WO2021208490A1 - 一种基于深度神经网络的血压测量方法及装置 - Google Patents

一种基于深度神经网络的血压测量方法及装置 Download PDF

Info

Publication number
WO2021208490A1
WO2021208490A1 PCT/CN2020/139658 CN2020139658W WO2021208490A1 WO 2021208490 A1 WO2021208490 A1 WO 2021208490A1 CN 2020139658 W CN2020139658 W CN 2020139658W WO 2021208490 A1 WO2021208490 A1 WO 2021208490A1
Authority
WO
WIPO (PCT)
Prior art keywords
blood pressure
sample
signal
neural network
deep neural
Prior art date
Application number
PCT/CN2020/139658
Other languages
English (en)
French (fr)
Inventor
闻博
李烨
苗芬
刘增丁
刘记奎
Original Assignee
中国科学院深圳先进技术研究院
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 中国科学院深圳先进技术研究院 filed Critical 中国科学院深圳先进技术研究院
Publication of WO2021208490A1 publication Critical patent/WO2021208490A1/zh

Links

Images

Classifications

    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/72Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235Details of waveform analysis
    • A61B5/7264Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/02Detecting, measuring or recording pulse, heart rate, blood pressure or blood flow; Combined pulse/heart-rate/blood pressure determination; Evaluating a cardiovascular condition not otherwise provided for, e.g. using combinations of techniques provided for in this group with electrocardiography or electroauscultation; Heart catheters for measuring blood pressure
    • A61B5/021Measuring pressure in heart or blood vessels
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/72Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235Details of waveform analysis
    • A61B5/7264Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device

