Patient-Specific Hemodynamics via Hybrid Physics-Neural Surrogates

New
License: arXiv preprint (CC BY 4.0), copyright held by authors · Updated 2026-11-11
My rating

The technical route of this work can be summarized in one sentence: keep the 0D reduced-order model as the skeleton, and hand back to neural networks only the closure relations that the reduction oversimplified, learning them from high-fidelity 3D data. It does not attempt to replace the physical model with a network. It repairs the parameter calibration stage of the physical model, preserving both ends at once, the millisecond-scale solve of the 0D model and the ability to interrogate every resistance and inductance value individually.

The paper is titled "Accelerated Patient-Specific Hemodynamic Simulations with Hybrid Physics-Based Neural Surrogates" and is authored by Natalia L. Rubio, Eric F. Darve and Alison L. Marsden, all of the Stanford University Department of Mechanical Engineering, with Darve additionally affiliated to the Stanford Institute for Computational and Mathematical Engineering. It is released as an arXiv preprint, identifier 2604.01549, with v1 submitted on 2 April 2026 and v2 revised on 29 August 2026, under a CC BY 4.0 license as stated on the preprint page. The preprint page lists no journal publication, so this entry records it as a preprint, a research manuscript released outside the peer-review pipeline rather than a finalized journal article.

1. Overview

The first reason to read this is that it demonstrates a route design that differs from end-to-end replacement. The data-driven reduced-order model path already exists, training machine learning models to fit data rather than deriving simplified models from the Navier-Stokes equations. Some of those approaches do embed physics, for instance graph neural networks that preserve the topology of the discretized vasculature, or physics-informed neural networks that use governing laws in the training loss. These models achieve higher accuracy than physics-based ROMs at comparable cost, but the paper identifies two hard prices: they require large training datasets and often struggle to generalize to new geometries. More decisively, such black-box techniques lack the interpretability that physics-based ROMs provide, and interpretability is a hard requirement for model interrogation and clinical adoption. This paper works around both costs. It does not retrain a model that predicts the flow field. It trains a network that predicts a small number of scalar parameters.

The second reason concerns how the error budget is allocated. Three-dimensional finite element simulations of cardiovascular flow typically take hours to days on a cluster, which excludes them from real-time response and from many-query scenarios such as uncertainty quantification and optimization. A 0D model solves in fractions of a second on a personal laptop or a phone. The engineering question was never which model to pick, but where to spend precision. The answer here is to pay the 3D cost once during offline calibration and run only the 0D solve online. That pattern transfers to any problem involving reduced-order model parameter calibration.

The third reason is the diagnostic value of the error analysis. The 0D model treats intravascular flow as fully developed idealized Poiseuille pipe flow and neglects junction dynamics entirely. The paper notes that this departs significantly from the true nonlinear nature of cardiovascular flow and introduces substantial error, particularly for vasculatures with extensive branching such as the pulmonary arteries, or with geometric complexity such as stenosis or curvature. Locating the error in junction geometry rather than in the flow field as a whole, and then inferring parameters from geometry, is exactly the framing of this work.

2. How It Works

On the 0D side, reduced-order models come from applying simplifying assumptions to the Navier-Stokes equations. The 0D model discretizes the vasculature into a network of vessel and junction elements, each assigned a set of lumped parameters governing its hemodynamics. The paper uses two such parameter models, the resistor-quadratic resistor-inductor model and the resistor-inductor model. The pressure drop sums three contributions: a linear resistance term capturing viscous effects proportional to flow rate, a quadratic resistance term capturing nonlinear pressure recovery and flow separation effects proportional to the square of flow rate, and an inductance term capturing inertial effects proportional to the time derivative of flow rate. The control is a traditional purely physics and heuristic 0D analog whose parameters follow from the Poiseuille-flow simplifications of the Navier-Stokes equations, generated by the ROM Simulation Tool of the open-source cardiovascular modeling suite SimVascular, referred to in the paper as standard, baseline or Poiseuille.

