Electrocardiogram classification based on convolutional neural network and transfer learning

Abstract

Deep learning is a branch of machine learning, and its methods are now being used to solve all kinds of problems. Deep learning algorithms can learn advanced features from massive data and automatically extract features, which makes deep learning surpass traditional machine learning algorithms. However, as deep learning algorithms rely on large amounts of data and run too slowly, transfer learning arises in response to this disadvantage. Transfer learning allows the use of existing knowledge in the relevant domain to solve a learning problem with only a small number of sample data in the target domain. Combining the two technologies of deep learning and transfer learning, on the one hand, advanced features of data samples can be automatically learned, and on the other hand, it can get rid of the dependence on sample data capacity. In this paper, the electrocardiogram (ECG) signal into spectrogram, and the model is trained with the ImageNet dataset, and then the trained model is transferred, because AlexNet model needs to be fixed image size, so the last pool layer is replaced by a spatial pyramid pooling layer, finally use Softmax classifier for PhysioNet challenge 2017 electrocardiogram data sets are classified, get a 92.84% accuracy and 83.26% F1.