Abnormal ECG detection based on an adversarial autoencoder.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Signal Processing

Background:

• Supervised learning models struggle with effective abnormal electrocardiogram (ECG) event detection in monitoring systems.
• Accurate anomaly detection is crucial for timely intervention and patient care.

Purpose of the Study:

• To propose and evaluate a novel ECG anomaly detection framework (ECG-AAE) for enhanced detection of abnormal ECG events.
• To leverage unsupervised learning with an adversarial autoencoder and temporal convolutional network (TCN) for anomaly detection.

Main Methods:

• Developed an ECG anomaly detection framework (ECG-AAE) comprising an autoencoder, discriminator, and outlier detector.
• Trained the ECG-AAE exclusively on normal ECG data, enabling it to learn normal signal characteristics.
• Utilized a temporal convolutional network (TCN) for feature extraction from normal ECG signals.

Main Results:

• The ECG-AAE achieved high performance on the MIT-BIH arrhythmia and CMUH datasets, with accuracy, precision, recall, F1-score, and AUC ranging from 0.9358 to 0.9854.
• Demonstrated superior performance compared to other popular outlier detection methods in identifying abnormal ECG signals.
• The framework successfully distinguished between normal and abnormal ECG signals based on reconstruction errors.

Conclusions:

• The proposed ECG-AAE framework offers an efficient and effective solution for detecting abnormal ECG events.
• Unsupervised anomaly detection using adversarial autoencoders and TCNs shows significant promise in clinical ECG monitoring.
• The ECG-AAE system has the potential to improve the reliability and accuracy of automated ECG analysis.