基于神经网络的毫米波雷达血压测量和波形重建

Neural Network-Based Blood Pressure Measurement and Waveform Reconstruction Using Millimeter Wave Radar

血压是评估心血管系统健康状况的重要生理指标,定期进行血压监测有助于及早诊断和干预心血管疾病。相对于传统电子血压计测量方法,毫米波雷达具有非接触等优势,在未来具备一定应用前景。本文提出了一种基于神经网络的毫米波雷达血压测量方法,旨在通过无接触的方式实现精确的血压值测量和波形重建。首先,使用扩展的微分交叉相乘算法和平均值滤波对雷达信号进行预处理,以有效提取雷达回波中的相位信息并去除相位常数。接着,利用小波滤波去除信号中的高频噪声和基线漂移,以获得高质量的脉搏波信号。随后,本文通过构建一个具备编码器-解码器结构的两阶段递进式特征融合与映射网络,分别建立脉搏波到血压值和血压波形的映射关系,以实现准确的血压测量和波形重建。所提模型在第一阶段使用MultiResUNet作为主干网络,实现对脉搏波多尺度特征的提取和融合,同时在多分辨率模块之间引入自注意力机制以挖掘特征向量间的长距离依赖关系,从而准确地重建血压波形;在第二阶段,模型利用第一阶段训练的编码器自动提取脉搏波深层特征,然后借助卷积神经网络和长短期记忆网络实现进一步的特征融合与映射,从而估计收缩压和舒张压。最后,使用雷达生命体征数据集对所提方法进行验证,得到收缩压的测量误差为3.49±5.75 mmHg,舒张压为2.40±3.59 mmHg,两者的测量性能均达到英国高压协会标准的A级要求。同时,所提模型对血压波形的重建误差为3.33 mmHg,偏差率为3.74%,亦证明了该方法在血压波形重建上的有效性。

Blood pressure is a critical physiological indicator for assessing cardiovascular health,and regular monitoring aids early diagnosis and intervention of cardiovascular diseases.Compared to traditional electronic sphygmomanometer measurement methods,millimeter wave radar provides non-contact measurement advantages,showing promising future applications.This study presents a neural network-based method using a millimeter-wave radar for accurate blood pressure measurement and waveform reconstruction in a non-contact manner.Initially,radar signals were preprocessed using an extended differentiate and cross-multiply algorithm and average filtering to effectively extract phase informa‐tion from the radar echoes and eliminate constant phase components.Wavelet filtering was then employed to eliminate high-frequency noise and baseline drift from the signals,obtaining high-quality pulse wave signals.Subsequently,a two-stage progressive feature fusion and mapping network with an encoder-decoder structure was constructed to estab‐lish the mapping between the pulse wave characteristics and blood pressure,enabling accurate blood pressure measure‐ment and waveform reconstruction.In the first stage,MultiResUNet was used as the backbone network to extract and fuse multi-scale features of the pulse wave while introducing a self-attention mechanism between multi-resolution blocks to explore long-distance dependencies between feature vectors,thus accurately reconstructing the blood pressure wave‐form.In the second stage,the model automatically extracted deep features of the pulse wave using the encoder trained in the first stage.The features were then further integrated and mapped using convolutional neural networks and long short term memory networks to estimate systolic and diastolic blood pressures.Finally,the proposed method was validated us‐ing the radar vital signs dataset,yielding a measurement error of 3.49±5.75 mmHg for systolic blood pressure and 2.40±3.59 mmHg for diastolic blood pressure,meeting the A-grade requirements of the British Hypertension Society standards.Additionally,the reconstruction error of the blood pressure waveform was 3.33 mmHg with a deviation rate of 3.74%,further demonstrating the effectiveness of this method in blood pressure waveform reconstruction.