# Opportunities and Challenges in Deep Learning Methods on Electrocardiogram Data: A Systematic Review

**Authors:** Shenda Hong, Yuxi Zhou, Junyuan Shang, Cao Xiao, Jimeng Sun  
**arXiv:** [2001.01550](https://arxiv.org/abs/2001.01550)  
**Date:** 2019-12-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2001.01550v1/

## Key points

- Objective: Conduct a systematic review of deep learning methods applied to ECG data
- Methods: Comprehensive search on Google Scholar, PubMed, and DBLP for papers published between January 1st, 2010 and September 30th, 2019 using deep learning and ECG-related terms
- Total of 124 papers included in the analysis, with 97 published in the past two years
- Various deep learning architectures used for different ECG analytics tasks
- Common architectures include CNN, RNN, CRNN, AE, GAN, FC, among others
- Some models incorporated traditional expert features or knowledge
- Dataset size varied along with sampling frequency and number of electrode leads
- Literature search process involved two stages: search and selection
- Final list consisted of 124 deep learning papers on ECG

## Summary

Objective: The objective of this study was to conduct a systematic review of deep learning methods applied to Electrocardiogram (ECG) data, focusing on model architecture and application tasks. Methods: A comprehensive search was conducted on Google Scholar, PubMed, and DBLP for papers published between January 1st, 2010 and September 30th, 2019. The search keywords included deep learning and ECG-related terms. The papers were analyzed based on their task, model architecture, and data characteristics. Unresolved challenges and limitations of existing models were also identified. Results: A total of 124 papers were included in the analysis, with 97 papers published in the past two years. Various deep learning architectures were utilized for different ECG analytics tasks such as disease detection/classification, annotation/localization, sleep staging, biometric human identification, denoising, etc. Common architectures included convolutional neural networks (CNN), recurrent neural networks (RNN), combination of CNN and RNN (CRNN), autoencoders (AE), generative adversarial networks (GAN), fully connected neural networks (FC), among others. Some models also incorporated traditional expert features or expert knowledge. The size of datasets varied along with sampling frequency and number of electrode leads. The literature search process involved two stages - search and selection - as shown in Figure 2. To avoid missing relevant papers, the search was expanded to include "anywhere in the article" rather than just the title. Papers that did not use deep learning were manually filtered out. The final list of deep learning papers on ECG consisted of 124 papers, significantly fewer than the initial search results. The analysis of each paper focused on three aspects: task, model architecture, and data characteristics.