Exploring the Association between Clinical Factors and Aortic Morphometry Using Neural Networks.

Introduction: aortic morphometry varies across its segments and may reflect diverse clinical influences. Objective: to identify which segments of the aorta show the greatest association with clinical variables using neural networks. Materials and methods: an analytical and cross-sectional study was conducted with 801 adults from the Harvard Dataverse repository (2018–2019) who underwent non-contrast chest CT. Aortic diameters from the sinus of Valsalva to the abdominal aorta, measured using artificial intelligence, were assessed, along with 20 clinical variables. A multilayer neural network was applied as a nonlinear statistical analysis tool. Results: in the performance analysis of the neural network model, the diameter with the lowest relative error was that of the middle descending aorta, whose relative error was 0.356 in the testing phase, obtaining a coefficient of determination of 0.613, indicating that the model explains 61.3% of the variability in diameter. Its most influential variables were age (importance of 0.168), creatinine (0.106), AST (0.082), and potassium (0.071). The mean square error was 0.324 in the training phase and 0.356 in the testing phase. Conclusions: the neural network model reveals a stronger nonlinear association between clinical variables and the diameter of the mid-descending aorta, highlighting its sensitivity to hemodynamic and metabo-lic influences and reinforcing its value as a structural marker in cardiovascular disease assessment.