Inferring spatial–temporal dynamics of ECG signals with deep neural networks for cardiovascular diseases diagnosis

• Published in: Biomedical signal processing and control 2024-11, Vol.97, p.106668, Article 106668
• Main Authors: Yu, Haitao, Lu, Yizhuo, Zheng, Shumei
• Format: Article
• Language: English
• Subjects: Cardiovascular diseases; Convolutional Recurrent Autoencoder; Diagnosis; ECG; Spatial-temporal dynamics
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2024.106668

•A convolutional recurrent autoencoder was proposed to estimate latent factors of ECG signals via unsupervised learning.•Cardiac dynamics on single-trials and spatial–temporal features of latent state was extracted in low-dimensional state space.•Different cardiovascular diseases showed distinct manifold forms allowing accurate diagnosis of cardiac disorders.•Experimental results demonstrated the efficiency of the proposed deep learning-based diagnosis scheme for CVDs. Electrocardiography (ECG) serves as a fundamental and pivotal non-invasive tool for assessing cardiac health and diagnosing cardiovascular diseases (CVD). Nonetheless, the inherent complexity and variability of clinically acquired multichannel ECG signals often challenge its effectiveness. In this study, we introduce an innovative deep learning scheme for CVD diagnosis, leveraging the spatial–temporal dynamics of ECG signals. We designed a Convolutional Recurrent Autoencoder (CRAE) framework that integrates Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (Bi-LSTM) networks. This architecture effectively captures essential spatial–temporal features and extracts meaningful latent dynamics from standard 12-lead ECGs, markedly surpassing the explainable variance ratio achieved by Principal Component Analysis (PCA). Incorporating a novel orthogonal constraint module, we optimized the pattern of latent dynamics to enhance separability among different CVD categories, significantly augmenting the discriminatory power of subsequent classifiers. Validated on the PTB-XL dataset, the latent dynamics of different ECG types demonstrated robust orthogonality in specific dimensions and displayed unique periodic activity trajectories in a visualized low-dimensional space. Employing a one-dimensional ResNet classifier (ResNet1d101) on the latent factors extracted by the CRAE yielded outstanding performance in CVD classification, with an accuracy of 91.19%, precision of 90.19%, recall of 86.10%, F1 score of 0.8798, and macro AUC of 0.9883, surpassing many existing CNN/RNN benchmarks and existing studies. These results underscore the potential of deep neural networks to effectively capture and represent the underlying spatial–temporal dynamics of ECGs for precise and robust CVD diagnosis, offering a promising tool for clinical application and research.