This proof of concept study demonstrates the capabilities of a virtually automatically generated digital twin framework for enhancing hemodynamic monitoring in critical care. By combining a deterministic cardiovascular model with patient-specific data through data assimilation techniques, the digital twin can act as a data denoiser, reconstruct physiological waveforms that are typically unavailable in critical care settings and generate clinically relevant biomarkers. Validation was performed using real data from patients under general anesthesia. The proposed framework efficient calibration and ability to follow the patient's state over time supports the possibility of real-time bedside applications.
Atrial fibrillation (AF) is the most prevalent cardiac arrhythmia and significantly increases the risk of stroke. Left atrial catheter ablation (LACA), utilising techniques like cryoablation and radiofrequency ablation, is commonly employed to restore sinus rhythm, which helps normalise atrial function and reduce blood stasis, thereby lowering the risk of thrombus formation and stroke. While these procedures are effective in restoring normal rhythm, cases of thrombus formation during ablation have been reported. Certain regions of the atrium appear more susceptible to thrombus development. To investigate this, computational fluid dynamics (CFD) and thrombogenesis models were used to simulate blood flow patterns, temperature distributions, non-Newtonian blood behaviour, and coagulation processes in both cryoablation and radiofrequency ablation scenarios. This study found no significant difference in flow characteristics or thrombogenicity between cryoablation and radiofrequency ablation. However, the right inferior pulmonary vein (RIPV) and the posterior wall of the left atrium (LA) were identified as two regions, commonly targeted during LACA, that are particularly prone to thrombus formation. Furthermore, the incorporation of non-Newtonian blood behaviour in the flow model revealed a correlation between exacerbated flow metrics and increased thrombogenicity, emphasising the relationship between non-Newtonian blood properties and stroke risk. This novel approach offers a promising framework for future studies, which could refine these models to better understand the mechanistic factors contributing to periprocedural stroke risk during LACA.
The left and right ventricles (LV, RV) of the heart are connected hemodynamically via the systemic and pulmonary circulations and joined mechanically by shared myocardial fibers and the interventricular septum, resulting in ventricular interdependence. Changes in volume, pressure, and contractility of one ventricle can affect the other by volume shifts between the two circulations and by displacement of the septum. Here, we present a spatially resolved, biventricular model of the heart that can be used to investigate the effects of LV-RV interactions. Changes in the shape of the LV, RV and septum are represented by a small number of time-dependent parameters. In combination with a closed-loop circulation, this results in a low-order, computationally efficient model system. The free wall of the LV is represented as part of a prolate spheroid, and the septum and the RV are modeled by non-axisymmetric deformations of the spheroidal shape. The model includes interventricular interactions occurring as a result of changes in preload and afterload, as well as those resulting from direct mechanical interactions mediated by the septum. Model parameters can be matched to fit realistic patient geometries from echocardiogram data. Ventricular function is analyzed in terms of preload recruitable stroke work (PRSW). With decreasing LV contractility, RV PRSW remains almost constant over a range of contractility, but decreases when contractility is severely impaired. In response to reduced RV contractility, LV PRSW decreases slightly. The model can be used to examine the circulatory and mechanical interactions of the ventricles in the context of cardiac disease.
Mitral valve (MV) regurgitation is a highly prevalent and deadly cardiac disease affecting over 2
Patients with non-valvular atrial fibrillation (AF) often experience cardiac arrhythmia, leading to altered blood flow and an increased risk of thrombus formation in the left atrial appendage (LAA), thereby significantly raising the likelihood of stroke. Understanding the fluid dynamics within the LAA is essential for identifying effective therapeutic strategies for AF patients. However, fluid simulations require verification and validation experiments to build the necessary credibility to be used to support clinical decisions. For doing so, in vitro experiments are key to create ground-truth data for computational models. In this study, an in vitro experimental setup was developed to investigate the fluid dynamics of the left atrium and LAA using a deformable, patient-specific silicone phantom fabricated from computed tomography (CT) data segmentation. The phantom included a geometry with LA, LAA and pulmonary veins and it was integrated into a mock circulatory loop, where physiological conditions were imposed. Particle Image Velocimetry (PIV) was performed using a dedicated LED-PIV system, focusing particularly on the LAA region. Velocity fields were acquired across multiple longitudinal and transversal planes, revealing zones of flow stagnation and vortex formation near the LAA apex. The data provide a robust experimental benchmark for validating computational fluid dynamics (CFD) models, advancing the understanding of LAA hemodynamics, and informing the design of therapeutic interventions for atrial fibrillation.
