [1.95, 4.20]	2.90 [2.30, 4.45]	2.65 [1.90, 3.80]	0.11
Parathyroid hormone, pg/mL	237.45 [134.38, 429.92]	225.70 [150.85, 379.80]	243.00 [125.90, 437.10]	0.915
25-hydroxyvitamin D, ng/mL	19.40 [15.60, 23.77]	17.90 [14.25, 22.50]	19.80 [15.70, 24.00]	0.249
Unsaturated iron-binding capacity, μmol/L	35.40 [25.85, 43.65]	29.60 [20.53, 36.00]	36.20 [27.70, 44.25]	0.006
Transferrin, g/L	2.00 [1.70, 2.40]	1.85 [1.50, 2.18]	2.10 [1.70, 2.50]	0.006
Iron, μmol/L	9.60 [6.60, 13.30]	9.30 [6.90, 14.40]	9.60 [6.60, 13.12]	0.867
LVEF	62.00 [55.00, 69.00]	57.00 [49.25, 62.75]	62.00 [56.00, 70.00]	<0.001

Continuous variables are described as mean (standard deviation, SD) for those with normal distribution or median (interquartile range) for those with non-normal distribution. Categorical variables are described as number (percent).

Predictor Selection

Twenty-seven candidate variables were included in the LASSO Cox regression analysis to identify potential predictors of AVF dysfunction. Based on the “one-standard-error” criterion for selecting the optimal penalty parameter (λ), five variables were retained: total protein, albumin, LVEF, history of hypertension, and history of heart disease (Fig. 2).

Fig. 2.

Selection of predictive variables using LASSO Cox regression. a Coefficient profiles of 27 candidate variables generated by LASSO Cox regression. Each curve represents the trajectory of a variable’s coefficient as the penalty parameter (log λ) increases. b Ten-fold cross-validation was used to identify the optimal λ value. The dotted vertical line indicates the λ corresponding to one standard error above the minimum mean cross-validated error, at which five variables were selected.

Univariate Cox proportional hazards regression analysis demonstrated that all five variables were significantly associated with AVF dysfunction risk. Specifically, higher levels of total protein (HR: 0.411; 95% CI: 0.289–0.585; p < 0.001), albumin (HR: 0.266; 95% CI: 0.152–0.467; p < 0.001), and LVEF (HR: 0.621; 95% CI: 0.521–0.739; p < 0.001) were associated with lower risk, while history of hypertension (HR: 3.279; 95% CI: 1.665–6.460; p = 0.001) and history of heart disease (HR: 2.533; 95% CI: 1.380–4.648; p = 0.003) were associated with higher risk (Table 2). The proportionality hazards assumption was met (p = 0.29).

Table 2.

Univariate and multivariate Cox regression analyses of predictive factors for AVF dysfunction and model performance

​	Univariable						Multivariable
	HR (95% CI)					p value	HR (95% CI)	p value
Total protein, 10 g/L	0.411 (0.289, 0.585)	​	​	​	​	<0.001	0.604 (0.372, 0.983)	0.042
Albumin, 10 g/L	​	0.266 (0.152, 0.467)	​	​	​	<0.001	0.468 (0.225, 0.969)	0.041
LVEF, 5%	​	​	0.621 (0.521, 0.739)	​	​	<0.001	0.627 (0.522, 0.753)	<0.001
Hypertension, N (%)
No	​	​	​	Reference	​	​	Reference	​
Yes	​	​	​	3.279 (1.665, 6.460)	​	0.001	2.234 (1.086, 4.598)	0.029
Heart disease, N (%)
No	​	​	​	​	Reference	​	Reference	​
Yes	​	​	​	​	2.533 (1.380, 4.648)	0.003	1.950 (1.024, 3.715)	0.042
C-index	0.707 (0.631, 0.782)	0.690 (0.612, 0.768)	0.704 (0.632, 0.776)	0.637 (0.574, 0.699)	0.580 (0.513, 0.647)	​	0.812 (0.753, 0.871)	​
C-index (internal bootstrap)	​	​	​	​	​	​	0.813 (0.753,0.873)	​

Multivariate Cox regression includes total protein, albumin, LVEF, history of hypertension, and history of heart disease. The concordance index (C-index) was used to evaluate model discrimination. Bootstrap validation was performed with 1,000 resamples.

