Deep learning algorithms for electrocardiogram classification

• Grant number: 19/25153-0
• Support Opportunities: Regular Research Grants
• Start date: August 01, 2021
• End date: January 31, 2024
• Field of knowledge: Engineering - Biomedical Engineering - Bioengineering
• Agreement: CONFAP - National Council of State Research Support Foundations
• Principal Investigator: Marco Antonio Gutierrez
• Grantee: Marco Antonio Gutierrez
• Host Institution: Instituto do Coração Professor Euryclides de Jesus Zerbini (INCOR). Hospital das Clínicas da Faculdade de Medicina da USP (HCFMUSP). Secretaria da Saúde (São Paulo - Estado). São Paulo , SP, Brazil
• Associated researchers: Baldoino Fonseca dos Santos Neto ; Carlos Alberto Pastore ; Idágene Aparecida Cestari ; José Eduardo Krieger ; Márcio de Medeiros Ribeiro ; Marina de Fátima de Sá Rebelo ; Nelson Samesima ; Ramon Alfredo Moreno
• Associated scholarship(s): 21/12935-0 - Classification of digitized ECG signals using deep learning techniques, BP.TT

Abstract

The use of telemedicine systems allows the spreading of cardiac healthcare delivery to populations in poor regions and remote areas. This is especially important in Brazil that is continental in size and with millions of patients using the public health system. This project aims to develop a tool for analysis and classification of electrocardiogram signals (EKG) as normal or abnormal. The set of algorithms to be developed may extend the classifiers to include categories to allow the diagnosis of arrhythmias and acute myocardial infarction. Since multiple disorders can be present concurrently on a single ECG, we will explore multi-label classification approaches. The training of the classifiers will use the Heart Institute database (InCor-HC-FMUSP), including more than 200,000 ECG records and associated reports. In addition, we will take into consideration the clinical and socio-demographic variables of the individuals, which may potentially influence the diagnosis. For feature extraction, we will consider supervised deep learning techniques. Initially, we will explore variants of convolutional neural networks to extract information directly from electrocardiographic tracing images. Then, in order to take into account temporal signal information, we will develop a tool for converting images into digital signals and explore variants of recurrent neural networks. In addition, we will consider a classification approach based on both manually extracted features of digital signals and features learned by deep learning techniques. (AU)