Content area
Full Text
Introduction
Electrocardiograms (ECGs) are electrophysiological signals produced by a potential difference, as measured by electrodes placed on the surface of the human skin. They are the most common, convenient, and inexpensive non-invasive tests of the cardiovascular system. Cardiologists leverage ECGs to diagnose various abnormal cardiac activity states by analyzing the corresponding waveform changes, with timely intervention known to improve the health of patients and the survival prognosis. Wearable ECG devices, which have user-friendly designs and compact sizes, can be worn by the subject 24/7, and the collected recordings can be automatically uploaded to a server1. Compared with hospital-based ECG machines, which collect ECGs over just a few minutes, wearable devices are more likely to collect non-rhythmic or transient data, such as premature ventricular beats, premature atrial beats, and other precious arrhythmia signals, in a few days of collection time. Typically, the non-rhythmic or transient class is also the rare class. For deep learning approaches, collecting more data on rare classes is a fundamental way to improve the classification performance of models for rare classes. At the same time, the prolonged use means that wearable ECG devices can collect abundant data, aligning with the advantages and requirements of deep learning methods for mining information to produce intelligent diagnostic models that enable clinical decision-making. Cardiovascular disease has far surpassed cancer as the leading cause of death worldwide2, and with increasing life expectancy, the aging of the population is gradually becoming a major problem in the field of healthcare across the world. To reduce the workload of cardiologists and realize intelligent monitoring and telemedicine, the need for clinically applicable ECG diagnostic approaches such as the one shown in Fig. 1a is increasingly urgent in modern society.
[See PDF for image]
Fig. 1
ECG diagnostic framework, dataset information, and mutually exclusive–symbiotic correlation of hierarchical multiple labels.
a One-stop intelligent diagnostic framework for 12-lead wearable ECG. b Example of hierarchical labels in our ECG database. c Conditional probability matrix representing the concurrence of multiple labels. d Mutually exclusive correlation matrix of multiple labels computed from the conditional probability matrix. e Mutually symbiotic correlation matrix of hierarchical labels computed from the conditional probability matrix. f Motivation and schematic diagram for local softmax. The subfigure shows a...