Computational Frontiers in Arteriovenous Fistula Maturation

Abstract

Arteriovenous (AV) fistulas, the preferred vascular access for hemodialysis, fail to mature in up to 60% of patients with kidney failure. This high failure rate is often attributed to adverse hemodynamic conditions, yet the exact mechanisms remain poorly understood. This review explores the application of computational fluid dynamics and machine learning to elucidate these mechanisms and predict clinical outcomes. Computational fluid dynamic models have been instrumental in characterizing the complex interplay between AV fistula geometry, such as anastomotic angle and curvature, and hemodynamic parameters, such as wall shear stress and oscillatory shear index. These studies consistently link disturbed flow patterns, including low wall shear stress and high oscillatory shear index, to regions prone to neointimal hyperplasia and stenosis. Concurrently, machine learning models have demonstrated significant promise in predicting AV fistula maturation, stenosis, and failure by leveraging diverse data sources, including clinical characteristics, ultrasound imaging, and acoustic bruit analysis. While powerful, the clinical utility of these computational models is often limited by small, single-center datasets, a lack of external validation, and simplifying assumptions that may not capture true physiological complexity. Future progress depends on integrating these complementary approaches, using larger and more diverse datasets, and validating models prospectively to create generalizable tools that can guide surgical planning and improve AV fistula maturation rates.

Introduction

Globally, 1.15 million individuals are diagnosed with kidney failure, with 845,000 initiating long-term hemodialysis annually.¹ The arteriovenous (AV) fistula is the long-term vascular access of choice because of better maturation rates and lower complication rates compared with grafts and catheters.² However, AV fistulas are limited by high maturation failure (up to 60%).^3–6 Successful maturation involves vascular remodeling, in which the vein dilates and thickens to withstand higher pressures and repeated cannulation.^7–12 Altered hemodynamics are integral to maturation, but maladaptive responses drive pathology, leading to stenosis and failure.¹³ Despite the recognized importance of hemodynamics, specific parameters determining maturation remain elusive. Advanced computational tools may bridge this gap. Computational fluid dynamics (CFDs) simulates blood flow in patient-specific or idealized AV fistula geometries, enabling detailed characterization of hemodynamic forces difficult to measure in vivo. In parallel, machine learning (ML) provides a data-driven approach to identify complex patterns and predict outcomes from large, multimodal datasets. This review reports on CFD and ML applications in AV fistula maturation and proposes future directions to translate these computational frontiers into tools to improve clinical outcomes.

Hemodynamics of AV Fistula Creation and Maturation

AV fistula creation connects a high-pressure arterial system to a low-pressure venous system, imposing acute hemodynamic stress on the vessels. As a result, the vein must dilate rapidly to accommodate the surge in flow. Successful maturation is driven by physiological increases in laminar wall shear stress (WSS).^13,14 This promotes a quiescent endothelial cell phenotype and increased nitric oxide production, leading to outward remodeling. Conversely, AV fistula failure is often linked to maladaptive inward remodeling driven by adverse hemodynamic conditions in the juxta-anastomotic region and at bends. These disturbed flow patterns promote endothelial dysfunction and pathological cellular proliferation, culminating in neointimal hyperplasia and stenosis. Nonetheless, current clinical practice lacks standardized guidance regarding the optimal surgical configuration on the basis of individualized hemodynamic parameters.⁷ This is where advanced computational studies are essential. Both CFD and ML can model complex, localized shear stress patterns that cannot be measured noninvasively, thereby bridging the gap between clinical outcomes and the underlying physics of blood flow.

CFD in AV Fistula Maturation

General Principles of CFD

CFD uses numerical analysis, grounded in the Navier–Stokes equations, to assess fluid flows. Because these equations are too complex to solve analytically, especially in biological contexts, these are solved algorithmically. Analysis begins with a geometric model segmented from medical imaging. The geometry is divided into a finite number of smaller domains known as meshing. Once meshed, appropriate boundary conditions are applied. A numerical scheme is chosen to solve the governing equations iteratively. Key hemodynamic parameters derived from simulations include WSS, a measure of the frictional force on the vessel wall, and the oscillatory shear index (OSI), which quantifies WSS deviation over a cardiac cycle. An OSI near 0 indicates laminar flow, while 0.5 signifies highly turbulent flow.

Modeling AV fistulas involves balancing physiological accuracy with computational cost through model assumptions. While often modeled as Newtonian fluid for simplicity, blood is a non-Newtonian, shear-thinning fluid, meaning its viscosity decreases at higher shear rates because of red blood cell disaggregation. Advanced simulations use the Carreau–Yasuda^15–20 equations to capture this behavior. Similarly, while a rigid-wall assumption is computationally efficient, it ignores natural distension and can overestimate WSS. More advanced fluid structure interaction models couple fluid dynamics with solid mechanics for greater accuracy at a higher computational cost. Furthermore, the anastomosis changes the arterial blood flow direction, dramatically increasing the Reynolds number, a dimensionless ratio of inertial to viscous forces. A low Reynolds number describes laminar flow, while larger values, typically more than 2000, describe turbulent flow. The increase in Reynolds number necessitates turbulence models (e.g., k-omega shear stress transport model,^21,22 Large Eddy Simulation²³) to accurately capture the chaotic flow behavior near the anastomosis.

Validation of CFD Output Parameters

Simulation accuracy hinges mainly on input quality, but in vivo validation is necessary to confirm the reliability of computational methods (Figure 1 and Table 1). Accuracy varies by targeted hemodynamic parameter and clinical test used. High-fidelity imaging (magnetic resonance imaging [MRI], magnetic resonance angiography, particle image velocimetry) reduces the margin of error (3%–10%) for flow velocity and pressure. Conversely, ultrasound (US) introduces higher variability and margin of error, sometimes leading to inconsistent predictions of flow volume and velocity.^{17,18,24–29} In addition, variability is produced from non-standardized image acquisition. MacDonald et al. found that differences in MRI sequences produced outputs that differed as much as 15% when using the same phantom.³⁰ Nonetheless, even models with high variability are a powerful tool for understanding the distribution of stresses on vessels.³¹ The standardization of in vivo measurements across all modalities would improve reproducibility, in which case we could assume that purely computational data such as WSS and OSI would be reliable predictors of AV fistula complications.

Figure 1.

Hemodynamic validation and characterization of patient-specific end-to-side radiocephalic AV fistula with computations fluid dynamics. (A) Flow rate waveforms (ml/s) in the vasculature before and after AV fistula creation. (B) Comparison of volumetric flow rate measured with FeMRA and CFD output. (C) Velocity streamlines measured at peak systole (a), mid-deceleration (b), and peak diastole (c), normalized TAWSS (d), and OSI (e). Adapted from ref. 17, with permission. AVF, arteriovenous fistula; CFD, computational fluid dynamic; FeMRA, ferumoxytol-enhanced MRA; MRA, magnetic resonance angiography; OSI, oscillatory shear index; TAWSS, time-averaged WSS; WSS, wall shear stress.

Table 1.

Summary of studies that have compared the output of computational fluid dynamics against various standards

Author	AV Fistula Type	CFDs Output	Standard	Findings
Kharboutly et al., 201024	Human, brachiocephalic, end-to-side (N=1)	Flow velocity fields	PIV	CFDs and PIV produced similar flow fields, with observed differences in magnitude not exceeding 10%
Niemann et al., 201028	In vitro model (molded latex), side-to-side (N=1)	Flow rates	US (plane wave)	CFDs and US flow rates differed by around 58% at the proximal vein and 23% at the distal vein, with relative errors increasing at high-volume flows
Van Canneyt et al., 201326	Human, radiocephalic, end-to-side (N=2)	Flow rate, flow velocity fields	Synthetic pulse wave generated from MRA	CFDs and synthetic pulse wave showed a maximum bias of 8.1% for centered, small sample volume, but the flow rate was highly inaccurate, with a maximum bias of 97.7%
Hoganson et al., 201427	Human, unspecified type (N=8)	Flow rate	US (pulse wave)	The mean difference between CFDs and US flow rates was 17%
Browne et al., 201525	Human, brachiocephalic, end-to-side (N=2)	Pressure gradient	Pressure catheter measurement	The error between CFDs and in vivo pressure drops was 8%–10%
Hyde-Linaker et al., 202217	Human, radiocephalic, end-to-side (N=1)	Flow rate	Phase-contrast MRA velocity	The mean percentage difference between CFDs and phase-contrast MRA flow rates in the feeding artery was 2.52%
Wang et al., 202418	Human, unspecified type, (N=1)	Flow velocity	US (color Doppler)	The standard CFDs model differed by 38% from US. The modified model, which had more symmetric branching at anastomoses and curved heel, differed by 16% from US

AV, arteriovenous; CFD, computational fluid dynamic; MRA, magnetic resonance angiography; PIV, particle image velocimetry; US, ultrasound.

Hemodynamic Parameters and Vascular Outcomes

One of the primary applications of CFD is to identify hemodynamic conditions associated with neointimal hyperplasia and stenosis (Table 2). A consistent finding across numerous studies is that regions of low (<0.4 Pa) and oscillating WSS (high OSI >0.2) are predisposed to neointimal hyperplasia development. Ene-Iordache and Remuzzi reported that low, oscillatory shear stress is located at sites of luminal reduction.¹⁵ Similar studies found abnormal hemodynamics, such as fluctuations in flow rate and changes to the flow pattern, in regions developing stenoses.^16,32 While abnormal flow pattern is predominately considered a risk factor for AV fistula complications in CFD research, there is growing interest in real-life application of dynamic assessment of hemodynamics using color Doppler, spectral doppler, or vector flow US or 4D MRI. Of note, Jia et al. were able to directly correlate neointimal hyperplasia to oscillating WSS in a canine model.³³ Kharboutly et al. postulated that calcification could be linked to WSS but found no significant correlation.³⁴ Looking at the effect that existing stenotic sites have on AV fistula maturation, Chen et al. found that hemodynamics became abnormal, even in nonaffected areas, in what seemed to be a negative feedback cycle that led to total AV fistula failure.³⁵ The variety of clinical studies, as opposed to studies on idealized AV fistula structures, is limited; and there is a benefit in understanding the biology at work by studying the effects of hemodynamics in conjunction with histological data.

Table 2.

Summary of studies that have used computational fluid dynamics for understanding the effect of hemodynamic parameters on vascular outcomes

Author	AV Fistula Type	Data Source	Hemodynamic Parameters	Vascular Pathology	Findings
Kharboutly et al., 200734	Human, brachiocephalic, end-to-side (N=1)	CTA	OSI	Calcification	There was no significant correlation between OSI and calcification
Ene-Iordache and Remuzzi, 201215	Computer model, radiocephalic, end-to-end (n=1); radiocephalic, end-to-side (n=1)	US	OSI, relative residence time, and WSS	Luminal narrowing	Zones of low/oscillatory shear stress were located at sites of luminal narrowing reported in previous studies. In the end-to-side model, this was prominent in the anastomoses and heel. In the end-to-end model, this was in the swing segment
Ene-Iordache et al., 201532	Human, radiocephalic, end-to-side, (N=1)	US, MRA	OSI, WSS	Neointimal hyperplasia	Multidirectional disturbed flow was concentrated on the anastomoses floor and swing segment
Jia et al., 201533	Canine, femoral, end-to-side (N=20)	US	Disturbed flow, WSS	Neointimal hyperplasia	Low and disturbed WSS along the inner wall of the juxta-anastomotic and juxta-ligation segments was found to be correlated with neointimal hyperplasia
Bozzetto et al., 201616	Human, brachiocephalic, end-to-side (n=2); radiocephalic, end-to-side (n=2)	MRA	OSI, WSS	Neointimal hyperplasia	There are marginal fluctuations in the velocity field and flow regimes in the juxta-anastomotic segment
Chen et al., 201635	In vitro model (silicone), end-to-side (N=1)	Pressure measurements in an in vitro model	Pressure and velocity	Degree of stenosis	Stenotic areas increased pressure across the anastomosis site, extending into nonstenosed areas and leading to further adverse vessel remodeling

AV, arteriovenous; CTA, computed tomography angiography; MRA, magnetic resonance angiography; OSI, oscillatory shear index; US, ultrasound; WSS, wall shear stress.

