Sparse Temporal AutoEncoder for ECG Anomaly Detection

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Cardiology

Background:

• Electrocardiogram (ECG) analysis is crucial for diagnosing cardiac conditions.
• Distinguishing normal from abnormal ECG signals is challenging due to signal complexity and variability.
• Current methods often require labeled abnormal data, limiting real-world applicability.

Purpose of the Study:

• To develop a novel unsupervised model for accurate ECG anomaly detection.
• To address the limitations of traditional ECG analysis methods.
• To create a robust tool for automated and intelligent cardiac diagnostics.

Main Methods:

• Proposed a Sparse Temporal Autoencoder (STAE) model.
• Utilized Temporal Convolutional Networks (TCNs) for hierarchical feature extraction from time and frequency domains.
• Integrated masked signal reconstruction and a hybrid sparse attention mechanism.

Main Results:

• Achieved the highest ROC-AUC of 0.872 on the PTB-XL dataset among unsupervised methods.
• Demonstrated a low inference time of 0.009 seconds.
• Validated the model's effectiveness in capturing critical temporal and spectral patterns.

Conclusions:

• The STAE model offers state-of-the-art performance for unsupervised ECG anomaly detection.
• The model's ability to train on normal data makes it suitable for practical deployment.
• STAE shows significant potential for advancing automated ECG analysis and cardiac diagnostics.