Файл:Deep Learning in information analysis of electrocardiogram signals for disease diagnostics Ignatov2015DeepLearningThesis.pdf

Ignatov Andrey Dmitrievich

Scientific advisor: Senior Researcher at CC RAS, DSc Konstantin V. Vorontsov

The Ministry of Education and Science of The Russian Federation Moscow Institute of Physics and Technology (State University) Faculty of Control and Applied Mathematics Department of «Intelligent Systems» under Dorodnicyn Computing Center of RAS

Abstract

With a rapid development of technology, new advanced methods for diseases detection become available. Thus, one possible way to diagnose the internal organs diseases is based on the information analysis of electrocardiogram signals. Since these signals contain information about various physiological processes taking place in the human body, it is possible to use them for health status evaluation. In this work, we propose a deep learning-based method for detecting various human diseases on the heart rate variability (HRV) data acquired from the electrocardiograms. We consider supervised and unsupervised approaches for training the models. In the first case, we use Convolutional Neural Networks and train them on the labeled HRV dataset. In the second case, we learn features from HRV data using Stacked Autoencoders and Restricted Boltzmann Machines, and then build a classifier on the obtained ones. The accuracy of the proposed models is evaluated on the dataset of labeled electrocardiograms with diagnosis in 14 different diseases. The final model demonstrates high performance with an average accuracy over 90% for most of the diseases. The source code of the models is written in Lua and available publicly.

Keywords: Deep Learning, Convolutional Neural Network, Stacked Autoencoder, Boltzmann Machine, electrocardiogram, heart rate variability.