Effect of Geometry on Hemodynamic Profile

CFD has been extensively used to investigate how variations in anastomotic angle, curvature, and planarity impact flow patterns (Figure 2 and Table 3). Importantly, these studies can be performed in the planning stage of surgeries to identify the best configuration for a given patient. Computational studies on anastomotic angle by Lee et al., Carroll et al., Cunnane et al., and Ene-Iordache et al. on end-to-side AV fistulas agree that more obtuse angles reduce WSS and flow disturbance, while acute angles have lower OSI and more helical flow but overall lower relative residence time.^22,36–38 Marcino et al., in a clinical study, found the optimal end-to-side configuration included larger angles and diameters, which agrees with the above. In the side-to-side configuration, the flow resembles end-to-side configurations with 90° anastomotic angles, but with a more uniform WSS configuration (Figure 3).³⁹ Krampf et al. varied inflow artery length and diameter in piggyback straight line onlay technique configurations, observing that flow volume was inverse to length, and proportional to diameter.⁴⁰ The geometry of the heel also has a noticeable effect on flow. Rounded heels tend to have a smoother flow and distributed WSS on adjacent vessel walls.¹⁸ In general, the effects of angle and diameter within a configuration have been well defined in small studies. However, the Kidney Disease Outcome Quality Initiative vascular access guidelines still do not recommend a particular configuration.⁷ Future work should evaluate the effects of geometry in vivo, conduct larger clinical studies, and compare different configurations directly with a focus on addressing the Kidney Disease Outcome Quality Initiative gaps.

Figure 2.

CFD for evaluating hemodynamics of AV fistulas. (A) Comparison of velocities (cm/s) for two patients, P1 and P2, with varying stimulated anastomotic angles. (B) Comparison of pressure (mm Hg) between two patients, P1 and P2, with varying stimulated anastomotic angles. The labels pA, dA, and V represent the proximal artery, distal artery, and vein, respectively. Note that different scales are used for the two patients to highlight the hemodynamic differences as a result of varying anastomosis angles. Adapted from ref. 39. (C) Comparison of WSS in patients who did not (left) and did (right) require intervention before successful AV fistula use. WSS was computed for both the PV and PA. Dotted lines in the side view color maps indicate the location of radial slices, which are taken approximately 5 mm from the anastomosis. O and I denote the locations of the outer and inner walls, respectively. (D) WSS for the no intervention group and the required intervention group, respectively, at three different time points: 1 day, 6 weeks, and 6 months. An asterisk (*) denotes a statistically significant difference, P < 0.05. Adapted from ref. 55, with permission. P1, patient 1; P2, patient 2; PA, proximal artery; PV, proximal vein.

Table 3.

Summary of studies that have used computational fluid dynamics for understanding the effect of geometry on hemodynamic profile

Author	AV Fistula Type	Data Source	Geometric Parameters	Hemodynamic Parameters	Findings
Hull et al., 201370	Computer model, various configurations (N=17)	Flow and structural parameters from literature	Configuration (side-to-side, end-to-side 45°, end-to-side 90°), cross-sectional area (3.5–18.8 mm2)	Pressure drop, venous inflow, velocity vector, WSS	Cross-sectional areas were found to be inversely proportional to pressure drops. 90° end-to-side had the lowest pressure drop, while 45° end-to-side had the highest. WSS was most uniform in side-to-side configurations
Ene-Iordache et al., 201338	Computer model, end-to-side (N=4)	Flow and structural parameters from literature	Angle (30°, 45°, 60°, 90°)	Relative residence time	Lower angles in end-to-side configurations resulted in lower overall relative resistance times
Lee et al., 201622	Computer model, end-to-side (N=3)	Flow and structural parameters from literature	Angle (45°, 90°, 135°)	WSS	Obtuse angles had the lowest overall WSS values
Carroll et al., 201836	Computer model, end-to-side (N=2)	Flow and structural parameters from literature	Angle (45°, 135°)	Flow disturbance	Obtuse angles had lower flow disturbance, characterized by stable flow velocities and antegrade flow
Cunnane et al., 201937	Computer model, end-to-side (N=16)	Flow and structural parameters from literature	Angle (10°, 35°, 50°, 70°), linear/nonlinear taper, and planar/nonplanar venous segment	Helical flow, OSI, vorticity	Lower angles had the lowest OSI. Any taper was found to suppress helical intensity
Krampf et al., 202040	Computer model, pSLOT (N=36)	Flow and structural parameters from literature	Diameter (1.5, 2, 2.5, 3, 4, 5 mm), inflow artery length (15, 20, 25, 30, 35, 40 cm)	Flow volume	Inflow artery length was inversely proportional to flow volume. Vessel diameter was found to be proportional to flow volume
Marcinnò et al., 202439	Human, radiocephalic, end-to-side (N=2)	US	Angle, vein diameter	Flow recirculation, WSS	When the vein outlet diameter is small, the angle has a negligible effect, and the WSS profiles are abnormal. Large diameters and larger angles reduce recirculation and WSS zones

AV, arteriovenous; OSI, oscillatory shear index; pSLOT, piggyback straight line onlay technique; US, ultrasound; WSS, wall shear stress.

Figure 3.

Common AV fistula geometries. End-to-side (A) and side-to-side (B) AV fistula geometries. Although AV fistula geometries used are generally consistent, each patient's AV fistula is unique and depends on factors such as patient anatomy, surgical technique, and surgical proficiency. Θ represents the anastomotic angle, which can affect hemodynamic flow and overall AV fistula maturation. Larger, obtuse anastomotic angles lead to reduced WSS and decrease in flow disturbance. Smaller, acute anastomotic angle has been shown to decrease OSI. Increased WSS has been associated with improved AV fistula outcomes through the promotion of favorable outward remodeling. Similarly, reduced OSI has been associated with decreased development of neointimal hyperplasia and improved AV fistula maturation rates. Overall, low WSS that is highly oscillatory (high OSI) has been correlated with areas of reduced luminal diameter. Created with Biorender.com.

Longitudinal Development and Modeling

CFD has also been used to identify potential baseline geometric and hemodynamic predictors of AV fistula maturation or failure in longitudinal studies (Table 4). These studies typically monitor flow rates before, at, and after the anastomosis and use these data, along with the appropriate geometry, to create simulations at distinct points along the AV fistula maturation timeline. If the AV fistula fails, CFD can show the hemodynamics preceding failure. Typically, sustained OSI correlates to failure,^20,21 while successful maturation is indicated by increased flow rates, the presence of helices or vortices in flow streams, and a constant-valued WSS.^41–43 In most cases, this is where analysis ends, and very few studies have attempted to predict failure. In this regard, shape optimization is promising; WSS is mapped directly to vessel remodeling, thereby creating an adaptive computer model. Using this method, Akherat et al. observed that shape optimization produced topologically similar results to in vivo vessels up to 32 weeks post-AV fistula creation.⁴⁴ Similarly, advanced numerical models,⁴⁵ deep-learning models, and new prediction schemes should continue to be developed, using outputs gathered from patient-inspired models as a guiding factor.