Atrial fibrillation (AF) is the most common arrhythmia. The risk of AF increases in postmenopausal women, suggesting beneficial effects of oestrogen. This study examines the effects of oestrogen at different concentrations corresponding to physiological states, including the menstrual cycle, pregnancy and menopause. We incorporate a hormone effect function, derived from experimental data, into an existing biophysical model of the atrial cardiomyocyte to simulate the effects of oestrogen. Our modelling results suggest that oestrogen may exert cardioprotective effects at higher physiological concentrations, such as those observed in the phases of the menstrual cycle and pregnancy. Specifically, it prolongs the action potential duration and facilitates the termination of isthmus-dependent re-entry in an AF substrate. These findings support the protective effect of oestrogen and provide insights into the role of oestrogen in AF dynamics, with implications for future studies to understand sex-specific arrhythmia risk and potential therapeutic strategies.
Cardiac diffusion tensor imaging (cDTI) is an MRI technique used to characterize myocardial microstructure. Key metrics including mean diffusivity (MD), fractional anisotropy (FA), and helix angle pitch (HAP) have shown relevance in detecting microstructural remodeling. Herein, we aim to establish healthy baseline metrics for this emerging technique. To do so we assess average global cDTI metrics in the left ventricle for a cohort of healthy subjects (N = 46) at 3 T with a single- shot echo-planar imaging (EPI) sequence. We hypothesize: (1) that there are no significant differences in cDTI metrics for groups within this population; and (2) that there are no significant differences between global, regional, and group-wide median values. To assess differences in global and regional cDTI metrics between groups, a one-way ANOVA test (for normal distributions) or Kruskal-Wallis test (for non-normal distributions) was performed to detect significance in cDTI metrics between four population groups: (1) Male < 40 years old (yo), Female < 40yo, Male ≥ 40yo, and Female ≥ 40yo. To evaluate differences in regional group-wide cDTI metrics compared to the global group-wide median, American Heart Association (AHA) segments were identified and unpaired t-tests with Bonferroni correction were used to detect significance between individual regions and the global group-wide median. We found no significant differences in global MD, FA, and HAP between age- and sex-based groups. Regional analysis revealed some significant differences compared to global group-wide cDTI metrics as well as in a few regions in group-specific comparisons. Overall, this study establishes baseline cDTI metrics in healthy subjects, providing a normative reference for assessing changes in cDTI metrics.
Cardiac digital twins have shown promise to personalize treatments. However, there are multiple challenges to incorporate patient-specific information from non-invasive data. For instance, recovering the activation sequence in atria from the standard electrocardiogram (ECG) remains elusive. Recent studies have tackled this task on the ventricles, where the ECG signal is much stronger. This work presents a novel methodology to recover the atrial electrical activity with physics-informed neural networks. Instead of focusing on the activation times, we predict the direction of propagation of the electrical wave at each point with a neural network. Then, by solving a linear system for the Poisson equation, we recover the activation times that satisfy the anisotropic eikonal equation. The proposed methodology is compared with a methodology that predicts directly the electrical propagation and does not enforce the propagation model. We compare it to a traditional physics-informed neural network formulation, where the eikonal equation is only weakly imposed. We validate our methodology in a biatrial synthetic case using realistic lead fields for ECG calculation. We then learn the activation sequence from patient data, recovering a physiological activation pattern. We believe this is a first step toward digital twinning of the atria.
