基于深度学习的脉搏波连续血压测量

发布时间：2018-09-05 12:57

【摘要】：当今社会,由于工作压力大、生活作息不规律等因素导致心血管疾病患者数量居高不下。如果能够对心血管相关参数加以研究,并分析得到它们之间的关系,就可以实现对心血管疾病的监控,起到预防且降低发病率的作用。血压是人体的一项重要生理参数,可以反应人体的心血管功能状况。而脉搏波信号包含很多人体生理病理信息,通过脉搏波特征参数测血压简单、成本低、精确度高、能够连续测量等优势,有广阔的发展前景。本文基于血压测量的理论基础,建立了两种血压连续测量模型。一种是按照传统做法通过回归分析构建血压模型;另一种以深度学习TensorFlow为框架,借助BP神经网络训练血压和特征参数的关系,构建神经网络的血压模型。实验结果证明,通过两种模型计算的血压误差都在3mmHg标准值之内,而第二种模型误差在2mmHg之内。符合国际标准值。本文的主要工作如下:首先,采用光电心率脉搏计(指尖式)完成了脉搏波信号的采集,用小波变换法和五点三次法完成信号的滤波;在特征点的提取上,提出一种混合算法用于识别特征点,即阈值差分法、小波变换法和微分法相结合的方法,结果表明,此算法能够准确得提取出特征点。其次,建立了基于线性回归的血压测量模型,即通过提取脉搏波的时域特征参数,分析血压与特征参数的相关性,利用逐步回归分析的方法得到血压模型,通过估算血压值,与标准血压相比,误差在3mm Hg以内。最后,提出一种以深度学习TensorFlow为框架,建立血压的BP神经网络模型,即以脉搏波特征参数作为BP神经网络的输入量,通过训练数据得到血压模型。通过TensorFlow中dropout对所构建的神经网络瘦身,去除一定量的特征参数,能够消除过拟合现象,减小误差,从而建立最优模型。为了清楚表达本文所构建的血压模型,运用可视化工具TensorBoard来呈现最终所构建的血压神经网络模型。通过估算血压,与标准血压相比,误差在2mm Hg以内,比传统做法更加精确。
[Abstract]:In today's society, the number of patients with cardiovascular disease is high due to the heavy work pressure and irregular living and rest. If the cardiovascular parameters can be studied and the relationship between them can be analyzed, the monitoring of cardiovascular disease can be realized, and the incidence of cardiovascular disease can be prevented and reduced. Blood pressure is an important physiological parameter, which can reflect the cardiovascular function of human body. Pulse wave signal contains a lot of physiological and pathological information of human body. It is simple to measure blood pressure by pulse wave characteristic parameters, low cost, high accuracy, and can be continuously measured, so it has a broad development prospect. Based on the theory of blood pressure measurement, two continuous blood pressure measurement models are established in this paper. One is to construct blood pressure model by regression analysis according to the traditional method; the other is to build BP model of neural network with the help of BP neural network to train the relationship between blood pressure and characteristic parameters in the framework of in-depth learning TensorFlow. The experimental results show that the blood pressure error calculated by the two models is within the 3mmHg standard value, while the second model error is within the 2mmHg. Conforms to the international standard value. The main work of this paper is as follows: firstly, the pulse wave signal is collected by using the photoelectric heart rate meter (fingertip type), the signal filtering is completed by wavelet transform and 5.3 times, and the feature points are extracted. A hybrid algorithm is proposed to identify feature points, that is, threshold difference method, wavelet transform method and differential method. The results show that the algorithm can extract feature points accurately. Secondly, a blood pressure measurement model based on linear regression is established, that is, by extracting the time domain characteristic parameters of pulse wave and analyzing the correlation between blood pressure and characteristic parameters, the blood pressure model is obtained by stepwise regression analysis, and the blood pressure value is estimated. Compared with the standard blood pressure, the error is within 3mm Hg. At last, a BP neural network model of blood pressure is established based on the framework of deep learning TensorFlow, that is, the characteristic parameters of pulse wave are used as input of BP neural network, and the blood pressure model is obtained by training data. Through dropout in TensorFlow, the neural network is reduced and a certain number of characteristic parameters are removed, which can eliminate the over-fitting phenomenon and reduce the error, thus the optimal model can be established. In order to express the blood pressure model clearly, the final BP neural network model is presented by using the visualization tool TensorBoard. By estimating blood pressure, compared with standard blood pressure, the error is within 2mm Hg, which is more accurate than traditional method.
【学位授予单位】：曲阜师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R443.5;TP181

【参考文献】