One technical detail deserves recording. Pressure recovery depends only on the speed of the flow and not on direction, so in cases of flow reversal it would be better represented by a quadratic form in Q squared. The paper states that this work includes no cases of significant flow reversal, and that preliminary studies reveal no significant difference between the even and odd quadratic forms, so the form with Q times its absolute value is retained for consistency with existing 0D literature. This is an explicit, documented modeling choice.

On the network side, neural networks map a geometric input vector to one of three outputs, the linear resistance, the quadratic resistance or the inductance. Separate networks handle junctions and vessels, so there are six in total. The number of hidden layers ranges from 2 to 4 and the hidden layer width, the size of the weight and bias matrices, ranges from 10 to 20 neurons. The same architectures are used across all anatomical cohorts. Training uses an adaptive moment estimator optimizer from the optax library to minimize a mean-squared error loss, and the networks are implemented in JAX for accelerated training. The paper also reports an inference-stage constraint: network outputs are restricted to lie within the range observed in the training set, and only at inference. Training pairs consist of geometric input features matched to optimal 0D parameters, where the optimal parameters are obtained by least squares fitting to high-fidelity 3D CFD simulations.

Geometric features are extracted with SimVascular, an open-source package supporting patient-specific cardiovascular modeling from medical images. Features are taken at centerline points associated with vessel or bifurcation inlets and outlets. Area is found by integrating over the contiguous cross-section normal to the centerline tangent; radius uses the maximum inscribed sphere radius, the largest sphere centered at that centerline point and fully contained in the vessel; angles come from comparing centerline tangents at inlets and outlets; path length is the centerline segment length between inlet and outlet, accumulated from consecutive points; and the flow ratio at a junction is computed from flows predicted by a standard 0D simulation.

Two geometric preprocessing steps matter. The first splits junctions with three or more outlets into a series of bifurcations. The second adjusts a pseudo entrance length. SimVascular's vessel-junction discretization draws junction boundaries close to the split site, where the flow is subject to significant localized 3D effects such as high-pressure impingement zones and recirculation. These effects are irrelevant in a 0D modeling context but create significant noise in the measured 3D pressure drop and therefore in the fitted resistance and inductance values, making the learning task harder. The paper therefore extends junction boundaries into the outlet vessels by a pseudo entrance length. A true entrance length depends on the flow Reynolds number, but these vasculatures span a wide Reynolds range over the cardiac cycle, so a single vessel-junction discretization must serve the whole simulation. One option is to pick the entrance length from the maximum Reynolds number over the cardiac cycle, but the inlet boundary condition may not be available when the 0D model is generated. The paper therefore chooses a flow-independent surrogate, ten times the outlet radius.

A loss variant also appears. The paper introduces a generation-weighted loss because the extensive branching of pulmonary geometries causes heavy over-representation of distal elements in the training set. Distal elements have relatively little effect on overall hemodynamics, so without weighting the networks prioritize accuracy on numerous but unimportant distal elements at the expense of the rarer but more impactful proximal ones. Weighting proximal elements more heavily counteracts this.

For evaluation, the forward 0D simulation uses svZeroDSolver and the resulting flow rate and pressure over a cardiac cycle are the primary outputs of interest. The metric measures how much the 0D solution differs from the high-fidelity 3D solution, which is taken as ground truth. Pressure errors are measured at the inlets.

3. Scope and Assumptions

The data-scale boundary is primary. The paper lists the relatively small amount of training data as the leading limitation. More high-fidelity 3D data would give a more accurate and statistically significant representation of the population, and would also allow fitting more flexible model forms that capture 3D fluid dynamics better. Such data could come from 3D simulations of patient-specific anatomies or from synthetically generated anatomies, which are available in larger numbers.

