Comparing Machine Learning Approaches for Predicting CFD-Derived Stroke Risk Indicators in Atrial Fibrillation Patients
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Non-valvular atrial fibrillation (AF) is associated with a five-fold increased risk of stroke, mainly due to impaired contractility of the left atrium (LA) leading to blood stasis and subsequent thrombus formation within the left atrial appendage (LAA). Current AF stroke risk stratification schemes, such as the CHA₂DS₂-VASc/ CHA₂DS₂-VA score, use comorbidities and do not capture mechanistic factors like blood flow dynamics and hypercoagulability. To address this, we developed a multiphase computational fluid dynamics (CFD) model of the LA, incorporating patient-specific geometries; modelling of the coagulation cascade; and non-Newtonian blood behaviour within the LAA.
Using 84 simulation cases generated via Latin Hypercube Sampling of physiological blood parameters and 21 patient-derived LA anatomies, we trained surrogate machine learning models, including Ridge regression, XGBoost, Gaussian Process Emulators (GPEs), and deep learning networks, to predict CFD outputs such as blood viscosity in and fibrin concentrations in the LAA. Deep learning achieved R² values up to 0.90, with the accuracy increasing when both physiological parameters and the raw CT image were included. Other models showed uneven performance with R 2 values below 0.7, highlighting the role of nonlinearities between parameters.
The study presents a novel CFD model that captures the transition from blood stasis to clot formation, representing the full thrombotic continuum underlying stroke risk in AF, and a deep learning approach to enable efficient prediction of mechanistic outputs of clinical value for stroke risk stratification in AF patients.
Author Summary
Atrial fibrillation is a common heart rhythm disorder that greatly increases the risk of stroke. In many patients, blood can pool inside a small pouch of the heart called the left atrial appendage, where clots may form and later travel to the brain. Current clinical tools used to estimate stroke risk mainly rely on a patient’s medical history and do not directly assess the mechanistic processes that lead to clot formation.
In this study, we developed a computer model that simulates how blood flows and clots inside the heart using patient-specific heart anatomies derived from medical imaging.
Our model combines blood flow, blood biochemistry, and the changing physical properties of blood during clot formation. We then used machine learning methods to predict these complex simulation results more efficiently. Deep learning models performed best, particularly when both clinical parameters and heart imaging data were included.
Our work provides a new way to study the full process linking abnormal blood flow to clot formation in atrial fibrillation. In the future, this approach could support more personalised and mechanistic assessment of stroke risk and help guide treatment decisions.
