Atrial Fibrillation Classification Using Transfer Learning

Abstract

The appearance of atrial fibrillation (AFIB) is increasing daily, mainly in the elderly. Therefore, an automatic detection system is required to avoid the severe impact of AFIB. A method of this type helps specialist medical professionals in the early diagnosis. Traditionally, the extraction of features was performed manually to recognize the Electrocardiogram (ECG) pattern. However, recent studies have that feature extraction and classification can be performed automatically using deep neural networks. Large-scale data volumes are required to train the effective DL models, which is not the case with the most biomedical physiological signal. Therefore, the identification and classification of AFIB are considered in terms of transfer learning. Our proposed model uses the features learned from the pre-trained model trained on millions of images and transfers their knowledge to our small data set. In this research, we have taken the ResNet50 model trained on millions of images and transferred their feature maps knowledge to our dataset. An accuracy of 99.27% is achieved using the proposed technology. © 2021 IEEE.