The integration of artificial intelligence into cardiac imaging is transforming the landscape of clinical diagnostics by enabling precise, efficient, and scalable solutions. This study presents an advanced deep learning approach leveraging the YOLOv8x model for real-time segmentation of heart chambers from a single horizontal long-axis cardiac magnetic resonance image (MRI). Left ventricular and left atrial volumes, and right atrial areas, were measured and validated against industry-standard software. To ensure scalability and accessibility, an end-to-end AWS pipeline was developed, seamlessly integrating image processing, model inference, and result visualization. This research underscores the potential of combining deep learning with cloud technologies to deliver robust and adaptable solutions for functional cardiac image analysis.
Cardiac diffusion tensor imaging (cDTI) is an emerging meth-od capable of characterizing the microstructural organization of both healthy and diseased myocardium. One of the challenging aspects of a cDTI study is the associated data processing due to various acquisition imperfections that can corrupt the acquired data. We sought to investigate the role of various data processing steps by evaluating an open-source cDTI data processing software, Cardiac Diffusion in Python (CarDpy). In order to achieve this goal, healthy volunteers (N = 40) were imaged. Imaging data was evaluated at six incremental postprocessing steps (POSTs) using the CarDpy pipeline. cDTI metrics such as mean diffusivity (MD), fractional anisotropy (FA), and helix angle range (HAR) were evaluated after each POST. Additionally, the uncertainties of MD (dMD), FA (dFA), and the primary eigenvector (de(1)) were evaluated to quantify the data precision. Statistical testing was performed after each POST in a comparison with the final POST. Empirical measurements of MD displayed stable trends across all POSTs, while a decrease in FA was observed with each incremental step. HAR remained stable after the integration of the POST that incorporated image registration into the data processing pipeline. Uncertainties decreased for all metrics as each incremental POST was added. dMD, dFA, and the de1 had minimal improvements after the POST that incorporated shotrejection into the data processing pipeline. Overall, this study provides an in-depth analysis pertaining to the impact of image processing on cDTI metrics and their corresponding uncertainties.
Motion tracking plays an important role in many domains including biomedical and mechanical engineering. Numerous methods have been proposed in the literature. While recent machine learningbased approaches provide fairly robust and accurate results, classical methods -combining statistical analysis of image intensity with a model of the underlying motion- remain widely used, as they offer greater control over the obtained results. Such approaches may handle highly complex motions; however, any artifact in the images (e.g.., partial voluming, local decrease of signal-to-noise ratio or even local signal void), may drastically affect the tracking. In order to reduce the impact of such artifacts, this paper extends a recently proposed motion tracking approach that relies on both a geometrical model of the tracked object and a model of the images themselves. The problem is thus formulated in terms of finding the displacement of the object such that the generated images, obtained with the image model, best match the acquired images. That way, if any artifact is present in the acquired images but also well represented in the image model, precise motion information can still be recovered from the acquired images. The performance of the proposed method is illustrated on tagged magnetic resonance images, for which acquired images are usually low-resolution, generating significant partial voluming. A simple model of such images is formulated. The method is applied to 2D synthetically generated image series representing various kinematics, with resolutions as low as those found in in vivo acquisitions, and compared to a classical tracking method. In order to avoid computing the cost function gradient, a derivative-free algorithm is used to solve the optimization problem. On the considered examples, the proposed method performs better than the classical tracking method.
High-fidelity patient-specific heart models, or cardiac digital twins, have the potential to revolutionize cardiovascular care by enabling precise diagnosis, personalized treatment planning, and accurate predictions of disease progression. However, the extensive computational cost of these models, particularly in finite element (FE) simulations, remains a significant barrier to their adoption in clinical practice. In this work, we address this challenge by developing a transformer-based data-driven surrogate model to predict left ventricle (LV) dynamics and global hemodynamics. The surrogate model achieves a speed improvement of more than 40,000 times compared to traditional FE simulations while maintaining an acceptable accuracy of 1.44 mm mean absolute error (MAE). In addition, its parallelization capability enables efficient generation of high-volume simulations, making it a compelling replacement for the FE model as a forward simulator in inverse analysis. These advances underscore the potential for the integration of cardiac digital twins into clinical workflows, advancing scalable patient-specific computational modeling in precision cardiology.
