AI‐Powered Multimodal Modeling of Personalized Hemodynamics in Aortic Stenosis

Figure 1. Overview of AI‐powered personalized modeling approach for AS. A) Schematic illustrating the calcification process of the AV leaflets in the aging heart and changes in the orifice area through disease progression. B) The overall workflow of developing a patient‐specific model from CT images to recapitulate clinical hemodynamic measurements of AS patients, where we propose the use of our DeepCarve‐fluid structure interaction (FSI) framework to replace conventional heart meshing techniques.

Figure 2. Overview of our AI‐powered geometry reconstruction algorithms for computational simulations. A) Visualization of our AI‐generated simulation meshes' element quality and spatial accuracy. B) Schematic of our multi‐frame meshing approach for computational flow simulations. C‐MAC: calcification meshing with anatomical consistency, FSI: fluid‐structure interaction, LV: left ventricle, AV: aortic valve.

Figure 3. Robust automated conversion to 3D printing geometry. A) Qualitative demonstration of the quality of the 3D printing mesh and its spatial consistency to the corresponding simulation mesh. Detailed steps of the conversion algorithms for the B) AV leaflets and C) the combined LV and aorta. C‐MAC: calcification meshing with anatomical consistency, LV: left ventricle, SDF: signed distance functions.

Figure 4. Overview of the computational model and patient‐specific flow patterns. A) Schematic of the FSI approach B) Illustration of an AI‐powered simulation ready mesh. Schematic of one‐way coupled, nodal displacement prescribed as a boundary condition (BC) to left ventricle (LV) wall. C‐I) Velocity map distributions in the anterior‐posterior plane for patients 1 to 7, respectively. Velocity fields are captured during the systolic temporal frame when the aortic valve (AV) velocity is at its peak in the cardiac cycle.

Figure 5. Comparison of patient‐specific hemodynamic findings with AI‐powered and conventional algorithms. A) Constructed valve and calcium geometry using DeepCarve and conventional meshing (CM). B) The visualization of particle pathlines based on fluid‐structure interaction (FSI) simulations. The cutting plane denotes the probe location for aortic valve (AV) velocity measurements. C) AV profiles of patients 1–7 during peak systole in the cardiac cycle. Comparison of FSI simulations with DeepCarve and conventional meshing/registration, shown by measured hemodynamics metrics of D) Vpeak, E) dPmax, F) dPmean, G) Left ventricular ejection fraction (LVEF), demonstrating that DeepCarve‐FSI allows for more accurate hemodynamic measurements and agrees well with the ground truth (clinical echo data). Each error bar represents means ± 1 SD (n= 3 consecutive heart cycles).

Figure 6. A platform for testing AI‐powered 3D printing mesh with soft robotic models. A) Overview of soft robotic system. B) As an example, the geometries of Patients 1 and 2 were 3D printed using our automated conversion. C) Representative echocardiographic velocity waveforms across the aortic valve from CW Doppler measurements during actuation by the soft robotic LV sleeve in three consecutive cardiac cycles. D) The percentage error of clinical metrics estimated from in vitro experiments and in silico simulations with FSI using conventional meshing/registration method (FSI‐CM) and DeepCarve (FSI‐DeepCarve). Clinical parameters include left ventricular ejection fraction (LVEF), mean pressure gradient (MP), peak pressure gradient (PP), dimensionless velocity index (DVI), and peak aortic velocity (PV).