基于BiLRCN和注意力机制的脉搏波血压测量

Pulse Wave Blood Pressure Measurement Based on BiLRCN and Attention Mechanism

为了提高无创血压测量的精度,提出了基于双向长期递归卷积网络(Bidirectional Long-term Recurrent Convolutional Network,BiLRCN)和注意力机制的脉搏波血压测量方法。通过2个卷积神经网络(Convolutional Neural Network,CNN)层提取出光电容积脉搏信号的高维度特征,将其作为双向长短期记忆(Bidirectional Long Short-Term Memory,BiLSTM)网络的输入,通过BiLSTM提取输入序列前后向的特征信息进行预测;根据注意力机制自动分配权重的特征,给予重要时刻脉搏特征数据较大的权重,并通过2个全连接层得到血压的测量值。将所提出的方法与CNN、长短期记忆(Long Short-Term Memory,LSTM)网络、BiLSTM网络、长期递归卷积神经网络(Long-term Recurrent Convolutional Network,LRCN)方法进行了收敛速度和血压测量的对比实验。实验结果表明,所提出的方法较LRCN均方误差降低了21.63%,平均绝对误差降低了67.5%,确定性相关系数提高了0.42%。所提出的方法的收敛速度更快、血压测量精度更高。

To improve the accuracy of noninvasive blood pressure measurement,a pulse wave blood pressure measurement method based on bidirectional long-term recurrent convolutional network(BiLRCN)and attention mechanism is proposed.The high-dimensional features of the photovolumetric pulse signal are extracted by two convolutional neural network(CNN),which is used as the input of bidirectional long short-term memory(BiLSTM)network,and the feature information in the forward and backward directions of the input sequence is extracted by BiLRCN for prediction.The attention mechanism is used to automatically assign the weighted features,which gives a larger weight to the important moments of the pulse feature data,and the two fully connected layers are used to obtain the blood pressure measurement value.The proposed method is compared with CNN,long short-term memory(LSTM)network,BiLSTM network,and long term recurrent convolutional neural network(LRCN)methods in terms of convergence speed and blood pressure measurement.The experimental results show that the proposed method decreases the mean square error by 21.63%,decreases the mean absolute error by 67.5%,and improves the coefficient of deterministic correlation by 0.42% compared to LRCN.The proposed method has faster convergence and higher accuracy of blood pressure measurement.