PaperlayerView exact revision on arXiv

Source-linked AI summary

Opportunities and Challenges of Deep Learning Methods for Electrocardiogram Data: A Systematic Review

Shenda Hong, Yuxi Zhou, Junyuan Shang, Cao Xiao, Jimeng Sun

arXiv:2001.01550v3eess.SPcs.CVcs.LG

TL;DR

Deep learning for ECG has produced promising results, but the field lacked a systematic review spanning both model architectures and application tasks. This paper reviews the literature across models, tasks, and data, finding generally better performance than traditional methods while identifying unresolved challenges.

  • Problem

    Existing reviews did not systematically cover deep learning methods for ECG from both model-architecture and application-task perspectives.

  • Method

    The authors systematically reviewed ECG deep-learning studies through a four-stage selection process and analyzed them by tasks, models, and data.

  • Results

    Deep learning methods generally achieved better performance than traditional ECG methods across the reviewed literature.

  • Takeaways & Limitations

    The review provides a systematic overview of deep learning methods for ECG applications and identifies potential future research opportunities.

  • Takeaways & Limitations

    ECG studies lack standardized collection procedures, making fair comparison across datasets difficult; high-quality data and annotations are also difficult to acquire.

Abstract

from arXiv · show

Background:The electrocardiogram (ECG) is one of the most commonly used diagnostic tools in medicine and healthcare. Deep learning methods have achieved promising results on predictive healthcare tasks using ECG signals. Objective:This paper presents a systematic review of deep learning methods for ECG data from both modeling and application perspectives. Methods:We extracted papers that applied deep learning (deep neural network) models to ECG data that were published between Jan. 1st of 2010 and Feb. 29th of 2020 from Google Scholar, PubMed, and the DBLP. We then analyzed each article according to three factors: tasks, models, and data. Finally, we discuss open challenges and unsolved problems in this area. Results: The total number of papers extracted was 191. Among these papers, 108 were published after 2019. Different deep learning architectures have been used in various ECG analytics tasks, such as disease detection/classification, annotation/localization, sleep staging, biometric human identification, and denoising. Conclusion: The number of works on deep learning for ECG data has grown explosively in recent years. Such works have achieved accuracy comparable to that of traditional feature-based approaches and ensembles of multiple approaches can achieve even better results. Specifically, we found that a hybrid architecture of a convolutional neural network and recurrent neural network ensemble using expert features yields the best results. However, there are some new challenges and problems related to interpretability, scalability, and efficiency that must be addressed. Furthermore, it is also worth investigating new applications from the perspectives of datasets and methods. Significance: This paper summarizes existing deep learning research using ECG data from multiple perspectives and highlights existing challenges and problems to identify potential future research directions.

1. Introduction

ECG is a widely used non-invasive cardiac diagnostic tool, but traditional automated analysis depends on expert-engineered features limited by data quality and human knowledge. Deep learning removes explicit feature extraction and has shown strong performance across ECG tasks, motivating a systematic review.

  • ECG records cardiac physiological activity and supports diagnosis of abnormalities including PAC, PVC, AF, MI, and CHF.
  • Traditional ECG analysis uses expert-engineered statistical, frequency-domain, and time-domain features before applying decision rules or machine learning.These features remain limited by data quality and human expert knowledge.
  • Deep learning automatically and implicitly extracts features from ECG data without an explicit human-engineered feature extraction step.
  • Deep learning features were reported as more informative than expert features, with superior performance on tasks including disease detection and sleep staging.
  • Previous reviews did not systematically examine deep learning for ECG across model architectures and application tasks, prompting this review.

2. Method

The review systematically searched and screened peer-reviewed deep-learning studies using ECG data, then analyzed included papers by tasks, models, and data. Its selection process used four stages and independent coarse-to-fine eligibility assessment.

  • The search covered Google Scholar, PubMed, and DBLP for peer-reviewed deep-learning studies using ECG data published from January 1, 2010, to February 29, 2020.
  • The selection framework comprised identification, screening, eligibility, and inclusion stages.
  • 1,621 papers were identified and 1,224 remained after duplicate removal; two independent reviewers then assessed eligibility using title-and-abstract and full-text screening.
  • Eligibility exclusions covered non-English papers, studies not focused on ECG data, studies without deep-learning methods, and studies lacking quantitative evaluations.
  • Included papers were analyzed by application task, deep-model architecture and expert-feature integration, and dataset characteristics.
  • The review summarized papers from model and task perspectives and discussed challenges that existing models cannot handle well.

