Arrhythmias classification by integrating stacked bidirectional LSTM and two-dimensional CNN

Classifying different types of arrhythmias based on ECG signal is an important research topic in healthcare. Traditional methods focus on extracting varieties of features from ECG and using them to build a classifier. However, ECG usually presents high inter- and intra-subjects variability both in morphology and timing, hence, it’s difficult for predesigned features to accurately depict the fluctuation patterns of each heartbeat. To this end, we propose a novel arrhythmias classification model by integrating stacked bidirectional long short-term memory network (SB-LSTM) and two-dimensional convolutional neural network (TD-CNN). Particularly, SB-LSTM mines the long-term dependencies contained in ECG from both directions to depict the overall variation trend of ECG, while TD-CNN exploits local characteristics of ECG to characterize the short-term fluctuation patterns of ECG. Moreover, we design a discrete wavelet transform (DWT) based ECG decomposition layer and a Sum Rule based intermediate classification result fusion layer, by which ECG can be analyzed from multiple time-frequency resolutions, and the classification results of our model can be more accurate. Experimental results based on MIT-BIH arrhythmia database shows that our model outperforms 3 baseline methods, achieving 99.5% of accuracy, 99.9% of sensitivity and 98.2% specificity, respectively.