Artificial intelligence and patient-specific computational modelling in transcatheter aortic valve replacement (TAVR): a narrative review of complementary roles along the clinical care pathway

Introduction

Transcatheter aortic valve replacement (TAVR) is a minimally invasive, catheter-based procedure for replacing a severely diseased native aortic valve with a bioprosthetic valve (1). Initially used in patients who were inoperable or at high surgical risk, TAVR has expanded into intermediate- and low-risk populations as evidence and device technology have evolved (2,3). Despite its less invasive nature, TAVR remains a complex intervention that depends on accurate diagnosis, appropriate patient selection, careful procedural planning, and effective post-procedural management. With the expansion of its indications, the demand for tools to support more complex and more precise patient-specific clinical decisions has also increased.

Artificial intelligence (AI) refers to computer systems that perform intelligent tasks based on external data inputs (4). Machine learning (ML), a major subset of AI, enables algorithms to identify patterns and learn statistical relationships from large datasets, which can then be applied to predict outcomes in new data. Specifically, deep learning (DL) is an advanced class of ML that uses multilayer neural networks to learn and interpret complex data through iterative processing (5). A key strength of DL is its ability to handle high-dimensional and heterogeneous data types, including medical images, time-series signals, and waveform data. Among DL architectures, convolutional neural networks (CNNs) are particularly well suited for image recognition and classification tasks because of their hierarchical feature extraction and translation-equivariant properties, making them especially valuable in medical image analysis (6,7). By contrast, patient-specific computational modelling uses individualized imaging data to reconstruct anatomy and simulate biomechanical or fluid dynamic processes, thereby enabling in silico evaluation of device-anatomy interaction and procedural mechanics. These approaches contribute differently to TAVR care, and their roles across the pathway are not interchangeable. AI-assisted methods have been used most often in screening, diagnosis, phenotyping, and outcome prediction, whereas computational modelling has been more prominent in pre-procedural planning and mechanistic simulation (8).

Although previous reviews have focused separately on AI-assisted imaging (9), risk prediction (10), or computational simulation (11) in TAVR, none has integrated these approaches within a unified carepathway framework. The present review addresses this gap by providing a structured synthesis of how AI, computational modelling, and emerging hybrid approaches contribute across all stages of TAVR care (Figure 1). It examines AI-dominant applications in early-stage evaluation, modelling-driven strategies in procedural planning, and hybrid frameworks at the interface between data-driven prediction and mechanistic simulation. These paradigms are not competing but complementary, with hybrid approaches representing promising yet still incompletely translated directions for future development. We present this article in accordance with the Narrative Review reporting checklist (available at https://rx.amegroups.com/article/view/10.21037/rx-26-0029/rc).

Figure 1 Overview of artificial intelligence and computational modelling across the TAVR workflow. AI, artificial intelligence; ECG, electrocardiography; ECHO, echocardiography; FEA, finite element analysis; TAVR, transcatheter aortic valve replacement; VR, virtual reality.

Methods

A structured literature search was performed in PubMed, Web of Science Core Collection, and Google Scholar, initially in September 2025 and updated on June 18, 2026. English-language studies published between January 1, 2012 and June 18, 2026 were considered. Search terms were grouped into two domains: procedure-related terms (transcatheter aortic valve replacement/implantation, TAVR/TAVI, and aortic valve disease) and technical terms (artificial intelligence, machine learning, deep learning, computational modelling, finite element analysis, computational fluid dynamics, and patient-specific simulation). A summary of the search strategy is provided in Table 1. The full search strategy and screening process are summarized in Appendix 1, Tables S1-S3, and Figure S1. Original studies were eligible if they evaluated AI-assisted, computational, or hybrid approaches relevant to any stage of the TAVR pathway. Editorials, case reports without sufficient methodological detail, and studies without clear relevance to TAVR were excluded. Titles, abstracts, and full texts were screened independently by two authors, and disagreements were resolved through discussion and consensus. The included evidence was then synthesized according to clinical stage and computational role across the TAVR pathway.

AI for detection and diagnosis of aortic valve disease (AVD)

AI and related digital methods are increasingly being explored to improve the screening and diagnosis of AVD, particularly through imaging, heart sound analysis, and electrocardiography (ECG). These applications can be broadly grouped into three clinically relevant functions: initial screening, diagnostic refinement, and phenotypic stratification (12). Echocardiography (Echo), heart sound analysis, and ECG-based models therefore serve different roles along the early evaluation pathway rather than addressing a single uniform diagnostic task.

AI-assisted echocardiography for diagnostic refinement and phenotypic stratification

Echo remains the diagnostic cornerstone for AVD (13). However, its interpretation is still influenced by inter-operator variability and subjective assessment (14). AI has therefore been introduced mainly to improve objectivity, reproducibility, and efficiency in routine evaluation. In this setting, AI-enhanced Echo is best viewed not as a replacement for Echo, but as a means of reducing operator-dependent uncertainty and extending assessment beyond conventional severity grading.

Among representative studies of AI-enabled diagnostic refinement in AVD, Sengupta et al. (15) illustrate how multimodal ML can refine conventional echocardiographic grading in aortic stenosis (AS). Using patient similarity analysis in an echocardiographic cohort, they derived phenotypic clusters and developed a classifier that showed consistent associations with computed tomography (CT)-quantified valve calcium and cardiac magnetic resonance (CMR)-measured left ventricular remodeling. In the combined cohort, the algorithm reclassified 57% of patients as having high-severity disease and improved prediction of 5-year aortic valve replacement outcomes. In practical terms, this type of model may help resolve diagnostic uncertainty when conventional assessment is discordant or equivocal, particularly before patients proceed to advanced imaging, procedural evaluation, or referral for valve intervention.

Echo-based applications address two distinct problems. One is the reproducibility of established measurements, especially left ventricular outflow tract (LVOT)-related parameters that influence continuity-equation grading. The other is the identification of biologically and prognostically distinct phenotypes that conventional severity categories do not fully capture. Distinguishing these aims clarifies the literature because these studies operate at different points in the diagnostic pathway. Measurement-focused work has concentrated on reproducibility. Because two-dimensional Echo grading relies on the continuity equation, variability in LVOT measurement, particularly LVOT diameter, can reduce the reproducibility of AS grading and contribute to misclassification (16). Krishna et al. (17) validated a fully automated AI-driven Echo platform (Us2.ai) for comprehensive AS assessment. Us2.ai achieved expert-level measurement accuracy while reducing operator dependence, particularly for LVOT-related measurements. Performance declined with poor image quality, and ectopic beats could not be excluded, but the overall correlations remained strong. For a TAVR-oriented workflow, the immediate value of such systems lies in standardizing upstream measurements before downstream planning decisions are made. The remaining question is whether these gains in reproducibility persist across centers, acquisition settings, image quality strata, and operator environments, and whether greater measurement consistency actually changes referral or planning decisions.