Definitions

  • the present invention relates to the field of computer technology, in particular to a blood pressure measurement method and device based on a deep neural network.
  • Pulse Wave Velocity Pulse Wave Velocity
  • the pulse wave characteristic parameter method mainly establishes a model with blood pressure by extracting the characteristics of the pulse wave waveform.
  • the measurement equipment is simple, but there is also a problem of poor robustness. For certain groups of people, such as patients with hypertension or arrhythmia, its corresponding Feature extraction is difficult, and it is often difficult to achieve accurate measurement.
  • the method also includes:
  • the sample training signal is input into the deep neural network, and the training operation is performed according to the migration learning to obtain the deep neural network blood pressure model.
  • performing a sample preprocessing operation on the sample signal to obtain the sample training signal includes:
  • the signal acquisition module is used to perform signal acquisition operations on the acquisition object to obtain pulse waves and ECG signals;
  • the sample training module is used to input sample training signals into the deep neural network based on the deep neural network, and perform training operations according to migration learning to obtain the deep neural network blood pressure model.
  • the signal filtering unit is used to filter the sample pulse wave and the sample ECG signal by using a band-pass filter to obtain the noise-reduced sample pulse wave and the noise-reduced sample ECG signal;
  • the segmentation processing unit is used to perform segmentation processing on the sample pulse waveform diagram and the sample electrocardiogram waveform diagram according to a preset segmentation method to obtain a segmented signal;
  • the signal sampling unit is used to perform sliding sampling operation on the sample training signal to obtain multiple network training samples
  • the network construction unit is used to construct a deep neural network-based convolutional neural network containing several computing units and use multiple network training samples to train the convolutional neural network to obtain a pre-trained network;
  • the blood pressure measurement method and device based on the deep neural network in the embodiment of the present invention receives a signal collection request, which at least carries identification information of the collection object; and then performs a signal collection operation on the collection object according to the signal collection request, Acquire the pulse wave and ECG signal of the collection object; then, perform preprocessing operations on the acquired pulse wave and ECG signal to obtain the preprocessed signal with noise reduction and low interference, to a certain extent to ensure that the subsequent accurate acquisition of blood pressure values Furthermore, based on the trained deep neural network blood pressure model, the pre-processed signal is input into the deep neural network blood pressure model for measurement operation, and the systolic and diastolic blood pressure corresponding to the pre-processed signal are obtained, which can ensure the acquisition of blood pressure value
  • the accuracy and completeness of the BP; output identification information, systolic blood pressure and diastolic blood pressure, the whole process of the present invention is simple to operate, low in computational complexity, and low cost; the blood pressure measurement
  • FIG. 2 is a preferred flow chart of the blood pressure measurement method based on the deep neural network of the present invention
  • Figure 6 is a schematic diagram of a blood pressure measurement device based on a deep neural network of the present invention.
  • FIG. 9 is another preferred schematic diagram of the blood pressure measurement device based on the deep neural network of the present invention.
  • a blood pressure measurement method based on a deep neural network is provided. See FIG. 1, which includes the following steps:
  • S1 Receive a signal collection request, the signal collection request carries at least the identification information of the collection object;
  • S3 Perform preprocessing operations on pulse waves and ECG signals to obtain preprocessed signals
  • S5 Output identification information, systolic blood pressure and diastolic blood pressure.
  • the signal acquisition request carries at least the identification information of the acquisition object; and then the signal acquisition operation is performed on the acquisition object according to the signal acquisition request to obtain the signal acquisition request.
  • the signal acquisition operation is performed on the acquisition object to acquire the pulse wave and ECG signal, which can be specifically performed on the fingertips of the acquisition object’s fingers, or the radial artery of the wrist, the carotid artery, etc.
  • the pulse wave and ECG signal are collected at the pulse of the aorta on the body surface.
  • the pulse wave and ECG signal are preprocessed to obtain the preprocessed signal.
  • the acquired pulse wave and ECG signal can be filtered and denoised, divided by heart beat, normalized, etc. Signal processing, to reduce impurities and interference in the signal, to a certain extent, to ensure the efficiency of the subsequent accurate acquisition of the blood pressure value of the collected object.
  • the preprocessed signal is input into the deep neural network blood pressure model for measurement operation, and the systolic and diastolic blood pressure corresponding to the preprocessed signal is obtained, which can be specifically passed
  • the signal waveform of the processed pre-processed signal is input as input to the deep neural network blood pressure model for a specific population obtained by migration learning training, and the continuous blood pressure of the collected object is calculated to obtain the contraction of the collected object. Pressure and diastolic blood pressure and output.
  • the sample signal in the database is obtained.
  • the sample signal carries at least the sample pulse wave and the sample ECG signal. Specifically, it can be obtained by reading a local public database, and obtaining a large number of pre-collected data such as hypertension in the public database. Sample pulse waves and sample ECG signals of patients and patients with arrhythmia facilitate the establishment of subsequent deep neural network blood pressure models.
  • this embodiment is based on a deep neural network, inputs sample training signals into the deep neural network, and performs training operations according to transfer learning.
  • the deep neural network may be used to perform processing on the processed sample training signals of people with arrhythmia. Transfer learning to obtain a deep neural network blood pressure model.
  • the model training model using the migration method has a low training cost, can output continuous blood pressure heart by heart, and has high measurement accuracy to ensure the accuracy of obtaining blood pressure values.
  • the deep neural network is essentially a convolutional neural network including several convolutional layers, pooling layers, batch normalization layers, and fully connected layers, and is established with systolic blood pressure and diastolic blood pressure as the output Model.
  • the sample preprocessing operation is performed on the sample signal, and the steps of obtaining the sample training signal include:
  • S701 Use a band-pass filter to filter the sample pulse wave and the sample ECG signal to obtain the noise-reduced sample pulse wave and the noise-reduced sample ECG signal;
  • S702 Obtain a sample pulse waveform diagram and a sample ECG waveform diagram corresponding to the noise-reduction sample pulse wave and the noise-reduction sample ECG signal;
  • S703 Perform segmentation processing on the sample pulse waveform diagram and the sample ECG waveform diagram according to the preset segmentation mode to obtain a segmented signal
  • S704 Perform normalization processing on the segmented signal to obtain a sample training signal.
  • the value of L is generally three times the sampling rate Fs.
  • the back end of the ECG is zero-filled, and the two ends of the PPG are zero-filled , So that the length of all samples is the same.
  • normalization processing is performed on the segmented signal, specifically by making its value in the interval [0, 1] to obtain the sample training signal.
  • this embodiment autonomously extracts feature modeling based on signal waveforms, does not require high waveform quality, does not need to detect dicrotic waves, and is easy to check.
  • the sample training signal is input into the deep neural network, and the training operation is performed according to transfer learning, and the steps of obtaining the deep neural network blood pressure model include:
  • S803 Use the pre-training network to perform migration learning on the sample signal to obtain a deep neural network blood pressure model.
  • performing a sliding sampling operation on the sample training signal to obtain multiple network training samples is to implement a sliding window, ensure the accuracy of feature extraction of the sample training signal, and reduce subsequent training operations. Repeating the calculation of convolution can reduce the computational complexity to a certain extent and improve the efficiency of subsequent acquisition of blood pressure values.
  • a convolutional neural network containing several computing units based on a deep neural network is constructed and multiple network training samples are used to train the convolutional neural network to obtain a pre-trained network, specifically by constructing a deep neural network, It is specifically a convolutional neural network that includes several computing units.
  • Each computing unit includes a convolutional layer, a batch normalization layer, and a pooling layer. After 3-5 unit calculations, the obtained features are input to the first layer. After connecting the layers, a pre-trained network can be obtained that can simultaneously output systolic and diastolic blood pressures SBP p and DBP p.
  • a pre-training network is used to perform migration learning on sample signals to obtain a deep neural network blood pressure model.
  • the pre-training network can be specifically used to perform migration learning for specific population data, even if most parameters are frozen (ie Keep unchanged), only fine-tune the last fully connected layer of the pre-trained network, that is, make the new learning rate 1/10 of the original learning rate to obtain the deep neural network blood pressure model, where the fine-tuning process is similar to sampling the sample signal The process of preprocessing operations.
  • a convolutional neural network containing several computing units based on a deep neural network is constructed, and multiple network training samples are used to train the convolutional neural network using a stochastic gradient descent optimization algorithm.
  • a deep neural network-based convolutional neural network containing several computing units is constructed and multiple network training samples are used to train the convolutional neural network using a stochastic gradient descent optimization algorithm.
  • the Loss function is jointly established
  • the training of the model can output systolic and diastolic blood pressure at the same time. While accurately acquiring the blood pressure value of the collected object, it can also facilitate the follow-up to provide an early warning of the risk of sudden rise and fall in blood pressure according to the change in the output blood pressure value, and remind the patient to take it in time Seek medical attention or protection.
  • the deep neural network of this embodiment is specifically a convolutional neural network including several computing units, each computing unit includes a convolutional layer, a batch normalization layer, and a pooling layer. After 3-5 units of calculation, After inputting the obtained features into a fully connected layer, the systolic and diastolic blood pressures SBP p and DBP p can be output at the same time.
  • the model is trained by jointly establishing the Loss function, where the loss function Loss is the mean square error, which is specifically expressed as follows:
  • SBP t and DBP t are the actual systolic and diastolic blood pressure, and n is the number of samples;
  • SBP p and DBP p can be expressed as follows:
  • A is the activation function
  • W and b are the parameters of the model.
  • the gradient of W is calculated from the Loss function, and the gradient is continuously updated with a preset specific learning rate (quantitatively and randomly changes the value of W), where the learning rate is generally 0.0001, and the learning rate can be set according to actual application requirements, and there is no specific limitation here; then, the Loss is minimized along the direction of the gradient, and finally the training is stopped when the Loss no longer changes, or the training is stopped when the number of training reaches a certain value , Usually the number of training does not exceed 10,000 times.
  • FIG. 12 is a comparison between the blood pressure and actual target systolic blood pressure measured by a group of collected subjects. It can be found that in each blood pressure interval, the error between the predicted value and the target value is small.
  • the method further includes:
  • the blood pressure value of the collection object is continuously measured according to the preset time period, so as to monitor the change of the blood pressure value of the collection object within the time period, that is, the blood pressure can be provided according to the change of the output blood pressure value.
  • Early warning of sudden rise and fall risk remind patients to seek medical treatment or protection in time, specifically by comparing historical systolic blood pressure and historical diastolic blood pressure with the current output systolic blood pressure and diastolic blood pressure respectively, that is, by calculating the historical systolic blood pressure and the current diastolic blood pressure.
  • the absolute value of the difference of systolic blood pressure is divided by the historical systolic pressure to obtain the rate of change of the systolic blood pressure.
  • the rate of change of diastolic blood pressure can be obtained.
  • the rate of change is greater than the expected
  • an early warning reminder is issued to the user terminal, where the early warning reminder may be in the form of timely information, email, or signal alarm, etc., and there is no specific limitation here.
  • the preset change threshold is set according to actual application requirements, and is usually set to 20%.
  • a blood pressure measurement device based on a deep neural network including:
  • the request receiving module 601 is configured to receive a signal collection request, and the signal collection request at least carries identification information of the collection object;
  • the signal acquisition module 602 is used to perform signal acquisition operations on the acquisition object to obtain pulse waves and ECG signals;
  • the preprocessing module 603 is used to perform preprocessing operations on pulse waves and ECG signals to obtain preprocessed signals;
  • the blood pressure measurement module 604 is used to input the preprocessed signal into the deep neural network blood pressure model based on the trained deep neural network blood pressure model for measurement operation, and obtain the systolic blood pressure and the diastolic blood pressure corresponding to the preprocessed signal;
  • the blood pressure measurement device based on the deep neural network in the embodiment of the present invention receives a signal collection request, which at least carries identification information of the collection object; and then performs a signal collection operation on the collection object according to the signal collection request to obtain the signal collection request.
  • pulse wave refers to the photoplethysmographic signal of the object to be collected
  • ECG signal refers to the single-lead electrocardiographic signal of the object to be collected.
  • the photoplethysmographic pulse wave of the finger of the object is collected, the signal length is 5 minutes, and the sampling frequency It is 250Hz.
  • the pulse wave and ECG signal are preprocessed to obtain the preprocessed signal.
  • the acquired pulse wave and ECG signal can be filtered and denoised, divided by heart beat, normalized, etc. Signal processing, to reduce impurities and interference in the signal, to a certain extent, to ensure the efficiency of the subsequent accurate acquisition of the blood pressure value of the collected object.
  • the preprocessed signal is input into the deep neural network blood pressure model for measurement operation, and the systolic and diastolic blood pressure corresponding to the preprocessed signal is obtained, which can be specifically passed
  • the signal waveform of the processed pre-processed signal is input as input to the deep neural network blood pressure model for a specific population obtained by migration learning training, and the continuous blood pressure of the collected object is calculated to obtain the contraction of the collected object. Pressure and diastolic blood pressure and output.
  • the device further includes:
  • the sample acquisition module 701 is used to acquire sample signals in the database, and the sample signals at least carry sample pulse waves and sample ECG signals;
  • the sample training module 703 is used to input sample training signals into the deep neural network based on the deep neural network, and perform training operations according to migration learning to obtain the deep neural network blood pressure model.
  • the sample signal in the database is obtained.
  • the sample signal carries at least the sample pulse wave and the sample ECG signal. Specifically, it can be obtained by reading a local public database, and obtaining a large number of pre-collected data such as hypertension in the public database. Sample pulse waves and sample ECG signals of patients and patients with arrhythmia facilitate the establishment of subsequent deep neural network blood pressure models.
  • the sample pulse wave and the sample ECG signal are preprocessed, specifically by extracting the same cycle of the ECG signal and the waveform of the pulse wave to obtain a specific form of tissue, which is used as a blood pressure model.
  • Input modeling can ensure the accuracy of subsequent acquisition of blood pressure values to a certain extent.
  • the deep neural network is essentially a convolutional neural network including several convolutional layers, pooling layers, batch normalization layers, and fully connected layers, and is established with systolic blood pressure and diastolic blood pressure as the output Model.
  • the sample processing module 702 includes:
  • the signal filtering unit 7021 is configured to use a band-pass filter to filter the sample pulse wave and the sample ECG signal to obtain the noise-reduced sample pulse wave and the noise-reduced sample ECG signal;
  • the waveform image acquisition unit 7022 is used to acquire the sample pulse waveform image and the sample electrocardiogram waveform image corresponding to the noise-reduction sample pulse wave and the noise-reduction sample ECG signal;
  • the normalization processing unit 7024 is used to perform normalization processing on the segmented signal to obtain a sample training signal.
  • sample pulse waveform diagram and the sample ECG waveform diagram corresponding to the noise-reduction sample pulse wave and the noise-reduction sample ECG signal and compare the sample pulse waveform diagram and the sample ECG waveform diagram according to a preset segmentation method Perform segmentation processing to obtain segmented signals; in order to be able to extract the waveforms of the ECG and pulse waves of the same period to obtain a specific form of tissue, which is used as the input of the subsequent blood pressure model for modeling, so as to obtain a continuous target for specific groups of people Deep neural network blood pressure model for blood pressure measurement.
  • the value of L is generally three times the sampling rate Fs.
  • the back end of the ECG is zero-filled, and the two ends of the PPG are zero-filled , So that the length of all samples is the same.
  • the sample training module 703 includes:
  • the signal sampling unit 7031 is used to perform a sliding sampling operation on the sample training signal to obtain multiple network training samples;
  • the network construction unit 7032 is used to construct a deep neural network-based convolutional neural network containing several computing units and use multiple network training samples to perform training operations on the convolutional neural network to obtain a pre-trained network;
  • the transfer learning unit 7033 is used to perform transfer learning on the sample signal using the pre-training network to obtain a deep neural network blood pressure model.
  • performing a sliding sampling operation on the sample training signal to obtain multiple network training samples is to implement a sliding window, ensure the accuracy of feature extraction of the sample training signal, and reduce subsequent training operations. Repeating the calculation of convolution can reduce the computational complexity to a certain extent and improve the efficiency of subsequent acquisition of blood pressure values.
  • a convolutional neural network containing several computing units based on a deep neural network is constructed and multiple network training samples are used to train the convolutional neural network to obtain a pre-trained network, specifically by constructing a deep neural network, It is specifically a convolutional neural network that includes several computing units.
  • Each computing unit includes a convolutional layer, a batch normalization layer, and a pooling layer. After 3-5 unit calculations, the obtained features are input to the first layer. After connecting the layers, a pre-trained network can be obtained that can simultaneously output systolic and diastolic blood pressures SBP p and DBP p.
  • a pre-training network is used to perform migration learning on sample signals to obtain a deep neural network blood pressure model.
  • the pre-training network can be specifically used to perform migration learning for specific population data, even if most parameters are frozen (ie Keep unchanged), only fine-tune the last fully connected layer of the pre-trained network, that is, make the new learning rate 1/10 of the original learning rate to obtain the deep neural network blood pressure model, where the fine-tuning process is similar to sampling the sample signal The process of preprocessing operations.
  • a convolutional neural network containing several computing units based on a deep neural network is constructed, and multiple network training samples are used to train the convolutional neural network using a stochastic gradient descent optimization algorithm.
  • a deep neural network-based convolutional neural network containing several computing units is constructed and multiple network training samples are used to train the convolutional neural network using a stochastic gradient descent optimization algorithm.
  • the Loss function is jointly established
  • the training of the model can output systolic and diastolic blood pressure at the same time. While accurately acquiring the blood pressure value of the collected object, it can also facilitate the follow-up to provide an early warning of the risk of sudden rise and fall in blood pressure according to the change in the output blood pressure value, and remind the patient to take it in time Seek medical attention or protection.
  • the deep neural network of this embodiment is specifically a convolutional neural network including several computing units, each computing unit includes a convolutional layer, a batch normalization layer, and a pooling layer. After 3-5 units of calculation, After inputting the obtained features into a fully connected layer, the systolic and diastolic blood pressures SBP p and DBP p can be output at the same time.
  • the model is trained by jointly establishing the Loss function, where the loss function Loss is the mean square error, which is specifically expressed as follows:
  • SBP t and DBP t are the actual systolic and diastolic blood pressure, and n is the number of samples;
  • SBP p and DBP p can be expressed as follows:
  • A is the activation function
  • W and b are the parameters of the model.
  • the gradient of W is calculated from the Loss function, and the gradient is continuously updated with a preset specific learning rate (quantitatively and randomly changes the value of W), where the learning rate is generally 0.0001, and the learning rate can be set according to actual application requirements, and there is no specific limitation here; then, the Loss is minimized along the direction of the gradient, and finally the training is stopped when the Loss no longer changes, or the training is stopped when the number of training reaches a certain value , Usually the number of training does not exceed 10,000 times.
  • the method further includes:
  • the history acquisition module 101 is used to acquire the historical systolic blood pressure and historical diastolic blood pressure of the collection object in the previous cycle;
  • the numerical value comparison module 102 is used to compare the historical systolic blood pressure and the historical diastolic blood pressure with the currently output systolic blood pressure and the diastolic blood pressure respectively;
  • the early warning prompt module 103 is configured to send an early warning prompt to the user terminal if the rate of change of the value is greater than the preset change threshold.
  • the blood pressure value of the collection object is continuously measured according to the preset time period, so as to monitor the change of the blood pressure value of the collection object within the time period, that is, the blood pressure can be provided according to the change of the output blood pressure value.
  • Early warning of sudden rise and fall risk remind patients to seek medical treatment or protection in time, specifically by comparing the historical systolic blood pressure and historical diastolic blood pressure with the current output systolic blood pressure and diastolic blood pressure respectively, that is, by calculating the historical systolic blood pressure and the current diastolic blood pressure.
  • the absolute value of the systolic blood pressure difference is divided by the historical systolic blood pressure to obtain the change rate of the systolic blood pressure. Similarly, the change rate of the diastolic blood pressure can be obtained.
  • the change threshold is set, an early warning reminder is issued to the user terminal, where the early warning reminder may be in the form of timely information, email, or signal alarm, etc., and there is no specific limitation here.
  • the preset change threshold is set according to actual application requirements, and is usually set to 20%.
  • the migration learning method based on the deep neural network of the present invention first establishes a pre-training model by using the ECG signal and the photoplethysmographic signal on the basis of the public database in the early stage; then, the migration is performed through the data for the specific population Learning, that is, updating a small number of parameters, on the one hand, it can avoid the subjective selection problem of manual feature extraction and the problem of feature point dependence, which is easy to implement, and on the other hand, it can be collected for specific populations, especially hypertension patients and arrhythmia patients. For objects that are difficult to monitor with the existing technology, convenient and accurate continuous blood pressure monitoring is realized.
  • the advantages of the blood pressure measurement method and device based on the deep neural network of the present invention are:
  • the disclosed technical content can be implemented in other ways.
  • the system embodiment described above is only illustrative.
  • the division of units may be a logical function division, and there may be other divisions in actual implementation.
  • multiple units or components may be combined or integrated into Another system, or some features can be ignored, or not implemented.
  • the displayed or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, units or modules, and may be in electrical or other forms.
  • the units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.
  • the functional units in the various embodiments of the present invention may be integrated into one processing unit, or each unit may exist alone physically, or two or more units may be integrated into one unit.
  • the above-mentioned integrated unit can be implemented in the form of hardware or software functional unit.