3. Results

The survey included 191 papers and examined their tasks, models, and data. Publication activity was concentrated after 2019 and within the medical information and biomedical engineering community.

  • 191 papers were included in the survey and analyzed according to tasks, models, and data.
  • 108 papers, approximately 57%, were published after 2019.
  • 112 papers came from the medical information and biomedical engineering community, compared with 25, approximately 13%, from artificial intelligence and data mining.

3.1. Task

The review covers a broad range of ECG tasks, including disease classification, waveform localization and annotation, sleep-related analysis, biometric identification, denoising, and other predictive or signal-processing applications.

  • Disease Detection and Classification: Disease detection maps input ECG data through neural-network layers to disease targets, supporting binary, multi-class, and multi-task classification.Examples include arrhythmia, atrial fibrillation, cognitive stress, decompensation, hypertension, sleep apnea, and diabetes detection.
  • Localization and Annotation: Deep learning methods localize clinically relevant ECG events and annotate waveform components such as fetal QRS complexes, P-waves, and T-waves.These tasks support identifying arrhythmia-related structures and locating origins or exits of cardiac abnormalities.
  • Sleep Staging: Sleep-related applications include obstructive sleep apnea detection, multi-class OSA-with-hypopnea classification, and sleep-stage identification from physiological signals.Reported architectures include sparse autoencoders, hidden Markov models, CNNs, and LSTMs.
  • Human Identification: Biometric studies use ECG signals for automatic human identification and authentication across multiple datasets and architectures.Approaches include CNNs, residual networks, PCANet, Eigen ECG networks, and bidirectional GRUs.
  • Denoising and Other Applications: Denoising methods address acquisition noise using encoder-decoder CNNs, fully convolutional denoising autoencoders, recurrent autoencoders, GANs, and wavelet-CNN combinations.The review also covers CNN, LSTM, and HMM combinations for other applications, including rhythm classification, risk prediction, waveform segmentation, and signal generation.

3.2. Models

The review covers CNNs, RNNs, CRNNs, autoencoders, GANs, fully connected networks, and other architectures for ECG analysis. Models increasingly combine learned representations with expert features and specialized strategies for sequential, noisy, imbalanced, and variable-length ECG data.

  • CNN: CNNs extract hierarchical ECG patterns with limited preprocessing; 1D CNNs operate on temporal signals, whereas 2D CNNs operate on transformed or image-like ECG representations.2D CNNs may use matrices derived from entropy, co-occurrence, morphology, RR intervals, or beat-to-beat correlations.
  • RNN: RNNs model ECG sequences by updating hidden states over time, making them suitable for temporal dependencies and inputs of varying lengths.GRU, LSTM, and BiLSTM variants address vanishing-gradient problems in classical RNNs.
  • CRNN: CRNNs combine CNN feature extraction with RNN temporal summarization, making them suitable for long, variable-length, multichannel ECG signals.CNNs extract local features, and RNNs summarize them across time into global features.
  • AE: Autoencoders learn reduced representations by encoding inputs and reconstructing them, while stacked, sparse, and contractive variants are widely used for ECG denoising.ECG denoising addresses baseline wandering, electrode contact noise, and motion artifacts; CNNs, RNNs, and CRNNs can serve as encoder and decoder modules.
  • GAN and expert-feature methods: GANs address ECG data imbalance through augmentation and have also been applied to denoising, while ensembles combining expert and raw morphological features outperform individual models.Expert features require additional extraction effort but can substantially benefit deep learning methods.
  • FC & Others: The reviewed architectures also include fully connected networks, restricted Boltzmann machines, U-nets, and other computer-vision-inspired models for classification, localization, and related tasks.Fully connected models are particularly used for extremely short ECG sequences, while modified U-nets support variable-length classification and R-peak detection.

3.3. Data

The reviewed ECG datasets vary substantially in leads, duration, source devices, subject backgrounds, and annotation types. Open-source datasets dominate, but inconsistent collection procedures and limited long-term annotated data constrain fair comparison and future progress.

