ST-Net: Synthetic ECG tracings for diagnosing various cardiovascular diseases. (August 2020)

• Record Type: Journal Article
• Authors: Deng, Yu; Gao, Zhongquan; Xu, Songhua; Ren, Pengyu; Wen, Yang; Mao, Ying; Li, Zongfang
• Is Part Of: Biomedical signal processing and control. Volume 61(2020)
• Publication Date: 2020-08
• Subjects: Cardiovascular disease -- Electrocardiography -- Signal synthesis -- Deep neural network; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2020.101997
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Diagnosis of various cardiovascular diseases with a single ECG tracing is challenging. ST-Net can diagnose four types of cardiovascular diseases with a single ECG tracing. ST-Net presents a state-of-the-art performance in diagnosing myocardial infarction. The inherent variance in ECG between individuals impact diagnostic accuracy. Abstract: Electrocardiography (ECG) is a prevalent approach to help diagnose cardiovascular disease (CVD) in clinical practice, but it is time-consuming for cardiologists and requires domain knowledge. Therefore, many researchers have attempted to automate that diagnostic procedure and some have developed wearable devices with a single ECG recording to detect CVD such as arrhythmia. A few have discussed feasible methods for wearable devices to increase the accuracy of diagnosing various CVDs, learning structural and morphological features in multiple ECG tracings but making a diagnosis solely with a single tracing. In this paper, we propose the Spark-trace Network (ST-Net) as a solution to the above issue. ST-Net encodes one of 12 real tracings to synthesize 11 new tracings that can capture the features of real ones. Then, ST-Net makes a diagnosis that relies on both real and synthetic tracings. ST-Net surpasses the baseline when classifying four types of CVDs and performs favorably when discriminating between myocardial infarction and normal rhythm, achieving 98.13% accuracy, 98.19% sensitivity, and 98.09% specificity on a five-fold test. Our network outperforms the state-of-the-art when diagnosing up to four types of CVDs on the Physikalisch-Technische Bundesanstalt (PTB) dataset. Additionally, we demonstrated the inherent variance in ECG tracings between individuals by comparing the diagnostic results with the Class-oriented dataset and Subject-oriented dataset.