Landscapes

  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Physics & Mathematics (AREA)
  • Biophysics (AREA)
  • Public Health (AREA)
  • Veterinary Medicine (AREA)
  • Physiology (AREA)
  • General Health & Medical Sciences (AREA)
  • Animal Behavior & Ethology (AREA)
  • Surgery (AREA)
  • Molecular Biology (AREA)
  • Pathology (AREA)
  • Biomedical Technology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Medical Informatics (AREA)
  • Evolutionary Computation (AREA)
  • Fuzzy Systems (AREA)
  • Signal Processing (AREA)
  • Psychiatry (AREA)
  • Mathematical Physics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Cardiology (AREA)
  • Vascular Medicine (AREA)
  • Measuring Pulse, Heart Rate, Blood Pressure Or Blood Flow (AREA)

Abstract

一种基于深度神经网络的血压测量方法及装置。该方法及装置通过接收信号采集请求,信号采集请求至少携带有采集对象的标识信息;对采集对象进行信号采集操作,获取脉搏波以及心电信号;对脉搏波和心电信号进行预处理操作,得到预处理信号;基于训练好的深度神经网络血压模型,将预处理信号输入至深度神经网络血压模型中进行测量操作,得到与预处理信号相对应的收缩压以及舒张压;输出标识信息、收缩压以及舒张压。