Beyond measurement refinement, ML has also been used to define clinically meaningful phenotypes that may not be captured by severity metrics alone. One clustering study identified three phenotypic profiles with distinct prognostic implications in patients with AS (18). Although AS severity was comparable across clusters, outcomes differed significantly during follow-up, and these associations were replicated in an independent validation cohort. Similar findings were reported in a later unsupervised clustering study describing a myocardial continuum in severe AS, in which patients from an external validation cohort could be assigned to the original phenogroups and showed stepwise differences in myocardial remodeling and mortality (19). Taken together, these findings indicate that ML-based phenotypic stratification may add prognostic structure beyond valve severity alone. They do not yet show when phenogroup-based reclassification should alter downstream management, such as referral timing, selection of advanced imaging, or pre-TAVR risk stratification. In clinical terms, phenotypic stratification will matter only if it changes planning or surveillance rather than simply redescribing biological heterogeneity.

Heart sound AI for initial screening

Among non-imaging approaches, AI-assisted heart sound analysis offers a potentially accessible front-end screening strategy, especially where Echo is not immediately available. One study (20) developed an AI-assisted stethoscope using the ReliefF algorithm for automated detection of left-sided valvular heart disease, with AUROC values of 0.85 and 0.76 in the training and independent testing sets, respectively. Chen et al. (21) also developed and validated a DL heart sound model (DenseNet121) for screening patients with AS, achieving an area under the curve (AUC) of 0.917 in a single-centre cohort. Overall, AI-assisted auscultation appears most valuable as a low-barrier first-line screening tool in settings where Echo is not readily available, but its clinical translation will ultimately depend on real-world robustness, prospective validation, and its ability to support appropriate referral for confirmatory imaging.

AI-assisted ECG for opportunistic detection and surveillance

AI-assisted ECG represents another promising non-imaging strategy for front-end screening. In parallel studies, Cohen-Shelly et al. (22) and Sawano et al. (23) used CNN-based models and reported strong diagnostic performance for early detection of AS and aortic regurgitation (AR), respectively. More recently, a large international study developed AI-assisted ECG models to diagnose existing moderate or severe regurgitant valvular heart diseases, including AR, while also predicting future disease risk, and externally validated these models in a transnational cohort (24). These findings suggest that AI-ECG may provide a scalable and readily deployable approach for identifying patients who warrant echo evaluation or surveillance. Although these models detect existing disease and also predict future risk to support longitudinal surveillance, the translational question is therefore whether AI-ECG can improve downstream Echo use while maintaining acceptable false-positive rates and clinically workable sensitivity-specificity trade-offs.

AI-assisted Echo, heart sound analysis, and ECG-based models may improve scalable AVD detection and diagnostic refinement, but their clinical value still depends on prospective validation and evidence that they improve referral or surveillance decisions.

AI for cardiac computed tomography angiography (CCTA)-based pre-TAVR planning

As a minimally invasive procedure, TAVR mandates comprehensive pre-procedural CCTA to provide an accurate anatomical assessment of the aortic root complex, which is essential for formulating patient-specific strategies (25,26). In routine practice, this assessment has traditionally relied on manual measurements of key structures, especially the aortic annulus, to guide prosthesis sizing. However, manual analysis is labor-intensive, time-consuming, and remains vulnerable to substantial inter- and intra-operator variability, with accuracy closely linked to anatomical expertise. Semi-automated platforms, such as FluoroCT (StreamLine Interventions Medical), standardize parts of this workflow, but they have not eliminated operator-dependent variability (9). This has driven the development of more advanced AI-assisted tools for automated anatomical quantification in pre-TAVR CT analysis. Accordingly, this subsection examines AI-assisted pre-TAVR CCTA analysis from three practical angles: segmentation and landmark detection, anatomical measurement, and prosthesis sizing support.

Automated segmentation and landmark detection

AI-assisted automated segmentation and landmark detection constitute the first layer of pre-TAVR CCTA analysis because they provide the structural basis for subsequent anatomical measurement and planning. At the algorithm level, ADPANet, a 3D U-Net-based framework, was developed to automatically detect and localize the aortic annulus plane on CCTA, demonstrating the feasibility of annular landmark identification despite the geometric complexity of the annulus (27). At the tool level, a multicentre fully automated DL system combined the FORSSMANN algorithm with a two-stage nnU-Net model to achieve coarse and fine segmentation of the aortic valvular complex, enabling simultaneous localization and quantification of annular dimensions, coronary ostial height, calcification burden, and aortic angulation (28). These studies indicate that pre-TAVR CCTA analysis is moving away from manual landmark placement toward faster and more standardized anatomical delineation. The main clinical value of this shift lies in the fact that measurement accuracy and downstream planning both depend on stable landmark definition.

Automated anatomical measurement and quantification

Beyond automated segmentation and landmark detection, a second major function is the extraction of clinically relevant anatomical measurements required for valve sizing and procedural preparation. The standardized pre-procedural workflows of these automated software tools comprise several key steps (Figure 2). A leading example is 3mensio (Pie Medical Imaging, Maastricht, The Netherlands) (29), a software platform widely adopted for pre-procedural TAVR planning, with subsequent validation showing accurate and reproducible assessment of aortic annular dimensions (30). A fully automated AI-driven planning software, 4TAVR, was recently shown to generate annular perimeter and area measurements with excellent agreement compared with expert operators, while also providing accurate assessment of the ascending aorta and other key pre-procedural anatomical parameters (31). Complementing this, a validation study comparing two AI-assisted pre-procedural planning methods against expert manual assessment demonstrated expert-level agreement for most TAVR planning measurements, supporting the feasibility of automated anatomical quantification in routine pre-procedural evaluation (32). However, not all anatomical targets have reached the same level of maturity. Current evidence supports more streamlined, less operator-dependent pre-TAVR evaluation, although performance remains target-specific, particularly for coronary ostial height and heavily calcified structures.

Figure 2 Standardized workflow of AI-assisted pre-TAVR CCTA analysis. AI, artificial intelligence; CCTA, cardiac computed tomography angiography; TAVR, transcatheter aortic valve replacement.

AI-assisted prosthesis sizing support

One important question is whether these outputs can meaningfully support accurate prosthesis sizing. Usability and accuracy data from annulus sizing platforms suggest that software-assisted planning can improve the correctness and consistency of valve size selection, including in less experienced users (33). Zou et al. (34) reported that a new deep-learning tool integrated into the CardioVerse system successfully analyzed 96.8% of cases and showed good-to-excellent agreement with manual and 3mensio-derived measurements across key aortic root parameters. These findings indicate that AI can make pre-TAVR assessment faster and more reproducible, even in anatomically challenging settings. The remaining question is more practical than technical: whether gains in speed and reproducibility actually reduce planning discrepancies, improve inter-operator consistency, and support safer coordination in real-world multidisciplinary TAVR programmes (35).

Overall, AI-assisted CCTA analysis can improve the efficiency and reproducibility of pre-TAVR planning, although robustness remains particularly important for calcified or bicuspid anatomy and high-stakes measurements such as coronary ostial height.

