With the development of atrial fibrillation (AF), patients with acute coronary syndrome (ACS) are characterized by a twofold increase in the 30-day mortality compared with patients with sinus rhythm. In this regard, there is great interest in developing models of risk stratification to identify adverse outcomes in these patients with a view to more careful monitoring of patients in this group.Material and methods. For the construction of predictive models, a statistical method was used for the classification trees and, the procedure for neural networks implemented in the STATISTICA package. For the construction of prognostic models, a sample was used, consisting of 201 patients with and without fatal outcome; condition of each patient was described by 42 quantitative and qualitative clinical indices. Each patient belonged to one of 3 groups according to the type of AF: new-onset AF in ACS patient, paroxysmal AF, documented in an anamnesis before the episode of ACS and the constant or persistent form of AF.Results. To determine predictors of models predicting the possible fatal outcome of a patient, the Spearman correlation coefficient was used. Examination of the correlations for each of the 3 groups separately allowed to reveal clinical indicators for each group – predictors of predictive models with predominantly moderate correlations to the categorical variable “lethal outcome”. After analyzing the prognostic ability of the developed models, a software module was created in the Microsoft Visual C # 2015 programming environment to determine lethal outcome possibility in patients with ACS in the presence of AF using classification trees and neural networks.Conclusion. It is shown that for patients with ACS in the presence of AF, it is possible to construct mathematically based prognostic models that can reliably predict the lethal outcome possibility in patients based on actual values of clinical indices. In this case, clinical indicators can be both quantitative and qualitative (categorical), breaking patients into certain categories. Similar applications, unlike risk scales, are mathematically justified and can form the basis of systems for supporting decision-making.
Classification Atrial Fibrillation Using Stacked Autoencoders Neural Networks