Description

一种基于深度神经网络的血压测量方法及装置 技术领域
本发明涉及计算机技术领域,具体而言,涉及一种基于深度神经网络的血压测量方法及装置。
背景技术
高血压是心血管疾病的主要危险因素,而心血管疾病是全球第一大死亡原因。到2015年,全球高血压人口估计为11.3亿,包括全世界四分之一的男性和五分之一的女性。然而高血压常常没有明显的症状,以至于大多数人都难以察觉,因为被称为“沉默杀手”。实际上,超过50%的高血压人群不了解自己的病情,只有不到20%的高血压得到了控制。另一方面,超过12至24小时或更长时间的动态血压监测是帮助诊断和控制高血压的准确有效的方法,这对连续血压的监测提出了更高的要求。
目前,常用的连续血压测量方法一般分为直接测量法和间接测量法两种。直接测量法通常指动脉插管法,是连续血压测量的金标准,然而其技术要求高,准备时间长,且对人体有创,通常仅用于临床上危重患者及手术过程中的血压测量。间接测量法通过检测动脉管壁的脉搏搏动,血管容积变化或者脉搏波特征等参数间接得到血压,又称无创测量法,主要包含动脉张力法、容积补偿法、脉搏波传输速度(时间)法、脉搏波特征参数法等。
另外,动脉张力法需要传感器精确定位于动脉被压迫部位的正上方,长时间测量时难以保持稳定;容积补偿法需要预置参考压力,测量装置较复杂,测量精度易受到预置参考压的影响,且需压迫血管,舒适性差;脉搏波传输时间法是近些年研究较多的方法,脉搏波传输时间(Pulse Wave Transmit Time,PWTT)指动脉脉搏从近心端到远心端分支动脉的时间延迟,其与收缩压的相关性较强,然而仅通过脉搏波传输时间或脉搏波传输速度(Pulse Wave Velocity,PWV)与血压的关系建立模型的精度不高,所建立的模型普适性也较差;脉搏波特征参数法主要通过提取脉搏波波形特征建立与血压的模型,测量设备简单,但是也存在鲁棒性差的问题,对于特定人群如高血压或心率失常病人等,其对应的特征提取困难,往往难以实现准确的测量。
发明内容
本发明实施例提供了一种基于深度神经网络的血压测量方法及装置,以至少解决现有基于深度神经网络的血压测量系统判断准确度低的技术问题。
根据本发明的一实施例,提供了一种基于深度神经网络的血压测量方法,包括以下步骤:
接收信号采集请求,信号采集请求至少携带有采集对象的标识信息;
对采集对象进行信号采集操作,获取脉搏波以及心电信号;
对脉搏波和心电信号进行预处理操作,得到预处理信号;
基于训练好的深度神经网络血压模型,将预处理信号输入至深度神经网络血压模型中进行测量操作,得到与预处理信号相对应的收缩压以及舒张压;
输出标识信息、收缩压以及舒张压。
进一步地,该方法还包括:
获取数据库中的样本信号,样本信号至少携带样本脉搏波以及样本心电信号;
对样本脉搏波以及样本心电信号进行预处理操作,得到样本训练信号;
基于深度神经网络,将样本训练信号输入至深度神经网络中,按照迁移学习进行训练操作,得到深度神经网络血压模型。
进一步地,对样本信号进行样本预处理操作,得到样本训练信号的步骤包括:
使用带通滤波器对样本脉搏波以及样本心电信号进行过滤操作,得到降噪样本脉搏波以及降噪样本心电信号;
获取与降噪样本脉搏波以及降噪样本心电信号相对应的样本脉搏波形图以及样本心电波形图;
按照预设的分段方式对样本脉搏波形图以及样本心电波形图进行分段处理,得到分段信号;
对分段信号进行归一化处理,得到样本训练信号。
进一步地,基于深度神经网络,将样本训练信号输入至深度神经网络中,按照迁移学习进行训练操作,得到深度神经网络血压模型的步骤包括:
对样本训练信号进行滑动取样操作,获得多个网络训练样本;
构建基于深度神经网络的包含若干计算单元的卷积神经网络并使用多个网络训练样本对卷积神经网络进行训练操作,得到预训练网络;
使用预训练网络对样本信号进行迁移学习,得到深度神经网络血压模型。
进一步地,构建基于深度神经网络的包含若干计算单元的卷积神经网络并使用多个网络训练样本使用随机梯度下降优化算法对卷积神经网络进行训练。
进一步地,该方法还包括:
获取采集对象的上一周期的历史收缩压以及历史舒张压;
将历史收缩压以及历史舒张压分别与当前输出的收缩压以及舒张压进行数值比较;
若数值的变化率大于预设变化阈值,则向用户终端发出预警提示。
根据本发明的另一实施例,提供了一种基于深度神经网络的血压测量装置,包括:
请求接收模块,用于接收信号采集请求,信号采集请求至少携带有采集对象的标识信息;
信号采集模块,用于对采集对象进行信号采集操作,获取脉搏波以及心电信号;
预处理模块,用于对脉搏波和心电信号进行预处理操作,得到预处理信号;
血压测量模块,用于基于训练好的深度神经网络血压模型,将预处理信号输入至深度神经网络血压模型中进行测量操作,得到与预处理信号相对应的收缩压以及舒张压;
数据输出模块,用于输出标识信息、收缩压以及舒张压。
进一步地,该装置还包括:
样本获取模块,用于获取数据库中的样本信号,样本信号至少携带样本脉搏波以及样本心电信号;
样本处理模块,用于对样本脉搏波以及样本心电信号进行预处理操作,得到样本训练信号;
样本训练模块,用于基于深度神经网络,将样本训练信号输入至深度神经网络中,按照迁移学习进行训练操作,得到深度神经网络血压模型。
进一步地,该样本处理模块包括:
信号过滤单元,用于使用带通滤波器对样本脉搏波以及样本心电信号进行过滤操作,得到降噪样本脉搏波以及降噪样本心电信号;
波形图获取单元,用于获取与降噪样本脉搏波以及降噪样本心电信号相对应的样本脉搏波形图以及样本心电波形图;
分段处理单元,用于按照预设的分段方式对样本脉搏波形图以及样本心电波形图进行分段处理,得到分段信号;
归一化处理单元,用于对分段信号进行归一化处理,得到样本训练信号。
进一步地,该样本训练模块包括:
信号取样单元,用于对样本训练信号进行滑动取样操作,获得多个网络训练样本;
网络构建单元,用于构建基于深度神经网络的包含若干计算单元的卷积神经网络并使用多个网络训练样本对卷积神经网络进行训练操作,得到预训练网络;
迁移学习单元,用于使用预训练网络对样本信号进行迁移学习,得到深度神经网络血压模型。
本发明实施例中的基于深度神经网络的血压测量方法及装置,通过接收信号采集请求,该信号采集请求至少携带有采集对象的标识信息;进而根据该信号采集请求对采集对象进行信号采集操作,获取该采集对象的脉搏波以及心电信号;然后,对获取到的脉搏波和心电信号进行预处理操作,以得到降噪低干扰的预处理信号,在一定程度上保证后续准确获取血压值的效率;进而,基于训练好的深度神经网络血压模型,将预处理信号输入至深度神经网络血压模型中进行测量操作,得到与预处理信号相对应的收缩压以及舒张压,能够保证血压值获取的准确性和完整性;输出标识信息、收缩压以及舒张压,本发明整个过程操作简单、计算复杂度低、成本低; 本发明基于深度神经网络的血压测量方法及装置获取血压值速度快、获取血压值准确度高。
附图说明
此处所说明的附图用来提供对本发明的进一步理解,构成本申请的一部分,本发明的示意性实施例及其说明用于解释本发明,并不构成对本发明的不当限定。在附图中:
图1为本发明基于深度神经网络的血压测量方法的流程图;
图2为本发明基于深度神经网络的血压测量方法的一优选流程图;
图3为本发明基于深度神经网络的血压测量方法的又一优选流程图;
图4为本发明基于深度神经网络的血压测量方法的另一优选流程图;
图5为本发明基于深度神经网络的血压测量方法的再一优选流程图;
图6为本发明基于深度神经网络的血压测量装置的示意图;
图7为本发明基于深度神经网络的血压测量装置的一优选示意图;
图8为本发明基于深度神经网络的血压测量装置的又一优选示意图;
图9为本发明基于深度神经网络的血压测量装置的另一优选示意图;
图10为本发明基于深度神经网络的血压测量装置的再一优选示意图;
图11为本发明基于深度神经网络的血压测量方法的单个降噪样本心电信号波形图ECG以及降噪样本脉搏波波形图PPG;
图12为本发明基于深度神经网络的血压测量方法的收缩压预测值与目标值对比曲线图。
具体实施方式
为了使本技术领域的人员更好地理解本发明方案,下面将结合本发明实施例中的附图,对本发明实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例仅仅是本发明一部分的实施例,而不是全部的实施例。基于本发明中的实施例,本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例,都应当属于本发明保护的范围。
需要说明的是,本发明的说明书和权利要求书及上述附图中的术语“第一”、“第二”等是用于区别类似的对象,而不必用于描述特定的顺序或先后次序。应该理解这样使用的数据在适当情况下可以互换,以便这里描述的本发明的实施例能够以除了在这里图示或描述的那些以外的顺序实施。此外,术语“包括”和“具有”以及他们的任何变形,意图在于覆盖不排他的包含,例如,包含了一系列步骤或单元的过程、方法、系统、产品或设备不必限于清楚地列出的那些步骤或单元,而是可包括没有清楚地列出的或对于这些过程、方法、产品或设备固有的其它步骤或单元。
实施例1
根据本发明一实施例,提供了一种基于深度神经网络的血压测量方法,参见图1,包括以下步骤:
S1:接收信号采集请求,信号采集请求至少携带有采集对象的标识信 息;
S2:对采集对象进行信号采集操作,获取脉搏波以及心电信号;
S3:对脉搏波和心电信号进行预处理操作,得到预处理信号;
S4:基于训练好的深度神经网络血压模型,将预处理信号输入至深度神经网络血压模型中进行测量操作,得到与预处理信号相对应的收缩压以及舒张压;
S5:输出标识信息、收缩压以及舒张压。
本发明实施例中的基于深度神经网络的血压测量方法,通过接收信号采集请求,该信号采集请求至少携带有采集对象的标识信息;进而根据该信号采集请求对采集对象进行信号采集操作,获取该采集对象的脉搏波以及心电信号;然后,对获取到的脉搏波和心电信号进行预处理操作,以得到降噪低干扰的预处理信号,在一定程度上保证后续准确获取血压值的效率;进而,基于训练好的深度神经网络血压模型,将预处理信号输入至深度神经网络血压模型中进行测量操作,得到与预处理信号相对应的收缩压以及舒张压,能够保证血压值获取的准确性和完整性;输出标识信息、收缩压以及舒张压,本发明整个过程操作简单、计算复杂度低、成本低;本发明基于深度神经网络的血压测量方法获取血压值速度快、获取血压值准确度高。
需要说明的是,本实施例中,对采集对象进行信号采集操作,获取脉搏波以及心电信号,具体可以是通过对采集对象的手指指尖处,或手腕桡动脉、颈部颈动脉等邻近体表的主动脉脉搏处进行脉搏波以及心电信号采 集。
其中,脉搏波是指采集对象的光电容积脉搏波信号;心电信号是指采集对象的单导联心电信号,其中,采集对象的手指的光电容积脉搏波,信号长度为5分钟,采样频率为250Hz。
在本实施例中,对脉搏波和心电信号进行预处理操作,得到预处理信号,具体可以是通过对获取到的脉搏波和心电信号进行滤波去噪、逐心拍分割、归一化等信号处理,以降低信号中的杂质和干扰,在一定程度上保证后续准确获取采集对象的血压值的效率。
在本实施例中,基于训练好的深度神经网络血压模型,将预处理信号输入至深度神经网络血压模型中进行测量操作,得到与预处理信号相对应的收缩压以及舒张压,具体可以是通过将处理完的预处理信号的信号波形作为输入,输入至利用迁移学习训练得到的针对特定人群的深度神经网络血压模型中,对采集对象的连续血压进行计算,得出该采集对象逐拍的收缩压和舒张压并输出。
