Machine Learning–Based Patient-Specific Prediction of Pressure Gradients for Patients with Coarctation of the Aorta

Objectives: Although patient-specific numerical modelling of aortic hemodynamics in patients with coarctation of the aorta (CoA) has large potential to supplement current clinical assessment and diagnosis, such methods are time and cost consuming, and require substantial user experience. As an alternative to conventional computational fluid dynamics (CFD), we propose a novel neural network (NN)-based approach that is user friendly and produces results almost instantly. However, since NN-based methods require large amounts of training data, a synthetic patient cohort based on a statistical shape model (SSM) is generated to fill this data gap.

Methods: A total of 159 (49 females) 3D magnetic resonance imaging datasets of patients were used. Based on this real cohort, an SSM of the aorta is built. Using this SSM, a large database of 4,000 unique synthetic stenosed aortic geometries is created. A suitable artificial NN is used to predict the hemodynamic outcome–based patient-specific geometrical and flow data. Hemodynamic prediction performance is evaluated by comparing NN and CFD computed hemodynamic outcome.

Result: Evaluation of the NN's capability to predict the hemodynamic outcome on the test cases showed good agreement between NN- and CFD-based results. For the clinically relevant parameter of pressure gradient across the aorta, a mean error of 6 mm Hg is observed between CFD- and NN-based results. Furthermore, it was shown that the NN's performance significantly benefits from including synthetic cases into the training data as compared with only using real patient data.

Conclusion: This study demonstrates that patient-specific data can be augmented with an SSM to train a NN that is capable of reliably computing aortic hemodynamics in CoA patients. NN based hemodynamic outcome prediction methods hold the potential to replace conventional numerical methods for in silico–based decision support and treatment for CoA patients, which could dramatically improve clinical feasibility of patient-specific modeling workflows.

Publication History

Article published online:
21 February 2021