Anatomically parameterised statistical shape modelling of LVAD-supported left ventricles for thrombosis risk...

Heart failure remains a major global burden, and limited donor availability has made left ventricular assist devices (LVADs) a critical alternative. However, most studies of left ventricular (LV) flow under LVAD support rely on single cases or small cohorts, contributing to large variability in inflow cannula designs and limiting population-level understanding of how LV morphology influences haemodynamics. This study addresses this gap using a parametrised statistical shape model (SSM) of the LV. A nnU-Net deep learning model was trained on 15 HeartMate 3 patients and used to segment 25 additional scans, yielding 40 LV geometries. An orthogonally constrained, anatomically parameterised SSM (SSM_OC-anat) was constructed using six metrics: LV volume (Vol), long- (LA) and short-axis (SA) lengths, sphericity (Sph), eccentricity (EI), and apical conicity ratio (ACR). Rigid-wall LV models were simulated at 5 L/min LVAD flow using the SST k-ω turbulence model with a non-Newtonian Carreau model. Simulations at ±1 SD along each parameter assessed relationships between morphology and haemodynamics using wall shear stress (WSS), turbulent kinetic energy (TKE), stagnation volume, and washout. Elongated LVs (LA + SD) showed the slowest washout, retaining 9.2% of blood after 10 s, indicating increased thrombosis risk. Maximum WSS was consistent across models (80-85 Pa at the cannula transition, ∼30 Pa at the tip). TKE was most sensitive to LV volume (∼1.5 J/kg variation), followed by eccentricity (0.5 J/kg). Larger LV volumes and SA were associated with improved washout and reduced stasis (∼2.5-3%), whereas increased LA or EI correlated with higher thrombosis risk. The developed framework establishes a robust foundation for understanding LV morphology in LVAD patients, supporting improved inflow cannula design, patient-specific assessment, and surgical planning.