例如,如表1所示,对于血压水平不同的各心率失常患者,本实施例可以测得的收缩压和舒张压,与实际值比较,均满足美国医疗器械促进协会(AMMI)推荐的平均误差小于等于5mmHg,标准误差小于等于8mmHg的标准。
表1对心率失常病人的血压预测结果
Figure PCTCN2020139658-appb-000001
Figure PCTCN2020139658-appb-000002
作为优选的技术方案中,参见图2,该方法还包括:
S6:获取数据库中的样本信号,样本信号至少携带样本脉搏波以及样本心电信号;
S7:对样本脉搏波以及样本心电信号进行预处理操作,得到样本训练信号;
S8:基于深度神经网络,将样本训练信号输入至深度神经网络中,按照迁移学习进行训练操作,得到深度神经网络血压模型。
在本实施例中,获取数据库中的样本信号,样本信号至少携带样本脉搏波以及样本心电信号具体可以是通过读取本地公共数据库,在该公共数据库中获取预先采集到的大量的如高血压病人及心率失常患者的样本脉搏波以及样本心电信号,便于后续深度神经网络血压模型的建立。
进一步地,本实施例通过对获取到样本脉搏波以及样本心电信号进行预处理操作,具体是通过提取同周期的心电信号和脉搏波的波形,以获取特定形式的组织,作为血压模型的输入进行建模,能够在一定程度上保证后续获取血压值的准确性。
进一步地,本实施例基于深度神经网络,将样本训练信号输入至深度神经网络中,按照迁移学习进行训练操作,具体可以是通过使用深度神经 网络对如心率失常人群的处理后的样本训练信号进行迁移学习,获得深度神经网络血压模型,利用迁移方法训练模型的模型训练成本小,能够输出逐心拍连续血压,测量准确度高,以保证获取血压值的准确性。
需要说明的是,本实施例中,深度神经网络,其实质为包括若干卷积层、池化层、批标准化层及全连接层的卷积神经网络,并以收缩压和舒张压为输出建立的模型。
作为优选的技术方案中,参见图3,对样本信号进行样本预处理操作,得到样本训练信号的步骤包括:
S701:使用带通滤波器对样本脉搏波以及样本心电信号进行过滤操作,得到降噪样本脉搏波以及降噪样本心电信号;
S702:获取与降噪样本脉搏波以及降噪样本心电信号相对应的样本脉搏波形图以及样本心电波形图;
S703:按照预设的分段方式对样本脉搏波形图以及样本心电波形图进行分段处理,得到分段信号;
S704:对分段信号进行归一化处理,得到样本训练信号。
具体地,使用带通滤波器对样本脉搏波以及样本心电信号进行过滤操作,得到降噪样本脉搏波以及降噪样本心电信号,本实施例具体是通过使用带通滤波器滤除样本脉搏波以及样本心电信号中的基线漂移和高频干扰,其中,降噪样本心电信号通常为0.5-35Hz,降噪样本脉搏波信号通常为0.3-6Hz,能够降低样本脉搏波以及样本心电信号中的杂质和干扰,能 够在一定程度上保证后续获取血压值的准确性。
进一步地,获取与降噪样本脉搏波以及降噪样本心电信号相对应的样本脉搏波形图以及样本心电波形图,并按照预设的分段方式对样本脉搏波形图以及样本心电波形图进行分段处理,得到分段信号;是为了能够提取同周期的心电和脉搏波的波形,以获取特定形式的组织,作为后续血压模型的输入进行建模,以得到能够针对特定人群进行连续血压测量的深度神经网络血压模型。
其中,参见图11,本实施例中预设的分段方式可以是根据信号的AB分割点坐标,获得将若干分段信号
Figure PCTCN2020139658-appb-000003
即具体是通过检测降噪样本心电信号的R波波峰与脉搏波波谷点,其中,使相邻R波波峰间中点为A(i),(i=1,2,...),同时,使对应周期的降噪样本脉搏波波谷点为B(i),(i=1,2,...);然后,根据相对应的AB点将信号分割为
Figure PCTCN2020139658-appb-000004
段;进而,为使每段信号等长,补零使其长度为L,获得分段信号
Figure PCTCN2020139658-appb-000005
如图11所示,其中,L的数值一般为采样率Fs的三倍,如图11所示,作为单个样本的心电ECG和脉搏波PPG波形,ECG后端补零,PPG两端补零,使得所有样本的长度一致。
进一步地,对分段信号进行归一化处理,具体是通过使其值位于[0,1]区间以获得该样本训练信号。
需要说明的是,本实施例根据信号波形自主提取特征建模,对波形质量要求不高,不需检测重搏波,易于检点。
作为优选的技术方案中,参见图4,基于深度神经网络,将样本训练 信号输入至深度神经网络中,按照迁移学习进行训练操作,得到深度神经网络血压模型的步骤包括:
S801:对样本训练信号进行滑动取样操作,获得多个网络训练样本;
S802:构建基于深度神经网络的包含若干计算单元的卷积神经网络并使用多个网络训练样本对卷积神经网络进行训练操作,得到预训练网络;
S803:使用预训练网络对样本信号进行迁移学习,得到深度神经网络血压模型。
需要说明的是,在本实施例中,对样本训练信号进行滑动取样操作,获得多个网络训练样本,是为了实现滑动窗口,保证对样本训练信号的特征提取的准确性,减少后续训练操作中重复卷积的计算,能够在一定程度上减少计算复杂度,提高后续获取血压值的效率。
在本实施例中,构建基于深度神经网络的包含若干计算单元的卷积神经网络并使用多个网络训练样本对卷积神经网络进行训练操作,得到预训练网络,具体是通过构建深度神经网络,其具体为包含若干计算单元的卷积神经网络,每个计算单元包含卷积层、批标准化层、池化层,在经过3-5个单元的计算之后,将获得的特征输入至1层全连接层后,能够得到可以同时输出收缩压和舒张压SBP p和DBP p的预训练网络。
具体地,使用预训练网络对样本信号进行迁移学习,得到深度神经网络血压模型,在本实施例中,具体可以是通过利用预训练网络针对特定人群数据进行迁移学习,即使大多数参数冻结(即保持不变),仅微调预训练网络的最后一层全连接层,即使得新学习率为原始学习率的1/10,以获 得深度神经网络血压模型,其中,微调过程类似对样本信号进行样本预处理操作的过程。
作为优选的技术方案中,构建基于深度神经网络的包含若干计算单元的卷积神经网络并使用多个网络训练样本使用随机梯度下降优化算法对卷积神经网络进行训练。
具体地,构建基于深度神经网络的包含若干计算单元的卷积神经网络并使用多个网络训练样本使用随机梯度下降优化算法对卷积神经网络进行训练,在本实施例中,通过联合建立Loss函数进行模型的训练,能够同时输出收缩压和舒张压,在准确获取到采集对象的血压值的同时,还能方便后续根据输出血压值的变化情况提供血压骤升骤降风险预警,提醒患者进行及时就医或防护。
其中,由于本实施例的深度神经网络具体为包含若干计算单元的卷积神经网络,每个计算单元包含卷积层、批标准化层、池化层,在经过3-5个单元的计算之后,将获得的特征输入至1层全连接层后,能够同时输出收缩压和舒张压SBP p和DBP p
进一步地,通过联合建立Loss函数进行模型的训练,其中,损失函数Loss为均方误差,具体表示如下:
Figure PCTCN2020139658-appb-000006
其中,SBP t和DBP t为实际的收缩压和舒张压,n为样本数量;
其中,SBP p和DBP p可以表示如下:
Figure PCTCN2020139658-appb-000007
其中,A为激活函数,W,b为模型的参数,对Loss函数求W的梯度,并以预先设置的特定的学习率不断更新梯度(定量随机改变W的值),其中,学习率一般为0.0001,而学习率可以根据实际应用需求进行设置,此处不作具体限制;然后,使Loss沿梯度下降的方向趋于最小,最终Loss不再变化时停止训练,或训练次数到达一定值时停止训练,通常训练次数不超过10000次。
进一步地,如图12所示,图12为一组采集对象测得的血压收缩压值与实际目标收缩压的对比,可以发现在各个血压区间内,预测值与目标值都误差较小。
作为优选的技术方案中,参见图5,该方法还包括:
S9:获取采集对象的上一周期的历史收缩压以及历史舒张压;
S10:将历史收缩压以及历史舒张压分别与当前输出的收缩压以及舒张压进行数值比较;
S11:若数值的变化率大于预设变化阈值,则向用户终端发出预警提示。
具体地,在本实施例,按照预先设置的时间周期对采集对象进行连续的血压值测量,以便于监测时间周期内的采集对象的血压值的变化,即能够根据输出血压值的变化情况提供血压骤升骤降风险预警,提醒患者进行及时就医或防护,具体可以是通过将历史收缩压以及历史舒张压分别与当 前输出的收缩压以及舒张压进行数值比较,即先通过计算历史收缩压与当前的收缩压的差值的绝对值,再利用该差值的绝对值除以历史收缩压,得到该收缩压的变化率,同理可获取舒张压的变化率,进而,当该变化率大于预设变化阈值时,则向用户终端发出预警提示,其中,该预警提示可以是及时信息、邮件或信号警报等形式,此处不作具体限制。
其中,预设变化阈值是根据实际应用需求进行设置的,通常设置为20%。
实施例2
根据本发明的另一实施例,提供了一种基于深度神经网络的血压测量装置,参见图6,包括:
请求接收模块601,用于接收信号采集请求,信号采集请求至少携带有采集对象的标识信息;
信号采集模块602,用于对采集对象进行信号采集操作,获取脉搏波以及心电信号;
预处理模块603,用于对脉搏波和心电信号进行预处理操作,得到预处理信号;
血压测量模块604,用于基于训练好的深度神经网络血压模型,将预处理信号输入至深度神经网络血压模型中进行测量操作,得到与预处理信号相对应的收缩压以及舒张压;
数据输出模块605,用于输出标识信息、收缩压以及舒张压。
本发明实施例中的基于深度神经网络的血压测量装置,通过接收信号采集请求,该信号采集请求至少携带有采集对象的标识信息;进而根据该信号采集请求对采集对象进行信号采集操作,获取该采集对象的脉搏波以及心电信号;然后,对获取到的脉搏波和心电信号进行预处理操作,以得到降噪低干扰的预处理信号,在一定程度上保证后续准确获取血压值的效率;进而,基于训练好的深度神经网络血压模型,将预处理信号输入至深度神经网络血压模型中进行测量操作,得到与预处理信号相对应的收缩压以及舒张压,能够保证血压值获取的准确性和完整性;输出标识信息、收缩压以及舒张压,本发明整个过程操作简单、计算复杂度低、成本低;本发明基于深度神经网络的血压测量装置获取血压值速度快、获取血压值准确度高。
需要说明的是,本实施例中,对采集对象进行信号采集操作,获取脉搏波以及心电信号,具体可以是通过对采集对象的手指指尖处,或手腕桡动脉、颈部颈动脉等邻近体表的主动脉脉搏处进行脉搏波以及心电信号采集。
其中,脉搏波是指采集对象的光电容积脉搏波信号;心电信号是指采集对象的单导联心电信号,其中,采集对象的手指的光电容积脉搏波,信号长度为5分钟,采样频率为250Hz。
在本实施例中,对脉搏波和心电信号进行预处理操作,得到预处理信号,具体可以是通过对获取到的脉搏波和心电信号进行滤波去噪、逐心拍分割、归一化等信号处理,以降低信号中的杂质和干扰,在一定程度上保 证后续准确获取采集对象的血压值的效率。
在本实施例中,基于训练好的深度神经网络血压模型,将预处理信号输入至深度神经网络血压模型中进行测量操作,得到与预处理信号相对应的收缩压以及舒张压,具体可以是通过将处理完的预处理信号的信号波形作为输入,输入至利用迁移学习训练得到的针对特定人群的深度神经网络血压模型中,对采集对象的连续血压进行计算,得出该采集对象逐拍的收缩压和舒张压并输出。
作为优选的技术方案中,参见图7,该装置还包括:
