Scholarlink

<p>The detection of cardiac arrhythmias is crucial for preventing severe cardiovascular diseases. As a result, deep learning algorithms have been developed in the state-of-the-art for the automatic detection of these conditions. While these models have produced promising results, they rely on having sufficient data to generalize well. This presents a challenge in the medical field, where data scarcity is often an issue. To address these challenges, this work proposes a hybrid architecture combining a deep autoencoder with a 1D CNN, trained on synthetic signals and evaluated on real signals. The objective is to demonstrate that synthetic signals can complement or even replace real signals when there are insufficient data. The MIT-BIH arrhythmia database, containing five classes, was used for this process. The dataset was split to train a generative adversarial network for generating synthetic signals, while the remaining data were used to evaluate the model on real signals. Signal preprocessing involved transforming the ECG data into 2D representations using different methods, including similarity maps, Gramian angular gields, Markov transition fields, and co-occurrence matrices (CoMs), each of which was evaluated with the proposed model to observe the performance of each one. The extracted features were compressed by an autoencoder and then passed through a 1D CNN for classification. The best results in real data evaluation were obtained with CoMs, achieving an accuracy of 99.7%, precision of 99.7%, recall of 99.7%, specificity of 99.9%, F1-score of 99.7%, Matthews correlation coefficient of 99.6%, and Cohen’s kappa of 99.3%.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

ECG classification using a CNN stacked with an autoencoder trained with synthetic signals

  • 期刊论文

  • 发表日期：

    2025年5月13日

查看全文
  • Enrique Quezada-Prospero,
  • Dante Mújica-Vargas,
  • Antonio Luna-Álvarez,
  • Viridiana Vela Rincón,
  • Andrés A. Arenas Muñiz

摘要

The detection of cardiac arrhythmias is crucial for preventing severe cardiovascular diseases. As a result, deep learning algorithms have been developed in the state-of-the-art for the automatic detection of these conditions. While these models have produced promising results, they rely on having sufficient data to generalize well. This presents a challenge in the medical field, where data scarcity is often an issue. To address these challenges, this work proposes a hybrid architecture combining a deep autoencoder with a 1D CNN, trained on synthetic signals and evaluated on real signals. The objective is to demonstrate that synthetic signals can complement or even replace real signals when there are insufficient data. The MIT-BIH arrhythmia database, containing five classes, was used for this process. The dataset was split to train a generative adversarial network for generating synthetic signals, while the remaining data were used to evaluate the model on real signals. Signal preprocessing involved transforming the ECG data into 2D representations using different methods, including similarity maps, Gramian angular gields, Markov transition fields, and co-occurrence matrices (CoMs), each of which was evaluated with the proposed model to observe the performance of each one. The extracted features were compressed by an autoencoder and then passed through a 1D CNN for classification. The best results in real data evaluation were obtained with CoMs, achieving an accuracy of 99.7%, precision of 99.7%, recall of 99.7%, specificity of 99.9%, F1-score of 99.7%, Matthews correlation coefficient of 99.6%, and Cohen’s kappa of 99.3%.