Autoencoder-based Arrhythmia Detection using Synthetic ECG Generation Technique

With a couple of million lives lost annually, cardiovascular disease (CVD) is the leading cause of death globally; about 80% of which are due to arrhythmia. Electrocardiogram (ECG) signals are important for arrhythmia diagnosis, researchers have used various ECG datasets in building arrhythmia detection systems to automate the manual time-consuming diagnostic process. However, existing datasets have class imbalance issues, and the traditional oversampling and undersampling techniques prove ineffective in handling the imbalance problem. We propose a novel approach to handling arrhythmia detection as an anomaly case to address this. In our proposed approach, we first use Generative Adversarial Networks (GANs) to synthetically generate normal training instances from the MIT-BIH arrhythmia dataset and then we use only the synthetically generated normal data to build the anomaly model using autoencoder (AE); employing the AE for unsupervised anomaly detection help in overcoming the GAN convergence issues. We evaluate the model using test data comprising both normal and abnormal samples that are not used by the GAN and compare its performance with other state-of-the-art works. The model achieved improved arrhythmia detection with an AUC-ROC of 0.6768 and an AUC-PR of 0.8537. While effectively tackling data scarcity and imbalance, this work also contributes valuable perspectives to enhance arrhythmia detection systems, providing a foundation for more reliable and adaptable solutions in healthcare.