  • Dataset availability: 150 of 191 reviewed works used open-source datasets, supporting follow-up research and reproduction.The review summarizes frequently used databases, including MIT-BIH, PhysioNet, PTB, and China Physiological Signal Challenge datasets.
  • Dataset heterogeneity: Databases differ in device source, lead count, recording duration, subject background, and annotation type, making results difficult to compare fairly across studies.High-quality data and expert annotations are difficult to acquire, and many studies still use the MIT-BIH Arrhythmia Database collected over 40 years ago.
  • Recording duration: Short-term ECG is cheaper and easier to collect, whereas long-term ECG can detect diseases with intermittent symptoms such as paroxysmal VF and AF.The review identifies a need for a new high-quality long-term ECG dataset with annotations to stimulate further studies.
  • Frequently used databases: The MIT-BIH Arrhythmia Database contains 48 half-hour records from 47 subjects and is fully annotated with beat-level and rhythm-level diagnoses.Each sequence has 11-bit resolution over a 10 mV range and a 360 Hz sampling frequency.
  • Frequently used databases: The PhysioNet Computing in Cardiology Challenge 2017 dataset contains 8,528 single-lead recordings lasting 9 seconds to just over 60 seconds, sampled at 300 Hz.The healthcare-device dataset includes normal, AF, other, and noise recordings, plus 3,658 private test recordings.
  • Frequently used databases: The PTB Diagnostic ECG Database contains 549 15-channel records from 290 subjects, including subjects with eight heart-disease types and healthy controls.Its sampling rate reaches 10 kHz.

4.2. Interpretability

Deep learning models are difficult to interpret because of their many parameters and complex architectures, a problem that is especially consequential for medical diagnosis. The review identifies surrogate-model and intrinsically interpretable architectures as two directions for addressing this challenge.

  • Challenge: Deep learning models are often black boxes, and unexplained diagnoses are unacceptable to medical experts.The review attributes this difficulty to numerous parameters and complex model architectures.
  • Existing approaches: Existing ECG approaches improve interpretability by adding expert features, visualizing multilevel attention weights, or generating attribution maps and lower-dimensional embeddings.These methods provide partial interpretation rather than fully resolving the black-box problem.
  • Future directions: One research direction uses a separate simple interpretable model to approximate a black-box model’s predictions and explain its results.This approach separates predictive modeling from interpretation.
  • Future directions: A second direction constructs interpretable deep models directly by borrowing tree-based connection concepts or adding attention mechanisms to hidden layers.The review presents attention as a mechanism that can be more readily understood by humans.

4. Discussion of Opportunities and Challenges

The review identifies unresolved challenges in ECG deep learning involving data comparability, deployment efficiency, integration with expert knowledge, imbalanced labels, multimodal inputs, and emerging applications. It also highlights opportunities for compression, hybrid feature use, multimodal modeling, and broader interdisciplinary studies.

  • 4.1. Data Collection: Inconsistent ECG collection procedures make fair comparisons between datasets difficult, while limited high-quality annotations encourage reuse of older databases.Studies vary in lead counts, recording durations, sources, and other collection characteristics; many rely on the MIT-BIH Arrhythmia Database collected over 40 years ago.
  • 4.3. Efficiency: Large deep models are difficult to deploy on portable healthcare devices, motivating compression methods such as distillation, quantization, and weight sharing.Knowledge distillation can transform large models into simpler ones with a minor decrease in accuracy.
  • 4.4. Integration with Traditional Methods: End-to-end training makes deep models difficult to integrate with traditional expert-feature-based methods after training.Proposed directions include designing architectures around expert knowledge or extracting latent embeddings for combination with expert features and traditional machine learning.
  • 4.2. Imbalanced Data: Rare but important disease labels create imbalanced datasets that are difficult for parameter-rich models to learn from effectively.Suggested responses include augmentation with synthetic data, specialized losses such as focal loss, and few-shot learning.
  • 4.6. Multimodal Data: Asynchronous multimodal measurements from ECG and other vital signs create an opportunity for models that jointly use heterogeneous data.Temperature, respiratory rate, and blood pressure may be collected alongside ECG but have different timelines and sampling frequencies.
  • 4.7. New Applications: Interdisciplinary applications include safe-driving assessment, emotion detection, mammalian ECG analysis, and age or gender estimation, but require adequate data support.These examples extend ECG deep learning beyond the principal healthcare analytics tasks reviewed.

5. Conclusion

The paper systematically reviews deep learning methods for ECG data across models, data, and tasks, finding generally better performance than traditional methods while identifying unresolved challenges and future opportunities.

  • 5. Conclusion: The review provides a systematic overview of ECG deep learning methods for real applications and highlights unresolved challenges and future research opportunities.Its analysis covers models, data, and tasks.
Loading 2001.01550v3…