A machine learning approach for predicting descending thoracic aortic diameter.

Background: To establish models for predicting descending thoracic aortic diameters and provide evidence for selecting the size of the stent graft for TBAD patients. Methods: A total of 200 candidates without severe deformation of aorta were included. CTA information was collected and 3D reconstructed. In the reconstructed CTA, a total of 12 cross-sections of peripheral vessels were made perpendicular to the axis of flow of the aorta. Parameters of the cross sections and basic clinical characteristics were used for prediction. The data was randomly split into the training set and the test set in an 8:2 ratio. To fully describe diameters of descending thoracic aorta, three predicted points were set based quadrisection, and a total of 12 models at three predicted points were established using four algorithms included linear regression (LR), support vector machine (SVM), Extra-Tree regression (ETR) and random forest regression (RFR). The performance of models was evaluated by mean square error (MSE) of the prediction value, and the ranking of feature importance was given by Shapley value. After modeling, prognosis of five TEVAR cases and stent oversizing were compared. Results: We identified a series of parameters which affect the diameter of descending thoracic aorta, including age, hypertension, the area of proximal edge of superior mesenteric artery, etc. Among four predictive models, all the MSEs of SVM models at three different predicted position were less than 2 mm Conclusion: The predictive models established by machine learning revealed the relationship between basic characteristics and diameters of different segment of descending aorta, which help to provide evidence for selecting the matching distal size of the stent for TBAD patients, thereby reducing the incidence of TEVAR complications.