Experimental Convolutional-Recurrent Network in ECG Rhythm for Atrial Fibrillation Classification

Abstract

Atrial Fibrillation (AF) is a type of irregular heart beating problem which could lead to complications such as heart abnormalities events, including mortality and sudden cardiac death. The presence of AF can be diagnosed by electrocardiogram (ECG), including no clear P-wave and irregular pattern of RR-interval. Typically, the characteristics have been determined from the magnitude or duration of ECG. Unfortunately, it remains a difficult task due to its episodic nature. An automatic classification for AF from ECG signals is valuable for healthcare. This paper proposes a deep learning (DL) approach using a combination of convolutional neural network (CNN) as feature extraction and recurrent network as a classifier based on ECG short rhythm. Also, grid-search-based hyperparameter optimization is used to obtain optimal hyperparameters of the model. CNN learns to extract features used in the classification task, and a recurrent network is suitable for sequential prediction to model the flow of time directly. Among 60 models of hyperparameter tuning, the experimental results and analysis indicate that CNN-bidirectional long short-term memory (BiLSTM) outperformed the general model of recurrent neural network (RNN) and gated recurrent unit (GRU) with 96.49% accuracy. The proposed model by employing ECG short rhythm shows promising results and an important approach that can be applied to classify sequential data for AF signal classification.