The geometric representation boundary follows. Junctions and vessels are represented by vectors of roughly 20 geometric features. The approach is lightweight and interpretable, but geometry could be represented in more detail by point cloud or shape analysis methods, enabling more accurate prediction of 0D parameters.

The hemodynamic premise is that the 0D model neglects junction dynamics and treats intravascular flow as fully developed Poiseuille flow. This work includes no cases of significant flow reversal.

The entrance-length choice is bounded. The surrogate of ten times the outlet radius is flow-independent, whereas a true entrance length depends on Reynolds number. The paper also identifies a countervailing effect: extending junction boundaries absorbs outlet branches and may reduce the resolution of the 0D system, limiting how well the model can represent 3D flow behavior. In some cases this may cancel the benefit of the adjustment, or simply mean the original junction boundaries were already far enough from the split site.

The dataset filtering rule is explicit. Models from patients with pulmonary hypertension are excluded, because in those anatomies the resistance of the modeled vasculature is negligible relative to the boundary conditions, so hemodynamics are almost entirely dominated by the boundary conditions, yielding artificially small differences between 3D and 0D solutions.

The clinical boundary: the conclusion positions the hybrid model as a real-time digital twin supporting clinical decision-making and research, but all evidence comes from numerical simulation datasets. No prospective clinical validation is included, nor any handling of image segmentation error or inter-patient anatomical variability in the pipeline.

4. Getting Started

First, set up the toolchain, which rests on three open-source components. SimVascular provides the full path from medical images to patient-specific cardiovascular modeling and includes the ROM Simulation Tool responsible for vessel-junction discretization and standard 0D model generation. svZeroDSolver performs the time-varying 0D blood flow solve. The Vascular Model Repository provides public anatomies with 3D simulations. Each component has its own documentation.

Second, obtain anatomical data. The dataset comes from prior work cited by the paper, whose authors the paper thanks for sharing it. The Vascular Model Repository catalogues anatomies by type, covering aortic, aortofemoral and pulmonary arteries. The appendix lists the naming of the anatomies forming each cohort and notes that these geometries are drawn from the set used in the cited work.

Third, generate the 3D simulations and optimal parameters. Simulations run with SimVascular's svSolver, using rigid walls, Windkessel or resistance outlet boundary conditions, and a prescribed patient-specific time-varying parabolic inflow. Least squares fitting of the 3D pressure drop for each vessel and junction element then yields the optimal parameters that serve as network training labels.

Fourth, run geometric preprocessing. Split three-way and larger junctions into sequences of bifurcations, then apply the entrance-length adjustment, extending junction boundaries downstream by ten times the outlet radius. This step determines the quality of the downstream learning task and is the easiest to overlook, yet it has the largest effect on the pulmonary cohort.

Fifth, extract features and pair them. From the centerline, compute area, maximum inscribed sphere radius, inlet-outlet angle and path length, plus the flow ratio at junctions, forming the vessel and junction feature vectors, and pair these with the optimal parameters to build the training set.

Sixth, train with the inference constraint. Minimize mean-squared error, or the generation-weighted variant, with Adam from optax in JAX, using 2 to 4 hidden layers of width 10 to 20 neurons. At inference, clip network outputs to the range observed in training.

Seventh, evaluate with five-fold cross-validation on geometries unseen during training. The metric is maximum inlet percent error over the cardiac cycle, reported as a mean with 95% confidence intervals across the five trials. Controls include standard Poiseuille parameters, learned vessel parameters only, learned junction parameters only, both, and optimal parameters obtained by direct fitting to the 3D solution.

A practical tip: the configuration comparison yields a clear recommendation. The quadratic resistance makes the framework less robust in a data-scarce landscape and does not significantly improve accuracy, degrading the mixed cohort from 7.1% to 59.5%. The paper states that all subsequent results use the more robust resistor-inductor model unless explicitly noted. Adopting that model directly is the safer choice for a reimplementation.

5. Comparison with Alternatives

Against purely data-driven reduced-order models, this is the central comparison. The data-driven route trains machine learning models to fit flow fields directly, with higher achievable accuracy, at the cost of large datasets, poor generalization to new geometries and lost interpretability. The hybrid route reduces the learning task from predicting an entire flow field to predicting a few scalar parameters, lowering both task difficulty and data demand by an order of magnitude while keeping every parameter individually inspectable. The paper's judgement is that under a clinical-adoption constraint, interpretability is a requirement rather than a bonus.

Against the original Poiseuille-based 0D model, standard parameters give mean errors of roughly 5.5% on the aortic cohort, 8.1% on aortofemoral, 30.4% on pulmonary and 12.8% on the mixed cohort. Learned parameters bring these to 2.6%, 3.7%, 7.9% and 7.1%. The distribution of the gap is informative: on geometrically simple aortas the Poiseuille assumption was never far off and the absolute gain is limited, while on broadly branching pulmonary arteries it fails badly and the gain is largest.

Against pure 3D finite element simulation, the 3D model is the most accurate but costs hours to days of cluster time and cannot enter real-time or many-query workflows. The hybrid model preserves the 0D solver speed while pulling accuracy back toward the 3D ground truth. The comparison is indirect: the paper does not report a direct hybrid-versus-3D error on a shared metric, but rather measures the 0D deviation against the 3D solution treated as truth.

Against the traditional use of 0D models as 3D boundary conditions, the most mature industrial pattern, where the anatomy of interest is modeled in 3D for a high-fidelity field while surrounding vasculature dynamics are captured in 0D to cut cost and include physiology outside the imaging region, the hybrid parameterization does not change that structure. It simply makes the 0D boundary condition more accurate.

Against multifidelity uncertainty quantification frameworks, where 0D solutions reduce the number of 3D solutions needed for statistical convergence, the value of the hybrid model lies in making the low-fidelity tier itself more accurate, so that the gap between tiers becomes better justified.

Against purely geometric feature engineering, the paper compares against the empirical stenosis correction built into the SimVascular ROM Simulation Tool, with an appendix measuring 0D error with and without it. The comparison shows that hand-added empirical corrections are the traditional route, while the learned route can cover such corrections and generalize further when data permits.

6. Benchmarks and Evidence

On cohorts and scale, the paper uses four cohorts drawn from a public dataset created in prior cited work and available at the Vascular Model Repository. They are 10 aortic anatomies, 17 aortofemoral anatomies, 5 pulmonary anatomies, and a mixed set containing all of them. Figure 4 reports the number of bifurcation and vessel elements per cohort and clarifies that the bifurcation count refers to inlet-outlet pairs, excluding artificial connector vessels and bifurcation outlets leading to connectors. Because the entrance-length adjustment can absorb entire vessels into junction regions, a vascular geometry may contain more bifurcations than vessels.

On main results, in five-fold cross-validation, learned parameters outperform baseline parameters in every cohort, and learning both vessel and junction parameters reduces error further. Mean maximum inlet percent error changes as follows: the aortic cohort goes from 5.5% to 2.6%, the aortofemoral cohort from 8.1% to 3.7%, the pulmonary cohort from 30.4% to 7.9%, and the mixed cohort from 12.8% to 7.1%. Analysis of the predicted parameters themselves confirms that network-predicted values are more accurate than those from standard Poiseuille assumptions.

The mixed cohort gains less. The paper explains that the network must predict accurately over a wider range of vessel and junction geometries without substantially more data, so additional training data and more complex networks are likely needed to reach the reductions seen in the anatomy-specific cohorts. This is the clearest evidence of the generalization boundary.

On configuration comparisons, using the resistor-quadratic resistor-inductor model instead of the resistor-inductor model raises the aortic error from 2.6% to 3.0% and the pulmonary error from 7.9% to 9.1%, while lowering the aortofemoral error from 3.7% to 2.8%. The paper notes these average errors all fall within one another's 95% confidence interval. In the mixed cohort, however, the quadratic model yields a large mean error of 59.5%, far above the 7.1% of the simpler model.

