Computer simulations of blood flow built from a patient's own CT or MR angiogram can predict outcomes that the scan's visible narrowing cannot. In symptomatic intracranial atherosclerosis, percent stenosis did not separate patients who had a recurrent stroke from those who did not, while simulated flow ratios did. In coronary disease, low wall shear stress at baseline (the frictional drag of blood on the vessel wall) predicted which plaques would grow two years later. Clinicians, however, reason over a patient's history, not a single visit. The natural question is "did the flow environment around this lesion get worse since the last scan?" Today nobody can answer it reliably. Each visit's simulation is rebuilt from scratch: a new vessel outline is traced, a new mesh is generated, new boundary conditions are set, and a new "healthy reference" segment is chosen. The difference between two visits can therefore come mostly from the pipeline rather than the patient.
We propose to treat longitudinal hemodynamics as a measurement problem and to apply multi-agent task decomposition where it is actually needed. Language-model agents would not replace a scripted simulation pipeline on clean cases. Their job is to enforce cross-visit consistency and to estimate how much of each reported change is pipeline noise. In two sentences: a team of specialized agents rebuilds and simulates each new scan so that it matches the previous visit's choices, and a dedicated auditor perturbs those choices to measure the noise floor. A clinical-reasoning agent then reasons only over hemodynamic changes that exceed this floor, and it states its uncertainty when they do not.
Core Claim
Hemodynamic change between serial scans is meaningful only when it exceeds the variability of the reconstruction and simulation pipeline. That variability is large and depends on the case: it is driven by segmentation thresholds, where the inlets and outlets are cut, mesh resolution, and the choice of reference segment. The coronary progression study had to normalize every index against adjacent healthy segments to obtain reproducible multi-patient values (Hemodynamic Study of Plaque Progression and Regression). The wearable-driven digital twin showed that a single snapshot can deviate from long-term exposure by 10% to 1000% (Longitudinal hemodynamic mapping framework).
We claim three things. First, an agent system can use a consistency auditor to estimate a per-lesion noise floor for each index $h$ (for example minimum time-averaged wall shear stress, helicity, or shear strain rate ratio). The auditor reruns the pipeline under plausible perturbations of these choices on each visit. Second, the system reports a change only when
$$|h_{t_2} - h_{t_1}| > z\sqrt{\sigma_{t_1}^2 + \sigma_{t_2}^2},$$
where $h_{t}$ is the index at visit $t$, $\sigma_{t}$ is the standard deviation of that index across the auditor's perturbation ensemble at visit $t$, and $z$ is a fixed threshold. With this rule, false reports of change on scan pairs with no true change fall sharply. Third, filtered changes of this kind are more useful to downstream longitudinal reasoning than either baseline-only hemodynamics or visual image differences.
The novelty is not "agents run CFD," which Foam-Agent 2.0 already does for engineering flows. It is not "an agent reads a longitudinal record," which ClinSeekAgent already does with database and chest X-ray tools. The novelty is the cross-visit correspondence and uncertainty contract between the simulation agents and the reasoning agent.
Research Program
The system has three layers.
The first is a simulation team adapted from the Foam-Agent decomposition (Foam-Agent; Foam-Agent 2.0). An Architect plans the case. A Reconstruction agent wraps conventional segmentation and surface-reconstruction tools of the kind described by Lei Liu et al. and Lerchl et al.. A Meshing agent, an Input Writer, a Runner that drives OpenFOAM, and a Reviewer that reads solver logs complete the team. We reuse Foam-Agent's log-driven repair loop because its ablation is the strongest evidence in the file that a dedicated reviewer matters: removing it dropped success from 83.6% to 37.3%. Foam-Agent itself flags meshing of irregular biological geometries as an open problem. We therefore expect the medical meshing step to be the main new failure mode, and we will measure it.
The second layer is new: a Correspondence agent and a Consistency auditor. When visit $t_2$ arrives, the Correspondence agent registers its vessel centerline to visit $t_1$. It then propagates the earlier decisions: which branches to keep, where to truncate inlets and outlets, which healthy segment serves as the normalization reference, and the lesion's coordinates along the centerline. The agent itself decides only in the ambiguous cases. These are branches that appear or disappear because of image quality, lesions that have shifted, and reference segments that have themselves become diseased. When it overrides a propagated decision, it must log a justification. The Consistency auditor builds a small perturbation ensemble at each visit, varying segmentation threshold, truncation point, and mesh density within ranges a careful human would consider acceptable. It returns $\sigma_t$ for each index. The auditor is a separate instance with no access to the clinical question. This is deliberate, following the evidence that independent critics break confirmation loops in code synthesis (Dong et al.) and causal analysis (LEGO).
The third layer is a clinical-reasoning agent in the ClinSeekAgent style. It sees the record and calls the hemodynamics team as a tool. It receives only per-lesion index trajectories with noise floors, never raw meshes or flow fields. This firewall is a design choice, not a convenience. The longitudinal MRI benchmark found that adding dense multi-view context degraded smaller models (How Good are Foundation Models in Longitudinal MRI Disease Progression Reasoning?), so the reasoning agent should receive compact, quantified evidence. The same benchmark found that vision-language models localize changes well but quantify them poorly, with under 16% accuracy. We assign quantification to the physics pipeline and qualitative synthesis to the language model.
We keep the physics deliberately simple: rigid walls, a few pulsatile cycles, and generalized boundary conditions, as in the clinical studies. The aim is to measure change reliably within a fixed physical model, not to improve the physical model. Long-horizon exposure mapping, as in the wearable digital twin, is out of scope. That work's static-geometry limitation is exactly what serial reconstruction addresses.
Baselines
The naive reference point has two parts. One is change in percent diameter stenosis between visits. The other is a carry-forward rule that reports baseline hemodynamics as the current state, analogous to the Last Context Image baseline that CRONOS had to beat (Lumiere NRMSE 8.38 against CRONOS's 7.55). Beating it means a statistically better progression AUC in the clinical analysis, and a nonzero detection rate at matched false-change rate in the synthetic analysis.
The status-quo baseline is a scripted pipeline that processes each visit independently with fixed parameters and normalizes to an adjacent healthy segment. It follows the protocol of the coronary progression study, whose strongest reported predictor was normalized minimum TAWSS with an odds ratio of 0.38 (p < 0.001) on 22 patients. For outcome prediction from CFD, the strongest reported discrimination in the file is a c-statistic of 0.776 for shear strain rate ratio and velocity ratio in intracranial atherosclerosis (Leng et al.). That figure comes from a single-visit design, and we use it as a reference scale, not as a direct comparator.
The strongest non-agentic competitor, and the one we take most seriously, is a scripted pipeline with propagated configuration. It registers each new scan to the previous one and copies truncation, branch selection, and reference segment deterministically, with no language model. It receives the same segmentation and solver tools, the same compute budget per case, and the same parameter tuning on the development split. Our system beats it only if it achieves a lower false-change rate at equal sensitivity, or completes more cases without human intervention, on the held-out set.
The strongest existing agentic method is Foam-Agent 2.0, given our medical reconstruction and meshing tools but none of the correspondence or auditing agents. It reports 88.2% executable success on its 110-case engineering benchmark. We found no reported result on Lacuna for any agentic CFD system on patient-specific vessels.
A single-agent ReAct baseline gets every tool and the same token budget.
Two ablations isolate the key ingredient. One removes the auditor while keeping correspondence. The other keeps the auditor but lets it share context with the clinical reasoner.
Why not just…?
Why not just use a scripted pipeline with registration-based propagation? This is the real threat. On clean scan pairs from the same scanner it may match or beat the agents at a fraction of the cost. We do not argue otherwise. We include it as the primary competitor and stratify results by case difficulty, judged by changes in scanner or protocol, new or lost branches, and a diseased reference segment. If the scripted pipeline matches us across all strata, the honest conclusion is that the auditor's perturbation ensemble, which a script can also run, is the contribution and the agents are not. In that case we would report the auditing method without the agents. The experiment is designed so that outcome is visible.
