A new transfer learning approach to detect cardiac arrhythmia from ECG signals

Deep Learning (DL) has turned into a subject of study in different applications, including medical field. Finding the irregularities in Electrocardiogram (ECG) is a critical part in patients’ health monitoring. ECG is a simple, non-invasive procedure used in the prediction and diagnosis of Cardiac Arrhythmia. This paper proposes a new transfer learning-based end to end approach to automate the cardiac arrhythmia classification. The proposed approach begins with gathering ECG Dataset and extracting beats after ECG beat segmentation. Developing a Model from scratch is time-consuming, so the concept of transfer learning is used. For transferring the knowledge to our ECG classification domain, the last layers of the model are fine-tuned such that model becomes more domain-specific to our target ECG data. Three pre-trained Convolutional Neural Networks (CNNs), AlexNet, Resnet18, GoogleNet are explored, and then, our model is designed by block wise fine-tuning each layer with different model training parameters. To update the weights and offsets, Adaptive moment estimation, Root means square propagation and Stochastic gradient descent with momentum (SGDM) are three different optimizers used. Investigating the results obtained by training fine-tuned models, we select the model which gives the system's best accuracy. MIT-BIH arrhythmia database is considered in this study. Performance of each Fine-tuned Model is evaluated by calculating Precision, Recall, Specificity, F -score and Accuracy. Moreover, our proposed fine-tuned Deep-CNN Model is effective and outperformed the existing models in the literature with accuracy of 99.56%.