基于深度学习的急性心肌梗死预测中动态心电的特征检测和分析

This project is focused on the investigation of feature extraction and multi-lead electrocardiogram (ECG) signals fusion for forecasting the acute myocardial infarction in the context of telemedicine. Fully considering the individual differences, we will investigate a set of T wave, ST segment feature detection and multi-lead signals fusion algorithms. The main research of this project includes: (1) the ECG signal de-noising method based on guided filtering will be designed after the baseline drift was reduced by the Bartworth high-pass filter; (2) the T wave morphology will be identified based on the deep neural network stacked with the sparse auto-encoders, and then the key points will be detected guiding by the morphology character. A convolution neural network will be designed for the ST segment morphology identification, and then the amplitude offset of ST segment will be calculated based on the detection of morphology; (3) information fusion about multi-lead T wave and ST segment feature will be realized by organizing deep belief network composing of restricted Boltzmann machines and softmax classifier for forecasting the acute myocardial infarction. This research has strong innovative as well as practical values. The result of this research can be directly applied to telemedicine systems.

本项目围绕提取预测急性心肌梗死的动态心电图典型特征和多导联信息融合问题进行研究。充分考虑人体个体差异的基本特点，借助心电信号大数据的优势，引入深度学习，研究心电信号中多导联T波和ST段特征检测和多导联信息融合算法。主要研究内容包括：（1）在巴特沃斯高通滤波器滤除基线漂移的基础上，构建指导信号，设计基于指导滤波的心电信号降噪算法；（2）基于稀疏自动编码器的深度神经网络识别T波形态，并在形态特征指导下准确检测T波。设计卷积神经网络提取ST段的形态特征，并依据特征波检测计算ST段振幅偏移量；（3）采用基于受限玻尔兹曼机的深度置信网络对多导联的T波和ST段形态特征和偏移量进行特征级信息融合，实现急性心肌梗死的预测。此研究成果可直接应用于远程医疗系统，具有较强的创新性和实用价值。

本项目围绕提取多导联心电信号特征实现急性心肌梗死诊断展开研究。主要包括：（1）在巴特沃斯高通滤波器滤除基线漂移的基础上，构建指导信号，设计基于指导滤波的心电信号降噪算法；（2）利用平稳和连续小波变换融合算法进行P、T特征波检测，实验表明该算法对P、T波检测的准确率大于99.8%，误差率小于0.27%；（3）与临床病理信息相结合，提取心电信号QRS波段和ST-T波段典型特征，设计基于形态特征的下壁心肌梗死检测算法，实验结果显示准确率达到98.33%；（4）基于压缩密集连接卷积神经网络提取多导联心电信号的形态特征和导联间结构特征，实现了11类急性心肌梗死的定位，准确率达99.92%。相关科研转化成果已经成功在多家医疗机构推广应用，取得了良好效果，具有较强的创新性和实用价值。