Bridging Computational Hemodynamics and Clinical Decision-Making in Severe Aortic Stenosis

Introduction: From Simulations to Practice

The recently published review by Keshavarz-Motamed et al ¹ offers a timely and comprehensive look at how personalized computational simulations and artificial intelligence (AI)-based hemodynamic models are reshaping our understanding of aortic stenosis (AS) and transcatheter aortic valve replacement (TAVR). Still, turning these technical advances and innovations into actionable clinical decisions remains a challenge, particularly in discordant forms of AS where hemodynamic assessment is complicated by overlaying comorbidities.

Figure 1.

Overview of the clinical domains where computational hemodynamics may inform decision making in aortic stenosis: diagnostic stratification, procedural planning, and post-aortic valve replacement (post-AVR) follow-up.

Clinical Context and Unmet Needs in AS

Severe AS remains a common disease with a poor prognosis if left untreated, and timely aortic valve replacement (AVR) is therefore crucial. The optimal timing of intervention remains a topic of ongoing research and debate. Current guidelines, both European Society of Cardiology/European Association for Cardio-Thoracic Surgery ² and American College of Cardiology (ACC)/American Heart Association (AHA) ³ recommend AVR with Class 1 indication for:

Symptomatic severe high-gradient (HG) AS.

Asymptomatic AS with left ventricular ejection fraction (LVEF) < 50%.

Severe AS undergoing cardiac surgery.

Symptomatic low-flow low-gradient (LF-LG) AS with LVEF < 50%.

Only 1 indication supports intervention for asymptomatic patients, which is why multiple key trials have sought to investigate early intervention in asymptomatic severe AS (AVATAR, ⁴ RECOVERY, ⁵ EVOLVED, ⁶ EARLY-TAVR, ⁷ while we are still waiting for other important trials expanding the indication for early intervention (EASY-AS trial (NCT04204915), PROGRESS trial (NCT04889872), Evolut EXPAND TAVR II trial (NCT05149775), DANAVR trial (NCT03972644)). Based on a published meta-analysis by Généreux et al, ⁸ early interventions do not seem to affect all-cause mortality or cardiovascular mortality, but may reduce heart failure (HF) hospitalization, and a composite of unplanned HF or cardiovascular hospitalization and stroke.

Severe AS is usually defined by an aortic valve area ≤1.0 cm² and a mean gradient ≥40 mmHg.^2,3 Yet many patients do not fit into these classic thresholds. Up to 25% to 40% of severe AS cases are classified as LF-LG or normal-flow low-gradient (NF-LG) forms, each with its own pathophysiology, prognosis, and treatment considerations.^9,10

Among the different flow/gradient subtypes, HG AS remains the most common and carries the most favorable outcomes after TAVR, with reported 1-year mortality < 10% and 5-year mortality between 30% and 45% depending on the population studied. ^{11 –} 13 Classical LF-LG AS, marked by reduced ejection fraction (EF), is associated with the worst survival: 10% to 20% mortality at 1 year, rising to 50% to 65% by 5 years.^14,15 Paradoxical LF-LG AS, despite preserved EF, shows intermediate mortality, with 10% to 15% mortality within 1 year and 40% to 55% by 5 years.^14,16,17 NF-LG AS patients continue to present a clinical dilemma, often described as a moderate-severe AS variant. ² They display better short-term outcomes compared to HG patients, but by 2.5 years show mortality patterns that match those of HG AS (∼30%). ¹⁸

Computational Hemodynamics: Clinical Roles Emerging

The review by Keshavarz-Motamed et al ¹ highlights the potential of computational hemodynamics, modeling ventricular-arterial coupling, shear stress, and flow disturbances. The modeling techniques show conceptual promise, but real-world application remains limited, and whether it can lead to actual improvements in patient outcomes remains to be proved. In HG, AS severity is clear, treatment decisions are established, and offer limited additional value for clinical decision making. The real challenge remains with the discordant AS subgroups. These patients have worse outcomes, more complex hemodynamics, and are often in a diagnostic gray area where gradients and valve area alone are not enough to guide treatment. Computational tools could help clarify disease severity, separate pseudo-severe from true severe AS, and better predict who will benefit from intervention, offering an individualized approach.