样本获取模块701,用于获取数据库中的样本信号,样本信号至少携带样本脉搏波以及样本心电信号;
样本处理模块702,用于对样本脉搏波以及样本心电信号进行预处理操作,得到样本训练信号;
样本训练模块703,用于基于深度神经网络,将样本训练信号输入至深度神经网络中,按照迁移学习进行训练操作,得到深度神经网络血压模型。
在本实施例中,获取数据库中的样本信号,样本信号至少携带样本脉搏波以及样本心电信号具体可以是通过读取本地公共数据库,在该公共数据库中获取预先采集到的大量的如高血压病人及心率失常患者的样本脉搏波以及样本心电信号,便于后续深度神经网络血压模型的建立。
进一步地,本实施例通过对获取到样本脉搏波以及样本心电信号进行 预处理操作,具体是通过提取同周期的心电信号和脉搏波的波形,以获取特定形式的组织,作为血压模型的输入进行建模,能够在一定程度上保证后续获取血压值的准确性。
进一步地,本实施例基于深度神经网络,将样本训练信号输入至深度神经网络中,按照迁移学习进行训练操作,具体可以是通过使用深度神经网络对如心率失常人群的处理后的样本训练信号进行迁移学习,获得深度神经网络血压模型,利用迁移方法训练模型的模型训练成本小,能够输出逐心拍连续血压,测量准确度高,以保证获取血压值的准确性。
需要说明的是,本实施例中,深度神经网络,其实质为包括若干卷积层、池化层、批标准化层及全连接层的卷积神经网络,并以收缩压和舒张压为输出建立的模型。
作为优选的技术方案中,参见图8,该样本处理模块702包括:
信号过滤单元7021,用于使用带通滤波器对样本脉搏波以及样本心电信号进行过滤操作,得到降噪样本脉搏波以及降噪样本心电信号;
波形图获取单元7022,用于获取与降噪样本脉搏波以及降噪样本心电信号相对应的样本脉搏波形图以及样本心电波形图;
分段处理单元7023,用于按照预设的分段方式对样本脉搏波形图以及样本心电波形图进行分段处理,得到分段信号;
归一化处理单元7024,用于对分段信号进行归一化处理,得到样本训练信号。
具体地,使用带通滤波器对样本脉搏波以及样本心电信号进行过滤操作,得到降噪样本脉搏波以及降噪样本心电信号,本实施例具体是通过使用带通滤波器滤除样本脉搏波以及样本心电信号中的基线漂移和高频干扰,其中,降噪样本心电信号通常为0.5-35Hz,降噪样本脉搏波信号通常为0.3-6Hz,能够降低样本脉搏波以及样本心电信号中的杂质和干扰,能够在一定程度上保证后续获取血压值的准确性。
进一步地,获取与降噪样本脉搏波以及降噪样本心电信号相对应的样本脉搏波形图以及样本心电波形图,并按照预设的分段方式对样本脉搏波形图以及样本心电波形图进行分段处理,得到分段信号;是为了能够提取同周期的心电和脉搏波的波形,以获取特定形式的组织,作为后续血压模型的输入进行建模,以得到能够针对特定人群进行连续血压测量的深度神经网络血压模型。
其中,参见图11,本实施例中预设的分段方式可以是根据信号的AB分割点坐标,获得将若干分段信号
Figure PCTCN2020139658-appb-000008
即具体是通过检测降噪样本心电信号的R波波峰与脉搏波波谷点,其中,使相邻R波波峰间中点为A(i),(i=1,2,...),同时,使对应周期的降噪样本脉搏波波谷点为B(i),(i=1,2,...);然后,根据相对应的AB点将信号分割为
Figure PCTCN2020139658-appb-000009
段;进而,为使每段信号等长,补零使其长度为L,获得分段信号
Figure PCTCN2020139658-appb-000010
如图11所示,其中,L的数值一般为采样率Fs的三倍,如图11所示,作为单个样本的心电ECG和脉搏波PPG波形,ECG后端补零,PPG两端补零,使得所有样本的长度一致。
进一步地,对分段信号进行归一化处理,具体是通过使其值位于[0,1]区间以获得该样本训练信号。
需要说明的是,本实施例根据信号波形自主提取特征建模,对波形质量要求不高,不需检测重搏波,易于检点。
作为优选的技术方案中,参见图9,该样本训练模块703包括:
信号取样单元7031,用于对样本训练信号进行滑动取样操作,获得多个网络训练样本;
网络构建单元7032,用于构建基于深度神经网络的包含若干计算单元的卷积神经网络并使用多个网络训练样本对卷积神经网络进行训练操作,得到预训练网络;
迁移学习单元7033,用于使用预训练网络对样本信号进行迁移学习,得到深度神经网络血压模型。
需要说明的是,在本实施例中,对样本训练信号进行滑动取样操作,获得多个网络训练样本,是为了实现滑动窗口,保证对样本训练信号的特征提取的准确性,减少后续训练操作中重复卷积的计算,能够在一定程度上减少计算复杂度,提高后续获取血压值的效率。
在本实施例中,构建基于深度神经网络的包含若干计算单元的卷积神经网络并使用多个网络训练样本对卷积神经网络进行训练操作,得到预训练网络,具体是通过构建深度神经网络,其具体为包含若干计算单元的卷积神经网络,每个计算单元包含卷积层、批标准化层、池化层,在经过3-5 个单元的计算之后,将获得的特征输入至1层全连接层后,能够得到可以同时输出收缩压和舒张压SBP p和DBP p的预训练网络。
具体地,使用预训练网络对样本信号进行迁移学习,得到深度神经网络血压模型,在本实施例中,具体可以是通过利用预训练网络针对特定人群数据进行迁移学习,即使大多数参数冻结(即保持不变),仅微调预训练网络的最后一层全连接层,即使得新学习率为原始学习率的1/10,以获得深度神经网络血压模型,其中,微调过程类似对样本信号进行样本预处理操作的过程。
作为优选的技术方案中,构建基于深度神经网络的包含若干计算单元的卷积神经网络并使用多个网络训练样本使用随机梯度下降优化算法对卷积神经网络进行训练。
具体地,构建基于深度神经网络的包含若干计算单元的卷积神经网络并使用多个网络训练样本使用随机梯度下降优化算法对卷积神经网络进行训练,在本实施例中,通过联合建立Loss函数进行模型的训练,能够同时输出收缩压和舒张压,在准确获取到采集对象的血压值的同时,还能方便后续根据输出血压值的变化情况提供血压骤升骤降风险预警,提醒患者进行及时就医或防护。
其中,由于本实施例的深度神经网络具体为包含若干计算单元的卷积神经网络,每个计算单元包含卷积层、批标准化层、池化层,在经过3-5个单元的计算之后,将获得的特征输入至1层全连接层后,能够同时输出收缩压和舒张压SBP p和DBP p
进一步地,通过联合建立Loss函数进行模型的训练,其中,损失函数Loss为均方误差,具体表示如下:
Figure PCTCN2020139658-appb-000011
其中,SBP t和DBP t为实际的收缩压和舒张压,n为样本数量;
其中,SBP p和DBP p可以表示如下:
Figure PCTCN2020139658-appb-000012
其中,A为激活函数,W,b为模型的参数,对Loss函数求W的梯度,并以预先设置的特定的学习率不断更新梯度(定量随机改变W的值),其中,学习率一般为0.0001,而学习率可以根据实际应用需求进行设置,此处不作具体限制;然后,使Loss沿梯度下降的方向趋于最小,最终Loss不再变化时停止训练,或训练次数到达一定值时停止训练,通常训练次数不超过10000次。
作为优选的技术方案中,参见图10,该方法还包括:
历史获取模块101,用于获取采集对象的上一周期的历史收缩压以及历史舒张压;
数值比较模块102,用于将历史收缩压以及历史舒张压分别与当前输出的收缩压以及舒张压进行数值比较;
预警提示模块103,用于若数值的变化率大于预设变化阈值,则向用户终端发出预警提示。
具体地,在本实施例,按照预先设置的时间周期对采集对象进行连续的血压值测量,以便于监测时间周期内的采集对象的血压值的变化,即能够根据输出血压值的变化情况提供血压骤升骤降风险预警,提醒患者进行及时就医或防护,具体可以是通过将历史收缩压以及历史舒张压分别与当前输出的收缩压以及舒张压进行数值比较,即先通过计算历史收缩压与当前的收缩压的差值的绝对值,再利用该差值的绝对值除以历史收缩压,得到该收缩压的变化率,同理可获取舒张压的变化率,进而,当该变化率大于预设变化阈值时,则向用户终端发出预警提示,其中,该预警提示可以是及时信息、邮件或信号警报等形式,此处不作具体限制。
其中,预设变化阈值是根据实际应用需求进行设置的,通常设置为20%。
需要说明的是,本发明基于深度神经网络的迁移学习方法,先通过前期在公共数据库的基础上利用心电信号和光电容积脉搏波信号建立预训练模型;然后,通过针对特定人群的数据进行迁移学习,即更新少量参数,能够在一方面避免人工特征提取的主观挑选问题和特征点依赖问题,易于实施,以及另一方面可针对特定人群采集的数据,尤其是高血压病人及心率失常患者等已有技术难以监测的对象,实现方便、准确的连续血压监测。
与现有的血压测量系统相比,本发明基于深度神经网络的血压测量方法及装置的优点在于:
1.过程操作简单、计算复杂度低、成本低;
2.血压值获取速度快;
3.血压值获取准确度高;
4.能够根据信号波形自主提取特征建模,对波形质量要求不高,不需检测重搏波,易于检点;
5.能够针对特定人群进行迁移学习,使得模型训练成本小,可以输出逐心拍连续血压,且测量准确度高;
6.能够提供血压骤升骤降的风险预警。
上述本发明实施例序号仅仅为了描述,不代表实施例的优劣。
在本发明的上述实施例中,对各个实施例的描述都各有侧重,某个实施例中没有详述的部分,可以参见其他实施例的相关描述。
在本申请所提供的几个实施例中,应该理解到,所揭露的技术内容,可通过其它的方式实现。其中,以上所描述的系统实施例仅仅是示意性的,例如单元的划分,可以为一种逻辑功能划分,实际实现时可以有另外的划分方式,例如多个单元或组件可以结合或者可以集成到另一个系统,或一些特征可以忽略,或不执行。另一点,所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口,单元或模块的间接耦合或通信连接,可以是电性或其它的形式。
作为分离部件说明的单元可以是或者也可以不是物理上分开的,作为单元显示的部件可以是或者也可以不是物理单元,即可以位于一个地方,或者也可以分布到多个单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
另外,在本发明各个实施例中的各功能单元可以集成在一个处理单元中,也可以是各个单元单独物理存在,也可以两个或两个以上单元集成在一个单元中。上述集成的单元既可以采用硬件的形式实现,也可以采用软件功能单元的形式实现。
集成的单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时,可以存储在一个计算机可读取存储介质中。基于这样的理解,本发明的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的全部或部分可以以软件产品的形式体现出来,该计算机软件产品存储在一个存储介质中,包括若干指令用以使得一台计算机设备(可为个人计算机、服务器或者网络设备等)执行本发明各个实施例方法的全部或部分步骤。而前述的存储介质包括:U盘、只读存储器(ROM,Read-Only Memory)、随机存取存储器(RAM,Random Access Memory)、移动硬盘、磁碟或者光盘等各种可以存储程序代码的介质。
以上所述仅是本发明的优选实施方式,应当指出,对于本技术领域的普通技术人员来说,在不脱离本发明原理的前提下,还可以做出若干改进和润饰,这些改进和润饰也应视为本发明的保护范围。