This study assesses the use of Eulerian volumetric velocity images of the ventricular blood flow for estimating material properties of a physiological fully-dimensional FSI model of the systolic phase of the heart contraction. An efficient partitioned, semi-implicit FSI algorithm is employed, coupling a fluid fractional step scheme with a hyperelastic solid model. Cardiac mechanics model parameters characterizing the active contraction and epicardial wall boundary conditions, accounting for the external tissue support, namely the myocardial tissue contractility and the epicardial stiffness, are estimated from synthetic (i) solid measurements only, (ii) fluid measurements only and (iii) combined fluid and solid measurements. The parameters are estimated efficiently using a reduced-order unscented Kalman filter. Ground truth values of the estimated parameters are accurately recovered if the data provides a sufficient temporal resolution. While the contractility estimation only benefits by aggregating fluid measurements to the solid data by a reduced sensitivity to measurement noise, the epicardial stiffness is resolved more accurately when using fluid measurements instead of solid measurements.
Myocardial infarction (MI) remains a leading cause of death, effecting nearly 1 million people annually. Cardiac simulations of MI in a clinical setting can aid in our understanding of the disease and guide optimal therapies. While the accuracy of current cardiac computational models of MI have improved significantly, execution times remain prohibitively slow for clinical use. In recent years computational technology has made major advances, in large part due to the rise of machine learning (ML). Using JAX, an open source ML-library, we developed a high speed cardiac simulation platform. This platform included an inverse modeling framework for estimation of active myofiber stress. The framework was developed using experimental and modeling data from a comprehensive dataset from a single ovine heart, including pressure volume loops and diffusion tensor magnetic resonance imaging data. We also conducted simulations to incorporate heart wall compressibility based on the levels of active contraction. When using a comparable mesh to our previous ABAQUS based approach, we showed a 100x speed up in simulation time. The substantial speed improvement also facilitates more sophisticated multiphysics cardiac simulations for both research and clinical settings. The simulation time was also compared to the FEniCS opensource code and showed considerable gains as well. Ongoing work includes enabling intrinsic electrophysiology capability to enable true electromechanical based simulations.
Cardiac fibers play an essential role in the electrical and mechanical function of the heart, making them a critical parameter in cardiac modeling. However, identifying the 3D arrangement of the fibers in clinical practice still encounters some limitations. Solving the inverse problem of inferring the fiber orientations from electrophysiological data has been reported as a potential method for identifying patient-specific fiber orientations. In this work, we employ.-Fibernet, a recently proposed Physics-Informed Neural Network (PINN) model, to reconstruct the fiber orientations on the right ventricle's endocardial surface from a single activation map. The model equips the traditional PINN framework with an ensemble of parallel neural networks to cope with the uncertainty in the fiber approximations. Each ensemble member estimates the activation times and fiber orientations. The best fiber orientations are finally selected using a specific method to reduce the uncertainty of predictions. We evaluated the performance of the model in a simple propagation pattern and a more realistic propagation pattern incorporating the influence of the Purkinje network. Our results indicate the robustness of.-Fibernet and its capacity to learn complex activation patterns and fiber distributions while trained with only a single activation map.
In embryonic stages of development, the mammalian heart is characterized by a rapid division of cells, and their organization into a functional beating organ. Although cardiomyocyte nuclei are known to be elongated and oriented along their length in the adult heart, how nuclear shape, orientation, and arrangement evolve during development is not well understood. To tackle this fundamental problem we examine the evolution of nuclear geometry in the mouse heart wall over four post-septation embryonic days, and in the newborn pup, using confocal microscopy and quantitative modeling. Over this period there are notable changes in nuclear count, density, spatial layout, and organization. For example, we observe a steady increase in nuclear elongation, but a decrease in their volume. Whereas nuclei are more densely packed towards the epicardium just after septation, in time they become more uniformly distributed throughout the heart wall. We hypothesize that such structural arrangements in the developing myocardium could have an important bearing on healthy heart function.
