Convolutional neural network for automatic detection of arrhythmias from ECG tracings

Abstract

Cardiovascular disorders are the leading cause of death in the world, and the electrocardiogram (ECG) is the main tool used to diagnose them. Ever since the shift from analog to digital ECG signals, automatic signal analysis techniques have become increasingly important in the medical diagnostic process. Their role in clinical practice is, to date, limited by the performance of existing models. Deep neural networks (DNNs) are models that are composed of a series of features that learn to perform tasks based on given examples.

This technology has recently been successful in a variety of fields, including biomedical image classification, and there are high expectations for how it could improve medical practice in numerous scenarios.

In this paper, a neural network model for implementing a binary classification algorithm will be presented, which is able to determine to which of its classes an instance of ECG signals belongs: regular signal or affected by arrhythmia.

For this purpose, the classifier was trained on 30-minute ECG sequences, derived from the public MIT-BIH Arrhythmia database.

The network far exceeds the common medical procedure aimed at recognizing

abnormalities in the rhythm of the ECG signal and compares well with deep learning techniques developed for the same purpose, reporting satisfactory performance. These results indicate that deep neural networks are suitable for the analysis of ECG signals; the use of this technology could set a new standard in clinical practice.

Objectives

The proposed solution involves a system that can predict the value of heart rate through a neural network. Training of the model is done from video signals, the subsequences of which are processed to build spatiotemporal maps that will be the input to the network. A further use of the BioDrone is to predict cardiac arrhythmias (LBBB, RBBB, PVC) that occur through alterations in the ECG signal.

Methodology

Various neural network architectures (ResNet50, GRU, Transformer, GAN) were tested. The use of the BioDrone for cardiac arrhythmia prediction involved training a convolutional neural network for classification purposes.

Conclusions

Teoresi’s expertise in the area of data analysis has enabled the identification of an efficient preprocessing strategy in both the ECG signal, through segmentation and augmentation, and in the video signal, in the creation of spatiotemporal maps that greatly facilitate data manipulation and reduce model training time. Expertise in Artificial Intelligence has enabled the development of classification and prediction models in line with the state of the art in this area.

Future developments

The neural network used could be exploited for different other types of applications. Possible future developments could be based on training the network through real life videos for extrapolation of clinical parameters (Health Rate) for the purpose of preventive diagnostics.