Research Projects

Bileaflet Mechanical Heart Valves (BMHVs) pose a significant thromboembolism risk due to turbulent blood flow, high shear stress, and stagnation, which promote clot formation. These clots can lead to life-threatening conditions like pulmonary embolism, stroke, or heart attack. Additionally, the foreign material of the valve and potential pannus formation (abnormal tissue growth) further contribute to this prothrombotic environment. Consequently, BMHV patients require lifelong anticoagulation, which has its own bleeding complications. Since thromboembolism risk varies greatly between patients, a "one-size-fits-all" valve design is insufficient, necessitating individualised analysis, and computational fluid dynamics (CFD) based numerical modelling is one of the approaches that can facilitate this analysis. However, the current conventional CFD models are computationally very expensive and often depend on oversimplified boundary conditions, limiting their applicability to patient-specific predictions and analysis. Therefore, this project aims to utilise artificial intelligence (AI) based Physics-Informed Neural Networks (PINNs) alongside CFD-FSI (Fluid-Structure Interaction) computational framework to identify and mitigate thrombus-prone regions in BMHVs. By analysing flow characteristics and correlating thrombus risk with specific geometric and design parameters, the present project will propose optimised valve geometries that will reduce thrombogenic potential while preserving favourable hemodynamic performance. In doing so, the proposed project is subdivided into three stages. Stage one will develop a hybrid PINN-CFD-FSI framework using an idealised BMHV, with PINNs will model outer chamber flows to provide boundary conditions for the leaflet domain, where conventional CFD–FSI will handle the complex FSI interaction. Platelet activation and shear stress will be modelled using an Euler-granular multiphase approach, and a specialised neural network will quantify thrombus risk. Stage two will generalise this prediction by training PINNs and NNs (Neural Networks) on various synthetic geometries, optimising a composite loss function that balances physical constraints and data-driven residuals. PINNs and neural networks will be trained on synthetic geometries with varied anatomical and design parameters (e.g., aortic root diameter, leaflet angles, hinge depth). A composite loss function will guide this training, balancing physical constraints with data-driven residuals through gradient-based optimisation. Stage three will focus on validating a novel leaflet design optimised to minimise high shear and blood stasis through CFD–FSI simulations. Neural networks, trained on post-processed data, will help predict thrombus-prone regions, enhancing the design's clinical utility. This project not only enables efficient and accurate simulation of valve dynamics but also lays the groundwork for patient-specific, thrombosis resistant valve designs.