A convolutional autoencoder framework for ECG signal analysis.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Electrocardiographic (ECG) signals are crucial for assessing heart activity and detecting anomalies.
• Automated analysis of ECG signals, especially for long-term monitoring and physician support, requires reliable systems.
• Deep learning, particularly autoencoders (AEs), offers advanced methods for analyzing time-varying signals like ECGs.

Purpose of the Study:

• To present a Convolutional Autoencoder (CAE)-based framework for ECG signal analysis and anomaly identification.
• To develop a system that supports physicians in diagnosing cardiological risks and pathologies through automated ECG analysis.
• To enhance the interpretability and robustness of the anomaly detection system.

Main Methods:

• Utilizing a Convolutional Autoencoder (CAE), a specialized deep neural network for signal data, for ECG analysis.
• Training the CAE framework on synthetic ECG data for anomaly detection.
• Testing and validating the framework on a 12-lead ECG benchmark dataset and real-world scenarios.

Main Results:

• The CAE-based framework demonstrated high accuracy in identifying anomalies in ECG signals.
• Achieved a ROC AUC of 97.82% on a simulated test set and 99.75% on a real test set.
• The system's explainability modules, based on reconstruction error and validated by expert annotations, proved effective.

Conclusions:

• The proposed CAE-based framework provides a powerful tool for automatic anomaly detection in ECG signals.
• The system effectively supports physicians in their decision-making process for diagnosing cardiac conditions.
• The integration of explainability and preprocessing modules enhances the clinical utility and reliability of the ECG analysis system.