文章摘要

文章信息/Info

Title:: Research on ECG signal feature point detection algorithm based on deep learning

Author(s):: LIANG Xiaohong; SONG Ningning; LIU Chengyou; TIAN Shuchang; ZHANG Huawei; QIN Hang; （Nanjing First Hospital, Nanjing Medical University, Nanjing 210006, China）

Keywords:: Electrocardiogram; Deep learning; Encoder-decoder structure; Convolution neural network; Bidirectional long short-term memory

摘要:: 为实现自动、准确、有效地心电数据分析，本研究提出了一种基于深度学习的心电信号智能分析模型ECG_SegNet，以识别P波、QRS波群和T波并检测波形的起止点和偏差。首先，在编码器路径引入标准空洞卷积模块，使模型能够提取更多的心电信号特征；然后，在编码结构加入双向长短期记忆网络以获得大量时间特征；此外，将编码器路径上各级特征集分别短接至解码器部分进行多尺度解码，以减少编码过程中的信息损失。最后，该模型分别在QT和LU数据库上进行训练和测试。在QT数据库上，P波、QRS波群、T波起止点检测的平均F1分别为99.53%、99.82%、99.41%；在LU数据库上，P波、QRS波群、T波起止点检测的平均F1分别为94.74%、98.88%、97.53%。结果表明，本研究在心电信号波形检测上具有良好的灵活性和准确性，是一种可靠的心电信号自动分析方法。

Abstract:: In order to realize automatic, accurate and effective analysis of ECG data,we proposed a deep learning based ECG intelligent analysis model ECG_SegNet to identify P waves, QRS complexes and T waves, and detect these waveforms′ onsets and offsets. Firstly, the standard dilated convolution module was introduced into the encoder path to extract more ECG signal features. Then the bidirectional long term and short term memory was added to the encoding structure, to obtain numerous temporal features. In addition, the feature sets of each level in the encoder path were connected to the decoder part for multi-scale decoding to mitigate the information loss in the encoding process. Finally, the model was trained and tested on QT and LU databases respectively. On the QT database, the average F1 of P wave, QRS complex and T wave detection was 99.53%, 99.82%, 99.41%, respectively. On the LU database, the average F1 of P wave, QRS complex and T wave detection was 94.74%, 98.88%, 97.53%, respectively. The results show that the model has good flexibility and reliability when applied to ECG signals detection, and it is a reliable method for analyzing ECG signals in real-time.