Autoencoder-based Arrhythmia Detection using Synthetic ECG Generation Technique.

Area of Science:

• Cardiology
• Medical Informatics
• Artificial Intelligence

Background:

• Cardiovascular disease (CVD) is a leading global cause of death, with arrhythmia accounting for a significant portion.
• Electrocardiogram (ECG) analysis is crucial for arrhythmia diagnosis, but existing datasets suffer from class imbalance.
• Traditional data augmentation techniques are often ineffective for addressing imbalance in ECG datasets.

Purpose of the Study:

• To propose a novel approach for arrhythmia detection by framing it as an anomaly detection problem.
• To address the challenges of data scarcity and class imbalance in ECG datasets for arrhythmia detection.
• To develop a more reliable and adaptable automated system for diagnosing arrhythmia.

Main Methods:

• Utilized Generative Adversarial Networks (GANs) to synthetically generate normal ECG instances from the MIT-BIH arrhythmia dataset.
• Employed an autoencoder (AE) for unsupervised anomaly detection, training solely on synthetically generated normal data.
• Evaluated the model on a separate test set containing both normal and abnormal ECG samples.

Main Results:

• The proposed model achieved an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 0.6768.
• The Area Under the Precision-Recall Curve (AUC-PR) was recorded at 0.8537.
• The approach effectively tackled data scarcity and imbalance issues inherent in arrhythmia datasets.

Conclusions:

• The novel anomaly detection approach using GANs and AEs offers improved arrhythmia detection performance.
• This method provides a robust solution for handling imbalanced ECG data, enhancing diagnostic reliability.
• The study lays the groundwork for more adaptable and dependable automated arrhythmia detection systems in healthcare.