Abstract:The prevalence of cardiovascular disease (CVD) has surged in recent years, making it the foremost cause of mortality among humans. The Electrocardiogram (ECG), being one of the pivotal diagnostic tools for cardiovascular diseases, is increasingly gaining prominence in the field of machine learning. However, prevailing neural network models frequently disregard the spatial dimension features inherent in ECG signals. In this paper, we propose an ECG autoencoder network architecture incorporating low-rank attention (LRA-autoencoder). It is designed to capture potential spatial features of ECG signals by interpreting the signals from a spatial perspective and extracting correlations between different signal points. Additionally, the low-rank attention block (LRA-block) obtains spatial features of electrocardiogram signals through singular value decomposition, and then assigns these spatial features as weights to the electrocardiogram signals, thereby enhancing the differentiation of features among different categories. Finally, we utilize the ResNet-18 network classifier to assess the performance of the LRA-autoencoder on both the MIT-BIH Arrhythmia and PhysioNet Challenge 2017 datasets. The experimental results reveal that the proposed method demonstrates superior classification performance. The mean accuracy on the MIT-BIH Arrhythmia dataset is as high as 0.997, and the mean accuracy and -score on the PhysioNet Challenge 2017 dataset are 0.850 and 0.843.

What problem does this paper attempt to address?

The main problem this paper attempts to address is improving the classification performance of electrocardiogram (ECG) signals in the detection and diagnosis of cardiovascular diseases. Specifically, existing neural network models often overlook the spatial dimension features of ECG signals, leading to inaccurate feature extraction and affecting classification results. Therefore, this paper proposes an autoencoder network architecture based on a low-rank attention mechanism (LRA-autoencoder), aiming to capture the latent features of ECG signals from a spatial perspective and enhance the feature distinction between different categories through singular value decomposition. ### Main Contributions: 1. **New Perspective**: This paper provides a new perspective to understand the correlation between ECG signal dimensions and spatial dimensions, effectively extracting the spatial dimension features of ECG signals and seamlessly integrating them into ECG signal classification applications. 2. **Performance Improvement**: By incorporating the correlation of ECG signal dimensions into the autoencoder architecture and optimizing data dimensions, the LRA-autoencoder model demonstrates superior performance compared to other pure autoencoder models in experimental results. ### Problems Addressed: - **Limitations of Existing Models**: Existing neural network models often overlook the spatial dimension features of ECG signals, leading to inaccurate feature extraction. - **Classification Performance Improvement**: By introducing a low-rank attention mechanism, the proposed LRA-autoencoder model can more accurately extract the features of ECG signals, significantly improving classification performance. ### Experimental Results: - Achieved an average accuracy of 0.997 on the MIT-BIH Arrhythmia dataset. - Achieved an average accuracy and F1 score of 0.850 and 0.843, respectively, on the PhysioNet Challenge 2017 dataset. In summary, this paper addresses the issue of existing models overlooking spatial dimension features when processing ECG signals by introducing a low-rank attention mechanism, thereby significantly improving the accuracy of cardiovascular disease detection and diagnosis.

ECG autoencoder based on low-rank attention

HADLN: Hybrid Attention-Based Deep Learning Network for Automated Arrhythmia Classification

Multi-label Classification of Abnormalities in 12-Lead ECG Using Deep Learning.

DeepECG: Image-based Electrocardiogram Interpretation with Deep Convolutional Neural Networks

An Automatic Cardiac Arrhythmia Classification System with Wearable Electrocardiogram

Automatic Attention Learning Using Neural Architecture Search for Detection of Cardiac Abnormality in 12-Lead ECG

A Neural Network for High-Precise and Well-Interpretable Electrocardiogram Classification

LSTM-Based Auto-Encoder Model for ECG Arrhythmias Classification

Comparison of Autoencoder Encodings for ECG Representation in Downstream Prediction Tasks

HA-ResNet: Residual Neural Network with Hidden Attention for ECG Arrhythmia Detection Using Two-Dimensional Signal

Arrhythmia classification of LSTM autoencoder based on time series anomaly detection

Abnormal ECG detection based on an adversarial autoencoder

Convolutional block attention autoencoder for denoising electrocardiograms

ASCNet-ECG: Deep Autoencoder based Attention aware Skip Connection network for ECG filtering

Masked Autoencoder for ECG Representation Learning

Multi-Channel Masked Autoencoder and Comprehensive Evaluations for Reconstructing 12-Lead ECG from Arbitrary Single-Lead ECG

RA-LENet:R-Wave Attention and Local Enhancement for Noise Reduction in ECG Signals

Advanced Neural Network Architecture for Enhanced Multi-Lead ECG Arrhythmia Detection through Optimized Feature Extraction

ECG-Based Heart Arrhythmia Diagnosis Through Attentional Convolutional Neural Networks

ECG signal classification for the detection of cardiac arrhythmias using a convolutional recurrent neural network

Deep learning in ECG diagnosis: A review