HR, hazard ratio; CI, confidence interval; LVEF, left ventricular ejection fraction.

To evaluate the predictive ability of each variable, the C-index was calculated. The C-index for predicting AVF dysfunction was 0.707 (95% CI: 0.631–0.782) for total protein, 0.690 (95% CI: 0.612–0.768) for albumin, 0.704 (95% CI: 0.632–0.776) for LVEF, 0.637 (95% CI: 0.574–0.699) for history of hypertension, and 0.580 (95% CI: 0.513–0.647) for history of heart disease (Table 2).

Development of Model and Predictive Nomogram

All five variables selected by the LASSO procedure were incorporated into a multivariate Cox proportional hazards regression model to quantify their associations with the risk of AVF dysfunction. The results showed that each of the five variables remained significantly associated with the outcome: total protein (HR: 0.604; 95% CI: 0.372–0.983; p = 0.042), albumin (HR: 0.468; 95% CI: 0.225–0.969; p = 0.041), LVEF (HR: 0.627; 95% CI: 0.522–0.753; p < 0.001), history of hypertension (HR: 2.234; 95% CI: 1.086–4.598; p = 0.029), and history of heart disease (HR: 1.950; 95% CI: 1.024–3.715; p = 0.042) (Table 2). The overall C-index of the model incorporating these five variables was 0.812 (95% CI: 0.753–0.871), indicating good discriminative ability (Table 2). Based on the multivariate Cox model, a nomogram was constructed to visually represent the predictive model (shown in Fig. 3).

Fig. 3.

Nomogram for predicting the risk of arteriovenous fistula (AVF) dysfunction in maintenance HD patients. The nomogram integrates five independent predictors – total protein, albumin, left ventricular ejection fraction (LVEF), history of hypertension, and history of heart disease – identified from multivariate Cox regression analysis. To estimate an individual patient’s risk of AVF dysfunction, locate the patient’s value on each variable axis, draw a vertical line to the “Points” axis to determine the score for each variable, and sum the points. The total score corresponds to the estimated probabilities of AVF dysfunction at 1, 2, and 3 years.

Internal Validation of the Predictive Nomogram

Internal validation using 1,000 bootstrap resamples demonstrated that the model had good predictive accuracy, with a corrected C-index of 0.813 (95% CI: 0.753–0.873) (Table 2). Calibration curve analysis showed good agreement between predicted and observed probabilities of AVF dysfunction at 1-, 2-, and 3-year follow-up intervals (shown in Fig. 4a). Time-dependent receiver operating characteristic analysis indicated that the areas under the curve for predicting AVF dysfunction at 1, 2, and 3 years were 0.818 (95% CI: 0.734–0.902), 0.836 (95% CI: 0.777–0.895), and 0.852 (95% CI: 0.796–0.908), respectively (shown in Fig. 4b).

Fig. 4.

Calibration and discrimination performance of the predictive nomogram. a Calibration curves for 1-, 2-, and 3-year predictions of AVF dysfunction. The x-axis represents the predicted probability, and the y-axis represents the observed probability. The 45-degree dashed line indicates perfect calibration. The close alignment of the curves with the diagonal line suggests good agreement between predicted and observed outcomes. b Time-dependent receiver operating characteristic (ROC) curves for 1-, 2-, and 3-year predictions. The areas under the curve (AUCs) were 0.818, 0.836, and 0.852, respectively, indicating strong discriminatory ability of the nomogram over time.

