A machine learning framework for rapid wall shear stress evaluation to assess early atherosclerotic risk in carotid sinus

Background Low wall shear stress (WSS) and high oscillatory shear index (OSI) are critical indicators of early atherosclerotic risk. However, conventional ultrasonic imaging cannot measure WSS directly, and high-fidelity Computational Fluid Dynamics (CFD) simulations require substantial computational time. Machine learning (ML) methods have demonstrated potential for accelerating hemodynamic evaluation across diverse specific scenarios. Purpose This study develops a machine-learning reduced-order model (ML-ROM) to rapidly predict transient WSS and OSI in the carotid sinus under varying inflow conditions, enabling early assessment of atherosclerotic risk for different physiological conditions. Methods We parameterized a baseline common carotid artery (CCA) waveform to generate 27 training and 8 testing inlet profiles covering sedentary to exercise-trained conditions (Figure 1a-c). These were used in high-resolution CFD simulations using an immersed boundary method (0.1 mm grid, 101 time-steps/cycle) of an MRA-based carotid model (Figure 1d). Proper Orthogonal Decomposition (POD) of the resulting velocity fields extracted dominant modes and temporal coefficients (Figure 1e). A neural network was trained to map the CCA inlet and internal carotid artery (ICA) outlet waveforms to POD coefficients (Figure f), enabling the subsequent reconstruction of velocity fields and WSS (Figure 1g-h). Results The ML-ROM demonstrated strong agreement with CFD results. For the eight test cases, the mean normalized mean absolute error (NMAE) of the predicted velocity fields was 2.8%. A representative result from a sedentary inlet case is shown in Figure 2. The derived WSS also showed excellent agreement, with errors below ±0.08 Pa. Similarly, the OSI was predicted accurately, with differences within ±0.04. The ML-ROM model successfully replicated the spatial distribution of low WSS and high OSI in the carotid sinus, key regions associated with early atherosclerosis. Conclusions The ML-ROM accurately predicts transient carotid hemodynamics across physiological conditions at greatly reduced computational cost, identifying regions associated with early atherosclerosis. In the future, by integrating multiple clinical data, such models are expected to evolve into hemodynamic assessment tools for prediction and diagnosis.