A novel temporal generative adversarial network for electrocardiography anomaly detection.

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• Automated Electrocardiogram (ECG) classification using deep learning shows promise for cardiac abnormality detection.
• Current deep learning models struggle with cardiac abnormalities absent or underrepresented in training data.
• There is a need for robust ECG anomaly detection methods capable of identifying novel or rare conditions.

Purpose of the Study:

• To propose a novel one-class classification based ECG anomaly detection generative adversarial network (GAN).
• To develop a generalized anomaly detector that reliably identifies unknown cardiac abnormalities from ECG signals.
• To improve the performance of automated ECG analysis for clinical diagnosis.

Main Methods:

• A generative adversarial network (GAN) architecture was developed for ECG anomaly detection.
• A Bi-directional Long-Short Term Memory (Bi-LSTM) layer was embedded within the GAN.
• Mini-batch discrimination was employed in the discriminator to synthesize ECG signals representing normal heartbeats.

Main Results:

• The proposed method achieved an accuracy of 95.5% and an Area Under the Curve (AUC) of 95.9% on the MIT-BIH arrhythmia database.
• The novel GAN approach outperformed several state-of-the-art semi-supervised learning based ECG anomaly detection algorithms.
• The method demonstrated robust detection of unknown anomaly classes, indicating its generalizability.

Conclusions:

• The developed one-class classification GAN with Bi-LSTM effectively detects cardiac abnormalities in ECG signals.
• This approach offers a reliable method for identifying unknown or rare arrhythmias, addressing limitations of current deep learning models.
• The proposed technique shows potential as a valuable diagnostic tool for cardiologists in identifying arrhythmias.