基于多分支融合神经网络的心电图导联重构方法

ECG signal reconstruction based on multi-layer feature fusion using neural network

设计一种新型的多分支信息融合神经网络结构,利用已知的I,Ⅱ,V2 3个导联心电信号来重构其它导联心电信号。基于卷积神经网络结构提取多个导联的特征然后进行线性相加融合,采用一种改进的双向长短期记忆网络结构来获得与时序相关的信息,从而实现心电图导联重构。使用Physikalisch Technische Bundesanstalt(PTB)数据库进行验证,导联重构方法具有0.944 4的相关系数和0.320 3的均方根误差,说明新型神经网络结构可以有效地实现心电图导联重构。

A novel neural network which can achieve multi-layer feature fusion is proposed for reconstructing the electrocardiogram(ECG) signals of other leads using the known ECG signals of leads I, Ⅱ and V2. The features of multiple leads are extracted by convolutional neural network for linear combination, and an improved bidirectional long short-term memory network structure is used to obtain temporal sequence correlation which is then fused with the features obtained by convolutional neural network for realizing ECG signal reconstruction. The proposed method is verified with Physikalisch Technische Bundesanstalt database. The results show that the signal reconstruction method has a correlation coefficient of 0.944 4 and a low root-mean-square error of 0.320 3, which demonstrates the effectiveness of the novel neural network structure for ECG signal reconstruction.