Arrhythmia detection based on semantic segmentation for Multi-lead ECG

Hanshuang Xie1, Huaiyu Zhu2, Yun Pan2
1Hangzhou Proton Technology Co Ltd, 2Zhejiang University


The automatic detection of arrhythmias plays an important role in the management and treatment of heart diseases. Most of the traditional methods divide ECG signals into several heartbeats according to QRS locations as the input data. The segmented heartbeats retain only morphological features while lacking rhythmic information; the classification results extremely rely on the correct identification of QRS, where it is difficult to improve the accuracy of the target problem, i.e., the arrhythmia detection. In this paper, we propose a multi-lead arrhythmia detection method based on semantic segmentation. We divide original ECG signals into 9 regions concerning arrhythmia heartbeat type . After preprocessing, original signals are divided into 30-second segments as the training data, and the area range of the QRS complex was limited 100ms forward and 100ms backward at the position of R peak is marked according to the heartbeat type . Our model consists of multiple parallel dilated convolutional neural network blocks. ECG signals of 5 open-access ECG databases are applied for model training and testing. As a result, the proposed method achieves an overall precision of 0.9792 and an overall recall of 0.9733 . Our model obtains outstanding results in arrhythmia detection, which could contribute to a better automated ECG diagnosis in clinical practice.