Keywords
Abstract
This study introduces the first method to model haemodynamics with artificial intelligence for the early stage prediction of aortic valve dysfunction progression. Current approaches in imaging the cardiovascular system do not accurately forecast the disease course in aortic stenosis which hampers timely intervention with best management strategies. We constructed a deep learning system that combines multimodal echocardiographic datasets to derive dynamic blood flow metrics and capture subtle valvular haemodynamic changes that steer structural alterations. We employed an aggregate dataset from 8,472 patients with different stages of aortic valve pathology from multiple centres, and trained and validated a hybrid convolutional-recurrent neural network architecture which achieved 91.7% accuracy in classifying stenosis severity and predicting progression within 24 months with an area under the curve 0.918. The AI model uncovered key hinges-of-change features like wall shear stress heterogeneity, systolic flow convergence angles, and velocities of leaflet excursion which conventional methods are blind to. Comparison against traditional echocardiographic measures yielded a net reclassification improvement of 27.4% arguing for the robustness of our model in predicting disease progression. The model also performed well across various patient subgroups including those with low ejection fraction, additional valvular regurgitation, and multiple other comorbidities.
The results show AI-based haemodynamic modelling improves the early detection of aortic valve malfunction, which may lead to customised monitoring and timely clinical interventions, thereby enhancing outcomes in patients with valvular heart disease.