The generation-weighted loss lowers mean error slightly in the aortic, aortofemoral and mixed cohorts, from 3.2% to 2.6%, 4.2% to 3.7% and 7.9% to 7.1%, with the most marked gain in the pulmonary cohort, from 16.7% to 7.9%. The paper attributes this to the extensive branching of pulmonary geometries causing heavy over-representation of distal elements, which have relatively small effect on overall hemodynamics.

The entrance-length adjustment has little effect on the aortic, aortofemoral and mixed cohorts, with the aortic mean moving from 3.3% to 2.6%, aortofemoral unchanged at 3.7% and the mixed cohort rising from 6.1% to 7.1%, while the pulmonary cohort improves sharply from 16.8% to 7.9%.

On runtime, the 0D simulation with svZeroDSolver takes no more than 2 seconds on the most complex pulmonary anatomies and no more than 1 second on simpler geometries, which the paper describes as real-time. By contrast, 3D finite element simulation generally requires hours to days.

On metric definition, the metric is maximum inlet percent error over the cardiac cycle, maximized across the cycle, computed per anatomy and averaged per cohort with 95% confidence intervals. A separate parameter-level comparison exists, defined as the difference between predicted and optimal parameters with relative errors normalized by the magnitude of the optimal parameter, excluding parameters whose fitted magnitude is below 5.

7. Role in Physics-AI Workflows

As a reference implementation of the hybrid modeling paradigm, the core contribution generalizes well beyond hemodynamics: any scenario with a known reduced-order model structure but difficult parameter calibration can follow the same template, keeping the interpretable physical skeleton and moving calibration from analytic formulas or manual tuning to learning from high-fidelity data. The same shape appears in structural mechanics reduction, thermal network modeling and circuit system identification.

As the low-fidelity tier in multifidelity methods, the hybrid model lifts the 0D tier from Poiseuille-assumption accuracy to data-driven accuracy while keeping millisecond to second-scale solves. In multifidelity uncertainty quantification, the error level of the low-fidelity tier determines the efficiency and the credible interval width of the whole method, so the leverage of this improvement exceeds its absolute accuracy gain.

As a case study in trading interpretability against accuracy, where domain constraints require per-parameter interrogation in clinical, regulatory and engineering review settings, black-box routes are excluded outright, and the hybrid route offers a path that preserves interrogability without giving up accuracy. The paper also delineates the cost envelope of that path, where data scale, geometric representation granularity and the flexibility of learnable model forms jointly bound achievable precision.

As the real-time layer of a digital twin, the mechanism is pay the 3D cost once offline, run only the 0D solve online. This division of labor applies to any application needing high-fidelity calibration with cheap online inference.

On connections to other entries, this work sits at the intersection of the reduced-order modeling and surrogate modeling threads on this site, and can be read alongside surrogate modeling and reduced-order modeling, and together with numerics and PDE solvers for the relation between reduction and discretization. On the biomedical side, the physical-model-plus-data-driven-parameters structure is comparable to tokamak plasma magnetic confinement control, while neural operators represent the alternative reduction route and fluid simulation toolkit situates the underlying solver landscape.

8. Limitations and Failure Modes

It does not apply in data-poor settings. The paper lists the small training set as the leading limitation, and the consequence is already visible in the mixed cohort, where generalization over a wider geometric range shrinks the gain sharply, and where the 59.5% collapse of the quadratic model is another manifestation of the same root cause.

It does not apply to clearly extrapolated geometric distributions. The five-fold cross-validation rotates within one anatomy type, and even the mixed cohort pools three types at a total of 32 anatomies. The paper offers no evidence for predicting geometric configurations absent from training, such as entirely new branching patterns or different vascular generation laws.

