CFD-derived radiomics from hemodynamic maps for quantitative assessment of left atrial flow in atrial fibrillation: a...

Atrial fibrillation (AF) alters left atrial (LA) hemodynamics and increases thromboembolic risk. Computational fluid dynamics (CFD) can characterize complex flow patterns. However, extracting objective biomarkers from three-dimensional velocity fields remains challenging. This proof-of-concept study evaluates whether radiomics-based quantification of spatial texture patterns can provide reproducible LA flow biomarkers for AF classification.

In this study thirty subjects (10 controls, and 20 AF patients) underwent contrast-enhanced cardiac computed tomography (CT). Five hemodynamic parameter maps (throughflow, wall-Parallelity Degree, normalized vorticity, local normalized helicity, and flow angle) across eight anatomical planes were generated from patient-specific CFD simulations. From 360 radiomics features, robust biomarkers were identified through nested leave-one-out cross-validation with three-stage feature selection: statistical filtering, effect-size thresholding, and stability assessment. Classification used the support vector machines with the radial basis function kernel. The model interpretability was evaluated using SHAP analysis.

Five robust features were identified. Four were Gray-Level Non-Uniformity features from throughflow regions, each achieving 100% selection frequency and bootstrap stability. The fifth was a Dependence Entropy feature from LAA helicity maps, with 80% selection frequency and 100% bootstrap stability. Effect sizes were large (ε ² = 0.46-0.60). The classifier achieved 93% accuracy, 95% sensitivity, 90% specificity, and an AUC of 0.92 (95% CI: 0.78-1.00). Permutation testing confirmed that performance significantly exceeded the chance level expected from random label assignment (p = 0.002). The features correlated significantly with LA enlargement (ρ = 0.45-0.64) and mitral regurgitation grade (ρ = 0.49-0.69), supporting biological plausibility.

Radiomics analysis of CFD-derived hemodynamic maps can effectively quantify complex LA flow patterns and identify reproducible flow biomarkers in AF. The identified features demonstrated high stability and significant correlations with established markers of atrial remodeling. They also showed discriminative performance substantially exceeding that of clinical variables alone. External validation in larger multicenter cohorts is needed. This framework establishes the feasibility of objective hemodynamic phenotyping with potential applications in AF risk stratification.