A Real-Time Framework for Arrhythmia Classification

Abstract

Identification of arrhythmia is crucial to recognize cardiovascular diseases early. There has been much research over the past to automate the detection process of arrhythmia. Though most of the recent approaches have given good accuracies, such approaches use high resources which limits the real-time classification in low-end devices. Most of the methods proposed in the literature are not explicitly evaluated to support real-time classification. In this work, we propose a novel hybrid classification framework for real-time arrhythmia classification by using a deep convolutional neural network and dynamic time warping (DTW) distance-based alignment measure. A new data structure based on circular arrays has also been proposed to calculate the alignment scores efficiently. The hybrid classifier is inspired by the similarity of the heart rhythm between adjacent consecutive beats. Incorporating such context of the unknown beat to the classification leads to a significant performance notably in prediction time. In order to evaluate the complete framework, forty-six ECG sequences in the MIT-BIH database streamed as a continuous time series data. Performance of the framework is tested on accuracy and speed over different hardware configurations. Our classification model achieved state-of-art performance with an average accuracy of over 97.05% for five supertypes of arrhythmia defined by AAMI standard. We have compared the performance of our CNN with VGG19 and AlexNet architecture based deep convolutional neural network models. The hybrid classifier prediction time outperformed the prediction time of deep learning methods proposed in literature achieving the same accuracy. As a future work, generalizing the proposed hybrid approach to other domains which uses sequence prediction can be investigated.