Why not just train a neural surrogate that maps geometry to wall shear stress and skip simulation? Point-cloud networks handle vessel geometry well enough for classification on IntrA, where PointNet++ reaches 88.51% aneurysm classification. However, the file reports no wall-shear-stress regression accuracy for them. The longitudinal setting also asks them to extrapolate to geometries that have grown beyond their training distribution, which is exactly where measurement must be trustworthy. A surrogate is complementary: the auditor could use one to make the perturbation ensemble cheap. We will test that variant against full simulation on the synthetic benchmark.
Why not just give one capable model all the tools? The file's evidence is indirect but consistent. The Foam-Agent reviewer ablations show large drops without an independent checker. The single-agent baseline in our experiments tests this directly, on the same tools and budget.
Why not just ask a vision-language model to compare the two angiograms? The longitudinal MRI benchmark reports under 16% volumetric quantification accuracy and Trend-F1 between 0.19 and 0.63 for directional progression. Flow quantities are not visible in the image at all. We still run the comparison on our question set, so the gap is measured rather than assumed.
Proposed Experiments
The first experiment, and the decisive one, is a controlled synthetic benchmark with known ground truth. We take the 103 vessel models from IntrA and apply known geometric changes: stenosis growth, aneurysm sac enlargement, and some zero-change controls. We revoxelize each result into image volumes under different noise levels, resolutions, and intensity profiles to mimic re-imaging. Ground-truth index changes come from high-resolution simulations on the unperturbed surfaces.
Metrics are absolute error in reported change, false-change rate on the zero-change pairs, sensitivity at a fixed false-change rate, calibration of the reported $\sigma$ against observed error, and fraction of cases completed without human intervention. The core claim fails if the propagated scripted pipeline equals the full system on false-change rate and completion. It also fails if the auditor's $\sigma$ is not calibrated, meaning the stated noise floor does not predict actual error.
The second experiment tests reproducibility on real data. We reconstruct the same real scan repeatedly, with resampled preprocessing, and on any available same-day repeat acquisitions. This yields test-retest coefficients of variation per index, per system.
The third experiment is clinical, and its power is limited. With a partner cohort of paired coronary CTA, using the progression, stable, and regression definitions of the coronary study, we compare three predictors of the categories at a third timepoint or follow-up: baseline-only normalized hemodynamics, raw delta hemodynamics, and noise-filtered delta hemodynamics. Given cohort sizes in the file of 22 to 32 patients, we treat this as feasibility evidence with confidence intervals, not proof.
The fourth experiment tests reasoning. We build question sets over serial vascular cases in the style of the longitudinal MRI benchmark, asking whether a lesion's shear environment worsened beyond noise and by how much. We compare the clinical-reasoning agent with and without the hemodynamics tool against a vision-language model given the images.
Throughout, we report token and compute cost per case alongside accuracy, since latency is a known barrier for multi-agent clinical systems (AMR).
References
- Hemodynamic Study of Plaque Progression and Regression Based on Coronary CTA Imaging using Computational Fluid Dynamics Method: Preliminary Results
- Establishing the longitudinal hemodynamic mapping framework for wearable-driven coronary digital twins
- Foam-Agent 2.0: An End-to-End Composable Multi-Agent Framework for Automating CFD Simulation in OpenFOAM
- ClinSeekAgent: Automating Multi-modal Evidence Seeking for Agentic Clinical Reasoning
- Foam-Agent: Towards Automated Intelligent CFD Workflows
- Biomechanical and Injury Tolerance Analysis of Tibial Movement Under Different Postures and Impact Loads
- Musculoskeletal spine modeling in large patient cohorts: how morphological individualization affects lumbar load estimation
- Self-Collaboration Code Generation via ChatGPT
- LEGO: A Multi-agent Collaborative Framework with Role-playing and Iterative Feedback for Causality Explanation Generation
- How Good are Foundation Models in Longitudinal MRI Disease Progression Reasoning?
- CRONOS: Continuous time reconstruction for 4D medical longitudinal series
- Computational Fluid Dynamics Modeling of Symptomatic Intracranial Atherosclerosis May Predict Risk of Stroke Recurrence
- IntrA: 3D Intracranial Aneurysm Dataset for Deep Learning
- Adaptive Memory and Reflection Multi-Agent System for Medical Question Answering