Machine learning in assessing the association between the size and structure of the ascending aortic wall in patients with aortic dilatation of varying severity

Abstract

Aim. To assess the association between pathological ascending aortic (AA) wall changes and its planimetric characteristics in non-syndromic non-familial (sporadic) aneurysm and dilation of the AA.Material and methods. The study included 174 patients with sporadic aneurysms and dilation of the AA, who underwent elective surgery between January 2010 and July 2015 and were divided into 2 groups: patients with AA aneurysm (AAA) (AA diameter >50 mm) and tricuspid aortic valve (AV) with significant aortic stenosis (AS) or regurgitation (AR) (AAA group, n=120), and persons with borderline AA dilatation (AA diameter 45-50 mm), associated with a bicuspid aortic valve (BAV) and significant AS (BD group, n=54). Standard paraclinical investigations and pathological examination of the VA wall were used. Statistical processing was carried out in the SPYDER 4.1.5 environment (Python 3.8), and included univariate correlation analysis, logistic regression analysis, as well as supervised machine learning (ML) methods (support vector machine, k-nearest neighbor method, random forest).Results. Logistic regression revealed positive associations between AA atherosclerosis and age, cystic medial necrosis (CMN) and sinus of Valsalva (SV) diameters. The support vector machine method demonstrated a tendency towards AA expansion at the SV level in individuals with CMN (accuracy, 60,5%), as well as towards expansion of the tubular AA in atherosclerosis (accuracy, 79,2%). During the random forest analysis, the first stage was to construct decision trees to predict three following outcomes: the presence of CMN, atherosclerosis, or normal aortic structure. The model accuracy was 64,2%. Next, the variables "CMN" and "atherosclerosis" were combined, and prediction was made for the outcomes "normal AA wall structure" and "pathological AA wall structure". The model accuracy was 73,5%.Conclusion. The use of ML opens up new opportunities for predicting aortopathy and a patient-centered approach to treatment. In AR, a more aggressive AA intervention is warranted. To predict aortopathies, thoracic aorta diameters indexed to body surface area should not be used. Aortic wall sampling (circular section) followed by a continuous pathological examination may be promising.

Keywords

• aneurysm
• thoracic aorta
• bicuspid aortic valve
• logistic models
• random forest
• risk factors