Detection of Cardiac Arrhythmias Using Neural Network-Based Heart Rate Variability Analysis

The paper discusses the application of neural networks for processing heart rate variability signals, which is necessary for more reliable diagnosis of cardiovascular diseases. The relevance of this task is due to the fact that computerized electrocardiogram processing systems, including neural networks, have firmly entered the practice of clinical research. The paper proposes, with the aim of increasing the reliability of diagnosis, to perform preprocessing of rhythmograms and to calculate a series of integral characteristics of cardiointervalograms using statistical and geometric methods. To solve the problem of diagnosing cardiac arrhythmias, a multilayer perceptron classifier of the form 11-10-4 was created. Preprocessing was performed to form input parameters, including the extraction of peaks and the formation of rhythmograms, as well as the calculation of heart rate variability parameters. The neural network receives 11 values for each ECG (electrocardiogram) signal as input: SDNN, SDANN, SDNNi, SDSD, NN50, pNN50, CV, Mo, AMo, VR, IN. Based on the ECG analysis results, it was classified into one of four classes: normal rhythm, atrial fibrillation, other types of arrhythmias, noise. Two algorithms were used for training: backpropagation and conjugate gradients. Testing the network on real ECGs showed a sufficiently high diagnostic accuracy (sensitivity not less than 85%, specificity 82%). The results of the work can be applied in automated ECG analysis systems, particularly in Holter monitors.