Table 4.

Summary of studies that have used computational fluid dynamics for longitudinal monitoring

Author	AV Fistula Type	Data Source	Timescale	Findings
He et al., 201341	Human, brachiocephalic, end-to-side (N=1)	MRI	7 mo post-AV fistula creation	Serial mapping revealed that WSS was highest 4 mo post-AV fistula creation and decreased by as much as 90% by 7 mo. Lumen cross-sectional area increased nonuniformly from months 4 to 7, while flow rates decreased
Remuzzi and Manini, 201445	Human, various configurations, (N=1)	Clinical data, US	6 wk post-AV fistula creation	A custom numerical model was developed, with the ability to predict flow rate, lumen diameter, and waveforms
Pike et al., 201742	Murine, jugular-carotid, end-to-side, (N=2)	MRI	3 wk post-AV fistula creation	Serial mapping revealed increased levels of disturbed flow, flow velocity, helices OSI, vortices, WSS, and spatial WSS gradient at 3 wk post-AV fistula creation
Javid Mahmoudzadeh Akherat et al., 201744	Human, brachiocephalic, end-to-side (N=1)	US, white blood viscosity, X-ray venography	8–32 wk post-AV fistula creation	WSS is a primary driver of vessel remodeling, with nonphysiological values forcing homeostasis through stenosis in nearly all cases. Shape optimization frameworks can predict failure modes in patients
Bozzetto et al., 201843	Human, radiocephalic, end-to-side (N=1)	MRI	6 wk post-AV fistula creation	At week 6 postsurgery, flow is more helicoidal and doubles in rate, while the vessel shows significant remodeling and a higher cross-sectional area compared with week 1
Soliveri et al., 202320	Human, radiocephalic, end-to-side (N=1)	MRI, US	1.5 yr post-AV fistula creation	During the first 6 mo, hemodynamic parameters (OSI, turbulent kinetic energy, spectral power index) were high and indicated turbulent flow, but returned to presurgery values after stenosis. Flow volume decreased drastically (70%) from its peak at 6 mo to 1.5 yr
Bartlett et al., 202421	Human, radiocephalic, end-to-side (N=1)	MRA, US	13 mo post-AV fistula creation	Stenosis developed gradually, but showed high recirculation and low volume by the time venoplasty was needed (months 10–13). Regions of high OSI and low HOLMES were located at sites of stenosis, showing some correlation to neointimal hyperplasia. Experimental USTIR magnitudes correlated with HOLMES, suggesting an ability to predict structural remodeling and neointimal hyperplasia using only US. WSS could not be correlated with vessel remodeling in the arterial segment

AV, arteriovenous; HOLMES, highly oscillatory low magnitude shear; MRA, magnetic resonance angiography; MRI, magnetic resonance imaging; OSI, oscillatory shear index; US, ultrasound; USTIR, ultrasound-derived turbulence intensity ratio; WSS, wall shear stress.

Interventions and Hemodialysis

CFD modeling is a valuable tool for simulating the effects of clinical interventions in silico (Figure 2 and Table 5). Studies by Northrup et al. and Somarathna et al. have used CFD to analyze the effects of sildenafil as a possible treatment for neointimal hyperplasia, demonstrating a visible increase in vessel cross sectional area.^46,47 The overall effectiveness of the treatment is less clear; core hemodynamic parameters, such as WSS and vorticity, were noticeably higher with treatment, while vessel parameters remained statistically similar to control groups, suggesting the possible involvement of other biological mechanisms. It is known that increase in nitric oxide synthase after AV fistula formation plays a role in the vascular remodeling and adaptive response to the increased pressure. In a study on murine models, Baltazar et al. showed that nitric oxide synthase expression correlates with more stable flow and lower WSS values.⁴⁸ Similarly, shear-induced platelet activation has been known to precede thrombosis, and CFD has been used to help model and monitor shear-induced platelet activation levels.⁴⁹ Others have modeled the effect of physical interventions on flow. Gunasekera et al. showed that stenting not only mechanically opens the lumen but also improves the downstream flow dynamics, reducing turbulence and normalizing adverse WSS patterns that could otherwise lead to in-stent restenosis.¹⁵ Although insightful, these models are limited in scope, and at times, the literature is unclear. Further work should be performed to coordinate CFD with histological data for more decisive conclusions on drug treatment and biological processes in AV fistula maturation.

Table 5.

Summary of studies that have used computational fluid dynamics for understanding the effect of hemodialysis and interventions

