Inferring left atrial thrombin concentration from 4D CT contrast dynamics by physics-informed neural networks & multi-fidelity coagulation cascade modeling

People with atrial fibrillation (AF), a common arrhythmia with a lifetime risk of 25%, have significantly higher rates of atrial thrombosis and are five times more likely to suffer a stroke than people with a regular heartbeat. Anticoagulant drug prescription to people with AF is based on clinical risk scores based on demographic factors with modest accuracy. These risk scores do not include patient-specific factors affecting thrombosis. Of note, they ignore the dynamics of the coagulation cascade under patient-specific left atrial flow. Here, we present a computational pipeline to predict the concentration of thrombin, a central coagulation enzyme responsible for clot fiber formation and platelet activation, from 4D CT clinical sequences of LAA contrast dynamics. First, a physics-informed neural network predicts blood residence time from contrast agent dynamics. Second, a computationally efficient multi-fidelity model of the coagulation cascade predicts thrombin concentration from residence time. This pipeline is tested on ground-truth data from CFD simulations in idealized, fixed-wall geometries and patient-specific, moving-wall left atrial meshes. Proof-of-principle of clinical application is shown on 4D CT acquisitions from AF patients.