We briefly present a new epicardial source model for the inverse problem of cardiac electrocardiography. It couples a surface equation on the epicardium with the three-dimensional torso, while incorporating heart-torso fluxes and possible anisotropic conductivities in the intracellular and extracellular domains. As this capability is new for a surface model, we investigate the impact of inserting epicardial fiber directions into the surface source model on inverse reconstructions, compared with isotropy-based reconstructions. Our results show that the epicardial fibers degrade the epicardial model based ECGi reconstructions, in particular the transmembrane voltage, whereas isotropic reconstructions demonstrate significantly higher accuracy for the extracellular potential and the transmembrane voltage. We believe that, due to interactions and compensations within the myocardial volume, the solutions of the bidomain model taken on a surface are more similar to isotropic solutions than strongly anisotropic ones. In this case, taking an isotropic assumption in a surface model based ECGi is a relevant choice.
In patients with non-valvular atrial fibrillation (AF), the left atrial appendage (LAA) is the primary site for thrombus formation. While oral anti-coagulation is the first-in-line treatment, many patients are ineligible due to bleeding risks, making left atrial appendage occlusion (LAAO) a viable alternative. LAAO involves implanting a device at the LAA entrance to block blood flow and reduce thrombus risk. However, suboptimal device implantation can lead to complications such as device-related thrombus. Given the anatomical variability of the LAA, selecting the optimal device configuration is challenging and requires patient-specific customization. Computational fluid dynamic simulations allow the comprehensive evaluation of hemodynamic patterns and are useful to assess the risk of blood stasis of distinct device configurations in a patient-specific manner. This study developed patient-specific in-silico hemodynamic models for 9 patients using two pacifier-type occluder configurations: one representing the clinical implantation and the other a virtual pre-planned implantation. Additionally, predictive algorithms were implemented for device optimality and transseptal puncture localization. The goal was to assess the risks associated with different device configurations and support clinicians in optimizing the pre-planning process for LAAO interventions. The findings indicate that proximal positions may be linked to a reduced risk of blood stasis, emphasizing the importance of considering both device orientation and left atrial (LA) morphology when evaluating blood stasis risk. Furthermore, TSP localization showed notable variability among patients.
Elevated left ventricular end-diastolic pressure (LVEDP) is an important index for the prognostication of several cardiovascular conditions, and is most reliably measured using cardiac catheterisation. This invasive surgical procedure carries potential risk of complications, high costs, and limits accessibility. We investigated the use of a multimodal machine learning model for non-invasive classification of elevated LVEDP, using a combination of apical two- and four-chamber 2D echocardiographic views, a 3D kinematic mesh of the left ventricle, and routine clinical measurements. The model achieved more sensitive results compared to clinical guidelines (sensitivity = 0.43 vs 0.25), but with lower accuracy (0.65 vs 0.76), specificity (0.74 vs 0.92), positive (0.38 vs 0.50) and negative (0.78 vs 0.80) predictive values. Among the different multimodal input types, the clinical measurements yielded the most significant influence on LVEDP classification, with a moderate influence of the apical four- and two-chamber views and a minimal effect of the 3D kinematic mesh. Different combinations of input types were explored, which highlighted how the interactions between different inputs influenced predictions and the potential for selecting specific combinations based on clinical objectives. This approach may offer a better understanding of markers for elevated filling pressure and demonstrate the predictive value of multimodal input data.
Patient-specific cardiovascular simulations have become an integral part of cardiovascular research. A primary factor hindering large patient cohort studies and clinical impact of cardiovascular simulations is their dependence on accurate patient-specific three dimensional geometric models which remain time-consuming and costly to construct from medical image data. Methods have been proposed to automate the model construction process, for either vascular or cardiac purposes. We propose a novel method, MeshGrow, that, automatically, reconstructs both the cardiac chambers as well as the aorta and its main sub-branches, and returns a simulation ready mesh with defined aortic valve connecting the two. We deploy two different methods of model construction for the cardiac chambers and vascular regions independently to address the specific challenges involved with each. The method is a two step approach; 1) meshing the cardiac structures and 2) growing the vasculature out from it. We present test results of our method on three CT image data and compare to ground truth manually constructed models. Additionally, we compare results with state-of-the-art methods. Results show that MeshGrow achieves higher metric scores than benchmark methods on test set. With this work, we demonstrate the advantage of anatomy specific modeling approaches for patient-specific cardiovascular simulation.
