Document resource
Introduction Transcranial Doppler (TCD) ultrasonography is the clinical standard for vasospasm surveillance following aneurysmal subarachnoid hemorrhage (aSAH), yet its interpretation depends on operator experience and patient-specific baselines that are often unavailable at initial presentation. We developed and are currently validating a machine learning pipeline capable of predicting patient-specific internal carotid artery (ICA) centerline blood flow velocity waveforms from demographic and physiologic parameters alone, establishing a computational baseline against which vasospasm-induced deviations can be detected.Materials & Methods A pipeline integrating XGBoost regression and a neural network was trained on TCD-derived ICA velocity waveforms from a reference cohort. Singular value decomposition (SVD) and principal component analysis reduced waveform dimensionality, enabling separate prediction of mean volumetric flow rate and normalized waveform shape. Full velocity waveforms were reconstructed via Womersley flow theory using generalized vessel radius and patient-specific cardiac period, hematocrit-derived viscosity, and demographic inputs (age, sex, height, weight, laterality, hypertension status, smoking history). Model outputs were validated in 9 patients across 19 ICA TCD acquisitions. Raw TCD signals were processed using FFT-based cycle detection and interpolated to 101 normalized time points for direct pointwise comparison with ML-predicted waveforms. Agreement was assessed using Pearson correlation, RMSE, and Bland-Altman analysis of peak systolic (PSV), mean (MFV), and end-diastolic velocities (EDV).Results The ML pipeline demonstrated strong waveform morphology agreement with cycle-averaged TCD profiles, with a mean Pearson r of 0.88 ± 0.11 (range 0.62-0.98; all p < 0.001) and mean RMSE of 15.0 ± 8.6 cm/s. Systematic amplitude underestimation of approximately 25% was observed across all metrics: measured versus predicted PSV was 64.3 ± 13.1 versus 48.2 ± 5.3 cm/s, MFV was 42.2 ± 10.0 versus 31.7 ± 3.6 cm/s, and EDV was 27.0 ± 9.9 versus 19.6 ± 3.6 cm/s. Bland-Altman analysis revealed a mean bias of −10.4 cm/s with limits of agreement of −33.5 to 12.6 cm/s.Conclusions This pipeline accurately predicts ICA waveform morphology from demographics alone, with consistent amplitude underestimation hypothesized to reflect model training on waveforms acquired under anesthesia or sedation, conditions known to suppress cardiac output and cerebral flow velocities. Recalibration using awake TCD data is planned and expected to resolve this systematic bias. This framework offers a novel non-invasive computational baseline for vasospasm surveillance and may reduce reliance on serial TCD comparisons in centers with limited neurosonology expertise.Disclosures J. Lim: None. M. Fung: None.Abstract P-018 Figure 1