Real-time risk stratification for coronary artery thrombosis using machine learning

Kawasaki Disease (KD) is an inflammatory pediatric condition which causes coronary artery aneurysms (CAAs) in 25% of untreated patients. CAAs may lead to thrombosis and subsequent myocardial infarction. While current treatment guidelines are based purely on normalized CAA diameter, previous work has shown that hemodynamics estimated using patient-specific computational simulations predicts risk better than CAA size. While standard clinical imaging cannot estimate relevant hemodynamic risk metrics, computational models are expensive and not feasible in clinical timeframes. This work describes a hybrid computational modeling and machine learning pipeline to provide real-time non-invasive risk stratification based on clinical imaging. Using data from patient-specific imaging, we develop mesh-based ray-tracing techniques for learning anatomical features of real aneurysms to develop a library of anatomically realistic synthetic aneurysms. The realism of these aneurysms has been tested on clinical experts, who had roughly equal rates of correctly identifying real and synthetic aneurysms. We then use computational fluid dynamics to compute hemodynamic metrics of interest using this entire library of real and synthetic aneurysms. Finally, we build a convolutional neural network model to provide real-time estimates of hemodynamics-based thrombosis risk.