Patient-specific computational modelling for TAVR

Accurate prosthesis sizing and deployment strategy are central to successful TAVR. Although CCTA-based anatomical assessment and conventional procedural judgment remain standard, they may be insufficient in complex or borderline anatomies, where undersizing can lead to device migration or paravalvular leak (PVL), whereas oversizing may increase the risk of conduction disturbance or coronary obstruction (36). Balloon predilation may provide additional intra-procedural information on annular compliance in selected patients, particularly those with bicuspid anatomy or heavy calcification (37), but this invasive approach introduces additional procedural complexity and potential hemodynamic instability. Against this background, patient-specific computational modelling has emerged as a non-invasive way to reduce uncertainty in prosthesis sizing, deployment, and risk assessment. Figure 3 conceptually illustrates how multimodal imaging, structural parameters, and biomechanical simulation may be integrated within a patient-specific planning framework. Table 2 summarizes key engineering and hybrid-modeling terms used in sections “AI for CCTA-based pre-TAVR planning” and “Patient-specific computational modelling for TAVR”. In this literature, simulation matters less because it reproduces procedural mechanics in the abstract and more because it helps clarify device-anatomy interactions when static imaging alone cannot resolve the uncertainty that drives clinical decision-making.

Figure 3 Conceptual workflow of patient-specific simulation for pre-TAVR planning in anatomically challenging scenarios. 3D, three-dimensinal; AR, aortic regurgitation; CCTA, cardiac computed tomography angiography; CTA, computed tomography angiography; TAVR, transcatheter aortic valve replacement.

Virtual balloon dilatation for sizing

Accurate valve sizing remains challenging in TAVR, especially in patients with complex anatomy or heavy calcification. In this setting, virtual balloon dilatation has been explored as a non-invasive approach to estimate annular compliance and device expansion before valve deployment. Li et al. (38) developed a TAVR simulation workflow incorporating balloon pre-dilation and post-dilation steps and reported that post-dilation may improve simulated stent morphology and reduce simulated PVL, but may also increase aortic stress, thereby underscoring the trade-off between sealing and tissue injury risk. In parallel, Zhang et al. (39) used finite element analysis (FEA) to investigate balloon burst during TAVR pre-dilatation in heavily calcified bicuspid anatomy and showed that balloon behavior depended strongly on calcification stiffness, contact conditions, balloon type, and failure criteria. These studies place virtual balloon simulation in its clearest role: it becomes most informative when annular compliance and device expansion cannot be inferred confidently from static anatomical measurements alone. In some settings, it functions mainly as a mechanistic tool for understanding expansion behavior; in others, it is being tested as a possible aid to procedural planning. The central unanswered question is therefore not simply whether these models are realistic, but when balloon-derived information is robust enough to guide valve choice, implantation strategy, or adjunctive maneuvers in anatomically complex cases.

Virtual valve implantation for strategy optimization

Beyond valve sizing, patient-specific computational modelling can also support virtual valve implantation for selecting a feasible deployment strategy. Obaid et al. (40) illustrated this in a high-risk patient with a pre-existing Starr-Edwards mitral prosthesis, where CT-based virtual implantation was used to compare valve options and implantation depths and to identify a safe deployment window for a Portico valve. This case highlights the value of virtual implantation as a geometric planning tool in highly constrained anatomies.

Virtual implantation techniques have also been used to explore position-sensitive deployment strategy and hemodynamic consequences. Extending this concept, Lantz et al. (41) developed an automated CT-based framework for prosthetic valve positioning and hemodynamic evaluation. Their results showed that non-optimal valve orientation or implantation angle could worsen transvalvular pressure drop and flow behavior, suggesting that virtual implantation may support not only geometric feasibility testing but also position-sensitive optimization. These studies show that virtual implantation is useful when procedural success depends on annular dimensions, the spatial interaction between the transcatheter valve and adjacent structures, as well as on the consequences of valve position itself. The current gap is less about whether virtual implantation can be performed and more about whether position-sensitive modelling remains reliable across valve platforms and complex anatomies and whether it changes operator strategy beyond what experienced teams can already infer from standard CT-based planning.

Patient-specific simulation for flow-related risk assessment

PVL is a classic example of how physics-based simulation adds value beyond static imaging in TAVR. Bianchi et al. (42) showed that coupled finite element and computational fluid dynamics simulations could reproduce post-TAVR PVL jets and evaluate procedural variables such as implantation depth and balloon over-expansion, with simulated regurgitation patterns corresponding well to echocardiographic findings. In bicuspid anatomy, Anam et al. further demonstrated that virtual implantation of newer-generation devices into the same patient anatomies reduced both PVL severity and thrombogenic potential compared with older devices (43). More recently, Kovarovic et al. showed through patient-specific in vitro and in silico models that even mild PVL may still be associated with increased thrombogenic risk because platelets can repeatedly re-enter confined PVL channels over successive cardiac cycles (44). Overall, this work argues against viewing PVL solely as a geometric or grading problem. Instead, patient-specific simulation indicates that leak behavior, flow recirculation, and thrombogenic exposure may contain clinically relevant information beyond conventional severity assessment. Simulation may therefore help characterize not only how much leak is present, but also what kind of flow-related risk that leak may carry—information that standard imaging or routine PVL grading may not fully capture.

At the same time, some patient-specific modelling is used primarily to explain biomechanical or flow-related mechanisms, whereas other models are moving toward decision support for device choice, implantation depth, and risk anticipation in anatomically challenging cases. At a more translational level, the PRECISE-TAVI trial showed that CT-derived patient-specific simulations could influence pre-procedural planning in one-third of patients with challenging anatomies and provide useful predictions of both PVL and conduction-related risk (45). This finding suggests that computational modelling may be moving beyond retrospective explanation toward clinically relevant planning support. The key issue now is not only validation in larger cohorts but whether simulation-informed planning consistently changes operator choice, improves adjudicated procedural outcomes, and remains robust across valve platforms and anatomically complex subgroups.

Overall, patient-specific computational modelling is most informative when procedural uncertainty depends on device-anatomy interaction, but wider use requires faster workflows, standardized assumptions, and better calibration of simulation outputs to clinical outcomes.

AI for post-TAVR complication prediction

Despite its minimally invasive nature, TAVR remains associated with several clinically important complications that continue to affect prognosis and quality of life (46). Among these, conduction disturbances are the most frequent, with permanent pacemaker implantation (PPM) occurring in approximately 8–20% of patients depending on valve type and implantation strategy (47). Peri-procedural stroke is less common but remains highly consequential, generally occurring in about 2–5% of cases (48). Major bleeding within 30 days remains non-negligible, occurring in roughly 5.5–6.6% of contemporary cohorts (49). Heart failure (HF) also contributes substantially to early rehospitalization, with 30-day HF readmission occurring in about 3.6% of patients and accounting for 31.4% of all 30-day readmissions after TAVR (50). For AI-assisted post-TAVR applications, the endpoint alone does not define the clinical task; the timing of prediction and the management option it informs are equally important. Baseline models, peri-procedural monitoring tools, and mechanistically informed approaches may target the same complication, yet they serve different decisions across the care pathway. Figure 4 summarizes this temporal landscape across pre-procedural risk stratification, intra-procedural monitoring, early post-discharge management, and longer-term prognostic assessment.