Author	AV Fistula Type	Data Source	Intervention	Dependent Variables	Findings
Ciandrini et al., 200954	Computer model (N=16)	Dialysis needles, US	Flow rates and pressure in extracorporeal circuits	Intravascular pressure gradient, Qa, Qb, and Qinv	Qinv has a significant correlation to Qa, but there is a critical value at which extracorporeal flow begins to cause vessel damage. When Qb is more than 30% of Qa, flow becomes turbulent, and features vortices and increased WSS, leading to long-term failure
Fulker et al., 201350	Computer model, end-to-side, radiocephalic (N=1)	Dialysis needles, flow, and structural parameters from literature	Needle back-eye, flow rate (200, 300, 400 ml/min), needle angle (10°, 20°, 30°), needle depth	Flow velocity field, WSS	The presence of any needle leads to abnormal flow conditions. Shallow needle angle and depth and lower flow rates decrease WSS. Back eye was found to minimally reduce WSS
Fulker et al., 201651	Computer model, end-to-side, radiocephalic (N=1)	Dialysis needles, flow, and structural parameters from literature	Needle orientation (antegrade, retrograde), and rotation angle	OSI, WSS	WSS decreased by 30% with a rotated venous needle, but remained above normal levels. Retrograde needle placement produced more stable flow and smaller areas of concentrated OSI
Fulker et al., 201752	Computer model, end-to-side radiocephalic (N=1)	Dialysis needles, flow, and structural parameters from literature, planar illumination	Flow rate (200, 300, 400 ml/min), plastic cannula placement (upper, middle, lower thirds)	Flow velocity field, WSS	Optimal cannula position occurs when the tip is furthest from the vessel wall and in the center of the vein. Optimal blood flow rates exist between 300 and 400 ml/min, where the risk of neointimal hyperplasia can be minimized
Sturm et al., 201771	Human, brachiocephalic, end-to-side (N=4)	US	Adjustable band	Flow rate, pressure	Banding can be effectively used to optimize flow rate in vivo by modulating internal pressures and distension of the vessel wall
Lee et al., 201972	Human, forearm loop (n=2); radiocephalic (n=2); brachiocephalic (n=1)	MRA	DYBAND	Eddy viscosity, flow rate	Eddy viscosity decreased, and the flow rate increased as the band was tightened. All patients had symptomatic improvement (flow rate reduced and pain decreased). This benefit was sustained with no AV fistula thrombosis or loss at a mean follow-up of 1 yr
Northrup et al., 202146	Murine, femoral, end-to-side (N=8)	MRI	PDE5A inhibitor (sildenafil)	Cross-sectional area, flow rate, OSI, vorticity, and WSS	Sildenafil-treated rats had significantly larger lumen cross-sectional areas and higher values for flow rates, OSI, vorticity, and WSS
Salikhova et al., 202249	Human, brachiocephalic, end-to-side (n=1); radiocephalic, end-to-side (n=1)	MRA	CSS, flow rate, WSS, vWF size	SIPAct	SPIAct was <2% in regions where WSS and CSS exceeded initial levels. Higher venous outflow lowers SPIAct (<2% at 750 ml/min, <0.02% at 1300 ml/min). Reducing vWF multimer size may raise SPIAct levels, but the results vary closely with observed flow rates
Bozzetto et al., 202253	Human, radiocephalic, end-to-side (N=6)	US	VasQ (nitinol-based external support device)	Anastomotic angle, flow stability	VasQ maintained the anastomotic angle and resulted in less flow disturbance. In the absence of the device, there was extreme dilation and a case of stenosis
Northrup et al., 202255	Human, brachiocephalic, (n=3); radiocephalic (n=3)	MRI	Angioplasty	Need for intervention at week 6 (on the basis of CFDs outputs at day 1)	1-d venous WSS was significantly higher in the no-intervention group versus the intervention group. Venous cross-sectional area in the intervention group decreased from day 1 to 6 wk, prompting intervention. Findings suggest a critical WSS threshold for lumen enlargement
Baltazar et al., 202348	Murine, jugular-carotid, end-to-side (N=22)	MRI	NOS	Flow velocity, vorticity, WSS	NOS overexpression led to lower vorticity and WSS and, overall, less disturbed flow than wild-type and knockout groups. In nearly all cases, retrograde flow was observed
Gunasekera et al., 202423	Human, end-to-side (N=1)	US	Stent placement	Flow velocity, pressure, WSS	Stent placement significantly decreased turbulent flow, lowered pressure drops, and minimized WSS patterning, suggesting healthy maturation
Somarathna et al., 202547	Murine, femoral, end-to-side (N=18)	MRI	PDE5A inhibitor (sildenafil)	Flow velocity, lumen size, neointimal hyperplasia, vorticity, WSS	PDE5A levels increased after AV fistula creation in all three test species. WSS and vorticity increased at the anastomosis, but there was no significant difference in average intima/media ratio when comparing sildenafil with control-treated rats. Overall, neointimal hyperplasia was not affected by PDE5A inhibition

AV, arteriovenous; CFD, computational fluid dynamic; CSS, continuous shear stress; DYBAND, dynamic banding; MRA, magnetic resonance angiography; MRI, magnetic resonance imaging; NOS, nitric oxide synthase; OSI, oscillatory shear index; PDE5A, phosphodiesterase type 5A; Qa, arteriovenous fistula flow; Qb, dialysis flow; Qinv, dialysis flow value for which pressure across the arteriovenous fistula changes direction; SIPAct, shear-induced platelet activation; US, ultrasound; WSS, wall shear stress.

Hemodialysis treatment requires repetitive cannulation with dialysis needles with cannula dwell times of 3–4 hours per session. The rheological effect of foreign bodies, such as dialysis needles, have been systematically studied using CFD; typically, shallow needle depths and angles, and optimal rates of dialysis relative to the AV fistula blood flow, have been found to have the smallest effect on WSS and overall flow.^50–53 Bedside monitoring procedures have also been developed using CFD as a backbone for estimating flow parameters in patients undergoing treatment.⁵⁴ In cases where surgical intervention was needed, WSS patterns were shown to fall below a critical value, at which lumen diameter decreases, causing stenosis of the AV fistula.⁵⁵ In silico studies using CFD allow for noninvasive preclinical evaluation and refinement of therapeutic strategies, potentially reducing the need for trial-and-error adjustments in patients. As new interventions and drugs are developed, studies should continue to cross-validate results with hemodynamics.

ML in AV Fistula Maturation

General Principles of ML

ML offers a complementary, data-driven approach to understanding and predicting AV fistula outcomes. By identifying complex relationships within large datasets, ML models can provide predictive insights that are not readily apparent when relying solely on mechanistic models. The studies in this field predominantly use supervised learning, where the algorithm maps input features to an output label, or collection of labels, on the basis of the dataset used for training. This can be a classification task (e.g., predicting a binary outcome, like maturation versus failure) or a regression task (e.g., predicting a continuous value, like 6-week flow rate).

Common algorithms include support vector machines (SVMs), which find an optimal hyperplane to separate data into classes; decision trees, which create a flow chart of simple rules for separating data; and ensemble methods such as random forest and eXtreme gradient boosting, which combine hundreds to thousands of decision trees to generate more robust and accurate predictions. More recently, deep learning, particularly with convolutional neural networks, has been applied to analyze and make predictions from medical imaging and physiological signals.

Prediction of AV Fistula Maturations versus Failure

A major application of ML to AV fistula research has been the early detection of dysfunction (Figure 4 and Table 6). Several groups have developed models on the basis of noninvasive, clinical measurements. Chiang et al.⁵⁶ and Chen et al.⁵⁷ used features derived from photoplethysmography and vascular wall motion to train SVM models that could classify AV fistula flow as normal or dysfunctional with high accuracy (88%–98%).

Figure 4.

