Prediction of stroke using single lead ECG signal: A deep learning approach

Stroke is the second most common cause of mortality globally. Its yearly death rate is approximately 5.5 million. Continuous monitoring of heart health is a critical aspect of modern healthcare. The primary objective centers on leveraging single-lead Electrocardiogram signals for accurate stroke detection through the identification of Atrial Fibrillation patterns. Ultimately, the main aim is to create a reliable system capable of integrating seamlessly into healthcare frameworks while prioritizing accuracy. A deep learning model for stroke prediction is proposed, which uses single lead Electrocardiogram signals collected for a duration of around 30 to 60 seconds for model training. Generative Adversarial Networks for effective data augmentation is proposed in order to overcome the imbalance in Normal and Atrial Fibrillation signals in the dataset. The proposed framework centers on the utilization of Long ShortTerm Memory networks, a type of recurrent neural network, for the classification of Electrocardiogram signals. A comparative study of Long Short-Term Memory models with different ways of data augmentation is done so as to select the best model. The key finding is that the Long Short-Term Memory model trained using Atrial Fibrillation signals generated via a Conditional Generative Adversarial Network for data augmentation, achieved a commendable accuracy of $\mathbf{9 6 \%}$. Remarkably, this accuracy closely matched to the accuracy of Long Short-Term Memory model trained by oversampling Atrial Fibrillation signals within the original dataset.