Figure 4 AI-assisted prediction and monitoring across post-TAVR complications and survival outcomes. AI, artificial intelligence; AUC, area under the curve; CNN, convolutional neural network; DL, deep learning; ECG, electrocardiography; FEA, finite element analysis; GBM, gradient boosting machine; HF, heart failure; ML, machine learning; MLP, multilayer perceptron; RF, random forest; TAVR, transcatheter aortic valve replacement; TCD, transcranial Doppler.

Conduction disturbances

Conduction disturbances remain among the most frequent complications after TAVR, and the risk of PPM is influenced by both procedural factors, such as implantation depth and prosthesis selection, and patient-specific baseline conduction substrate (51). To improve the prediction of PPM, Truong et al. (52) developed a random forest model by integrating dynamic post-TAVR ECG parameters (delta PR, delta QRS: interval differences between post- and pre-TAVR measurements), achieving an AUC of 0.81. Because this model relies on post-procedural electrical change, it fits most naturally into early surveillance after TAVR rather than into pre-procedural triage. In parallel, Jia et al. (53) validated an AI-assisted visual strategy by training a CNN directly on pre-procedural 12-lead ECG images to predict conduction disturbances after TAVR, achieving an AUC of 0.764 using pre-TAVR ECG images alone and 0.779 after incorporating additional clinical features. This model addresses a different clinical moment: it estimates baseline susceptibility before valve implantation rather than tracking evolving electrical injury afterward. Its immediate clinical use, however, remains constrained by single-center development and the absence of external validation.

A complementary mechanistically informed approach was proposed by Galli et al. (54), who incorporated FEA-derived contact metrics, particularly maximum contact pressure (CPmax) and contact pressure index (CPI), into the prediction of conduction abnormalities after TAVR. These metrics quantify the intensity and spatial extent of local frame-tissue interaction near the conduction system and therefore move beyond purely correlative prediction. The value of this hybrid strategy lies in linking device-anatomy interaction to a clinically relevant endpoint such as PPM, rather than treating conduction risk as a black-box association alone. Wider use, however, remains limited by modelling complexity, computational burden, and the lack of real-time scalability. The point of divergence across these studies is not whether PPM risk can be modeled, but when that information becomes clinically useful. Taken together, one model speaks to pre-procedural vulnerability, another to evolving post-procedural injury, and the hybrid approach to a possible mechanical pathway of damage. Their usefulness therefore depends less on discrimination in isolation than on whether each prediction window can be translated into a specific clinical decision, such as prosthesis selection, implantation strategy, telemetry duration, or early pacemaker surveillance.

Cerebrovascular events (CVEs)

CVEs are less frequent than conduction disturbances but remain among the most devastating complications after TAVR, particularly because many clinically overt events cluster in the early peri-procedural period (55). Their pathogenesis is multifactorial, involving patient-related, valvular, and procedure-related contributors (56). To improve risk stratification, Okuno et al. (57) utilized a DL model to integrate heterogeneous clinical and imaging variables for predicting CVE after TAVR, achieving an AUC of 0.79. This level of performance indicates that DL may be able to combine mixed predictors for an infrequent but high-impact endpoint, although the absence of broader multicenter validation still limits confidence in generalizability.

Work in this area has followed two fairly separate tracks: one aims to estimate cerebrovascular risk before or around the procedure, whereas the other monitors embolic activity during the procedure itself. These approaches are related, but they do not answer the same question and should not be interpreted as if they were interchangeable. Beyond prediction, Baig et al. (58) introduced an AI-assisted robotic transcranial Doppler system for continuous intraoperative monitoring. This system dynamically quantified high-intensity transient signals (HITS), used as surrogate markers of cerebral emboli, and identified valve positioning and implantation as the most embolic phase of the procedure. The monitoring data were also consistent with fewer HITS in procedures performed with cerebral protection, suggesting a more objective way to assess protective strategies in real time. This is a different use case from pre-/peri-procedural risk modelling. One approach estimates who is more likely to experience CVE; the other follows embolic activity as the procedure unfolds and may help evaluate specific phases of intervention or the performance of protective devices. The field still lacks stronger evidence showing that either kind of output reliably changes management and improves adjudicated cerebrovascular outcomes, rather than only sharpening surrogate embolic readouts.

Post-procedural bleeding

Bleeding remains a clinically important adverse event after TAVR because of its association with mortality and reduced quality of life, and may be partly modifiable through more refined risk stratification (59,60). Navarese et al. (61) used ML and multivariable regression to derive and externally validate the highly parsimonious six-item PREDICT-TAVR score for 30-day Valve Academic Research Consortium-2 (VARC-2) bleeding events using large prospective multicenter registries. This score incorporates hemoglobin and serum iron levels, oral anticoagulation, dual antiplatelet therapy, femoral artery diameter, creatinine clearance, and demonstrated good predictive performance, with an AUC of 0.80 in derivation and 0.78 in external validation. Its main advantage is practical rather than architectural: the model is simple, externally validated, and based on variables that are already available in routine care.

A different modelling strategy is illustrated by BLeNet (62), a DL model built from 56 heterogeneous baseline, imaging, procedural, and post-procedural variables to predict longer-term bleeding. BLeNet outperformed traditional Cox proportional hazards (Cox-PHs) and random survival forest (RSF) models, achieving a 3-year AUC of 0.84 and a C-index of 0.81, compared with 0.72 for Cox-PH and 0.70 for RSF. The contrast here is not simply a matter of better versus worse numbers. These models were built with different priorities: one emphasizes short-term bedside usability, whereas the other accepts greater complexity to model bleeding risk over a longer horizon. Although interpretation libraries may improve post hoc inspection of model behavior, they should not be mistaken for full clinical interpretability or implementation readiness. Bleeding prediction may be nearer to practical use than several other post-TAVR AI applications, but the relevant standard is still whether the model changes care. Better discrimination matters only if it affects antithrombotic strategy, follow-up intensity, or both, rather than merely producing a more efficient retrospective risk ranking.

HF

HF is a major driver of early 30-day rehospitalization after TAVR and therefore represents a clinically relevant target for post-discharge risk stratification and targeted preventive strategies (63). To address this challenge, the TAVR-HF readmission score (64) was developed using an ML-derived AutoScore framework, analyzing a national cohort of 92,363 patients. The final model demonstrated discriminative ability for 30-day HF readmission (AUC 0.76). The score uses only seven clinically weighted variables, including HF subtype, electrolyte abnormalities, and chronic comorbidities, and is therefore well suited to rapid point-of-care screening. That same simplicity also defines its ceiling. Because the model was derived from administrative claims and coding-based hospitalization data, it gains scale and broad representativeness at the cost of procedural, imaging, and physiologic granularity. This trade-off is central to interpreting the score. Claims-based tools are relatively easy to deploy, but they are less able to capture anatomy, procedural complexity, or residual hemodynamic burden, all of which may matter for individualized post-TAVR HF prediction. External validation remains necessary, and moderate discrimination alone does not establish that the score improves discharge planning, transitional care, or readmission reduction in routine practice (65). Thus, its current value lies mainly in rapid discharge-stage screening, whereas more individualized HF prediction will likely require integration of procedural, imaging, and hemodynamic information.