Application of a deep learning model for screening stenosis in AV fistulas. The figure illustrates the four-stage process for developing a deep learning model to screen for significant stenosis (≥50%) in native AV fistula requiring PTA. The first stage involves the recording of AV fistula shunt sounds using a wireless electronic stethoscope and subsequent audio file acquisition. The second stage, data preprocessing, takes the audio files and applies padding/truncation to normalize their length, followed by a STFT and mel-filter to convert the time-domain data into a mel spectrogram (a visual representation of sound's frequency and magnitude over time). In the third stage, data augmentation and DCNN models, the mel spectrograms are augmented to increase the dataset size, and three different DCNN models (DenseNet169, EfficientNetB5, and ResNet50) are trained to classify the mel spectrograms as representing either <50% AV fistula stenosis (post-PTA) or ≥50% AV fistula stenosis (pre-PTA). The final stage, result analysis, evaluates the performance of the DCNN models using a ROC curve and confusion matrices to determine their effectiveness in predicting hemodynamically significant AV fistula stenosis. Adapted from ref. 59, with permission. DCNN, deep convolutional neural network; PTA, percutaneous transluminal angioplasty; ROC, receiver operating characteristic; STFT, short-time Fourier transform.

Table 6.

Summary of studies that have employed machine learning for understanding arteriovenous fistula maturation

Author	AV Fistula Type	ML Model	Prediction Target	Features	Findings	Validation	Ground Truth
Chiang et al., 201956	Human, unspecified type (N=153)	NBC, SVM, kNN	Binary classification of degree of stenosis (>30%) and blood flow volume (<600 ml/min)	BP, oxygen saturation, and photoplethysmography measurements	SVM performed best, achieving 87.84% accuracy for degree of stenosis and 88.61% for blood flow volume assessment	Ten-fold cross-validation	US
Ota et al., 202058	Human, radiocephalic (n=19), brachiocephalic (n=1)	CNN, GRU, long short-term memory	Five-class classification of AV fistula bruit sounds (normal, hard, high, intermittent, whistling) related to stenosis	Mel-frequency log spectrograms from recorded AV fistula auscultation sounds	Achieved 70%–93% accuracy, 0.75–0.92 AUC depending on sound type. Deep learning can provide an objective auscultation index. Grad-CAM is used for interpretability	Train/test split	Expert classification
Grochowina et al., 202069	Human, unspecified types (N=38)	SVM, kNN, random forest	Six-class classification of AV fistula condition (A: patent to F: failed)	Selected frequency domain features from phono-angiography (acoustic bruit signals)	kNN performed best with 81% accuracy. Developed prototype device “NefDiag”	Leave-one-out cross-validation (patient-based)	Expert classification, US
Chen et al., 202157	Human, unspecified type (N=45)	SVM	Qa dysfunction (based on flow thresholds versus US measurement)	Harmonic ratios of vascular wall motion signals through pulse radar sensor	SVM classifier accurately predicted flow dysfunction compared with US. High specificity (100%) and sensitivity (max 95.2% at 750 ml/min threshold)	Ten-fold cross-validation	US
Peralta et al., 202162	Human, unspecified type (N=13,369)	XGBoost	Classify AV fistula failure within 3 mo through four-class risk classification; on the basis of vascular access changes, procedures to re-establish AV fistula maturation, and AV fistula complications	Demographics, biochemistry, vital signs, dialysis parameters, AV fistula parameters, and comorbidities	XGBoost algorithm obtained an AUROC of 0.80 (95% CI, 0.79 to 0.81). SHAP analysis is used to determine the importance of input variables for risk prediction	Training/test split (70/30), 30-fold cross-validation	Patient health information, demographics, vital signs
Park et al., 202259	Human, brachiobasilic (n=2), radiocephalic (n=16), brachiocephalic (n=22)	DenseNet201, ResNet50, EfficientNetB5	Classification of AV fistula stenosis (positive for ≥50%, negative for <50%)	Mel-frequency log spectrograms from recorded AV fistula auscultation sounds	ResNet50 performed best (AUROC of 0.99, precision of 0.95, recall of 0.98, and F1 score of 0.96). Grad-CAM is used for interpretability	Train/valid/test split (70/10/20)	DSA
Heindel et al., 202263	Human, radiocephalic (N=704)	Lasso, elastic net, logistic regression, random forest, pruned tree, boosted tree	Predict AV fistula maturation (unassisted AV fistula usage within 1 yr, two-needle cannulation for ≥90 d without preceding intervention)	4–6 wk US characteristics: volume flow, vessel diameter, stenosis, UAB criteria met (≥500 ml/min of flow and ≥4 mm venous diameter), KDOQI criteria met (≥600 ml/min and ≥6 mm venous diameter)	Lasso performed best and had the fewest covariates (three covariates; AUROC of 0.79, AUPRC of 0.72, accuracy of 0.73, sensitivity of 0.67, specificity of 0.77, PPV of 0.69, NPV of 0.75)	Train/test split (70/30), ten-fold cross-validation	Patient health information, demographics, vital signs, US
Park et al., 202360	Human, brachiobasilic (n=2), radiocephalic (n=16), brachiocephalic (n=22)	ResNet50, logistic regression, SVM, decision tree	Quantify AV fistula stenosis and predict 6-mo primary patency	Mel-frequency log spectrograms from recorded AV fistula auscultation sounds	ResNet50 predicted the degree of stenosis best with metrics: pre-AV (and post-AV) fistula creation (MAE of 0.028 [0.043], MSE of 0.001 [0.003], RMSE of 0.036 [0.051], and R2 of 0.956 [0.940]). When predicting 6-mo primary patency, it achieved an AUROC of 0.870, as well as precision, recall, and F1 scores of 1.0	Train/valid/test split (70/10/20)	DSA
Heindel et al., 202564	Human, radiocephalic (N=914)	Cox, random survival forest, logistic regression, elastic net	Predict AV fistula failure (predict the number of AV fistula interventions preoperation and postoperation within the first year)	Demographics, comorbidities, access history, anatomic features, and postoperative US	Random survival forest with baseline covariates (AUROC of 0.75, IBS of 0.22), the Cox model with 4–6 wk US covariates (AUROC of 0.77, IBS of 0.20), and the Cox model with 12 wk US covariates performed the best (AUROC of 0.76, IBS of 0.20)	Train/test split (75/25)	Patient health information, demographics, US
Poushpas et al., 202373	Human, brachiobasilic (n=2); brachiocephalic (n=7); radiocephalic (n=2)	Long short-term memory	Classification of AV fistula stenosis at the efferent vein (PSV <50, PSV >400 cm/s, or diameter <3.5 mm), the anastomosis (PSV >400 cm/s), and the afferent artery (PSV >400 cm/s, or bi/triphasic pulsatile flow, or volume flow of <300 ml/min)	Tensor decomposition from the full-frame B-mode US cine loops and cropped B-mode cine loops	Cropped B-mode US cine loop tensor decomposition long short-term memory model performed best with an AUROC of 0.82 and a PPV of 0.96	Train/valid/test split (60/20/20)	US
Doneda et al., 202466	Human, radiocephalic, end-to-end (n=21), side-to-side (n=117); brachiocephalic, end-to-side (n=16); brachiobasilic, end-to-side (n=2)	kNN, logistic regression, decision tree, SVM	Classify AV fistula as mature or failed. Blood flow volumes: upper arm brachial, radial, and ulnar; middle forearm, brachial diameter at middle upper arm; middle arm radial and cephalic diameters proximal with respect to the anastomosis; ulnar diameter at middle forearm	Age, sex, AV fistula type, presence of hypertension, presence of diabetes, and AV fistula nonmaturation	kNN network performed the best with an accuracy of 0.97	N-fold cross-validation (N=156)	Patient health information, demographics, vital signs, US
Chung et al., 202461	Human, unspecified type (N=84)	CNN, CRNN, ViT-GRU	Classify AV fistula into mature or fail (requires intervention within 10 d of creation)	Mel-frequency log spectrograms from recorded AV fistula auscultation sounds	ViT-GRU performed best on out-of-fold predictions with an F1 score of 0.71 and an AUROC of 0.77, but performed worst on the testing dataset. CNN generalized best on out-of-fold predictions and the testing dataset with an AUROC of >0.70 for both sets	Train/test split (73/27)	US
Yang et al., 202065	Human, brachiobasilic (n=4); brachiocephalic (n=4); radiocephalic (n=8)	KNN, SVM	Classify AV fistula into mature or fail (6-mo primary patency)	Photoplethysmography measurements	SVM with feature set 1 performed the best with an accuracy of 0.77, a sensitivity of 0.78, a specificity of 0.77, a precision of 0.82, and an F1 of 0.80	Five-fold cross-validation	US
Shah et al., 202567	Human, radiocephalic (n=223); unspecified type (n=124)	Neural boosted, decision tree, fit stepwise, generalized regression, bootstrap forest, normalized logistic, boosted tree, kNN, SVM, random forest	Classify AV fistula as mature or fail (based on blood flow rate). Investigate the link between heart failure, AV fistula blood flow, and death	Age, flow at first measurement of >1000 ml/min, time to highest Qa, and heart failure	Random forest with bootstrap aggregation performed best overall (AUROC 0.94, acc 0.84, sens 0.86, spec 0.83, pre 0.70, and F1 0.77)	Ten-fold cross-validation, completed five times	US
Shu et al., 202568	Human, unspecified type (N=974)	Random forest, XGBoost, kNN, decision tree, SVM, logistic regression, ANN	Classify AV fistula as stenosed (>50% diameter thickening)	Demographics, comorbidities, biochemistry, AV fistula parameters	XGBoost had the best performance with an AUROC of 0.83. The most important model features were the number of surgeries, prothrombin time activity, lymphocyte count, duration of AV fistula use, triglyceride levels, vitamin B12 levels, and C-reactive protein levels	Train/test split (70/30). Five-fold cross-validation. SHAP interpretation	Patient health information, demographics, vital signs, US

