Blood flow predictions in data-poor regimes: A physics-informed Bayesian approach

Computation modeling blood flow properties can aid diagnosis and treatment of cardiovascular and cerebrovascular diseases. However, high-fidelity predictions are computationally expansive, and blood flow measurement via transcranial Doppler ultrasound or imaging alone lack the sufficient resolution to be used directly or else used to train a machine learning surrogate model. Such limitations make it vital to develop a computationally inexpensive model that provides prediction based on sparse computational/clinical data. We present a physics-informed Gaussian process regression technique to predict the blood flow properties from a very few sparse measurements. The presented algorithm is computationally inexpensive, and it has the potential to be used in clinical settings. We demonstrate our methodology on examples such as a Y-shaped bifurcation, abdominal aorta, and brain vasculature.