Long-term survival prediction

AI has shown promise in long-term survival risk prediction following TAVR, with some studies reporting performance beyond established surgical risk scores. In a study focused on 5-year mortality after TAVR, Penso et al. (66) reported that a multilayer perceptron neural network, trained on over 80 clinical and echocardiographic variables, achieved an AUC of 0.79. Notably, the analysis identified organic mitral regurgitation as the single most influential predictor of long-term mortality, raising the possibility that multimodal AI models capture prognostic signals that conventional summary scores do not fully incorporate. Similarly, Agasthi et al. (67) applied a gradient boosting machine (GBM) model to predict 1-year mortality, leveraging a multimodal dataset that included clinical, electrocardiographic, computed tomographic, and echocardiographic parameters from 1,055 patients. Their GBM model achieved an AUC of 0.72, outperforming conventional scores such as the TAVR2-SCORE (0.56) and the CoreValve risk model (0.53).

These studies point in the same direction: broader and more heterogeneous inputs may strengthen prognostic stratification after TAVR. Even so, statistical outperformance should not be conflated with clinical replacement. The more consequential question is whether these models reveal modifiable pathways of risk, clinically useful thresholds, or patient subgroups that justify different follow-up strategies, rather than simply sorting mortality risk more finely. At present, heterogeneity in endpoint definitions, limited reporting of clinical utility, and the scarcity of prospective multicenter validation remain major barriers to routine deployment (10). At this stage, AI-based survival models are best regarded as promising additions to long-term risk assessment rather than as established substitutes for conventional prognostic frameworks.

Overall, AI-enabled prediction and monitoring tools show promise across major complications and survival prediction, but clinical maturity remains uneven; future work should prioritize externally validated, calibrated, and interpretable models that connect prediction timing to actionable decisions and clinically meaningful outcomes.

Emerging hybrid AI-mechanistic frameworks for TAVR

This section examines early efforts to combine data-driven AI with mechanistic modelling in TAVR. Most current studies do not describe established clinical workflows; instead, they explore surrogate or acceleration-oriented strategies designed to reduce the computational cost of high-fidelity simulation. Although the evidence base remains limited and methodologically heterogeneous, two broad directions have emerged. Some studies aim to approximate biomechanical or flow-related states more efficiently, whereas others combine mechanistic and clinical variables to improve complication prediction. In both settings, however, the same limitation persists: many accelerated outputs remain surrogate in nature, and their clinical value cannot be assumed unless they are linked to adjudicated patient outcomes. One proof-of-concept study showed that a DL-finite element framework could be trained on simulated valve deformation data to predict post-implantation biomechanical states directly from valve geometry, thereby reducing the need for repeated full finite element solves while preserving mechanically relevant outputs (68). A more TAVR-specific hybrid strategy was reported by Galli et al. (54), who combined patient-specific computer modelling with supervised ML to predict post-TAVR conduction abnormalities. By integrating mechanistic variables such as valve-induced contact pressure together with anatomical and procedural features, the best-performing model achieved an area under the curve of 0.84. These studies suggest that hybrid frameworks may be useful either as computational surrogates for biomechanical outputs or as clinically oriented models enriched by mechanistic information, but they should be interpreted according to the clinical question they address rather than treated as interchangeable approaches.

Hybrid modelling has also been explored as an acceleration strategy for flow simulation. Physics-informed neural networks have been used to approximate flow through a TAVR device, and although the dominant flow components could be captured, discrepancies with CFD remained, particularly in finer vortex structures (69). This suggests that computational acceleration is feasible, but not yet a substitute for high-fidelity haemodynamic simulation. Recent methodological work supports the same conclusion, emphasizing that reliable TAVR simulation still depends on realistic material modelling, accurate geometric reconstruction, and explicit representation of calcification and boundary conditions (70). Credibility analyses have further shown that deployment-related outputs may be predicted with relatively low error, whereas haemodynamic quantities such as effective orifice area and pressure gradient remain substantially more uncertain, indicating that model acceleration is currently more reliable for some outputs than for others (71).

The clinical meaning of these hybrid outputs, therefore, remains unsettled. Many current computational TAVR studies still report mechanistic or surrogate endpoints-such as stress, leaflet deformation, or flow-related metrics-rather than prospectively validated clinical outcome measures, and a calibrated surrogate-to-outcome validation pathway is still lacking (72). Consistent with this, recent scoping work has identified substantial variation in modelling pipelines, limited standardization, and inconsistent validation across patient-specific numerical TAVR studies, while highlighting automation, validation, interpretability, and workflow integration as major translational barriers (11).

Overall, hybrid AI-mechanistic frameworks remain early-stage approaches that may accelerate biomechanical or flow simulation and enrich complication prediction with simulation-derived variables. Their clinical translation will require standardized modelling pipelines, uncertainty-aware validation, and stronger calibration of surrogate outputs against adjudicated patient outcomes.

Discussion

This review suggests that AI and patient-specific computational modelling play distinct, rather than interchangeable, roles across the TAVR pathway. Their uneven distribution across stages of care appears to reflect the type of clinical uncertainty being addressed, rather than simple differences in technological maturity alone. AI-assisted approaches have shown the greatest maturity in tasks that benefit from scalable pattern recognition, automated quantification, and multimodal prediction, including early screening and phenotypic stratification of AVD, CCTA-based anatomical assessment, and post-TAVR risk stratification. By contrast, computational modelling has remained most informative in situations where clinically relevant uncertainty depends on explicit device-anatomy interaction, such as sizing ambiguity, deployment strategy, PVL, and conduction-related mechanical risk. Hybrid frameworks are conceptually attractive because they attempt to combine statistical scalability with mechanistic fidelity, but their translational role remains provisional because they must bridge methods that are optimized for different data structures, validation pathways, and decision contexts. A purely maturity-based explanation is therefore insufficient, although differences in data availability, annotation structure, and implementation burden likely also contribute to the current asymmetry.

Across the field, the challenge extends beyond proof of concept: mature AI tools require clinical integration, whereas many computational and hybrid approaches still require validation. Across the current literature, many systems have reported promising performance in retrospective and/or single-centre settings, yet their broader clinical value remains uncertain because robustness, generalizability, and reproducibility have not been established consistently across centres, imaging protocols, valve platforms, and patient populations. Current advances remain largely separate rather than integrated across the TAVR pathway, from screening and automated CCTA to simulation for complex anatomy and risk-guided monitoring or follow-up.

A further priority is to narrow the persistent gap between mechanistic or surrogate outputs and validated clinical endpoints. Variables such as contact pressure, leaflet stress, deformation, and flow-derived metrics are increasingly informative, but their relationships to procedural complications, valve durability, and longer-term outcomes remain insufficiently calibrated. This issue is especially important for hybrid frameworks, many of which currently accelerate or approximate computational outputs whose clinical relevance has not yet been prospectively established. Translation will require uncertainty-aware validation, calibration against clinical endpoints, and evidence that model outputs improve clinical decisions.

