Open AccessNeural network analysis of mortality risk predictors in patients after acute coronary syndrome
Author(s) -
Д. А. Швец,
А. Ю. Карасёв,
М. В. Смоляков,
С. В. Поветкин,
В. И. Вишневский
Publication year - 2020
Publication title -
rossijskij kardiologičeskij žurnal
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.141
H-Index - 14
eISSN - 2618-7620
pISSN - 1560-4071
DOI - 10.15829/1560-4071-2020-3-3645
Subject(s) - medicine , acute coronary syndrome , ejection fraction , cardiology , coronary artery disease , percutaneous coronary intervention , machine learning , heart failure , myocardial infarction , computer science
Aim. To study the possibilities of neural network analysis of clinical and instrumental data to predict the mortality risk in patients after acute coronary syndrome (ACS). Material and methods. The study involved 400 patients after ACS which who observed for 62 months. The criterion for the complicated course of coronary artery disease (CAD) is the cardiovascular death. Group 1 consisted of 310 patients with uncomplicated course of CAD; group 2 — 90 patients with complicated course of CAD. To predict mortality risk, the machine learning method and neural network analysis was used. Machine learning was carried out with the inclusion of clinical, laboratory and instrumental (electrocardiography, echocardiography) parameters (49 in total). To solve the classification problems, two types of neural network architectures were used: Multilayer Perceptron (MLP) and Convolutional Neural Networks (CNN). The ratio in the examples for learning and validation was 340/60. The method of learning with a teacher was used on the available data in which the outcomes were known, and the neural network parameters were adjusted so as to minimize the error. Results. The following factors made the highest contribution to the mortality risk after ACS: age; history of MI and acute cerebrovascular accident; atrial fibrillation, class 2-3 heart failure; no history of percutaneous coronary intervention; stage 3 chronic kidney disease; reduced left ventricle ejection fraction. Most of the deaths occurred in the 2nd and 4th years of follow-up, which may be due to the low effectiveness of secondary prevention of CAD. CNN architecture had higher accuracy (sensitivity — 68%; specificity — 84%; area under curve=0,74). An advantage of CNN is its ability to analyze patterns over time using recurrent neural networks. Conclusion. Neural network analysis of clinical, laboratory and instrumental data allows configuring network parameters for mortality risk prediction. CNN predicts 5-year mortality risk after ACS with a sensitivity of 68% and a specificity of 84%.