ANN, artificial neural network; AUC, area under the curve; AUPRC, area under the precision-recall curve; AUROC, area under the receiver operating characteristic curve; AV, arteriovenous; CI, confidence interval; CNN, convolutional neural network; CRNN, convolutional recurrent neural network; DSA, digital subtraction angiography; Grad-CAM, gradient-weighted class activation mapping; GRU, gated recurrent unit; IBS, integrated Brier score; KDOQI, Kidney Disease Outcome Quality Initiative; kNN, k-nearest neighbor; MAE, mean absolute error; ML, machine learning; MSE, mean squared error; NBC, Naïve Bayes classifier; NPV, negative predictive value; PPV, positive predictive value; PSV, peak systolic velocity; Qa, arteriovenous fistula flow; RMSE, root mean squared error; SHAP, SHapley Addition exPlanations; SVM, support vector machine; UAB, University of Alabama at Birmingham; US, ultrasound; ViT-GRU, vision transformer–gated recurrent unit; XGBoost, eXtreme gradient boosting.

Features derived from auscultation of the AV fistula bruit have also proven to be a rich data source. Studies by Ota et al.,⁵⁸ Park et al.,^59,60 and Chung et al.⁶¹ have successfully used convolutional neural networks to analyze the bruit. The raw audio signal is converted into a mel spectrogram, a visual representation of how the spectrum of frequencies in the sound changes over time, which is an ideal input for image-based deep learning. These models have been able to detect significant stenosis with area under the receiver operating characteristic curve (AUROC) values as high as 0.99, suggesting that a simple digital stethoscope, paired with such an algorithm, could become a powerful screening tool.

Other models have focused on predicting failure of AV fistula maturation using baseline patient characteristics. Peralta et al.,⁶² using a clinical database of more than 13,000 patients, developed an eXtreme gradient boosting model to predict AV fistula failure within 3 months with an AUROC of 0.80. The key predictors identified included vessel characteristics and patient-level factors, such as age and inflammatory markers, highlighting the multifactorial nature of AV fistula failure.

Predicting whether a newly created AV fistula will successfully mature is a critical clinical challenge. Heindel et al.^63,64 trained several ML models on baseline clinical data from more than 500 patients to predict 1-year maturation. Their best-performing model, a Lasso logistic regression, achieved an AUROC of 0.79. This type of model identifies a sparse set of the most important predictors, which in the case of Heindel et al. included preoperative vein diameter, patient age, and absence of peripheral artery disease. Yang et al.⁶⁵ used a different approach, extracting features from photoplethysmography signals to predict 6-month primary patency. Their SVM model achieved an accuracy of 77% and an F1 score, a metric integrating precision and recall in ML, of 0.80. This highlights a key distinction in approaches: Using static, preoperative data allow for presurgical risk stratification, while using dynamic, postoperative data may provide more accurate, updated predictions as the AV fistula evolves.

Current Challenges and Future Directions

Model Fidelity and Assumptions

Many CFD studies rely on simplifying assumptions, such as treating blood as a Newtonian fluid or vessel walls as rigid. While computationally convenient, these assumptions can affect the output accuracy.^{17,19–21,43} Only 20% of the CFD studies referenced in this review have incorporated more complex fluid structure interactions or non-Newtonian fluid properties. The clinical relevance of findings from idealized or animal models must also be interpreted with caution.