This review also has several limitations that should constrain the interpretation of its conclusions. First, as a narrative review, it was designed to map the breadth and organization of the evidence landscape rather than to compare the effectiveness of specific tools or provide formal comparative quality appraisal; accordingly, no risk-of-bias assessment was performed. This design limits the extent to which the present review can support claims about the relative superiority of AI-based, computational, or hybrid approaches for any given clinical task. Second, the included studies were highly heterogeneous in imaging inputs, computational workflows, validation strategies, outcome definitions, and levels of clinical maturity. This heterogeneity restricts direct comparison of reported performance metrics and makes apparent differences across studies difficult to interpret as true differences in clinical utility. Third, grouping studies into AI-based, computational modelling, and hybrid categories improved conceptual clarity but necessarily simplified areas of overlap between these approaches. As a result, the resulting framework should be understood as an analytic aid for interpreting the field rather than as a rigid classification of mutually exclusive methods. Fourth, publication bias and the rapid evolution of the field may have influenced the apparent balance of evidence, particularly for emerging hybrid frameworks and translational simulation studies. Consequently, the present synthesis is better interpreted as a structured account of how the field is currently organized than as definitive proof of which class of tools performs best or is ready for routine clinical implementation.

Conclusions

AI, patient-specific computational modelling, and emerging hybrid frameworks are not converging toward a single dominant digital paradigm in TAVR. Rather, each appears best suited to a different form of clinical uncertainty along the care pathway. AI is currently most mature for scalable screening, automated anatomical quantification, and post-procedural risk stratification, whereas patient-specific computational modelling remains most informative for mechanism-sensitive procedural planning, particularly in anatomically complex cases. Hybrid frameworks are promising for balancing speed and fidelity, but their clinical role remains provisional until simulation- or surrogate-derived outputs are more consistently linked to adjudicated outcomes. Progress in precision TAVR will therefore depend less on any single technology than on the careful integration of complementary computational approaches into standardized, interpretable, and trustworthy decision-support workflows.

Acknowledgments

During the preparation of this manuscript, ChatGPT (OpenAI, San Francisco, CA, USA) and DeepSeek (DeepSeek AI, Hangzhou, China) were used to assist with language polishing and editorial refinement, including improving grammar, clarity, readability, and organization of the text. These tools were not used to generate or analyse data, conduct statistical analyses, or make final scientific decisions. All AI-assisted content, references, interpretations, and revisions were critically reviewed, verified, and approved by the authors, who take full responsibility for the accuracy and integrity of the manuscript.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://rx.amegroups.com/article/view/10.21037/rx-26-0029/rc

Peer Review File: Available at https://rx.amegroups.com/article/view/10.21037/rx-26-0029/prf