Once the decision for valve replacement is made, the next challenge is technical planning. This involves selecting the appropriate valve type, sizing it correctly, and choosing between different self-expanding and balloon-expandable TAVR devices based on the patient's anatomy. In patients with small aortic roots undergoing surgical AVR (SAVR), surgical root enlargement might be needed to accommodate a larger valve and avoid patient-prosthesis mismatch (PPM), ¹⁹ which can significantly affect outcomes in both SAVR and TAVR.^20–22

Follow-up after bioprosthetic aortic valve implantation is largely shaped by consensus and observational data rather than trial-based evidence. Major guidelines, including the 2020 ACC/AHA, recommend lifelong clinical monitoring and periodic echocardiography, particularly after the first 5 years, but these recommendations are not backed by randomized studies. ³ The rationale for surveillance stems from data showing that both structural valve deterioration (SVD) and bioprosthetic valve thrombosis (BPVT) can occur silently and unpredictably.^23,24 BPVT tends to appear early and may be asymptomatic at diagnosis, making routine imaging critical for detection and timely treatment. ²⁵ Meanwhile, the risk of SVD increases with time, especially in younger patients, patients with PPM at baseline, or with certain valve types.^26,27

The Valve Academic Research Consortium-3 criteria ²⁸ and related consensus frameworks, that is, Heart Valve Collaboratory ²⁹ now provide standardized definitions and surveillance protocols. But like the guidelines, these are based on expert agreement, registry data, and observational data, and not randomized controlled trials (RCTs). ³⁰

Given this landscape, there is a growing interest in computational tools not only for procedural planning but also for modeling individual thrombosis risk. This could allow follow-up to be tailored to patient-specific factors, replacing routine one-size-fits-all follow-up.

Clinical Implications and Limitations

The most relevant use of computational hemodynamics is in LF-LG and NF-LG AS. These cases often fall outside traditional diagnostic thresholds, and it is not always clear whether symptoms or dysfunction of the LV is valve-related.^9,15–17 Keshavarz-Motamed et al describe how detailed flow simulations capturing 3D intraventricular flow, vortex patterns, and coronary perfusion can help resolve some of this ambiguity. Techniques such as vector flow mapping and echo-particle image velocimetry can show disturbed vortex formation and impaired flow-contraction timing in LF-LG AS. Combined with pressure–volume loop analysis, these can add information about myocardial work and efficiency beyond standard measures such as stroke volume or EF. This might help separate true severe from pseudo-severe AS or help differentiate between symptoms related to inherent myocardial disease rather than from the valve itself.

For procedural planning, the review outlines how models integrating pressure–volume loop behavior with valve deployment mechanics can support decisions on valve type, sizing, and positioning. This becomes relevant in borderline anatomy or when there is concern for PPM. The review also presents AI-based modeling tools are also discussed in terms of simulating TAVR outcomes under different deployment scenarios. Postimplantation, patient-specific modeling could also have a role in follow-up. Simulations of wall stress and altered flow could identify those at higher risk of BPVT, and AI-based risk models could guide how often and how intensively to monitor patients. Recent work comparing different modeling frameworks shows that Doppler-based input (C3VI-DE) can yield better agreement with catheter-derived reference data than computed tomography-based models (C3VI-CT), particularly in patients with complex valvular and ventricular interactions undergoing TAVR. ³¹

There are some limitations to consider. Hemodynamics frameworks still depend on high-quality imaging, detailed input parameters, and are not yet integrated into daily clinical workflows, and there is considerable time spent moving imaging data to models. Raw computer power is also an issue, as some of the models are time-consuming, with computational time ranging from hours to days, and outcome validation is limited. For now, use is mostly confined to research or specialized centers with the necessary infrastructure.

Computational models could be favorable in areas where conventional assessment is limited, particularly in discordant AS, procedural planning, and risk stratification after TAVR, but their routine use depends on more than technical feasibility (Figure 1). Most models have not been prospectively tested,^32,33 and few have demonstrated added value beyond current standards in terms of patient outcomes or clinical decision making.

Additionally, it is possible that utilizing computational models in asymptomatic severe AS would enable us to better identify those patients who might benefit from early intervention.

One important limitation of computational hemodynamics is that it does not account for comorbidities, which are highly prevalent in patients with AS and even more so in those with discordant forms. ³⁴ Frailty is a key example. It influences treatment decisions, recovery, and long-term prognosis, ³⁵ yet it cannot be measured through flow simulations or pressure–volume models. In many cases, symptoms such as fatigue or dyspnea reflect not only valve disease but also frailty-related decline and sarcopenia. ³⁶ This is especially relevant in older patients with discordant AS, where imaging may suggest severe valvular obstruction, but the actual benefit of intervention is uncertain. Other common conditions, such as left ventricular hypertrophy, diastolic dysfunction, coronary artery disease, pulmonary hypertension, chronic kidney disease, diabetes, atrial fibrillation, and chronic obstructive pulmonary disease, further complicate both diagnosis and outcomes. ³⁷ These factors are often responsible for persistent symptoms or limited recovery after valve replacement, yet most fall outside the scope of any current computational framework. While modeling may help define the hemodynamic severity of AS, it cannot replace clinical evaluation or account for the broader context in which treatment decisions are made.

Future Directions

Future studies should embed these computational frameworks in prospective workflows, using standardized input from clinical imaging and producing outputs that are clinically interpretable. Their use should be focused on specific decision points where existing methods lack precision, not applied broadly without context. The next step is not further development in isolation, but pragmatic research that ties computational modeling to measurable clinical benefit.