E-177 Machine learning model for the prediction of patient-specific waveforms of blood flowthrough the internal carotid artery. (23rd July 2022)

• Record Type: Journal Article
• Authors: Fillingham, P; Levitt, M; Kurt, M; Lim, D; Federico, E; Keen, J; Aliseda, A
• Is Part Of: Journal of neurointerventional surgery. Volume 14(2022)Supplement 1
• Page Start: A173
• Page End: A173
• Publication Date: 2022-07-23
• Subjects: Nervous system -- Surgery -- Periodicals; Cerebrovascular disease -- Surgery -- Periodicals; 617.48
• DOI: 10.1136/neurintsurg-2022-SNIS.288
• Languages: English
• ISSNs: 1759-8478

Abstract

Abstract : Introduction: Computational fluid dynamics (CFD) is an excellent tool for studying cerebral aneurysms as the formation, growth, and rupture of these aneurysms are heavily influenced by complex hemodynamic variables that cannot be measured in vivo, but can be calculated using CFD. The results of CFD simulations are highly dependent on small changes in the prescribed boundary conditions, however. This makes the use of patient specific information crucial to the accuracy of the results, yet many CFD studies of cerebral aneurysms rely on stereotypical boundary conditions, without concern for important factors for blood flow such as age, weight, smoking history, and high blood pressure. The reliance on standard values has limited the scope, impact and accuracy of CFD study. The gold standard for obtaining patient specific blood flow measurements is endovascular Doppler manometry, which allows for direct measurement of time resolved pressure and velocity data, but requires specialized equipment and is not suited for wide scale cross site study. There is a need to be able to accurately predict patient specific blood flow waveforms without the need for specialized equipment and in a way that allows for retrospective and cross site study. Objective: In this work we develop a machine learning model to predict the blood flow waveform in the internal carotid artery (ICA) based on patient data, trained on state of the arc Doppler manometry measurements. Method: N=35 (train n=30, test n=5) patient specific Doppler manometry ICA waveform data was used to train a three-step machine learning model using only the following data as inputs: sex, age, weight, height, heartrate, ICA side, smoking history, and hypertension history. The three-step process is as follows 1. Prediction of time averaged flowrate through the ICA using a gradient boosted regressions tree. 2. Prediction of waveform shape through a convolutional neural net trained to predict the normalized first eight principal components of each individual waveform. 3. Scaling of predicted waveform shape to match predicted flowrate through numerical solution of the canonical Womersley flow profile of pulsatile flow in a pipe. The results obtained through the model were then compared to standard practice of determining waveforms for CFD boundary conditions obtained by scaling the stereotypical wave form based on vessel diameter. Results: The final flow model is able calculate ICA blood flow waveforms far more accurately than the standard method for prescribing boundary conditions, outperforming for all but one subject. The relative RMSE for centerline velocity from the training set was 0.085 m/s while the RMSE for the test set was 0.099 m/s, (mean velocity in the ICA is ~0.29 m/s) while the RMSE on centerline velocity for the standard methodology was 0.181 m/s. This ~50% reduction in error represents a significant improvement in the prediction of patient specific waveforms. Conclusions: We have successfully developed a machine learning model for predicting ICA blood flow waveforms that drastically improves on the standard practice for prescribing boundary conditions for CFD of cerebral aneurysms. Disclosures: P. Fillingham: None. M. Levitt: None. M. Kurt: None. D. Lim: None. E. Federico: None. J. Keen: None. A. Aliseda: None.