As shown in Figure 5, the decision curve analysis showed that the nomogram yielded a higher net benefit than the treat-all or treat-none approaches when the threshold probability ranged from 1% to 46% at 1 year, 2%–66% at 2 years, and 2%–53% at 3 years, indicating good clinical applicability across multiple decision thresholds. These findings suggest that the model could support decision-making by identifying patients who are most likely to benefit from early interventions targeting AVF dysfunction.

Fig. 5.

Decision curve analysis (DCA) for the predictive model. The DCA curves evaluate the net clinical benefit of the predictive model for AVF dysfunction at 1-, 2-, and 3-year time points across a range of threshold probabilities. The x-axis represents the threshold probability, and the y-axis represents the net benefit. The model demonstrates superior net benefit compared to the “treat-all” and “treat-none” strategies within a wide range of thresholds, indicating potential for clinical decision-making.

Discussion

This study retrospectively analyzed risk factors for AVF dysfunction in patients undergoing MHD and developed a predictive model using multivariate Cox regression analysis. Total protein, albumin, LVEF, history of hypertension, and history of heart disease were identified as independent risk factors. A nomogram based on these predictors was constructed and internally validated, demonstrating good discrimination and calibration, as well as potential clinical utility for individualized risk assessment. For example, a patient with a history of both hypertension and heart disease, a total protein level of 70 g/L, an albumin level of 40 g/L, and an LVEF of 40 would receive approximately 1.72 points for hypertension, 1.43 points for heart disease, 2.70 points for total protein, 2.44 points for albumin, and 8.00 points for LVEF, totaling approximately 16.29 points. According to the nomogram, this patient would have an estimated 41% probability of AVF dysfunction at 1 year, 61% at 2 years, and 67% at 3 years.

Comparison with Existing AVF Dysfunction Prediction Models

A variety of models have been developed to predict AVF dysfunction in patients undergoing maintenance HD, each with distinct strengths and limitations. Early models such as that proposed by Lok et al. [19] incorporated factors including age, coronary artery disease, peripheral vascular disease, and race to estimate the risk of AVF maturation failure. While helpful for risk stratification, the model showed poor generalizability, as demonstrated by its failure to validate in a cohort of 694 Asian patients [20]. Wongmahisorn [21] proposed a simplified score using readily available clinical variables such as diabetes mellitus, history of central venous catheter placement, and access-related interventions. Despite good internal performance and ease of use, it excluded patients with unsuccessful cannulation, limiting its application to early AVF outcomes. Similarly, Hasuike et al. [22] identified risk factors including low blood flow rate, use of arteriovenous grafts, and elevated coagulation markers. Although informative, the model’s reliance on specialized biomarkers and single-center data reduces its practicality and external relevance.

More recent models have employed alternative statistical and data-driven approaches. Logistic regression-based models, such as those developed by Eslami et al. [23] and Kumar et al. [24], demonstrated moderate to high discriminative performance but were limited by small sample sizes and lack of external validation. Machine learning models have also been introduced to improve prediction accuracy. Peralta et al. [25] analyzed data from over 13,000 patients across European dialysis centers and achieved an areas under the curve of 0.80. However, its applicability to other populations, particularly in Asia, remains uncertain. Grochowina et al. [26] proposed a novel acoustic-based diagnostic system (NefDiag), which showed acceptable accuracy but was tested in a small cohort and exhibited limitations in noisy environments. In contrast, our model integrates five routinely measured clinical variables – total protein, albumin, LVEF, history of hypertension, and history of heart disease – and demonstrates strong discriminative performance (C-index: 0.813) following internal validation. Our study addresses these limitations by using a Chinese cohort and incorporating routinely available clinical indicators, thereby enhancing the clinical applicability of the prediction model. It offers a simple, accessible, and practical tool for early identification of patients at high risk of AVF dysfunction in real-world clinical settings.