Funding: This study was supported by the Zhejiang Provincial Natural Science Foundation of China (No. LTGY24H180019), the Medical and Health Science and Technology Project of Zhejiang Province (Nos. 2024KY160), the Jie Bang Gua Shuai Project of Wenzhou Science and Technology Bureau (Nos. ZY2024019 and ZY2023022), and Key R&D Program of “Jianbing Lingyan+X” in Zhejiang Province (No. 2026C04038).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://rx.amegroups.com/article/view/10.21037/rx-26-0029/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Mack MJ, Leon MB, Thourani VH, et al. Transcatheter Aortic-Valve Replacement with a Balloon-Expandable Valve in Low-Risk Patients. N Engl J Med 2019;380:1695-705. [Crossref] [PubMed]
2. Ando T, Ashraf S, Villablanca P, et al. Meta-Analysis of Effectiveness and Safety of Transcatheter Aortic Valve Implantation Versus Surgical Aortic Valve Replacement in Low-to-Intermediate Surgical Risk Cohort. Am J Cardiol 2019;124:580-5. [Crossref] [PubMed]
3. Leon MB, Mack MJ, Pibarot P, et al. Transcatheter or Surgical Aortic-Valve Replacement in Low-Risk Patients at 7 Years. N Engl J Med 2026;394:773-83. [Crossref] [PubMed]
4. Khera R, Oikonomou EK, Nadkarni GN, et al. Transforming Cardiovascular Care With Artificial Intelligence: From Discovery to Practice: JACC State-of-the-Art Review. J Am Coll Cardiol 2024;84:97-114. [Crossref] [PubMed]
5. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015;521:436-44. [Crossref] [PubMed]
6. Dey D, Slomka PJ, Leeson P, et al. Artificial Intelligence in Cardiovascular Imaging: JACC State-of-the-Art Review. J Am Coll Cardiol 2019;73:1317-35. [Crossref] [PubMed]
7. Corral-Acero J, Margara F, Marciniak M, et al. The 'Digital Twin' to enable the vision of precision cardiology. Eur Heart J 2020;41:4556-64. [Crossref] [PubMed]
8. Gupta A, Krumholz HM. The Evolving Landscape of TAVR: A JACC Themed Issue. J Am Coll Cardiol 2025;85:875-7. [Crossref] [PubMed]
9. Sun S, Yeh L, Imanzadeh A, et al. The Current Landscape of Artificial Intelligence in Imaging for Transcatheter Aortic Valve Replacement. Curr Radiol Rep 2024;12:113-20. [Crossref] [PubMed]
10. Shojaei S, Mousavi A, Kazemian S, et al. Artificial Intelligence in Risk Stratification and Outcome Prediction for Transcatheter Aortic Valve Replacement: A Systematic Review and Meta-Analysis. J Pers Med 2025;15:302. [Crossref] [PubMed]
11. Grossi B, Perri LM, Raona V, et al. Predicting procedural outcomes in transcatheter aortic valve implantation: a scoping review of numerical patient-specific simulations. Prog Biomed Eng (Bristol) 2025;8:012001. [Crossref] [PubMed]
12. Sengupta PP, Kluin J, Lee SP, et al. The future of valvular heart disease assessment and therapy. Lancet 2024;403:1590-602. [Crossref] [PubMed]
13. Liu Q, Hahn RT. The Role of Multimodality Imaging in Transcatheter Aortic Valve Replacement. Curr Cardiol Rep 2019;21:84. [Crossref] [PubMed]
14. Bunting KV, Steeds RP, Slater K, et al. A Practical Guide to Assess the Reproducibility of Echocardiographic Measurements. J Am Soc Echocardiogr 2019;32:1505-15. [Crossref] [PubMed]
15. Sengupta PP, Shrestha S, Kagiyama N, et al. A Machine-Learning Framework to Identify Distinct Phenotypes of Aortic Stenosis Severity. JACC Cardiovasc Imaging 2021;14:1707-20. [Crossref] [PubMed]
16. Manna D, Eliasson M, Bech-Hanssen O, et al. Reproducibility of Echocardiographic Measures of Aortic Stenosis Severity and its Impact on Grading of Severity. J Am Soc Echocardiogr 2024;37:370-372.e2. [Crossref] [PubMed]
17. Krishna H, Desai K, Slostad B, et al. Fully Automated Artificial Intelligence Assessment of Aortic Stenosis by Echocardiography. J Am Soc Echocardiogr 2023;36:769-77. [Crossref] [PubMed]
18. Kwak S, Lee Y, Ko T, et al. Unsupervised Cluster Analysis of Patients With Aortic Stenosis Reveals Distinct Population With Different Phenotypes and Outcomes. Circ Cardiovasc Imaging 2020;13:e009707. [Crossref] [PubMed]
19. Bohbot Y, Raitière O, Guignant P, et al. Unsupervised clustering of patients with severe aortic stenosis: A myocardial continuum. Arch Cardiovasc Dis 2022;115:578-87. [Crossref] [PubMed]
20. Zhou Z, Xie K, Huang Y, et al. Automatic Diagnosis of Left Valvular Heart Disease Based on Artificial Intelligence Stethoscope. JACC Adv 2025;4:101993. [Crossref] [PubMed]
21. Chen Z, Fu B, Wang J, et al. Heart sound model based on DenseNet121 architecture for diagnosis of aortic stenosis: A prospective clinical trial. Chinese Journal of Clinical Thoracic and Cardiovascular Surgery 2023;30:514-21.
22. Cohen-Shelly M, Attia ZI, Friedman PA, et al. Electrocardiogram screening for aortic valve stenosis using artificial intelligence. Eur Heart J 2021;42:2885-96. [Crossref] [PubMed]
23. Sawano S, Kodera S, Katsushika S, et al. Deep learning model to detect significant aortic regurgitation using electrocardiography. J Cardiol 2022;79:334-41. [Crossref] [PubMed]
24. Liang Y, Sau A, Zeidaabadi B, et al. Artificial intelligence-enhanced electrocardiography to predict regurgitant valvular heart diseases: an international study. Eur Heart J 2025;46:4823-37. [Crossref] [PubMed]
25. Otto CM, Nishimura RA, Bonow RO, et al. 2020 ACC/AHA Guideline for the Management of Patients With Valvular Heart Disease: Executive Summary: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Circulation 2021;143:e35-71. [Crossref] [PubMed]
26. Vahanian A, Beyersdorf F, Praz F, et al. 2021 ESC/EACTS Guidelines for the management of valvular heart disease: Developed by the Task Force for the management of valvular heart disease of the European Society of Cardiology (ESC) and the European Association for Cardio-Thoracic Surgery (EACTS). Rev Esp Cardiol (Engl Ed) 2022;75:524. [Crossref] [PubMed]
27. Cho Y, Park S, Hwang SH, et al. Aortic Annulus Detection Based on Deep Learning for Transcatheter Aortic Valve Replacement Using Cardiac Computed Tomography. J Korean Med Sci 2023;38:e306. [Crossref] [PubMed]
28. Wang M, Niu G, Chen Y, et al. Development and validation of a deep learning-based fully automated algorithm for pre-TAVR CT assessment of the aortic valvular complex and detection of anatomical risk factors: a retrospective, multicentre study. EBioMedicine 2023;96:104794. [Crossref] [PubMed]
29. de Vaan J, Verstraeten L, de Jaegere P, et al. The 3mensio Valves™ multimodality workstation. EuroIntervention 2012;7:1464-9. [Crossref] [PubMed]
30. Stortecky S, Heg D, Gloekler S, et al. Accuracy and reproducibility of aortic annulus sizing using a dedicated three-dimensional computed tomography reconstruction tool in patients evaluated for transcatheter aortic valve replacement. EuroIntervention 2014;10:339-46. [Crossref] [PubMed]
31. Toggweiler S, Wyler von Ballmoos MC, Moccetti F, et al. A fully automated artificial intelligence-driven software for planning of transcatheter aortic valve replacement. Cardiovasc Revasc Med 2024;65:25-31. [Crossref] [PubMed]
32. Corbin D, Santaló-Corcoy M, Tastet O, et al. Validation Study of Two Artificial Intelligence-Based Preplanning Methods for Transcatheter Aortic Valve Replacement Procedures. J Soc Cardiovasc Angiogr Interv 2024;3:101289. [Crossref] [PubMed]
33. Spanke J, Nübel J, Hölschermann F, et al. Usability and accuracy of two different aortic annulus sizing software programs in patients undergoing transcatheter aortic valve replacement. J Cardiovasc Imaging 2024;32:1. [Crossref] [PubMed]
34. Zou H, Jiang Y, Huang H, et al. Automated and quantitative assessment of aortic root based on cardiac computed tomography angiography using a new deep-learning tool: a comparison study. Quant Imaging Med Surg 2024;14:8414-28. [Crossref] [PubMed]
35. Tsakirian F, Afendoulis D, Mavroudis A, et al. Applications of Artificial Intelligence in Transcatheter Aortic Valve Replacement: A Review of the Literature. Life (Basel) 2025;15:1724. [Crossref] [PubMed]
36. Bui HT, Khair N, Yeats B, et al. Transcatheter Heart Valves: A Biomaterials Perspective. Adv Healthc Mater 2021;10:e2100115. [Crossref] [PubMed]
