Own experience of neural analysis in predicting long-term survival of patients with chronic heart failure

Abstract

The aim – predicting the risk of fatal outcome within 3 years in patients with CHF and reduced left ventricular ejection fraction (LVEF) with neural network.
Materials and methods. The retrospective analysis of 490 medical histories of patients who were hospitalized in the heart failure department between 2011 and 2018 years with CHF II–IV functional class according to NYHA with LVEF ≤ 40 % on the background of coronary heart disease was conducted. Patients with clinical signs of heart failure II NYHA functional class – 455 (92.8 %) patients and with clinical signs of CHF III NYHA functional class – 35 (7.2 %) patients. The analysis was conducted for 490 patients: 228 (46.5 %) patients had a fatal event within three years, 262 (53.5 %) patients survived three years. 
Results and discussion. Factor features was selected for building a neural network. The information about 8 factor characteristics (ACE inhibitor (X1), atrium fibrillation (X2), renal dysfunction (X3), age (X4), arterial pressure (X5), (X6), LV diastolic volume index ( X7), LV myocardial mass index (X8)) was used for building of neural network models. Three-layer Multiplayer Perceptron (MLP) model with one hidden layer (logistic activation function) was built in the final. This neural network is a type of multi-layer neural network (multi-network perceptron). Each neuron of this system uses a nonlinear activation function. The sensitivity and specificity of this neural network model was evaluated. The education of this neural network did by the method of backpropagation. The quality of the classification of this model was evaluated with the test set of cases. The area under the curve of the operating characteristics of the neural MLP model for predicting the risk of a fatal event exceeds that area of the 9-factor logistic regression model (p<0.001).
Conclusions. The use of neural network analysis increase the accuracy of predicting a fatal outcome for a 3-year period compared to a multifactor logistic regression model. According to the construction of the neural network MLP model, a strong connection of the risk of a fatal event over the course of 3 years with the presence of atrium fibrillation, renal dysfunction, age, LV diastolic volume index, LV myocardial mass index and ACE inhibitor in the treatment was revealed. The quality of the neural network MLP model is high (area under the curve AUC=0.842). Youden Index (Y_crit=0.3049), the sensitivity of the MLP model is 76.8 % (95 % CI 70.7–82.1 %), the specificity of the model is 81.3 % (95 % CI 76.0–85.8 %), predictive significance +PV – 78.1 % (95 % CI 73.3–82.3 %), predictive significance –PV – 80.1 % (95 % CI 75.9–83.7 %). The accuracy of predicting the risk of a fatal outcome within 3 years for the neural network MLP model is higher than for the 9-factor logistic regression model (р<0.001).