Claims (10)

  1. 一种基于深度神经网络的血压测量方法,其特征在于,包括以下步骤:
    接收信号采集请求,所述信号采集请求至少携带有采集对象的标识信息;
    对所述采集对象进行信号采集操作,获取脉搏波以及心电信号;
    对所述脉搏波和心电信号进行预处理操作,得到预处理信号;
    基于训练好的深度神经网络血压模型,将所述预处理信号输入至所述深度神经网络血压模型中进行测量操作,得到与所述预处理信号相对应的收缩压以及舒张压;
    输出所述标识信息、所述收缩压以及所述舒张压。
  2. 根据权利要求1所述的基于深度神经网络的血压测量方法,其特征在于,在所述基于训练好的深度神经网络血压模型的步骤之前,所述方法还包括:
    获取数据库中的样本信号,所述样本信号至少携带样本脉搏波以及样本心电信号;
    对所述样本脉搏波以及样本心电信号进行所述预处理操作,得到样本训练信号;
    基于深度神经网络,将所述样本训练信号输入至所述深度神经网络中,按照迁移学习进行训练操作,得到所述深度神经网络血压模型。
  3. 根据权利要求2所述的基于深度神经网络的血压测量方法,其特征在于,所述对所述样本信号进行样本预处理操作,得到样本训练信号的步骤包括:
    使用带通滤波器对所述样本脉搏波以及样本心电信号进行过滤操作,得到降噪样本脉搏波以及降噪样本心电信号;
    获取与所述降噪样本脉搏波以及降噪样本心电信号相对应的样本脉搏波形图以及样本心电波形图;
    按照预设的分段方式对所述样本脉搏波形图以及样本心电波形图进行分段处理,得到分段信号;
    对所述分段信号进行归一化处理,得到所述样本训练信号。
  4. 根据权利要求2所述的基于深度神经网络的血压测量方法,其特征在于,所述基于深度神经网络,将所述样本训练信号输入至所述深度神经网络中,按照迁移学习进行训练操作,得到所述深度神经网络血压模型的步骤包括:
    对所述样本训练信号进行滑动取样操作,获得多个网络训练样本;
    构建基于深度神经网络的包含若干计算单元的卷积神经网络并使用所述多个网络训练样本对所述卷积神经网络进行训练操作,得到预训练网络;
    使用所述预训练网络对所述样本信号进行迁移学习,得到所述深度神经网络血压模型。
  5. 根据权利要求4所述的基于深度神经网络的血压测量方法,其特征在于,构建基于深度神经网络的包含若干计算单元的卷积神经网络并使用多个网络训练样本使用随机梯度下降优化算法对所述卷积神经网络进行训练。
  6. 根据权利要求1所述的基于深度神经网络的血压测量方法,其特征在于,在所述输出所述标识信息、所述收缩压以及所述舒张压的步骤之后,所述方法还包括:
    获取所述采集对象的上一周期的历史收缩压以及历史舒张压;
    将所述历史收缩压以及历史舒张压分别与当前输出的所述收缩压以及舒张压进行数值比较;
    若数值的变化率大于预设变化阈值,则向用户终端发出预警提示。
  7. 一种基于深度神经网络的血压测量装置,其特征在于,包括:
    请求接收模块,用于接收信号采集请求,所述信号采集请求至少携带有采集对象的标识信息;
    信号采集模块,用于对所述采集对象进行信号采集操作,获取脉搏波以及心电信号;
    预处理模块,用于对所述脉搏波和心电信号进行预处理操作,得到预处理信号;
    血压测量模块,用于基于训练好的深度神经网络血压模型,将所述预处理信号输入至所述深度神经网络血压模型中进行测量操作,得到与所述 预处理信号相对应的收缩压以及舒张压;
    数据输出模块,用于输出所述标识信息、所述收缩压以及所述舒张压。
  8. 根据权利要求7所述的基于深度神经网络的血压测量装置,其特征在于,所述装置还包括:
    样本获取模块,用于获取数据库中的样本信号,所述样本信号至少携带样本脉搏波以及样本心电信号;
    样本处理模块,用于对所述样本脉搏波以及样本心电信号进行所述预处理操作,得到样本训练信号;
    样本训练模块,用于基于深度神经网络,将所述样本训练信号输入至所述深度神经网络中,按照迁移学习进行训练操作,得到所述深度神经网络血压模型。
  9. 根据权利要求8所述的基于深度神经网络的血压测量装置,其特征在于,所述样本处理模块包括:
    信号过滤单元,用于使用带通滤波器对所述样本脉搏波以及样本心电信号进行过滤操作,得到降噪样本脉搏波以及降噪样本心电信号;
    波形图获取单元,用于获取与所述降噪样本脉搏波以及降噪样本心电信号相对应的样本脉搏波形图以及样本心电波形图;
    分段处理单元,用于按照预设的分段方式对所述样本脉搏波形图以及样本心电波形图进行分段处理,得到分段信号;
    归一化处理单元,用于对所述分段信号进行归一化处理,得到所述样 本训练信号。
  10. 根据权利要求8所述的基于深度神经网络的血压测量装置,其特征在于,所述样本训练模块包括:
    信号取样单元,用于对所述样本训练信号进行滑动取样操作,获得多个网络训练样本;
    网络构建单元,用于构建基于深度神经网络的包含若干计算单元的卷积神经网络并使用所述多个网络训练样本对所述卷积神经网络进行训练操作,得到预训练网络;
    迁移学习单元,用于使用所述预训练网络对所述样本信号进行迁移学习,得到所述深度神经网络血压模型。
PCT/CN2020/139658 2020-04-13 2020-12-25 一种基于深度神经网络的血压测量方法及装置 WO2021208490A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010285645.XA CN111493850A (zh) 2020-04-13 2020-04-13 一种基于深度神经网络的血压测量方法及装置
CN202010285645.X 2020-04-13

Publications (1)

Publication Number Publication Date
WO2021208490A1 true WO2021208490A1 (zh) 2021-10-21

Family

ID=71848453

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/139658 WO2021208490A1 (zh) 2020-04-13 2020-12-25 一种基于深度神经网络的血压测量方法及装置

Country Status (2)

Country Link
CN (1) CN111493850A (zh)
WO (1) WO2021208490A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115130509A (zh) * 2022-06-29 2022-09-30 哈尔滨工业大学(威海) 基于条件式变分自编码器的心电信号生成方法
CN115270848A (zh) * 2022-06-17 2022-11-01 合肥心之声健康科技有限公司 一种ppg与ecg自动转换智能算法、存储介质和计算机系统
CN115836846A (zh) * 2022-12-14 2023-03-24 北京航空航天大学 一种基于自监督迁移学习的无创血压估计方法
CN116383617A (zh) * 2023-04-21 2023-07-04 复旦大学 一种基于脉搏波波形特征的智能血压检测方法及系统
CN116803340A (zh) * 2023-07-19 2023-09-26 北京理工大学 一种基于多源数据融合和图神经网络的无创血压检测方法

