Deep vein thrombosis (DVT) most often begins in the pockets behind venous valve leaflets, where blood flow stagnates to initiate the clotting process. However, the initiating physics at the scale of individual valve leaflets remains unclear. This means DVT prevention relies almost entirely on lifestyle changes and risk is not proactively measured. However, recent evidence shows that venous valve geometry and leaflet elasticity reshape local fluid dynamics in patterns that can concentrate platelets, red cells and procoagulant proteins. We hypothesise that the drivers of these dynamics can be captured by advanced mechanistic flow models and mapped onto clinical DVT phenotypes that can be then rapidly inferred using machine learning models.
Fluid–structure interaction (FSI) models can capture bidirectional coupling between pulsatile venous flow and deformable valve leaflets, including shear stresses on the endothelium. When augmented with transport–reaction models for platelets, fibrin and von Willebrand factor, such frameworks can generate testable hypotheses about triggers of thrombosis and predict how age- or disease-related leaflet stiffening alters risk. This project will thus create a workflow that couples blood flow, structural valve dynamics and nascent thrombosis in anatomically accurate venous segments (Fluid-Structure-Biochemisitry Interaction, FSBI, models).
Given that FSBI models are computationally very expensive, Machine Learning (ML) surrogates like Gaussian Process Emulators (GPEs) and deep neural networks trained on mechanistic outputs from the FSBI models will be used to create ML Digital Twins for rapid, personalised DVT prediction. By training on database of real patient datasets (e.g. compression ultrasound with Doppler; CT/MR venography) augmented with FSBI simulations, these ML surrogates can deliver millisecond predictions with quantified uncertainty, enabling probabilistic risk maps. Therefore, the proposed Digital Twin will open a path to evaluate DVT blood signatures that are invisible to standard scanning protocols, shifting care from a reactive model to a proactive one.