Several ML studies also rely on data assumptions, like assuming that there is zero or minimal overlap between two classes.^{56,57,59,62,65–68} The spatial delineation of the data used in these studies is not reported, but some employ validation techniques, like cross-validation,^{56,57,61,63,66,67,69} to prove that their models perform well on unseen cases, indicating that the general underlying data structure can be distinguished by an ML model. Certain studies maintained a training, validation, and testing split, but did not perform this training multiple times to assess model robustness.^{58–60,62,63,67}

Data Limitations and Generalizability

A pervasive issue is the reliance on small, single-center datasets. Many CFD studies are based on a single patient, and while ML models often use larger cohorts, they are rarely validated on external datasets. This lack of external validation severely limits the generalizability of the models. Studies conducted by Park et al.^59,60 aimed to improve generalizability by using data augmentation techniques to artificially expand their datasets. While this does benefit training by providing data that have been altered enough from the original data to be seen as distinct by the ML model, the augmented and original data are still from the same source, and the model may still be unable to generalize. The work by Peralta et al.,⁶² which used a large, multicenter database, is a notable exception and represents a crucial direction for future research. Furthermore, class imbalance is a common problem in ML studies, where models trained on datasets with few instances of failure may perform poorly in predicting failure. None of the studies analyzed in this review addressed class imbalances in their work. The importance of prospective validation cannot be overstated; a model's true utility is only demonstrated when it can accurately predict outcomes in a real-world setting using new patient data.

Clinical Integration and Interpretability

For computational models to be clinically useful, they must be interpretable. While ML models can achieve high predictive accuracy, many black box architectures, such as deep neural networks, do not reveal the reasoning behind a certain prediction. Techniques like SHapley Addition exPlanations analysis, used by Shu et al.,⁶⁸ are important steps toward improving model interpretability by assigning an importance value to each feature for every prediction. Park et al.^59,60 used a visualization technique, gradient-weighted class activation mapping, to highlight the portions of the auscultation signals that determined stenosis classification. Gradient-weighted class activation mapping provides a heatmap overlay on the input image, highlighting the most influential parts of the signal, as well as weighted activations.

Similarly, CFD simulations produce vast amounts of data, and distilling these into clinically meaningful metrics that can guide clinical decisions remains a challenge.

There are several other impediments to widespread integration of an ML model for AV fistula maturation prediction. Small, localized, and incomplete datasets affect the accuracy of ML models. Lack of external validation in large testing cohorts prevents clinical adoption. Artificial intelligence tools also need to be integrated into existing electronic medical records without slowing down clinical workflows. The need for massive datasets also conflicts with strict data privacy regulations in health care systems. The capital investment required for infrastructure, training, and maintaining these artificial intelligence systems is also significant.

Model Synergy

The most promising future direction lies in the synergy between mechanistic and data-driven modeling. CFD can generate high-fidelity, patient-specific hemodynamic features (e.g., WSS distribution, OSI maps) that are impossible to measure in vivo. These can then be used as highly informative inputs for ML models, combining the physical rigor of CFD with the pattern-recognition power of ML. One promising vision for this integration is the concept of a digital twin: a patient-specific, dynamic computational model of their vascular access. This model would be created presurgically, used to simulate different anastomotic configurations to select the optimal approach, and then continuously updated postoperatively with new data (e.g., from US or wearable sensors). This living model could then be used to predict the trajectory of maturation, alert clinicians to early signs of dysfunction, and test potential interventions before they are performed on the patient. Although these digital twins have great potential for improving patient outcomes, they are far from clinical integration. Factors such as data collection, computational time, and model accuracy need to be considered and optimized before digital twins can be clinically feasible. We hypothesize that these models can be used in preoperative and postoperative settings. Preoperative prediction models would use baseline patient characteristics (e.g., demographics, comorbidities, vessel flow rates) to determine the potential for successful AV fistula maturation versus other alternative long-term vascular access options (e.g., AV grafts). These models would have the additional input of early hemodynamic measurements to allow physicians to successfully predict the timing of AV fistula maturation as well as determine which patients may benefit from adjunctive procedures to assist with secondary AV fistula maturation (e.g., balloon-assisted maturation, stent placement). The probability of AV fistula maturation success could help patients make educated decisions regarding how the management of permanent vascular access may align with their personal goals, values, and lifestyle.

Conclusion

Computational modeling, through both mechanistic CFD and data-driven (ML) approaches, has significantly advanced our understanding of the complex factors governing AV fistula maturation. CFD has been pivotal in confirming the role of adverse hemodynamics—specifically low and oscillatory WSS—in the pathogenesis of neointimal hyperplasia and stenosis. ML has demonstrated the potential to predict a range of clinical outcomes using readily available clinical and noninvasive data.

The future of this field lies in the synergy of these two domains. Integrating patient-specific CFD-derived hemodynamic parameters as inputs in ML models could create more powerful predictive tools. Progress will require a concerted effort to build larger, multicenter, and multimodal datasets; to validate models prospectively in diverse patient populations; and to develop interpretable outputs that can be seamlessly integrated into clinical decision-making. By overcoming these challenges, computational frontiers can be pushed forward, ultimately transforming the management of vascular access and improving outcomes for patients with kidney failure.

Disclosures

Disclosure forms, as provided by each author, are available with the online version of the article at http://links.lww.com/JSN/F816.

Author Contributions

Conceptualization: Allan John R. Barcena, Edward Castillo, Steven Y. Huang, Anna E. Marks, Marites P. Melancon, Amanda Nowacki, Leonardo Ramirez-Mireles.

Data curation: Allan John R. Barcena, Edward Castillo, Steven Y. Huang, Anna E. Marks, Marites P. Melancon, Amanda Nowacki, Leonardo Ramirez-Mireles.

Funding acquisition: Marites P. Melancon.

Project administration: Marites P. Melancon.

Supervision: Marites P. Melancon.

Writing – original draft: Allan John R. Barcena, Amanda Nowacki, Leonardo Ramirez-Mireles.

Writing – review & editing: Allan John R. Barcena, Edward Castillo, Steven Y. Huang, Anna E. Marks, Marites P. Melancon, Amanda Nowacki, Leonardo Ramirez-Mireles.

Funding

NIH NIDDK - Houston Area Incubator for Kidney, Urologic and Hematologic Research Training (Award Number: 1TL1DK147564-01). National Institutes of Health - National Heart Lung and Blood Institute (Award Number: 5R01HL159960-04).