Families Citing this family (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111493850A (zh) * 2020-04-13 2020-08-07 中国科学院深圳先进技术研究院 一种基于深度神经网络的血压测量方法及装置
CN112336325B (zh) * 2020-10-12 2024-02-23 乐普(北京)医疗器械股份有限公司 一种融合标定光体积描计信号数据的血压预测方法和装置
CN112336326B (zh) * 2020-10-21 2023-09-05 华南师范大学 容积脉搏波信号处理方法、血压测量装置和存储介质
CN112836647B (zh) * 2021-02-05 2023-07-21 合肥工业大学 一种基于ppg信号的无创式甘油三酯估计系统
CN112914527B (zh) * 2021-02-08 2023-06-30 合肥工业大学 一种基于脉搏波光电容积描记术的动脉血压信号采集方法
CN113057611B (zh) * 2021-03-19 2024-01-12 北京京东拓先科技有限公司 用于输出信息的方法、装置、设备以及存储介质
CN115336991A (zh) * 2021-05-14 2022-11-15 广州视源电子科技股份有限公司 一种血压测量装置及电子血压计
CN113397511B (zh) * 2021-05-24 2023-09-05 深圳大学 血压测量方法及装置
CN113456042A (zh) * 2021-06-30 2021-10-01 浙江师范大学 一种基于3d cnn的无接触面部血压测量方法
CN113647905A (zh) * 2021-07-13 2021-11-16 广东省科学院健康医学研究所 深度神经网络预测血压的训练方法、计算机装置和存储介质
CN113598734A (zh) * 2021-07-28 2021-11-05 厦门大学 一种基于深度神经网络模型的无袖带血压预测方法
CN114692668A (zh) * 2021-08-13 2022-07-01 康泰医学系统(秦皇岛)股份有限公司 一种压力值确定系统、方法、装置及可读存储介质
CN115500803A (zh) * 2022-09-29 2022-12-23 联想(北京)有限公司 一种信息确定方法及电子设备
CN115670902B (zh) * 2022-10-10 2024-07-12 山东大学 一种心肺复苏的智能决策系统及方法
CN117982116A (zh) * 2022-11-01 2024-05-07 华为技术有限公司 数据处理方法、装置及计算机可读存储介质
WO2024140490A1 (zh) * 2022-12-27 2024-07-04 北京津发科技股份有限公司 光信号处理方法和系统、连续血压测量方法、电子设备及介质
CN116028809A (zh) * 2022-12-27 2023-04-28 北京津发科技股份有限公司 一种连续血压的测量模型训练、测量方法及装置

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109965862A (zh) * 2019-04-16 2019-07-05 重庆大学 一种无袖带式长时连续血压无创监测方法
CN110251105A (zh) * 2019-06-12 2019-09-20 广州视源电子科技股份有限公司 一种无创血压测量方法、装置、设备及系统
US20190298193A1 (en) * 2016-02-03 2019-10-03 Angilytics Inc. Low pressure actuation blood pressure monitoring
KR20200022135A (ko) * 2018-08-22 2020-03-03 주식회사 셀바스헬스케어 심혈관 분석기
CN111493850A (zh) * 2020-04-13 2020-08-07 中国科学院深圳先进技术研究院 一种基于深度神经网络的血压测量方法及装置

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018175098A1 (en) * 2017-03-24 2018-09-27 D5Ai Llc Learning coach for machine learning system
JP6911498B2 (ja) * 2017-05-01 2021-07-28 オムロン株式会社 学習装置、学習方法、及び学習プログラム
CN108498089B (zh) * 2018-05-08 2022-03-25 北京邮电大学 一种基于深度神经网络的无创连续血压测量方法
CN109645980A (zh) * 2018-11-14 2019-04-19 天津大学 一种基于深度迁移学习的心律异常分类方法
CN110797119B (zh) * 2019-09-23 2022-09-20 深圳甲田科技有限公司 健康风险智能监测装置和迁移学习方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190298193A1 (en) * 2016-02-03 2019-10-03 Angilytics Inc. Low pressure actuation blood pressure monitoring
KR20200022135A (ko) * 2018-08-22 2020-03-03 주식회사 셀바스헬스케어 심혈관 분석기
CN109965862A (zh) * 2019-04-16 2019-07-05 重庆大学 一种无袖带式长时连续血压无创监测方法
CN110251105A (zh) * 2019-06-12 2019-09-20 广州视源电子科技股份有限公司 一种无创血压测量方法、装置、设备及系统
CN111493850A (zh) * 2020-04-13 2020-08-07 中国科学院深圳先进技术研究院 一种基于深度神经网络的血压测量方法及装置

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
A SHYAM; RAVICHANDRAN VIGNESH; S.P PREEJITH; JOSEPH JAYARAJ; SIVAPRAKASAM MOHANASANKAR: "PPGnet: Deep Network for Device Independent Heart Rate Estimation from Photoplethysmogram", 2019 41ST ANNUAL INTERNATIONAL CONFERENCE OF THE IEEE ENGINEERING IN MEDICINE AND BIOLOGY SOCIETY (EMBC), IEEE, 23 July 2019 (2019-07-23), pages 1899 - 1902, XP033624901, DOI: 10.1109/EMBC.2019.8856989 *
WU DAN: "Continuous and Noninvasive Blood Pressure Measurement Based on Deep Neural Network and Its Applications", CHINESE DOCTORAL DISSERTATIONS DATABASE-UNIVERSITY OF CHINESE ACADEMY OF SCIENCES, SHENZHEN INSTITUTES OF ADVANCED TECHNOLOGY, 1 October 2017 (2017-10-01), XP055859980 *
YAN CONG; LI ZHENQI; ZHAO WEI; HU JING; JIA DONGYA; WANG HONGMEI; YOU TIANYUAN: "Novel Deep Convolutional Neural Network for Cuff-less Blood Pressure Measurement Using ECG and PPG Signals", 2019 41ST ANNUAL INTERNATIONAL CONFERENCE OF THE IEEE ENGINEERING IN MEDICINE AND BIOLOGY SOCIETY (EMBC), IEEE, 23 July 2019 (2019-07-23), pages 1917 - 1920, XP033625013, DOI: 10.1109/EMBC.2019.8857108 *

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115270848A (zh) * 2022-06-17 2022-11-01 合肥心之声健康科技有限公司 一种ppg与ecg自动转换智能算法、存储介质和计算机系统
CN115270848B (zh) * 2022-06-17 2023-09-29 合肥心之声健康科技有限公司 一种ppg与ecg自动转换智能算法、存储介质和计算机系统
CN115130509A (zh) * 2022-06-29 2022-09-30 哈尔滨工业大学(威海) 基于条件式变分自编码器的心电信号生成方法
CN115836846A (zh) * 2022-12-14 2023-03-24 北京航空航天大学 一种基于自监督迁移学习的无创血压估计方法
CN116383617A (zh) * 2023-04-21 2023-07-04 复旦大学 一种基于脉搏波波形特征的智能血压检测方法及系统
CN116383617B (zh) * 2023-04-21 2023-09-22 复旦大学 一种基于脉搏波波形特征的智能血压检测方法及系统
CN116803340A (zh) * 2023-07-19 2023-09-26 北京理工大学 一种基于多源数据融合和图神经网络的无创血压检测方法

Also Published As

Publication number Publication date
CN111493850A (zh) 2020-08-07

Similar Documents

Publication Publication Date Title
WO2021208490A1 (zh) 一种基于深度神经网络的血压测量方法及装置
Tanveer et al. Cuffless blood pressure estimation from electrocardiogram and photoplethysmogram using waveform based ANN-LSTM network
Şentürk et al. Repetitive neural network (RNN) based blood pressure estimation using PPG and ECG signals
CN106413534B (zh) 连续血压测量装置、测量模型建立方法和系统
CN106974631B (zh) 一种基于脉搏波波形和心电信号的血压测量方法及装置
CN106821356B (zh) 基于Elman神经网络的云端连续血压测量方法及系统
Oweis et al. QRS detection and heart rate variability analysis: A survey
CN108577883A (zh) 一种冠心病筛查装置、筛查系统以及信号特征提取方法
Li et al. Design of a continuous blood pressure measurement system based on pulse wave and ECG signals
CN111248879B (zh) 一种基于多模态注意力融合的高血压老年人活动分析方法
CN107928654A (zh) 一种基于神经网络的脉搏波信号血压检测方法
LU502768B1 (en) Intelligent blood pressure prediction method based on multi-scale residual network and ppg signal
CN111887858A (zh) 基于跨模态映射的心冲击图信号心率估计方法
CN111839488B (zh) 基于脉搏波的无创连续血压测量装置和方法
CN111887828B (zh) 围术期患者非接触式生理信息监测装置、计算机设备和存储介质
Pan et al. Improved blood pressure estimation using photoplethysmography based on ensemble method
Singla et al. Cuff-less blood pressure measurement using supplementary ECG and PPG features extracted through wavelet transformation
He et al. A new approach for daily life Blood-Pressure estimation using smart watch
Rajput et al. Automated detection of hypertension using wavelet transform and nonlinear techniques with ballistocardiogram signals
Jegan et al. Biosensor-based feature extraction and physiological parameters measurement for biomedical applications
Pu et al. Cuff-less blood pressure estimation from electrocardiogram and photoplethysmography based on VGG19-LSTM network
Ghosh et al. Introduction of boosting algorithms in continuous non-invasive cuff-less blood pressure estimation using pulse arrival time
CN117100239A (zh) 基于cnn+lstm和ppg信号的血压检测系统
CN111513689A (zh) 一种脉搏波中主动脉打开特征点识别方法
Pachauri et al. Wavelet transform based arterial blood pressure waveform delineator

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20931576

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20931576

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 20931576

Country of ref document: EP

Kind code of ref document: A1

122 Ep: pct application non-entry in european phase

Ref document number: 20931576

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 11.12.2023)