37. Liu X, He Y, Zhu Q, et al. Supra-annular structure assessment for self-expanding transcatheter heart valve size selection in patients with bicuspid aortic valve. Catheter Cardiovasc Interv 2018;91:986-94. [Crossref] [PubMed]
38. Li J, Meng Z, Yan W, et al. Computational study of the balloon dilation steps on transcatheter aortic valve replacement. Front Bioeng Biotechnol 2023;11:1333138. [Crossref] [PubMed]
39. Zhang X, Wu H, Pan Z, et al. Mechanism of balloon burst during transcatheter aortic valve replacement pre-dilatation: Image observation and validation by finite element analysis. Comput Biol Med 2024;168:107714. [Crossref] [PubMed]
40. Obaid DR, Smith D, Gilbert M, et al. Computer simulated “virtual TAVR” to guide TAVR in the presence of a previous starr-edwards mitral prosthesis. Journal of Cardiovascular Computed Tomography 2019;13:38-40.
41. Lantz J, Collins JD, Leng S, et al. A numerical framework for preprocedural prosthetic valve positioning and hemodynamic evaluation. Biomech Model Mechanobiol 2025;25:3. [Crossref] [PubMed]
42. Bianchi M, Marom G, Ghosh RP, et al. Patient-specific simulation of transcatheter aortic valve replacement: impact of deployment options on paravalvular leakage. Biomech Model Mechanobiol 2019;18:435-51. [Crossref] [PubMed]
43. Anam SB, Kovarovic BJ, Ghosh RP, et al. Assessment of Paravalvular Leak Severity and Thrombogenic Potential in Transcatheter Bicuspid Aortic Valve Replacements Using Patient-Specific Computational Modeling. J Cardiovasc Transl Res 2022;15:834-44. [Crossref] [PubMed]
44. Kovarovic BJ, Rotman OM, Parikh PB, et al. Mild Paravalvular Leak May Pose an Increased Thrombogenic Risk in Transcatheter Aortic Valve Replacement (TAVR) Patients-Insights from Patient Specific In Vitro and In Silico Studies. Bioengineering (Basel) 2023;10:188. [Crossref] [PubMed]
45. Hokken TW, Wienemann H, Dargan J, et al. Clinical value of CT-derived simulations of transcatheter-aortic-valve-implantation in challenging anatomies the PRECISE-TAVI trial. Catheter Cardiovasc Interv 2023;102:1140-8. [Crossref] [PubMed]
46. Walther T, Hamm CW, Schuler G, et al. Perioperative Results and Complications in 15,964 Transcatheter Aortic Valve Replacements: Prospective Data From the GARY Registry. J Am Coll Cardiol 2015;65:2173-80. [Crossref] [PubMed]
47. Reyes Mitre A, de la Garza HL, Guerreiro CE, et al. Minimizing Permanent Pacemaker Implantation After TAVR: Current Strategies, Monitoring Pathways, and Future Directions. J Clin Med 2025;14:7700. [Crossref] [PubMed]
48. Guedeney P, Barthelemy O, Montalescot G. TAVR Procedure-Related Stroke: When the Safety Net Fails. JACC Cardiovasc Interv 2024;17:678-80. [Crossref] [PubMed]
49. van Nieuwkerk AC, Aarts HM, Hemelrijk KI, et al. Bleeding in Patients Undergoing Transfemoral Transcatheter Aortic Valve Replacement: Incidence, Trends, Clinical Outcomes, and Predictors. JACC Cardiovasc Interv 2023;16:2951-62. [Crossref] [PubMed]
50. Zahid S, Din MTU, Khan MZ, et al. Trends, Predictors, and Outcomes of 30-Day Readmission With Heart Failure After Transcatheter Aortic Valve Replacement: Insights From the US Nationwide Readmission Database. J Am Heart Assoc 2022;11:e024890. [Crossref] [PubMed]
51. Nuche J, Ellenbogen KA, Mittal S, et al. Conduction Disturbances After Transcatheter Aortic Valve Replacement: An Update on Epidemiology, Preventive Strategies, and Management. JACC Cardiovasc Interv 2024;17:2575-95. [Crossref] [PubMed]
52. Truong VT, Beyerbach D, Mazur W, et al. Machine learning method for predicting pacemaker implantation following transcatheter aortic valve replacement. Pacing Clin Electrophysiol 2021;44:334-40. [Crossref] [PubMed]
53. Jia Y, Li Y, Luosang G, et al. Electrocardiogram-based prediction of conduction disturbances after transcatheter aortic valve replacement with convolutional neural network. Eur Heart J Digit Health 2024;5:219-28. [Crossref] [PubMed]
54. Galli V, Loncaric F, Rocatello G, et al. Towards patient-specific prediction of conduction abnormalities induced by transcatheter aortic valve implantation: a combined mechanistic modelling and machine learning approach. Eur Heart J Digit Health 2021;2:606-15. [Crossref] [PubMed]
55. Armijo G, Nombela-Franco L, Tirado-Conte G. Cerebrovascular Events After Transcatheter Aortic Valve Implantation. Front Cardiovasc Med 2018;5:104. [Crossref] [PubMed]
56. Patel KP, Rathod KS, Lansky AJ, et al. Cerebrovascular Events in Patients Undergoing Transcatheter Aortic Valve Replacement: A Review. Stroke 2024;55:2754-64. [Crossref] [PubMed]
57. Okuno T, Overtchouk P, Asami M, et al. Deep learning-based prediction of early cerebrovascular events after transcatheter aortic valve replacement. Sci Rep 2021;11:18754. [Crossref] [PubMed]
58. Baig AA, Manion C, Khawar WI, et al. Cerebral emboli detection and autonomous neuromonitoring using robotic transcranial Doppler with artificial intelligence for transcatheter aortic valve replacement with and without embolic protection devices: a pilot study. J Neurointerv Surg 2024;16:1167-73. [Crossref] [PubMed]
59. Avvedimento M, Real C, Nuche J, et al. Incidence, Predictors, and Prognostic Impact of Bleeding Events After TAVR According to VARC-3 Criteria. JACC Cardiovasc Interv 2023;16:2262-74. [Crossref] [PubMed]
60. van Nuland PJA, van Ginkel DJ, Overduin DC, et al. The impact of stroke and bleeding on mortality and quality of life during the first year after TAVI: A POPular TAVI subanalysis. Catheter Cardiovasc Interv 2024;104:1107-18. [Crossref] [PubMed]
61. Navarese EP, Zhang Z, Kubica J, et al. Development and Validation of a Practical Model to Identify Patients at Risk of Bleeding After TAVR. JACC Cardiovasc Interv 2021;14:1196-206. [Crossref] [PubMed]
62. Jia Y, Luosang G, Li Y, et al. Deep Learning in Prediction of Late Major Bleeding After Transcatheter Aortic Valve Replacement. Clin Epidemiol 2022;14:9-20. [Crossref] [PubMed]
63. Kolte D, Khera S, Sardar MR, et al. Thirty-Day Readmissions After Transcatheter Aortic Valve Replacement in the United States: Insights From the Nationwide Readmissions Database. Circ Cardiovasc Interv 2017;10:e004472. [Crossref] [PubMed]
64. Zahid S, Agrawal A, Salman F, et al. Development and Validation of a Machine Learning Risk-Prediction Model for 30-Day Readmission for Heart Failure Following Transcatheter Aortic Valve Replacement (TAVR-HF Score). Curr Probl Cardiol 2024;49:102143. [Crossref] [PubMed]
65. Theriot ML, Bergen EC, Lugo JG, et al. Reducing Readmissions After Transcatheter Aortic Valve Replacement Through Enhanced Transitional Care. JACC Case Rep 2025;30:105711. [Crossref] [PubMed]
66. Penso M, Pepi M, Fusini L, et al. Predicting Long-Term Mortality in TAVI Patients Using Machine Learning Techniques. J Cardiovasc Dev Dis 2021;8:44. [Crossref] [PubMed]
67. Agasthi P, Ashraf H, Pujari SH, et al. Artificial Intelligence Trumps TAVI2-SCORE and CoreValve Score in Predicting 1-Year Mortality Post-Transcatheter Aortic Valve Replacement. Cardiovasc Revasc Med 2021;24:33-41. [Crossref] [PubMed]
68. Balu A, Nallagonda S, Xu F, et al. A Deep Learning Framework for Design and Analysis of Surgical Bioprosthetic Heart Valves. Sci Rep 2019;9:18560. [Crossref] [PubMed]
69. Oldenburg J, Borowski F, Schmitz K-P, et al. Computation of flow through TAVI device by means of physics informed neural networks. Current Directions in Biomedical Engineering 2022;8:741-4.
70. Mutlu O, Saribay M, Yavuz MM, et al. Material modeling and recent findings in transcatheter aortic valve implantation simulations. Comput Methods Programs Biomed 2024;255:108314. [Crossref] [PubMed]
71. Catalano C, Scuoppo R, Turgut T, et al. Credibility assessment of patient-specific modeling in transcatheter aortic valve implantation. I. A population-based validation of patient-specific modeling. APL Bioeng 2025;9:046101. [Crossref] [PubMed]
72. Rauseo E, Bevis L, Chen X, et al. Understanding the mechanisms of TAVI durability through computational modelling: a multidisciplinary review. Eur Heart J Digit Health 2026;7:ztag020. [Crossref] [PubMed]