Artificial neural networks in prediction of atrial fibrillation in men with coronary artery disease
Aim. The aim of the study was to select, based on mathematical apparatus of artificial neural networks (ANN), the most sen- sitive parameters for creating an ANN model aimed at prediction of atrial fibrillation (AF) in men with coronary artery disease (CАD).Material and Methods. The study focused on data of men from the register of coronary angiography with angiographically proven coronary artery disease: the main group comprised 180 men with AF; the comparison group comprised 713 men of comparable age without AF. The ANN mathematical model, a multilayer perceptron with one hidden layer, was used to assess the risk of AF. The initial group of patients was divided into three samples: the training, test, and control samples.Results. Patients with AF were significantly less likely to be employed in budget organizations (55.0% vs 63.7%, p = 0.040) and more often showed higher (III–IV) heart failure NYHA classes (49.2% vs 21.1%, p < 0.001), higher body mass index (BMI) (30.2 [27.4; 33.2] kg/m2 vs 29.0 [26.1; 32.3] kg/m2, p = 0.002), and higher echocardiographic indices of the left ventricular (LV) myocardial mass (163.7 [144.5; 192.4] g/m2 vs 143.9 [126.1; 169.0] g/m2, p < 0.001), left (25.8 [24.1; 29.1] mm/m2 vs 25.6 [23.9; 27.5] mm/m2, p = 0.020) and right ventricular dimensions, and the left atrial diameter (23.6 [21.7; 25.7] mm/m2 vs 21.1 [19.7; 22.7] mm/m2, p < 0.001). The group of AF patients had higher rate of hemodynamically significant mitral regurgitation (48.2% vs 14.1%, p < 0.001). In this group of patients, the index of aortic root dimensions (7.7 [16.4; 19.0] mm/m2 vs 18.3 [17.8; 20.0] mm/m2, р = 0.002) and LV ejection fraction (EF) were lower (49 [42; 56]% vs 56 [47; 60]%, p < 0.001); coronary calcification (23.2% vs 15.7%, p = 0.024 ) and proximal lesions of the right coronary artery (RCA) (28.3% vs 22.7%, p = 0.025) were detected more often. The final model, which included 10 parameters, had the diagnostic accuracy of 85%, sensitivity of 85%, and specificity of 86%.Conclusion. Atrial fibrillation in men with coronary artery disease can be predicted by ANN model that takes into account the presence of significant mitral regurgitation, extra-budgetary employment, severity of heart failure, coronary calcification, proximal lesion of RCA, BMI, echocardiographic indexes of left heart, aortic root dimensions, and LV EF.