Document resource
Introduction The ability to predict patient-specific blood flow waveforms from demographic data has significant implications for both computational fluid dynamics (CFD) boundary condition prescription and non-invasive hemodynamic surveillance. Our group has been working to validate a machine learning model that predicts internal carotid artery (ICA) centerline velocity waveforms from patient demographics, trained on endovascular Doppler manometry data. However, clinically relevant hemodynamic phenomena — including vasospasm following subarachnoid hemorrhage, aneurysm growth, and post-treatment flow redistribution — occur across multiple intracranial vessel segments beyond the ICA. The current model is limited by its single-vessel training set and relatively small cohort (N=35). To address this, we have developed a scalable TCD data extraction pipeline and propose expanding the model architecture to encompass the middle cerebral artery (MCA), anterior cerebral artery (ACA), and basilar artery.Materials and Method The three-step prediction pipeline will be retrained for each target vessel segment using expanded training data drawn from TCD-derived cycle-averaged waveforms from multi-site cohorts, supplementing the existing (N=35) training set. Vessel-specific SVD decomposition of training waveforms will be performed, and the dense neural network (ICADNN) retrained to predict the 6 leading normalized principal components of waveform shape for each segment independently. The XGBoost volumetric flow rate regressor will be similarly retrained with vessel identity incorporated as a categorical input, with Womersley reconstruction parameters adjusted for segment-specific radius and flow characteristics. Model performance will be benchmarked against the original ICA model using RMSE on centerline velocity and compared against the stereotypical waveform scaling standard, consistent with the methodology of Fillingham et al.Results Preliminary validation of the existing ICA model against non-invasive TCD measurements demonstrated strong waveform morphology agreement and consistent systematic amplitude underestimation, attributed to model training on waveforms acquired under anesthesia rather than in the awake state. These findings establish proof-of-concept for TCD-based validation of the pipeline and motivate the amplitude recalibration and multi-vessel expansion proposed here. Full model retraining and validation across MCA, ACA, and basilar artery segments are ongoing, with vessel-specific performance metrics to be reported.Conclusions Preliminary ICA validation supports the feasibility of the ML waveform prediction framework, and expansion to multiple intracranial vessel segments using larger multi-site TCD cohorts would establish a comprehensive tool for patient-specific hemodynamic characterization across the cerebrovascular tree. This has direct applications in CFD boundary condition prescription for aneurysm modeling, non-invasive vasospasm surveillance following subarachnoid hemorrhage, and longitudinal hemodynamic monitoring in patients with intracranial vascular disease.Disclosures J. Lim: None. M. Fung: None.