Abstract:

Arrhythmia is an irregular heartbeat and can be diagnosed via electrocardiography (ECG). Since arrhythmia can be a fatal health problem, developing automatic detection and diagnosis systems is vital. Although there are accurate machine learning models in the literature to solve this problem, most models assume all arrhythmia types present in training. However, some arrhythmia types are not seen frequently, and there are not enough heartbeat samples from these rare arrhythmia classes to use them for training a classifier. In this study, the arrhythmia classification problem is defined as an anomaly detection problem. We use ECG signals as inputs of the model and represent them with 2-D images. Then, by using a transfer learning approach, we extract deep image features from a Convolutional Neural Network model (VGG16). In this way, it is aimed to get benefit from a pre-trained deep learning model. Then, we train a ν-Support Vector Machines model with only normal heartbeats and predict if a test sample is normal or arrhythmic. The test performance on rare arrhythmia classes is presented in comparison with binary SVM trained with normal and frequent arrhythmia classes. The proposed model outperforms the binary classification with 90.42 % accuracy.

Keywords: