Radiel Health - Uncertainty Aware Prediction for Hemodynamic Parameters in Vascular Surgery

Abstract

We present an uncertainty-aware neural surrogate for predicting hemodynamic parameters on vascular geometries. Our architecture combines Graph Attention Networks (GAT) for local geometric encoding with General Powerful Scalable (GPS) layers for global context, and a Bayesian Neural Network (BNN) task head for epistemic uncertainty quantification. On bifurcation geometries with varying Reynolds numbers and angles, the model achieves 𝑅2=0.962 for Wall Shear Stress (WSS) magnitude with a 12% relative error, while providing calibrated 95% credible intervals. This addresses a core challenge in deploying machine learning surrogates for clinical decision support: knowing when not to trust the model.