Risk Factors for AVF Dysfunction in MHD Patients

As the main contributor to plasma colloid osmotic pressure, low total protein levels may lead to hypovolemia, blood stasis, and a hypercoagulable state, increasing thrombosis risk [27]. Protein deficiency may also impair immune function, promoting inflammation and intimal hyperplasia. Our results on albumin align with previous studies [25, 28], confirming hypoalbuminemia as a significant risk factor. While albumin exerts anticoagulant effects by inhibiting fibrin polymerization and platelet aggregation, low levels conversely promote coagulation and endothelial activation and are linked to endothelial inflammation in HD patients [29, 30]. Adequate blood flow (>600 mL/min) is essential for effective HD, typically requiring blood flow above 600 mL/min. As native venous flow is insufficient, achieved via AVF creation [31]. Reduced LVEF reflects impaired cardiac output, which leads to slower circulation, blood stasis, and platelet activation, all of which are conditions favoring thrombosis. Our findings support prior evidence linking low LVEF to AVF dysfunction [32]. Hypertension promotes endothelial injury, intimal hyperplasia, and vascular stenosis, increasing AVF failure risk [33, 34]. Elevated pulse pressure and neurohormonal activation further exacerbate vascular damage through inflammation and increased permeability [35]. Heart disease shares common pathways with CKD, including systemic inflammation and endothelial dysfunction, which enhance vasoconstriction and thrombosis [36]. Comorbid dyslipidemia and vascular calcification increase shear stress and medial thickening, further predisposing to AVF dysfunction [37].

Female sex and older age are commonly associated with higher risk of AVF dysfunction, likely due to smaller vessel size and vascular aging, respectively [38–42]. Diabetes may also contribute though endothelial damage and hypercoagulability [43]. However, these associations were not significant in our cohort – possibly due to the predominance of male participants, limited representation of older adults, or insufficient statistical power. Further studies with larger and more diverse populations are warranted to validate these findings.

Limitation

However, there are several limitations that should be acknowledged. First, the patients included in this study were recruited from a single clinical center, which may limit the generalizability of the findings. Second, the absence of initial AVF blood flow data may have constrained the predictive performance of the model. Third, external validation is necessary to establish the broader applicability of our results.

Conclusion

In summary, total protein, albumin, LVEF, history of hypertension, and history of heart disease were identified as independent risk factors for AVF dysfunction in patients undergoing maintenance HD. The predictive model developed based on these variables demonstrated good performance and clinical applicability. This model may serve as a practical tool for early risk stratification, enabling clinicians to implement timely and targeted interventions aimed at reducing AVF dysfunction and enhancing dialysis effectiveness.

Statement of Ethics

The study was approved by the Institutional Review Board (IRB) of The Second Affiliated Hospital of Shenzhen University (Approval No. BYL20231005). Written informed consent was obtained from the parent/legal guardian of participants prior to the study.

Conflict of Interest Statement

The authors have no conflicts of interest to declare.

Funding Sources

This study was supported by the High-level Medical Team Project in Baoan, Shenzhen (No. 202401), Shenzhen Baoan’s district medical research project (2022JD139) and Healthcare and Medical Research Projects of Shenzhen Bao’an District Medical Association (BAYXH2023008).

Author Contributions

Xiaolu Sui and Weixue Xiong designed the study, analyzed the data and wrote the manuscript. Qianli Fu, Jinzhu Huang, and Jinling Li collected the data. Tingfei Xie, Yunpeng Xu, Jiahui Chen, and Yanzi Zhang analyzed and interpreted the data. Jihong Chen contributed to the design of the study, and wrote the manuscript, approved the version to be published and agreed to be accountable for all aspects of the work. All authors read and approved the final manuscript.

Funding Statement

This study was supported by the High-level Medical Team Project in Baoan, Shenzhen (No. 202401), Shenzhen Baoan’s district medical research project (2022JD139) and Healthcare and Medical Research Projects of Shenzhen Bao’an District Medical Association (BAYXH2023008).

Data Availability Statement

The data that support the findings of this study are not publicly available due to their containing information that could compromise the privacy of research participants but are available from the corresponding author upon reasonable request.