It does not apply to downstream tasks needing full field information. The hybrid model outputs lumped parameters and pressure and flow waveforms, not local velocity fields, vortex structures or wall shear stress distributions. If the downstream task is local hemodynamic assessment, such as vortex patterns after stent deployment or the effect of valve geometry on local flow, the 3D model remains necessary.

It does not apply where entrance-length adjustment is unhelpful. The paper notes that extending junction boundaries absorbs outlet branches and reduces 0D resolution, which limited the benefit for the aortofemoral cohort and may cancel it.

It does not apply to significant flow reversal. The quadratic term uses a form with Q times its absolute value, while a squared form is more appropriate when reversal occurs, and this work does not cover that regime.

The parameter-level failure mode is diagnosable. The paper's parameter error comparison provides a diagnostic layer independent of the end-to-end metric. If predicted parameters deviate from optimal values, the cause lies in geometric representation or data coverage; if the parameters are accurate while end-to-end error remains high, the problem lies in the discretization or the solver rather than in learning. This layered diagnosis is the most practically useful part of the paper's methodology.

Specific populations are excluded. Models from pulmonary hypertension patients are deliberately removed because hemodynamics in those anatomies is almost entirely boundary-condition dominated, artificially compressing the difference between 3D and 0D solutions. Applicability to that population is therefore unverified and should not be extrapolated.

The clinical pipeline has gaps. Between CT or MR images and a usable hemodynamic conclusion lie segmentation, geometric reconstruction and error propagation steps that this work bypasses by starting from anatomical geometry and high-fidelity simulations. Those upstream stages are outside its validated scope.

9. Version and Source Notes

On publication and dating, the paper appears as an arXiv preprint, identifier 2604.01549, category cs.CE, with v1 submitted on 2 April 2026 and v2 revised on 29 August 2026, under a CC BY 4.0 license as stated on the preprint page. The preprint page lists no journal volume or DOI, so this entry records it as a preprint with year 2026.

On title, the arXiv title is "Accelerated Patient-Specific Hemodynamic Simulations with Hybrid Physics-Based Neural Surrogates". The phrase used in the original entry, "Patient-Specific Hemodynamics via Hybrid Physics-Neural Surrogates", is a rewording rather than the formal title, and citations should follow the arXiv title.

On authorship and institutions, the three authors are Natalia L. Rubio, Eric F. Darve and Alison L. Marsden, all in the Stanford Department of Mechanical Engineering, with Darve also affiliated to the Stanford Institute for Computational and Mathematical Engineering. Under the CRediT statement, Rubio contributes conceptualization, data curation, formal analysis, investigation, methodology, software, validation, visualization and original drafting; Darve contributes supervision, review and editing; Marsden contributes supervision, review and editing and funding acquisition.

Funding comes from the National Science Foundation Graduate Research Fellowship Program and grant 2310909, National Institutes of Health grant R01LM013120, and a Stanford Graduate Fellowship. The paper thanks Dr. Martin Pfaller and his co-authors for sharing the dataset used in the cited work.

Cohort and sample sizes are four cohorts: 10 aortic, 17 aortofemoral, 5 pulmonary anatomies and a mixed set containing all of them. The appendix lists the anatomy names forming each cohort.

On error figures, the paper body and abstract give two related sets of numbers that must be distinguished. The abstract and conclusion use the summary formulation of at least 50% reduction together with the specific pulmonary figure of 30.4% to 7.9%. The results section reports five-fold cross-validation mean maximum inlet percent error per cohort, from 5.5% to 2.6% for aortic, 8.1% to 3.7% for aortofemoral, 30.4% to 7.9% for pulmonary and 12.8% to 7.1% for the mixed cohort. Any quoted figure should state whether it is the summary or the per-cohort version.

On model configuration, the comparison between the two resistance models carries a key qualifier. In the anatomy-specific cohorts the average errors of the two models fall within one another's 95% confidence intervals and the difference is not significant; the marked difference appears only in the mixed cohort, at 59.5% against 7.1%. The paper consequently adopts the simpler model for all subsequent results. Statements that one model is simply better or worse, without the cohort qualifier, are inaccurate.

On runtime, the 0D solve with svZeroDSolver takes no more than 2 seconds on the most complex pulmonary anatomies and no more than 1 second on simpler geometries. The original entry's claim about 2 seconds for the most complex pulmonary model is directionally right but omits the 1-second figure for simpler geometries, and that timing covers the 0D forward solve alone, excluding network inference and geometric feature extraction.

On content that drifted into the original entry: the closing part of its summary claimed integration of measured strain and physical constraints, long-term deployment on large-scale infrastructure, and an AI alternative for structural health monitoring. None of that belongs to this paper, which involves no strain measurement acquisition, no long-term infrastructure deployment and no structural health monitoring. The parallel English passage needs the same correction. The digital twin framing does have a basis in the paper, whose conclusion explicitly mentions digital twins and clinical decision support, so that part can stand.

Other corrections to the original entry: its description of networks predicting 0D resistance and inductance parameters from geometry is correct, but it omits the scale of 32 anatomies across three cohorts, the two key preprocessing steps of bifurcation splitting and entrance-length adjustment, the five-fold cross-validation protocol, and the finding that the more robust model is the simpler resistor-inductor one rather than the quadratic variant.

On data and code availability, the paper uses three open-source components: SimVascular for patient-specific modeling including the ROM Simulation Tool, svZeroDSolver for the 0D solve, and the Vascular Model Repository for public anatomies and 3D simulation datasets. The paper itself provides no separate code repository address, and the networks are implemented in JAX with optax. Practical reuse therefore comes mainly from that open-source ecosystem and the public datasets.

References

  1. Rubio N. L., Darve E. F. and Marsden A. L., "Accelerated Patient-Specific Hemodynamic Simulations with Hybrid Physics-Based Neural Surrogates", arXiv preprint 2604.01549 (cs.CE, v1 2 April 2026, v2 29 August 2026, CC BY 4.0), https://arxiv.org/abs/2604.01549
  2. Full-text HTML version of arXiv 2604.01549v2, covering the abstract, results, discussion, limitations, appendices and the neural architecture table, https://arxiv.org/html/2604.01549v2
  3. SimVascular project and ROM Simulation Tool documentation, covering patient-specific modeling, vessel-junction discretization and standard 0D parameter generation, https://simvascular.github.io/documentation/rom_simulation.html
  4. SimVascular/svZeroDSolver repository, the 0D blood flow solver used for every 0D forward simulation in the paper, https://github.com/SimVascular/svZeroDSolver
  5. SimVascular/svSolver repository, the solver used for the high-fidelity 3D simulations in the paper, https://github.com/SimVascular/svSolver
  6. Vascular Model Repository, the public source of geometry and 3D simulation data for the four anatomical cohorts in the paper, https://www.vascularmodel.com/index.html
  7. Appendix 9.1 of the paper, giving hidden layer counts and widths for the six networks, and Appendix 9.2, listing the anatomy names per cohort and the exclusion of pulmonary hypertension models, referenced in item 1
  8. Figures 5, 6 and 7 of the paper, showing per-cohort mean MPE with 95% confidence intervals, the quadratic versus linear resistance comparison, and the ablations on the generation-weighted loss and entrance-length adjustment, referenced in item 1
  9. Table 2 of the paper, reporting mean and relative error between predicted and optimal parameters, excluding parameters with magnitude below 5, referenced in item 1
  10. Appendix 9.3 of the paper, comparing 0D error with and without the empirical stenosis correction of the SimVascular ROM Simulation Tool, referenced in item 1
  11. Acknowledgments and CRediT statement of the paper, citing NSF grant 2310909, NIH grant R01LM013120 and a Stanford Graduate Fellowship, with the author contribution breakdown, referenced in item 1
Tags: