WO2022077889A1 - 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 - Google Patents

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 Download PDF

Info

Publication number
WO2022077889A1
WO2022077889A1 PCT/CN2021/088023 CN2021088023W WO2022077889A1 WO 2022077889 A1 WO2022077889 A1 WO 2022077889A1 CN 2021088023 W CN2021088023 W CN 2021088023W WO 2022077889 A1 WO2022077889 A1 WO 2022077889A1
Authority
WO
WIPO (PCT)
Prior art keywords
blood pressure
dimensional tensor
data
dimensional
tensor
Prior art date
Application number
PCT/CN2021/088023
Other languages
English (en)
French (fr)
Inventor
曹君
张碧莹
王思瀚
吴泽剑
Original Assignee
乐普（北京）医疗器械股份有限公司
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 乐普（北京）医疗器械股份有限公司 filed Critical 乐普（北京）医疗器械股份有限公司
Priority to US18/248,755 priority Critical patent/US20230371831A1/en
Priority to EP21878943.6A priority patent/EP4226856A1/en
Publication of WO2022077889A1 publication Critical patent/WO2022077889A1/zh

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/02—Detecting, 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
    • A61B5/024—Detecting, measuring or recording pulse rate or heart rate
    • A61B5/02416—Detecting, measuring or recording pulse rate or heart rate using photoplethysmograph signals, e.g. generated by infrared radiation
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/02—Detecting, 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
    • A61B5/021—Measuring pressure in heart or blood vessels
    • A61B5/02141—Details of apparatus construction, e.g. pump units or housings therefor, cuff pressurising systems, arrangements of fluid conduits or circuits
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/02—Detecting, 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
    • A61B5/021—Measuring pressure in heart or blood vessels
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/30—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment

Definitions

  • the invention relates to the technical field of signal processing, in particular to a blood pressure prediction method and device for fusing calibration photoplethysmograph signal data.
  • Photoplethysmography is a non-invasive detection method that detects changes in blood volume in living tissue by means of photoelectric means.
  • the beating of the heart will make the blood flow per unit area in the blood vessel to form a periodic change, and the corresponding blood volume will also change accordingly, so that the PPG signal reflecting the amount of light absorbed by the blood will also change periodically, and the periodic change of the PPG signal It is closely related to heart beat and blood pressure changes.
  • Diastolic and systolic blood pressure data can be obtained by using a well-trained artificial intelligence blood pressure prediction network composed of a convolutional neural network (CNN) model and an artificial neural network (ANN) model for the PPG signal. .
  • CNN convolutional neural network
  • ANN artificial neural network
  • the purpose of the present invention is to provide a blood pressure prediction method, device, electronic device, computer program product and computer-readable storage medium that integrates the calibration photoplethysmograph signal data for the defects of the prior art, and combines the calibration PPG signal data with the The real-time PPG signal data is fused, and then the CNN+ANN prediction network, which is well-trained for predicting relative blood pressure data, is used to predict the blood pressure of the fusion data to obtain the relative blood pressure data.
  • the predicted blood pressure data is obtained by inversely inferring the absolute blood pressure data; in this way, the prediction accuracy of the artificial intelligence blood pressure prediction network can be improved.
  • a first aspect of the embodiments of the present invention provides a blood pressure prediction method that fuses calibrated photoplethysmographic signal data, the method includes:
  • the time length of the real-time PPG signal data is the same as the time length of the calibration PPG signal data;
  • the input data preparation processing of the convolutional neural network CNN model is performed, and the CNN input four-dimensional tensor is generated;
  • the multi-layer convolution pooling calculation is performed on the CNN input four-dimensional tensor, and the CNN output four-dimensional tensor is generated;
  • the input data preparation processing of the artificial neural network ANN model is performed, and the ANN input two-dimensional tensor is generated;
  • the regression calculation is performed on the ANN input two-dimensional tensor, and the ANN output two-dimensional tensor is generated;
  • the calibrated diastolic blood pressure data and the ANN output two-dimensional tensor, perform blood pressure data calculation to generate a blood pressure two-dimensional tensor;
  • the preset prediction type information is the first type, calculate the mean blood pressure data according to the blood pressure two-dimensional tensor to generate diastolic blood pressure prediction data and systolic blood pressure prediction data; when the prediction type information is the second type , perform blood pressure data extraction processing on the blood pressure two-dimensional tensor, and generate a diastolic blood pressure prediction data sequence and a systolic blood pressure prediction data sequence.
  • CNN input four-dimensional tensor specifically include:
  • signal data sampling processing is performed on the real-time PPG signal data to generate a real-time PPG data sequence
  • signal data sampling processing is performed on the calibration PPG signal data to generate a calibration PPG data sequence
  • the data length L 1 is the same as the data length L 2 of the calibration PPG data sequence
  • the real-time PPG data sequence is divided into data segments in sequence to obtain the total number of real-time segments and one-dimensional real-time PPG tensors; and then divide the total number of real-time segments into the real-time PPG one-dimensional tensors
  • the quantity is fused into a two-dimensional tensor to generate a real-time PPG two-dimensional tensor;
  • the calibration PPG data sequence is divided into data segments in order to obtain the total number of calibration segments and the calibration PPG one-dimensional tensors; Fusion into a two-dimensional tensor to generate a calibrated PPG two-dimensional tensor;
  • two-dimensional tensor fusion processing is performed to generate a fusion two-dimensional tensor
  • the shape of the fused two-dimensional tensor is B 3 ⁇ W 5 ;
  • the fused 2-D tensor includes the B 3 fused 1-D tensors;
  • the fused 1-D tensor consists of the corresponding
  • the described real-time PPG one-dimensional tensor and the demarcated PPG one-dimensional tensor are spliced together in the order in which the real-time PPG one-dimensional tensor precedes and the demarcated PPG one-dimensional tensor precedes;
  • the shape of the fusion two-dimensional tensor is increased from a two-dimensional tensor shape to a four-dimensional tensor shape, and the CNN input four-dimensional tensor is generated;
  • the shape of the CNN input four-dimensional tensor is B 4 ⁇ H 1 ⁇ W 6 ⁇ C 1 ;
  • the CNN model to perform multi-layer convolution pooling calculation on the CNN input four-dimensional tensor to generate a CNN output four-dimensional tensor, specifically including:
  • the CNN input four-dimensional tensor is used as the first input four-dimensional tensor, and then the first input four-dimensional tensor is sent to the first layer of the convolutional network layer of the CNN model, and the first layer of convolution pooling is performed Calculate, generate the first output four-dimensional tensor; then use the first output four-dimensional tensor as the second input four-dimensional tensor, and then send the second input four-dimensional tensor into the second layer volume of the CNN model Build the network layer, perform the second layer convolution pooling calculation, and generate the second output four-dimensional tensor; until the end, use the penultimate output four-dimensional tensor as the last input four-dimensional tensor, and then use the last input
  • the four-dimensional tensor is sent into the last layer of the convolutional network layer of the CNN model, and the last layer of convolution pooling is calculated to generate the four-dimensional tensor output by the CNN;
  • the first input four-dimensional tensor is sent to the first layer of the convolutional network layer of the CNN model, and the first layer of convolution pooling calculation is performed to generate the first output four-dimensional tensor, specifically including:
  • the first input four-dimensional tensor is sent to the first convolutional layer of the first convolutional network layer, and the first convolution calculation is performed to generate a first convolutional four-dimensional tensor;
  • the four-dimensional tensor is accumulated, and sent to the first pooling layer of the first layer of convolutional network layer, and the first pooling calculation is performed to generate the first output four-dimensional tensor.
  • the input data preparation processing of the artificial neural network ANN model is performed, and the ANN input two-dimensional tensor is generated, specifically including:
  • the shape of the CNN output four-dimensional tensor is reduced from the four-dimensional tensor shape to the two-dimensional tensor shape, and the ANN input two-dimensional tensor is generated;
  • the shape of the ANN input two-dimensional tensor is B 6 ⁇ W' 1 ;
  • the ANN input two-dimensional tensor includes the B 6 ANN input one-dimensional tensors; so
  • the shape of the ANN input one-dimensional tensor is 1 ⁇ W' 2 ;
  • the ANN model is used to perform regression calculation on the ANN input two-dimensional tensor, and the ANN output two-dimensional tensor is generated, specifically including:
  • the ANN input two-dimensional tensor is used as the first input two-dimensional tensor, and then the first input two-dimensional tensor is sent to the first fully connected layer of the ANN model, and the first layer is fully connected Calculate, generate the first output two-dimensional tensor; then use the first output two-dimensional tensor as the second input two-dimensional tensor, and then send the second input two-dimensional tensor into the ANN model
  • the second layer of full connection layer, the second layer of full connection calculation is performed, and the second output two-dimensional tensor is generated; until the end, the penultimate output two-dimensional tensor is used as the last input two-dimensional tensor, and then the The last input two-dimensional tensor is sent to the last fully connected layer of the ANN model, and the last layer of fully connected calculation is performed to generate the ANN output two-dimensional tensor;
  • the ANN model includes multiple layers of the fully connected layers;
  • the shape of the ANN output two-dimensional tensor is B 7 ⁇ W 9 ;
  • the ANN output two-dimensional tensor includes the B 7 ANN output one-dimensional tensors
  • the ANN outputs a one-dimensional tensor including relative data of diastolic blood pressure and relative data of systolic blood pressure.
  • the blood pressure data calculation is performed to generate the blood pressure two-dimensional tensor, which specifically includes:
  • the preset relative relationship information is a difference relationship
  • use the calibrated diastolic pressure data to perform diastolic pressure increment processing on all the diastolic pressure relative data in the two-dimensional tensor output by the ANN; and use the calibrated systolic pressure pressure data, perform systolic blood pressure increment processing on all the systolic blood pressure relative data in the ANN output two-dimensional tensor; and then use the ANN output two-dimensional tensor that has completed the increment processing as the blood pressure two-dimensional tensor;
  • the shape of the blood pressure two-dimensional tensor is B 8 ⁇ W 10 ;
  • the two-dimensional blood pressure tensor includes the B 8 blood pressure one-dimensional tensors;
  • the blood pressure one-dimensional tensor includes diastolic blood pressure data and systolic blood pressure data;
  • the diastolic blood pressure data is the sum of the corresponding relative diastolic blood pressure data and the calibrated diastolic blood pressure data;
  • the systolic blood pressure data is the sum of the corresponding relative systolic blood pressure data and the calibrated systolic blood pressure data.
  • the preset prediction type information when the preset prediction type information is the first type, calculate the mean blood pressure data according to the two-dimensional tensor of blood pressure to generate diastolic blood pressure prediction data and systolic blood pressure prediction data; when the prediction type information When it is the second type, the blood pressure data extraction process is performed on the blood pressure two-dimensional tensor, and the diastolic blood pressure prediction data sequence and the systolic blood pressure prediction data sequence are generated, specifically including:
  • the prediction type information is the first type, calculate the average value of all the diastolic blood pressure data in the blood pressure two-dimensional tensor to generate the diastolic blood pressure prediction data; calculate all the diastolic blood pressure data in the blood pressure two-dimensional tensor the average value of the systolic blood pressure data to generate the systolic blood pressure prediction data;
  • the prediction type information is the second type, sequentially extract the diastolic blood pressure data in the blood pressure two-dimensional tensor to form the diastolic blood pressure prediction data sequence; extract the systolic blood pressure in the blood pressure two-dimensional tensor data to form the systolic blood pressure prediction data series.
  • a second aspect of the embodiments of the present invention provides a blood pressure prediction device that fuses calibrated photoplethysmographic signal data, including:
  • the acquisition module is used for acquiring calibration photoplethysmograph PPG signal data and corresponding calibration diastolic blood pressure data and calibration systolic blood pressure data; obtaining real-time PPG signal data; wherein, the time length of the real-time PPG signal data and the calibration PPG signal data the same length of time;
  • the artificial intelligence computing module is used to prepare and process the input data of the convolutional neural network CNN model according to the calibrated PPG signal data and the real-time PPG signal data, and generate a CNN input four-dimensional tensor;
  • the CNN input four-dimensional tensor performs multi-layer convolution pooling calculation, and generates a CNN output four-dimensional tensor; and then according to the CNN output four-dimensional tensor, the input data preparation processing of the artificial neural network ANN model is performed, and the ANN input two-dimensional tensor is generated. Then utilize the ANN model to carry out regression calculation to the ANN input two-dimensional tensor, and generate the ANN output two-dimensional tensor;
  • the blood pressure prediction module is used to calculate the blood pressure data according to the calibrated diastolic blood pressure data, the calibrated systolic blood pressure data and the ANN output two-dimensional tensor, and generate a blood pressure two-dimensional tensor; when the preset prediction type information is the first When the type is one type, the mean blood pressure data is calculated according to the two-dimensional tensor of blood pressure, and the predicted data of diastolic blood pressure and the predicted data of systolic blood pressure are generated; when the predicted type information is the second type, the two-dimensional tensor of blood pressure is calculated. , perform blood pressure data extraction processing, and generate diastolic blood pressure prediction data sequence and systolic blood pressure prediction data sequence.
  • a third aspect of the embodiments of the present invention provides an electronic device, including: a memory, a processor, and a transceiver;
  • the processor is configured to be coupled with the memory to read and execute the instructions in the memory, so as to implement the method steps described in the first aspect above;
  • the transceiver is coupled to the processor, and the processor controls the transceiver to transmit and receive messages.
  • a fourth aspect of the embodiments of the present invention provides a computer program product, where the computer program product includes computer program code, and when the computer program code is executed by a computer, causes the computer to execute the method described in the first aspect.
  • a fifth aspect of the embodiments of the present invention provides a computer-readable storage medium, where computer instructions are stored in the computer storage medium, and when the computer instructions are executed by a computer, the computer is made to execute the method described in the first aspect above instruction.
  • Embodiments of the present invention provide a blood pressure prediction method, device, electronic device, computer program product, and computer-readable storage medium that fuses calibrated photoplethysmographic signal data.
  • the calibrated PPG signal data and corresponding calibrated blood pressure data of the test object are obtained. (calibrated diastolic blood pressure data, calibrated systolic blood pressure data), and then use the well-trained CNN+ANN artificial intelligence blood pressure prediction network for predicting relative blood pressure data to perform blood pressure prediction operations on the real-time PPG signal and the calibrated PPG signal of the test object.
  • Blood pressure data (relative data of diastolic blood pressure, relative data of systolic blood pressure), and then calculate the absolute blood pressure data of the calibrated blood pressure data and relative blood pressure data according to the relative relationship information to obtain the final blood pressure data; thus improving the prediction accuracy of the artificial intelligence blood pressure prediction network Spend.
  • FIG. 1 is a schematic diagram of a blood pressure prediction method for fusing calibration photoplethysmometer signal data according to Embodiment 1 of the present invention
  • FIG. 2a is a schematic structural diagram of a convolutional neural network provided in Embodiment 1 of the present invention.
  • FIG. 2b is a schematic structural diagram of an artificial neural network according to Embodiment 1 of the present invention.
  • FIG. 3 is a block diagram of a blood pressure prediction device that fuses calibrated photoplethysmographic signal data according to Embodiment 2 of the present invention
  • FIG. 4 is a schematic structural diagram of an electronic device according to Embodiment 3 of the present invention.
  • Convolutional neural network CNN is one of the core networks in the field of feature recognition that we know.
  • the feature extraction calculation is performed on the one-dimensional PPG signal data.
  • the feature data that conforms to the characteristics of the PPG signal is reserved for other networks. study.
  • the CNN model mentioned in this article is a convolutional neural network model that has been trained through feature extraction, and is specifically composed of multiple convolutional network layers.
  • Each convolutional network layer consists of 1 layer of convolutional layer and 1 layer of pooling. layer composition.
  • the convolutional layer is responsible for the feature extraction calculation for the input data of the CNN model
  • the pooling layer is for downsampling the extraction results of the convolutional layer
  • the output data of each convolutional network layer will be used as the next layer of volume.
  • the input data of the network layer is responsible for the feature extraction calculation for the input data of the CNN model
  • the input data and output data format of each convolutional network layer are in the form of 4-dimensional tensors: input 4-dimensional tensor (shape is X i4 ⁇ X i3 ⁇ X i2 ⁇ X i1 ), output 4-dimensional tensor (shape is X i4 ⁇ X i3 ⁇ X i2 ⁇ X i1 ) is X o4 ⁇ X o3 ⁇ X o2 ⁇ X o1 ), where: X i4 , X i3 , X i2 , and X i1 are the 4-, 3-, 2-, and 1-dimensional parameters of the input four-dimensional tensor, X o4 , X o3 , X o2 and X o1 are the 4-, 3-, 2-, and 1-dimensional parameters of the output four-dimensional tensor, respectively.
  • the shape of the output data is relative to the shape of the input data
  • the change rule of the dimension parameters is: 1) X o4 relative to X i4 , 4-dimensional parameters (in this embodiment, corresponding to PPG The total number of pieces of signal data) will not change; 2) X o3 , X o2 relative to X i3 , X i2 , 3- and 2-dimensional parameters will change, and the change is related to the size of the convolution kernel of each convolution layer and the sliding
  • the setting of the step size is related to the pooling window size and sliding step size of the pooling layer; 3) X o1 relative to X i1 , the change of the 1-dimensional parameter is related to the selected output space dimension in the convolution layer ( the number of convolution kernels).
  • ANN Artificial neural network
  • ANN refers to a complex network structure formed by a large number of neurons (nodes) connected to each other.
  • ANN simulates neuron activity with a mathematical model, and is an information processing system based on imitating the structure and function of the neural network of the brain.
  • a common application of ANN is to perform regression calculations on data.
  • the ANN model mentioned in this paper is an artificial neural network model that has been trained by regression; specifically, the ANN model consists of multiple fully connected layers. Among them, each fully connected layer includes multiple nodes, and each node is connected to all nodes in the previous layer, which is used to synthesize the node data extracted earlier to perform a node calculation, and calculate the result.
  • the node calculation here is also called fully-connected calculation, including node connection calculation and activation calculation.
  • the common activation function is to use rectified linear units ( Rectified Linear Unit, ReLU) function is mostly used, and other functions can also be used to complete the calculation.
  • ReLU Rectified Linear Unit
  • the input data and output data format of each fully connected layer are in the form of 2-dimensional tensors: input two-dimensional tensor (shape Y i2 ⁇ Y i1 ), output two-dimensional tensor (shape Y o2 ⁇ Y ) o1 ), wherein: Y i2 and Y i1 are the 2-dimensional and 1-dimensional parameters of the input 2-D tensor, and Y o2 and Y o1 are the 2-dimensional and 1-dimensional parameters of the output 2-D tensor.
  • the shape of the output data is relative to the shape of the input data
  • the change rule of the dimension parameters is: 1) Y o2 is relative to Y i2 , 2-dimensional parameters (in this embodiment, corresponding to the PPG signal The total number of data segments) will not change; 2) Y o1 is relative to Y i1 , and the change of the 1-dimensional parameter is related to the total number of nodes in the fully connected layer.
  • the Y o1 of the last fully connected layer in the embodiment of the present invention is specifically 2.
  • the shape of the two-dimensional tensor finally output by the ANN model in the embodiment of the present invention is: the total number of fragments ⁇ 2, the two-dimensional tensor includes the total number of fragments and one-dimensional tensors, each one-dimensional tensor corresponds to one fragment, and each one-dimensional tensor corresponds to one fragment.
  • the dimension tensor includes two blood pressure data related to the corresponding segment: diastolic relative data and systolic relative data.
  • the output result of the CNN model needs to be converted from the output four-dimensional tensor shape (X o4 ⁇ X o3 ⁇ X o2 ⁇ X o1 ) to the ANN model
  • the conventional artificial intelligence blood pressure prediction network directly divides the real-time PPG signal data (the real-time PPG signal data collected from the test object and needs to be used for blood pressure prediction) into segments (the reason why the PPG signal data is segmented into multiple
  • the PPG signal data fragment which is limited by the input length of the CNN model
  • the CNN model is used to extract the features of the PPG signal to obtain the feature data, and then the feature data is input into the ANN model for regression calculation.
  • the absolute systolic and diastolic blood pressure data corresponding to the PPG signal data segment. In practical applications, we found that approximate PPG signal data may be collected from different test objects.
  • the calculated absolute systolic and diastolic blood pressure data of different test objects are almost the same.
  • the reality is that we found that due to individual differences between people (eg, gender, age, height, weight, arm span, BMI, body temperature, whether to cite strong coffee, whether to exercise, etc.), for different For humans, even if the collected PPG signal data is similar, it may correspond to different blood pressure states.
  • the embodiment of the present invention makes some adjustments to the CNN model and the ANN model of the above-mentioned artificial intelligence blood pressure prediction network respectively: the adjusted CNN model receives real-time PPG signal data while receiving real-time PPG signal data. , and also receives the calibrated PPG signal data; and the final output of the CNN model is not the absolute characteristic data of the real-time PPG signal data, but the real-time PPG signal data to the calibrated PPG signal data, according to a specific relative relationship, the calculated relative characteristic data; here , the specific relative relationship can be difference, ratio, log ratio, exponential ratio, etc.
  • the input data of the adjusted ANN model changed from absolute feature data to relative feature data, and the output data also changed from absolute blood pressure data to relative blood pressure data.
  • the so-called calibrated PPG signal data is a piece of reference data specially collected for the test object whose time length is consistent with the real-time PPG signal data.
  • the data corresponds to the calibrated systolic and diastolic blood pressure data.
  • the above-mentioned artificial intelligence blood pressure prediction network (CNN+ANN) that combines the calibrated PPG signal data and real-time PPG signal data, when predicting the real-time PPG signal data collected from different objects, even if the real-time data is similar, it can also be calibrated because of their respective The differences in the PPG signal data lead to obvious relative differences, and naturally the differences in the final predicted blood pressure data will also be more obvious. As a result, the above-mentioned problem of inaccurate measurement of the conventional prediction network is solved, and the prediction accuracy of the artificial intelligence blood pressure prediction network is further improved.
  • Embodiment 1 of the present invention provides a blood pressure prediction method that fuses calibrated photoplethysmograph signal data, using an artificial intelligence blood pressure prediction network (CNN+ANN) that predicts relative blood pressure data, and fuses calibrated PPG signal data and real-time PPG signal data. Then, according to the calibrated blood pressure data and relative blood pressure data, according to a specific relative relationship, the absolute blood pressure data is reversed, and the predicted blood pressure data corresponding to the real-time PPG signal is obtained.
  • the prediction accuracy of the artificial intelligence blood pressure prediction network can be improved.
  • FIG. 1 is a schematic diagram of a blood pressure prediction method by fusing calibration photoplethysmometer signal data provided in Embodiment 1 of the present invention.
  • the method mainly includes the following steps:
  • Step 1 Acquire calibration photoplethysmograph PPG signal data and corresponding calibrated diastolic blood pressure data and calibrated systolic blood pressure data.
  • the device can obtain the calibrated PPG signal data and the calibrated diastolic blood pressure data and the calibrated systolic blood pressure data with the calibrated PPG signal data from the local storage medium; In the medium, the calibrated PPG signal data, the calibrated diastolic blood pressure data and the calibrated systolic blood pressure data are obtained.
  • the device is specifically a terminal device or a server that implements the method provided by the embodiment of the present invention
  • the calibrated PPG signal data is a piece of PPG signal data pre-collected for the test object
  • the corresponding calibrated diastolic blood pressure data and calibrated systolic blood pressure data are in The actual blood pressure data obtained by measuring the blood pressure of the test object through the blood pressure measuring device while collecting the calibrated PPG signal data.
  • the length of the acquired calibrated PPG signal data is 10 seconds
  • the corresponding calibrated diastolic blood pressure data is 74 mmHg (millimetre Hg, mmHg)
  • the calibrated systolic blood pressure data is 113 mmHg.
  • Step 2 obtain real-time PPG signal data
  • the time length of the real-time PPG signal data is the same as the time length of the calibration PPG signal data.
  • the device can perform real-time PPG signal acquisition and data sampling processing on the test object through its own PPG signal acquisition device to obtain real-time PPG signal data; the device can also perform real-time PPG signal acquisition on the test object through the PPG signal acquisition device connected to itself. PPG signal acquisition and data sampling processing to obtain real-time PPG signal data.
  • the acquired real-time PPG signal data length is 10 seconds, which is consistent with the calibrated PPG signal data length.
  • Step 3 according to the calibration PPG signal data and the real-time PPG signal data, carry out the input data preparation processing of the convolutional neural network CNN model, and generate the CNN input four-dimensional tensor;
  • step 31 perform signal data sampling processing on real-time PPG signal data to generate a real-time PPG data sequence, and perform signal data sampling processing on calibration PPG signal data to generate a calibration PPG data sequence;
  • the data length L 1 of the real-time PPG data sequence is the same as the data length L 2 of the calibration PPG data sequence ;
  • sampling frequency is stored in the local storage medium of the device
  • Step 32 According to the preset segment length, the real-time PPG data sequence is divided into data segments in order to obtain the total number of real-time segments and real-time PPG one-dimensional tensors; and then fuse the total number of real-time segments and real-time PPG one-dimensional tensors into In a two-dimensional tensor, generate a real-time PPG two-dimensional tensor;
  • the total number of real-time segments int(L 1 /segment length); int() is the rounding function;
  • the shape of the real-time PPG one-dimensional tensor is 1 ⁇ W 1 ;
  • W 1 is the one-dimensional parameter of the real-time PPG one-dimensional tensor , and
  • W 1 segment length;
  • the shape of the real-time PPG 2D tensor is B 1 ⁇ W 2 ;
  • segment lengths are stored in the device's local storage medium
  • the value of the segment length is determined by the input data length of the CNN model; the segment division method adopts the sequential segment division method, so the data between adjacent segments does not overlap;
  • Step 33 according to the length of the segment, perform data segment division processing on the calibration PPG data sequence in order to obtain the total number of calibration segments and the calibration PPG one-dimensional tensors; and then fuse the total number of calibration segments and the calibration PPG one-dimensional tensors into a two-dimensional tensor. In the tensor, generate a calibrated PPG two-dimensional tensor;
  • the total number of calibrated fragments int(L 2 /segment length);
  • the shape of the calibrated PPG one-dimensional tensor is 1 ⁇ W 3 ;
  • the shape of the calibrated PPG 2D tensor is B 2 ⁇ W 4 ;
  • Step 34 performing two-dimensional tensor fusion processing on the real-time PPG two-dimensional tensor and the calibrated PPG two-dimensional tensor according to the order of real-time data first and calibration data last, to generate a fusion two-dimensional tensor;
  • the shape of the fused 2D tensor is B 3 ⁇ W 5 ;
  • the fusion two-dimensional tensor includes B 3 fusion one-dimensional tensors;
  • the fusion one-dimensional tensor consists of the corresponding real-time PPG one-dimensional tensor and the calibrated PPG one-dimensional tensor, according to
  • the real-time PPG one-dimensional tensor is in the front, and the calibrated PPG one-dimensional tensor is in the order of the latter, which are spliced together;
  • the data of the real-time PPG two-dimensional tensor [5,250] is ⁇ (Z 1,1 ,...Z 1,250 ),(Z 2,1 ,...Z 2,250 ),(Z 3,1 ,...Z 3,250 ),( Z 4,1 ,...Z 4,250 ),(Z 5,1 ,...Z 5,250 ) ⁇ , where (Z 1,1 ,...Z 1,250 ), (Z 2,1 ,... Z 2,250 ), (Z 3,1 ,...Z 3,250 ), (Z 4,1 ,...Z 4,250 ), (Z 5,1 ,...Z 5,250 ) are the first, second, third, fourth, and fifth real-time PPG in the real-time PPG two-dimensional tensor in turn One-dimensional tensor data;
  • the data to calibrate the PPG two-dimensional tensor [5,250] is ⁇ (D 1,1 ,...D 1,250 ),(D 2,1 ,...D 2,250 ),(D 3,1 ,...D 3,250 ),(D 4 ,1 ,...D 4,250 ),(D 5,1 ,...D 5,250 ) ⁇ , where (D 1,1 ,... D 1,250 ), (D 2,1 ,... D 2,250 ), (D 3,1 ,... D 3,250 ), (D 4,1 ,...D 4,250 ), (D 5,1 ,...D 5,250 ) are the 1st, 2nd, 3rd, 4th, and 5th calibration PPG one-dimensional tensors in order Tensor data;
  • the data format of the fusion two-dimensional tensor should be:
  • Step 35 according to the four-dimensional tensor input data format of the CNN model, the shape of the fusion two-dimensional tensor is increased from the two-dimensional tensor shape to the four-dimensional tensor shape, and the CNN input four-dimensional tensor is generated;
  • the shape of the CNN input four-dimensional tensor is B 4 ⁇ H 1 ⁇ W 6 ⁇ C 1 ;
  • the shape of the fused 2D tensor is upgraded from a 2D tensor shape to a 4D tensor shape, and the process only resets the tensor shape without destroying the actual data order in the tensor.
  • Step 4 using the CNN model, perform multi-layer convolution pooling calculation on the CNN input four-dimensional tensor, and generate the CNN output four-dimensional tensor;
  • it includes: taking the CNN input four-dimensional tensor as the first input four-dimensional tensor, and then sending the first input four-dimensional tensor into the first layer of the convolutional network layer of the CNN model, performing the first layer of convolution pooling calculation, and generating The first output four-dimensional tensor; then the first output four-dimensional tensor is used as the second input four-dimensional tensor, and then the second input four-dimensional tensor is sent to the second convolutional network layer of the CNN model, and the second layer is rolled
  • the pooling calculation generates the second output 4D tensor; until the end, the penultimate output 4D tensor is used as the last input 4D tensor, and then the last input 4D tensor is sent to the last layer of the CNN model
  • the convolutional network layer performs the last layer of convolution pooling calculation to generate a CNN output four-dimensional tensor;
  • the first input four-dimensional tensor is sent to the first layer of the convolutional network layer of the CNN model, the first layer of convolution pooling is calculated, and the first output four-dimensional tensor is generated.
  • the first input four-dimensional tensor Send it to the first convolutional layer of the first convolutional network layer, perform the first convolution calculation, and generate the first convolutional four-dimensional tensor; send the first convolutional four-dimensional tensor to the first convolutional network layer
  • the first pooling layer of performs the first pooling calculation to generate the first output four-dimensional tensor.
  • Figure 2a is a schematic structural diagram of the convolutional neural network provided in Embodiment 1 of the present invention, and the total number of layers is four, then,
  • the CNN input four-dimensional tensor is used as the first input four-dimensional tensor; then the first input four-dimensional tensor is sent to the first convolutional layer of the first convolutional network layer of the CNN model to perform the first convolution calculation to generate the first convolutional layer.
  • the first output four-dimensional tensor is used as the second input four-dimensional tensor; then the second input four-dimensional tensor is sent to the second convolutional layer of the second convolutional network layer of the CNN model for the second convolution calculation to generate The second convolutional four-dimensional tensor; and then the second convolutional four-dimensional tensor is sent to the second pooling layer of the second convolutional network layer for the second pooling calculation, and the second output four-dimensional tensor is generated;
  • the second output four-dimensional tensor is used as the third input four-dimensional tensor; then the third input four-dimensional tensor is sent to the third convolutional layer of the third convolutional network layer of the CNN model for the third convolution calculation to generate The third convolutional four-dimensional tensor; then the third convolutional four-dimensional tensor is sent to the third pooling layer of the third convolutional network layer for the third pooling calculation, and the third output four-dimensional tensor is generated;
  • the third output four-dimensional tensor is used as the fourth input four-dimensional tensor; then the fourth input four-dimensional tensor is sent to the fourth convolutional layer of the fourth convolutional network layer of the CNN model for the fourth convolution calculation to generate The fourth convolutional four-dimensional tensor; then the fourth convolutional four-dimensional tensor is sent to the fourth pooling layer of the fourth convolutional network layer for the fourth pooling calculation, and the fourth output four-dimensional tensor is generated; here
  • the fourth output 4D tensor is the final output CNN output 4D tensor.
  • the shape of the input data will change after each layer of convolutional layer or pooling layer, but it still maintains the 4-dimensional tensor form, in which the 4-dimensional parameters (total number of fragments) will not occur.
  • Change the change of 3-dimensional and 2-dimensional parameters is related to the size of the convolution kernel of each convolution layer and the setting of the sliding step, as well as the pooling window size and sliding step of the pooling layer;
  • the variation is related to the selected output spatial dimension (the number of convolution kernels) in the convolutional layer.
  • the setting of the number of layers in the network and the setting of various parameters of each layer must be constantly revised according to experience and experimental results.
  • Step 5 according to the CNN output four-dimensional tensor, carry out the input data preparation processing of the artificial neural network ANN model, and generate the ANN input two-dimensional tensor;
  • step 51 reduce the shape of the CNN output four-dimensional tensor from the four-dimensional tensor shape to the two-dimensional tensor shape, and generate the ANN input two-dimensional tensor;
  • the shape of the ANN input two-dimensional tensor is B 6 ⁇ W' 1 ;
  • the ANN input two-dimensional tensor includes B 6 ANN input one-dimensional tensors;
  • the shape of the ANN input one-dimensional tensor is 1 ⁇ W' 2 ;
  • reducing the shape of the CNN output four-dimensional tensor from the four-dimensional tensor shape to the two-dimensional tensor shape is actually reducing the shape of the CNN output four-dimensional tensor from four-dimensional to two-dimensional, and the process only resets the tensor shape. does not destroy the actual data order within the tensor;
  • the shape of the dimension tensor should be 5 ⁇ 2560, which is represented here as the ANN input two-dimensional tensor [5,2560]; correspondingly, the ANN input two-dimensional tensor includes five ANN input one-dimensional tensors, and the ANN input one
  • the shape of the dimension tensor should be 1 ⁇ 2560, which is represented here as the ANN input one-dimensional tensor [2560];
  • each ANN input one-dimensional tensor two data are added: the calibrated diastolic blood pressure data and the calibrated systolic blood pressure data, so as to improve the calculation accuracy of the ANN model; after the data addition is completed, the ANN input one-dimensional tensor
  • the shape changes from 1 ⁇ W' 2 to 1 ⁇ W' 3 ;
  • the shape of the corresponding ANN input two-dimensional tensor changes from B 6 ⁇ W' 1 to B 6 ⁇ W 8 ;
  • the shape of the ANN input two-dimensional tensor before data addition is 5 ⁇ 2560
  • the shape of the ANN input one-dimensional tensor before data addition is 1 ⁇ 2560
  • the ANN input one-dimensional tensor after data addition is completed
  • the shape of the tensor becomes 1 ⁇ 2562
  • the shape of the corresponding ANN input 2D tensor becomes 5 ⁇ 2562, which is represented here as the ANN input 2D tensor [5, 2562].
  • Step 6 using the ANN model, perform regression calculation on the ANN input two-dimensional tensor, and generate the ANN output two-dimensional tensor;
  • it includes: using the ANN input two-dimensional tensor as the first input two-dimensional tensor, and then sending the first input two-dimensional tensor into the first fully connected layer of the ANN model, performing the first layer of fully connected calculation, and generating The first output two-dimensional tensor; then the first output two-dimensional tensor is used as the second input two-dimensional tensor, and then the second input two-dimensional tensor is sent to the second fully connected layer of the ANN model, and the first The second-layer full connection calculation generates the second output two-dimensional tensor; until the end, the penultimate output two-dimensional tensor is used as the last input two-dimensional tensor, and then the last input two-dimensional tensor is sent to the ANN The last layer of the model is fully connected, and the last layer of fully connected calculation is performed to generate the ANN output two-dimensional tensor;
  • FIG. 2b is a schematic structural diagram of the artificial neural network provided in the first embodiment of the present invention.
  • the ANN model includes four fully connected layers, then
  • the ANN input two-dimensional tensor is used as the first input two-dimensional tensor, and then the first input two-dimensional tensor is sent to the first fully connected layer of the ANN model, and the first layer of fully connected calculation is performed to generate the first output.
  • two-dimensional tensor ;
  • the first output two-dimensional tensor is used as the second input two-dimensional tensor, and then the second input two-dimensional tensor is sent to the second fully connected layer of the ANN model, and the second layer of fully connected calculation is performed to generate the first Two output 2D tensors;
  • the second output two-dimensional tensor is used as the third input two-dimensional tensor, and then the third input two-dimensional tensor is sent to the third fully connected layer of the ANN model, and the third layer of fully connected calculation is performed to generate the first Three output 2D tensors;
  • the third output two-dimensional tensor is used as the fourth input two-dimensional tensor, and then the fourth input two-dimensional tensor is sent to the fourth fully connected layer of the ANN model, and the fourth layer of fully connected calculation is performed to generate the first Four-output two-dimensional tensor; the fourth output two-dimensional tensor here is the final output ANN output two-dimensional tensor;
  • the ANN model consists of a fully connected layer.
  • Each node of the fully connected layer is connected to all nodes of the previous layer, which is used to synthesize the features extracted in the front.
  • Each fully connected layer can be set The number of nodes in this layer and the activation function (there are more ReLUs, and can also be changed to others).
  • the number of nodes in the last fully connected layer of the ANN model in this embodiment is 2, and the shape of the corresponding final output two-dimensional tensor is specifically B 7 ⁇ 2.
  • (R db1 ,R sb1 ),(R db2 ,R sb2 ),(R db3 ,R sb3 ),(R db4 ,R sb4 ),(R db5 ,R sb5 ) are the first in the ANN output two-dimensional tensor in turn .
  • the 2nd, 3rd, 4th, and 5th ANN output data of one-dimensional tensors (R sbi is the relative data of systolic blood pressure, R dbi is the relative data of diastolic blood pressure, and the value of i ranges from 1 to 5).
  • Step 7 according to the calibrated diastolic blood pressure data, the calibrated systolic blood pressure data and the ANN output two-dimensional tensor, calculate the blood pressure data, and generate the blood pressure two-dimensional tensor;
  • the preset relative relationship information is a difference relationship
  • use the calibrated diastolic blood pressure data to output all the diastolic blood pressure relative data in the two-dimensional tensor from the ANN, and perform diastolic blood pressure increment processing; and use the calibrated systolic blood pressure data to analyze the ANN.
  • the shape of the blood pressure two-dimensional tensor is B 8 ⁇ W 10 ;
  • the two-dimensional blood pressure tensor includes B 8 blood pressure one-dimensional tensors;
  • the blood pressure one-dimensional tensor includes diastolic blood pressure data and systolic blood pressure data;
  • systolic blood pressure data is the sum of the corresponding relative systolic blood pressure data and the calibration systolic blood pressure data.
  • the relative relationship information includes at least a difference relative relationship (difference relationship).
  • the diastolic blood pressure data can be obtained by adding the diastolic blood pressure data relative data in the two-dimensional tensor output by the ANN to the calibrated diastolic blood pressure data, and the diastolic blood pressure data can be obtained;
  • the systolic blood pressure data can be obtained by adding the calibrated systolic blood pressure data to the data.
  • ANN outputs a two-dimensional tensor [5,2] as ⁇ (7,23),(6,22),(7,21),(11,20),(9,18) ⁇ ,
  • the calibrated diastolic blood pressure data is 74 mmHg
  • the calibrated systolic blood pressure data is 113 mmHg
  • Step 8 when the preset prediction type information is the first type, perform mean blood pressure data calculation according to the two-dimensional tensor of blood pressure, and generate diastolic blood pressure prediction data and systolic blood pressure prediction data; when the prediction type information is the second type, The two-dimensional tensor of blood pressure is extracted and processed to generate diastolic blood pressure prediction data sequence and systolic blood pressure prediction data sequence;
  • it includes: when the prediction type information is the first type, calculating the average value of all diastolic blood pressure data in the blood pressure two-dimensional tensor to generate diastolic blood pressure prediction data; calculating the average value of all systolic blood pressure data in the blood pressure two-dimensional tensor to generate systolic blood pressure prediction data; when the prediction type information is the second type, the diastolic blood pressure data in the blood pressure two-dimensional tensor is sequentially extracted to form a diastolic blood pressure prediction data sequence; the systolic blood pressure data in the blood pressure two-dimensional tensor is extracted to form a systolic blood pressure prediction data sequence.
  • the embodiment of the present invention supports two types of prediction data output: the first type, outputting a pair of mean blood pressure prediction data (diastolic blood pressure prediction data and systolic blood pressure prediction data); the second type, extracting blood pressure data by segments, composed of Dynamic blood pressure data series (diastolic blood pressure prediction data series and systolic blood pressure prediction data series).
  • the prediction type information is stored in the local storage medium of the device, and in this embodiment of the present invention, the content of the prediction type information is read to select which type is specifically used to output prediction data.
  • the prediction type information is the first type
  • the data of the blood pressure two-dimensional tensor [5,2] is ⁇ (81,136),(80,135),(81,134),(85,133),(83,131) ⁇
  • Systolic blood pressure prediction data (136+135+134+133+131)/5 ⁇ 134(mmHg).
  • the prediction type information is the second type
  • the data of the blood pressure two-dimensional tensor [5,2] is ⁇ (81,136),(80,135),(81,134),(85,133),(83,131) ⁇ , then:
  • the data content of the diastolic blood pressure prediction data sequence [5] is (81, 80, 81, 85, 83);
  • the data content of the systolic blood pressure prediction data series [5] is (136, 135, 134, 133, 131).
  • FIG. 3 is a block diagram of a blood pressure prediction device that fuses and calibrated photoplethysmography signal data according to Embodiment 2 of the present invention.
  • the device may be the terminal device or server described in the foregoing embodiment, or may An apparatus for implementing the method provided by the embodiment of the present invention in a terminal device or server, for example, the apparatus may be an apparatus or a chip system of the foregoing terminal device or server.
  • the device includes:
  • the acquisition module 31 is used to acquire the calibrated photoplethysmograph PPG signal data and the corresponding calibrated diastolic blood pressure data and calibrated systolic blood pressure data; obtain real-time PPG signal data; wherein, the time length of the real-time PPG signal data and the time length of the calibrated PPG signal data same;
  • the artificial intelligence computing module 32 is used to prepare and process the input data of the convolutional neural network CNN model according to the calibrated PPG signal data and the real-time PPG signal data, and generate the CNN input four-dimensional tensor; then use the CNN model to perform the CNN input four-dimensional tensor.
  • Multi-layer convolution pooling calculation generate CNN output four-dimensional tensor; then according to CNN output four-dimensional tensor, prepare the input data of artificial neural network ANN model, generate ANN input two-dimensional tensor; then use ANN model to ANN Input two-dimensional tensor for regression calculation, and generate ANN output two-dimensional tensor;
  • the blood pressure prediction module 33 is configured to output a two-dimensional tensor according to the calibrated diastolic blood pressure data, the calibrated systolic blood pressure data and the ANN, calculate the blood pressure data, and generate a blood pressure two-dimensional tensor; when the preset prediction type information is the first type, according to Two-dimensional tensor of blood pressure, calculate the mean blood pressure data, generate diastolic blood pressure prediction data and systolic blood pressure prediction data; when the prediction type information is the second type, perform blood pressure data extraction processing on the blood pressure two-dimensional tensor, and generate diastolic blood pressure prediction Data series and systolic blood pressure prediction data series.
  • a blood pressure prediction device that fuses and calibrates photoplethysmographic signal data provided by an embodiment of the present invention can perform the method steps in the above method embodiments, and its implementation principles and technical effects are similar, and will not be repeated here.
  • each module of the above apparatus is only a division of logical functions, and may be fully or partially integrated into a physical entity in actual implementation, or may be physically separated.
  • these modules can all be implemented in the form of software calling through processing elements; they can also all be implemented in hardware; some modules can also be implemented in the form of calling software through processing elements, and some modules can be implemented in hardware.
  • the acquisition module may be a separately established processing element, or may be integrated into a certain chip of the above-mentioned device to be implemented, in addition, it may also be stored in the memory of the above-mentioned device in the form of program code, and a certain processing element of the above-mentioned device may be used.
  • each step of the above-mentioned method or each of the above-mentioned modules can be completed by an integrated logic circuit of hardware in the processor element or an instruction in the form of software.
  • the above modules may be one or more integrated circuits configured to implement the above methods, such as: one or more Application Specific Integrated Circuit (ASIC), or one or more digital signal processors ( Digital Signal Processor, DSP), or, one or more Field Programmable Gate Array (Field Programmable Gate Array, FPGA), etc.
  • ASIC Application Specific Integrated Circuit
  • DSP Digital Signal Processor
  • FPGA Field Programmable Gate Array
  • the processing element may be a general-purpose processor, such as a central processing unit (Central Processing Unit, CPU) or other processors that can call program codes.
  • these modules can be integrated together and implemented in the form of a system-on-a-chip (SOC).
  • SOC system-on-a-chip
  • the above-described embodiments it may be implemented in whole or in part by software, hardware, firmware, or any combination thereof.
  • software it can be implemented in whole or in part in the form of a computer program product.
  • the computer program product includes one or more computer instructions.
  • the computer program instructions When the computer program instructions are loaded and executed on a computer, the procedures or functions described in accordance with the embodiments of the present invention are generated in whole or in part.
  • the aforementioned computer may be a general purpose computer, a special purpose computer, a computer network, or other programmable device.
  • the above-mentioned computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the above-mentioned computer instructions may be transmitted from a website site, computer, server or data center via wired communication. (such as coaxial cable, optical fiber, digital subscriber line (Digital Subscriber Line, DSL)) or wireless (such as infrared, wireless, Bluetooth, microwave, etc.) to another website site, computer, server or data center transmission.
  • the above A computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that contains an integration of one or more available media.
  • the aforementioned available media can be magnetic media, (eg, floppy disks, hard disks, magnetic tape), optical media (eg, DVD), or semiconductor media (eg, solid state disk (SSD)), and the like.
  • FIG. 4 is a schematic structural diagram of an electronic device according to Embodiment 3 of the present invention.
  • the electronic device may be the aforementioned terminal device or server, or may be a terminal device or server that is connected to the aforementioned terminal device or server and implements the method of the embodiment of the present invention.
  • the electronic device may include: a processor 41 (eg, a CPU), a memory 42 , and a transceiver 43 ; the transceiver 43 is coupled to the processor 41 , and the processor 41 controls the transceiver 43 to transmit and receive.
  • Various instructions may be stored in the memory 42 for implementing various processing functions and implementing the methods and processing procedures provided in the above embodiments of the present invention.
  • the electronic device involved in the embodiment of the present invention further includes: a power supply 44 , a system bus 45 and a communication port 46 .
  • a system bus 45 is used to implement communication connections between components.
  • the above-mentioned communication port 46 is used for connection and communication between the electronic device and other peripheral devices.
  • the system bus mentioned in FIG. 4 may be a peripheral component interconnect standard (Peripheral Component Interconnect, PCI) bus or an Extended Industry Standard Architecture (Extended Industry Standard Architecture, EISA) bus or the like.
  • PCI peripheral component interconnect
  • EISA Extended Industry Standard Architecture
  • the system bus can be divided into address bus, data bus, control bus and so on. For ease of presentation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.
  • the communication interface is used to realize the communication between the database access device and other devices (eg client, read-write library and read-only library).
  • the memory may include random access memory (Random Access Memory, RAM), and may also include non-volatile memory (Non-Volatile Memory), such as at least one disk storage.
  • the above-mentioned processor can be a general-purpose processor, including a central processing unit CPU, a network processor (NP), a graphics processor (Graphics Processing Unit, GPU), etc.; it can also be a digital signal processor DSP, an application-specific integrated circuit ASIC, field programmable gate array FPGA or other programmable logic device, discrete gate or transistor logic device, discrete hardware components.
  • embodiments of the present invention further provide a computer-readable storage medium, where instructions are stored in the storage medium, and when the storage medium runs on a computer, the computer executes the methods and processing procedures provided in the foregoing embodiments.
  • Embodiments of the present invention further provide a chip for running instructions, where the chip is used to execute the methods and processing procedures provided in the foregoing embodiments.
  • An embodiment of the present invention further provides a program product, the program product includes a computer program, the computer program is stored in a storage medium, at least one processor can read the computer program from the storage medium, and the at least one processor executes the above implementation The methods and processing procedures provided in the example.
  • Embodiments of the present invention provide a blood pressure prediction method, device, electronic device, computer program product, and computer-readable storage medium that fuses calibrated photoplethysmographic signal data. First, the calibrated PPG signal data and corresponding calibrated blood pressure data of the test object are obtained.
  • a software module can be placed in random access memory (RAM), internal memory, read only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other in the technical field. in any other known form of storage medium.
  • RAM random access memory
  • ROM read only memory
  • electrically programmable ROM electrically erasable programmable ROM
  • registers hard disk, removable disk, CD-ROM, or any other in the technical field. in any other known form of storage medium.

Landscapes

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

Abstract

融合标定光体积描计信号数据的血压预测方法以及与方法对应的装置、电子设备、计算机程序产品和计算机存储介质。方法包括：获取实时PPG信号数据、标定PPG信号数据、标定舒张压数据和收缩压数据；根据实时、标定PPG信号数据，利用CNN+ANN模型进行相对血压数据计算，生成ANN输出二维张量；根据标定舒张压、收缩压数据和ANN输出二维张量进行血压数据计算，生成血压二维张量；预测类型信息为第一类型时，根据血压二维张量进行均值计算，生成舒张压预测数据和收缩压预测数据；预测类型信息为第二类型时，对血压二维张量进行数据提取处理，生成舒张压预测数据序列和收缩压预测数据序列。方法可以提高人工智能血压预测网络的预测准确度。

Description

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
本申请要求于2020年10月12日提交中国专利局、申请号为202011085575.X、发明名称为“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”的中国专利申请的优先权。
技术领域
本发明涉及信号处理技术领域，特别涉及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。
背景技术
光体积描计(Photoplethysmography,PPG)法，是借助光电手段在活体组织中检测血液容积变化的一种无创检测方法。心脏搏动会使得血管内单位面积的血流量形成周期性变化，与之对应的血液体积也相应发生变化，从而使得反映血液吸收光量的PPG信号也会发生周期性变化，且PPG信号的周期性变化与心脏搏动、血压变化是密切相关的。对PPG信号使用训练成熟的由卷积神经网络(Convolutional Neural Network，CNN)模型和人工神经网络(Artificial Neural Network,ANN)模型组成的人工智能血压预测网络，可以获得血压的舒张压和收缩压数据。
发明内容
本发明的目的，就是针对现有技术的缺陷，提供一种融合标定光体积描计信号数据的血压预测方法、装置、电子设备、计算机程序产品及计算机可读存储介质，将标定PPG信号数据与实时PPG信号数据进行融合，再使用训 练成熟的用于预测相对血压数据的CNN+ANN预测网络对融合数据进行血压预测得到相对血压数据，再根据体现相对血压与标定血压数据关系的相对关系信息，进行绝对血压数据反推得到预测血压数据；这样，可以提高人工智能血压预测网络的预测准确度。
为实现上述目的，本发明实施例第一方面提供了一种融合标定光体积描计信号数据的血压预测方法，所述方法包括：
获取标定光体积描计PPG信号数据和对应的标定舒张压数据与标定收缩压数据；
获取实时PPG信号数据；所述实时PPG信号数据的时间长度与所述标定PPG信号数据的时间长度相同；
根据所述标定PPG信号数据和所述实时PPG信号数据，进行卷积神经网络CNN模型的输入数据准备处理，生成CNN输入四维张量；
利用所述CNN模型，对所述CNN输入四维张量进行多层卷积池化计算，生成CNN输出四维张量；
根据所述CNN输出四维张量，进行人工神经网络ANN模型的输入数据准备处理，生成ANN输入二维张量；
利用所述ANN模型，对所述ANN输入二维张量进行回归计算，生成ANN输出二维张量；
根据所述标定舒张压数据、所述标定收缩压数据和所述ANN输出二维张量，进行血压数据计算，生成血压二维张量；
当预设的预测类型信息为第一类型时，根据所述血压二维张量，进行均值血压数据计算，生成舒张压预测数据和收缩压预测数据；当所述预测类型信息为第二类型时，对所述血压二维张量，进行血压数据提取处理，生成舒张压预测数据序列和收缩压预测数据序列。
优选的，所述根据所述标定PPG信号数据和所述实时PPG信号数据，进行卷积神经网络CNN模型的输入数据准备处理，生成CNN输入四维张量，具 体包括：
按照预设的采样频率，对所述实时PPG信号数据进行信号数据采样处理生成实时PPG数据序列，对所述标定PPG信号数据进行信号数据采样处理生成标定PPG数据序列；所述实时PPG数据序列的数据长度L 1与所述标定PPG数据序列的数据长度L 2相同；
按照预设的片段长度，对所述实时PPG数据序列，按顺序进行数据片段划分处理，得到实时片段总数个实时PPG一维张量；再将所述实时片段总数个所述实时PPG一维张量融合到一个二维张量中，生成实时PPG二维张量；
其中，所述实时片段总数＝int(L 1/片段长度)；所述int()为取整函数；所述实时PPG一维张量的形状为1×W 1；所述W 1为实时PPG一维张量的1维参数，且W 1＝片段长度；所述实时PPG二维张量的形状为B 1×W 2；所述B 1为所述实时PPG二维张量的2维参数，且B 1＝实时片段总数；所述W 2为所述实时PPG二维张量的1维参数，且W 2＝W 1；
按照所述片段长度，对所述标定PPG数据序列，按顺序进行数据片段划分处理，得到标定片段总数个标定PPG一维张量；再将所述标定片段总数个所述标定PPG一维张量融合到一个二维张量中，生成标定PPG二维张量；
其中，所述标定片段总数＝int(L 2/片段长度)；所述标定PPG一维张量的形状为1×W 3；所述W 3为标定PPG一维张量的1维参数，且W 3＝片段长度；所述标定PPG二维张量的形状为B 2×W 4；所述B 2为所述标定PPG二维张量的2维参数，且B 2＝标定片段总数；所述W 4为所述标定PPG二维张量的1维参数，且W 4＝W 3；
对所述实时PPG二维张量和所述标定PPG二维张量，按照实时数据在前、标定数据在后的顺序，进行二维张量融合处理，生成融合二维张量；
其中，所述融合二维张量的形状为B 3×W 5；所述B 3为所述融合二维张量的2维参数，且B 3＝B 2＝B 1；所述W 5为所述融合二维张量的1维参数，且W 5＝W 2+W 4；所述融合二维张量包括所述B 3个融合一维张量；所述融合一维张量由对应的 所述实时PPG一维张量和所述标定PPG一维张量，按所述实时PPG一维张量在前、所述标定PPG一维张量在后的顺序，拼接而成；
按所述CNN模型的四维张量输入数据格式，将所述融合二维张量的形状从二维张量形状升维到四维张量形状，生成所述CNN输入四维张量；
其中，所述CNN输入四维张量的形状为B 4×H 1×W 6×C 1；所述B 4为所述CNN输入四维张量的4维参数，且B 4＝B 3；所述H 1为所述CNN输入四维张量的3维参数，且H 1＝2；所述W 6为所述CNN输入四维张量的2维参数，且W 6＝W 5/2；所述C 1为所述CNN输入四维张量的1维参数，且C 1＝1。
优选的，所述利用所述CNN模型，对所述CNN输入四维张量进行多层卷积池化计算，生成CNN输出四维张量，具体包括：
将所述CNN输入四维张量做为第一输入四维张量，再将所述第一输入四维张量送入所述CNN模型的第一层卷积网络层，进行第一层卷积池化计算，生成第一输出四维张量；接着将所述第一输出四维张量做为第二输入四维张量，再将所述第二输入四维张量送入所述CNN模型的第二层卷积网络层，进行第二层卷积池化计算，生成第二输出四维张量；直至最后，将倒数第二个输出四维张量做为最后一个输入四维张量，再将所述最后一个输入四维张量送入所述CNN模型的最后一层卷积网络层，进行最后一层卷积池化计算，生成所述CNN输出四维张量；
其中，所述CNN模型包括多层所述卷积网络层；所述卷积网络层包括卷积层和池化层；所述CNN输出四维张量的形状为B 5×H 2×W 7×C 2；所述B 5为所述CNN输出四维张量的4维参数，且B 5＝B 4；所述H 2为所述CNN输出四维张量的3维参数；所述W 7为所述CNN输出四维张量的2维参数；所述C 2为所述CNN输出四维张量的1维参数。
进一步的，所述将所述第一输入四维张量送入所述CNN模型的第一层卷积网络层，进行第一层卷积池化计算，生成第一输出四维张量，具体包括：
将所述第一输入四维张量，送入所述第一层卷积网络层的第一卷积层， 进行第一卷积计算，生成第一卷积四维张量；将所述第一卷积四维张量，送入所述第一层卷积网络层的第一池化层，进行第一池化计算，生成所述第一输出四维张量。
优选的，所述根据所述CNN输出四维张量，进行人工神经网络ANN模型的输入数据准备处理，生成ANN输入二维张量，具体包括：
按所述ANN模型的二维张量输入数据格式，将所述CNN输出四维张量的形状从四维张量形状降维到二维张量形状，生成所述ANN输入二维张量；
其中，所述ANN输入二维张量的形状为B 6×W' 1；所述B 6为所述ANN输入二维张量的2维参数，且B 6＝B 5；所述W' 1为所述ANN输入二维张量的1维参数，且W' 1＝H 2*W 7*C 2；所述ANN输入二维张量包括所述B 6个ANN输入一维张量；所述ANN输入一维张量的形状为1×W' 2；所述W' 2为所述ANN输入一维张量的1维参数,且W' 2＝W' 1；
在每个所述ANN输入一维张量的末端，增加所述标定舒张压数据和所述标定收缩压数据；所述ANN输入一维张量的形状变为1×W' 3；所述W' 3为所述ANN输入一维张量新的1维参数，且W' 3＝W' 2+2＝W' 1+2＝H 2*W 7*C 2+2；所述ANN输入二维张量的形状变为B 6×W 8；所述W 8为所述ANN输入二维张量新的1维参数，且W 8＝W' 3＝H 2*W 7*C 2+2。
优选的，所述利用所述ANN模型，对所述ANN输入二维张量进行回归计算，生成ANN输出二维张量，具体包括：
将所述ANN输入二维张量做为第一输入二维张量，再将所述第一输入二维张量送入所述ANN模型的第一层全连接层，进行第一层全连接计算，生成第一输出二维张量；接着将所述第一输出二维张量做为第二输入二维张量，再将所述第二输入二维张量送入所述ANN模型的第二层全连接层，进行第二层全连接计算，生成第二输出二维张量；直至最后，将倒数第二个输出二维张量做为最后一个输入二维张量，再将所述最后一个输入二维张量送入所述ANN模型的最后一层全连接层，进行最后一层全连接计算，生成所述ANN输出 二维张量；
其中，所述ANN模型包括多层所述全连接层；所述ANN输出二维张量的形状为B 7×W 9；所述B 7为所述ANN输出二维张量的2维参数，且B 7＝B 6；所述W 9为所述ANN输出二维张量的1维参数，且W 9＝2；所述ANN输出二维张量包括所述B 7个ANN输出一维张量；所述ANN输出一维张量包括舒张压相对数据和收缩压相对数据。
优选的，所述根据所述标定舒张压数据、所述标定收缩压数据和所述ANN输出二维张量，进行血压数据计算，生成血压二维张量，具体包括：
当预设的相对关系信息为差值关系时，使用所述标定舒张压数据，对所述ANN输出二维张量中的所有所述舒张压相对数据，进行舒张压增值处理；并使用所述标定收缩压数据，对所述ANN输出二维张量中的所有所述收缩压相对数据，进行收缩压增值处理；再将完成增值处理的所述ANN输出二维张量做为所述血压二维张量；
其中，所述血压二维张量的形状为B 8×W 10；所述B 8为所述血压二维张量的2维参数，且B 8＝B 7；所述W 10为所述血压二维张量的1维参数，且W 10＝2；所述血压二维张量包括所述B 8个血压一维张量；所述血压一维张量包括舒张压数据和收缩压数据；所述舒张压数据为对应的所述舒张压相对数据与所述标定舒张压数据相加的和；所述收缩压数据为对应的所述收缩压相对数据与所述标定收缩压数据相加的和。
优选的，所述当预设的预测类型信息为第一类型时，根据所述血压二维张量，进行均值血压数据计算，生成舒张压预测数据和收缩压预测数据；当所述预测类型信息为第二类型时，对所述血压二维张量，进行血压数据提取处理，生成舒张压预测数据序列和收缩压预测数据序列，具体包括：
当所述预测类型信息为所述第一类型时，计算所述血压二维张量中的所有所述舒张压数据的平均值，生成所述舒张压预测数据；计算所述血压二维张量中的所有所述收缩压数据的平均值，生成所述收缩压预测数据；
当所述预测类型信息为所述第二类型时，依次提取所述血压二维张量中的所述舒张压数据，组成所述舒张压预测数据序列；提取所述血压二维张量中的所述收缩压数据，组成所述收缩压预测数据序列。
本发明实施例第二方面提供了一种融合标定光体积描计信号数据的血压预测装置，包括：
获取模块用于获取标定光体积描计PPG信号数据和对应的标定舒张压数据与标定收缩压数据；获取实时PPG信号数据；其中，所述实时PPG信号数据的时间长度与所述标定PPG信号数据的时间长度相同；
人工智能计算模块用于根据所述标定PPG信号数据和所述实时PPG信号数据，进行卷积神经网络CNN模型的输入数据准备处理，生成CNN输入四维张量；接着利用所述CNN模型，对所述CNN输入四维张量进行多层卷积池化计算，生成CNN输出四维张量；再根据所述CNN输出四维张量，进行人工神经网络ANN模型的输入数据准备处理，生成ANN输入二维张量；接着利用所述ANN模型，对所述ANN输入二维张量进行回归计算，生成ANN输出二维张量；
血压预测模块用于根据所述标定舒张压数据、所述标定收缩压数据和所述ANN输出二维张量，进行血压数据计算，生成血压二维张量；当预设的预测类型信息为第一类型时，根据所述血压二维张量，进行均值血压数据计算，生成舒张压预测数据和收缩压预测数据；当所述预测类型信息为第二类型时，对所述血压二维张量，进行血压数据提取处理，生成舒张压预测数据序列和收缩压预测数据序列。
本发明实施例第三方面提供了一种电子设备，包括：存储器、处理器和收发器；
所述处理器用于与所述存储器耦合，读取并执行所述存储器中的指令，以实现上述第一方面所述的方法步骤；
所述收发器与所述处理器耦合，由所述处理器控制所述收发器进行消息 收发。
本发明实施例第四方面提供了一种计算机程序产品，所述计算机程序产品包括计算机程序代码，当所述计算机程序代码被计算机执行时，使得所述计算机执行上述第一方面所述的方法。
本发明实施例第五方面提供了一种计算机可读存储介质，所述计算机存储介质存储有计算机指令，当所述计算机指令被计算机执行时，使得所述计算机执行上述第一方面所述的方法的指令。
本发明实施例提供一种融合标定光体积描计信号数据的血压预测方法、装置、电子设备、计算机程序产品及计算机可读存储介质，首先获取测试对象的标定PPG信号数据和对应的标定血压数据(标定舒张压数据、标定收缩压数据)，再使用训练成熟的用于预测相对血压数据的CNN+ANN人工智能血压预测网络，对测试对象的实时PPG信号和标定PPG信号进行血压预测运算得到相对血压数据(舒张压相对数据、收缩压相对数据)，再根据相对关系信息对标定血压数据和相对血压数据进行绝对血压数据计算得到最终的血压数据；由此提高了人工智能血压预测网络的预测准确度。
附图说明
图1为本发明实施例一提供的一种融合标定光体积描计信号数据的血压预测方法示意图；
图2a为本发明实施例一提供的卷积神经网络的结构示意图；
图2b为本发明实施例一提供的人工神经网络的结构示意图；
图3为本发明实施例二提供的一种融合标定光体积描计信号数据的血压预测装置的模块结构图；
图4为本发明实施例三提供的一种电子设备的结构示意图。
具体实施方式
为了使本发明的目的、技术方案和优点更加清楚，下面将结合附图对本发明作进一步地详细描述，显然，所描述的实施例仅仅是本发明一部份实施例，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它实施例，都属于本发明保护的范围。
在对本发明实施例进行详细阐述之前，先对上文提及的人工智能血压预测网络中的CNN和ANN的结构和数据格式进行简要描述。
卷积神经网络CNN，是我们已知的特征识别领域的核心网络之一。应用血压特征识别领域中，是对一维PPG信号数据进行特征提取计算，在对输入的原始PPG信号数据进行卷积运算和池化运算之后，保留符合PPG信号特性的特征数据以供其他网络进行学习。文中提及的CNN模型是一种已经通过特征提取训练完成后的卷积神经网络模型，具体由多层卷积网络层组成，每层卷积网络层由1层卷积层和1层池化层组成。其中，卷积层负责对CNN模型的输入数据进行特征提取计算，池化层则是对卷积层的提取结果进行降采样,每层卷积网络层的输出数据将会做为下一层卷积网络层的输入数据。此处，每层卷积网络层的输入数据和输出数据格式均为4维张量形式：输入四维张量(形状为X i4×X i3×X i2×X i1)、输出四维张量(形状为X o4×X o3×X o2×X o1)，其中：X i4、X i3、X i2、X i1分别为输入四维张量的4、3、2和1维参数，X o4、X o3、X o2、X o1分别为输出四维张量的4、3、2和1维参数。每经过一层卷积网络层计算，输出数据的形状相对于输入数据的形状而言，维度参数的变化规律是：1)X o4相对于X i4，4维参数(在本实施例中对应PPG信号数据的片段总数)不会发生变化；2)X o3、X o2相对于X i3、X i2，3维和2维参数会发生变化，该变化与每一个卷积层的卷积核大小以及滑动步长的设定有关，也与池化层的池化窗口大小和滑动步长有关；3)X o1相对于X i1，1维参数的变化与卷积层中选定的输出空间维数(卷积核的个数)有关。
人工神经网络ANN，是指由大量的神经元(节点)互相连接而形成的复杂 网络结构，是对人脑组织结构和运行机制的某种抽象、简化和模拟。ANN以数学模型模拟神经元活动，是基于模仿大脑神经网络结构和功能而建立的一种信息处理系统。ANN常见应用是对数据进行回归计算。文中提及的ANN模型是一种已经通过回归训练完成后的人工神经网络模型；具体的，该ANN模型由多层全连接层组成。其中，每层全连接层包括多个结点，每一个结点都与上一层的所有结点相连，用来把前边提取到的结点数据综合起来进行一次结点计算，并将计算结果做为当前结点的值留待下一层全连接层的结点进行连接获取；这里的结点计算也叫全连接计算，包括节点连接计算和激活计算，常见的激活函数以使用整流线性单元(Rectified Linear Unit，ReLU)函数的情况居多，也可以使用其他函数来完成计算。此处，每层全连接层的输入数据与输出数据格式均为2维张量形式：输入二维张量(形状为Y i2×Y i1)、输出二维张量(形状为Y o2×Y o1)，其中：Y i2、Y i1为输入二维张量的2维和1维参数，Y o2、Y o1为输出二维张量的2维和1维参数。每经过一层全连接层计算，输出数据的形状相对于输入数据的形状而言，维度参数的变化规律是：1)Y o2相对于Y i2，2维参数(在本实施例中对应PPG信号数据的片段总数)不会发生变化；2)Y o1相对于Y i1，1维参数的变化与全连接层的结点总数有关，本发明实施例的最后一层全连接层的Y o1具体为2。本发明实施例的ANN模型最终输出的二维张量的形状为：片段总数×2，该二维张量包括片段总数个一维张量，每个一维张量对应一个片段，每个一维张量包括两个与对应片段相关的血压数据：舒张压相对数据和收缩压相对数据。
在本发明实施例中，因为需要将CNN模型的输出数据向ANN模型进行输入，所以需要将CNN模型的输出结果从输出四维张量形状(X o4×X o3×X o2×X o1)向ANN模型的输入二维张量形状(Y i2×Y' i1)进行降维处理；其中，Y' i1为输入二维张量的1维参数，Y i2＝X o4、Y' i1＝X o3*X o2*X o1；这里输入二维张量包括Y i2个输入一维张量。本发明实施例为进一步提高ANN模型的计算精度，在每个输入一维张量的末尾添加两个数据：标定舒张压数据与标定收缩压数据；从 而使得输入一维张量的形状从1×Y' i1变为1×(Y' i1+2)；对应的输入二维张量的形状也从Y i2×Y' i1变为Y i2×Y i1，其中Y i1＝Y' i1+2＝X o3*X o2*X o1+2。
在对本发明实施例进行详细阐述之前，再对本发明实施例的人工智能血压预测网络中的CNN和ANN的应用特点进行简要描述。
常规的人工智能血压预测网络，直接将实时PPG信号数据(从测试对象处采集到的实时的、需要进行血压预测的PPG信号数据)分段(之所以将PPG信号数据进行分段变成多个PPG信号数据片段，是受CNN模型的输入长度限制)后做为CNN模型的输入数据，并利用CNN模型进行PPG信号特征提取得到特征数据，再将特征数据输入到ANN模型中进行回归计算得到各个PPG信号数据片段对应的绝对收缩压、舒张压数据。在实际应用中，我们发现从不同的测试对象上可能会采集到近似的PPG信号数据，如果使用上述模型，那么计算得到的不同测试对象的绝对收缩压、舒张压数据就几乎一致。然而实际情况是，我们发现由于人与人之间的个体差异(例如，性别、年龄、身高、体重、臂展宽度、BMI、体温、是否引用咖啡浓茶、是否运动后等因素)，对于不同人而言，即便采集到的PPG信号数据近似，也可能对应不同的血压状态。
于是，为提高人工智能血压预测网络的预测准确度，本发明实施例对上述人工智能血压预测网络的CNN模型和ANN模型分别做了一些调整：调整后的CNN模型在接收实时PPG信号数据的同时，还接收标定PPG信号数据；且CNN模型最终输出的不是实时PPG信号数据的绝对特征数据，而是实时PPG信号数据对标定PPG信号数据、按特定相对关系、运算出的相对特征数据；此处，特定相对关系可以是差值、比值、对数比值、指数比值等关系。调整后的ANN模型的输入数据从绝对特征数据变为相对特征数据，输出数据也从绝对血压数据变成了相对血压数据。这里，所谓标定PPG信号数据，是对测试对象专门采集的一条时间长度与实时PPG信号数据一致的参考数据，且在采集标定PPG信号数据的同时也通过血压测量设备获取到了与这条标定PPG信 号数据对应的标定收缩压、舒张压数据。在从ANN模型的输出数据中获得收缩压、舒张压相对数据之后，可以根据特定相对关系(例如，差值、比值、对数比值、指数比值等关系)，结合标定收缩压、舒张压数据进行绝对血压的反向推导，最终得出与实时PPG信号数据对应的预测收缩压、舒张压数据。
使用上述融合了标定PPG信号数据和实时PPG信号数据的人工智能血压预测网络(CNN+ANN)，对从不同对象采集来的实时PPG信号数据进行预测时，即使实时数据近似，也可因为各自标定PPG信号数据存在的差异性，得出明显的相对差异性，自然最后预测的血压数据差异性也会比较明显。由此，就解决了上述常规预测网络测量失准的问题，也进一步提高了人工智能血压预测网络的预测准确度。
本发明实施例一提供一种融合标定光体积描计信号数据的血压预测方法，使用预测相对血压数据的人工智能血压预测网络(CNN+ANN)，对融合了标定PPG信号数据与实时PPG信号数据的输入数据进行预测运算，得到相对血压数据；再根据标定血压数据、相对血压数据，按特定相对关系，进行绝对血压数据的反推，得到与实时PPG信号对应的预测血压数据。使用本发明实施例一提供的方法，可以提高人工智能血压预测网络的预测准确度。
如图1为本发明实施例一提供的一种融合标定光体积描计信号数据的血压预测方法示意图所示，本方法主要包括如下步骤：
步骤1，获取标定光体积描计PPG信号数据和对应的标定舒张压数据与标定收缩压数据。
具体的，设备可以从本地存储介质中，获取标定PPG信号数据以及与标定PPG信号数据的标定舒张压数据和标定收缩压数据；设备还可以从与自身连接的其他终端设备、服务器或数据库的存储介质中，获取标定PPG信号数据、标定舒张压数据和标定收缩压数据。
此处，设备具体为实现本发明实施例提供的方法的终端设备或者服务器；标定PPG信号数据是对测试对象预先采集的一条PPG信号数据；对应的标定 舒张压数据和标定收缩压数据，是在采集该标定PPG信号数据的同时通过血压测量设备对测试对象进行血压测量获得的实际血压数据。
例如，获取的标定PPG信号数据长度为10秒，对应的标定舒张压数据为74毫米汞柱(millimetre Hg,mmHg)、标定收缩压数据为113mmHg。
步骤2，获取实时PPG信号数据；
其中，实时PPG信号数据的时间长度与标定PPG信号数据的时间长度相同。
具体的，设备可以通过自身的PPG信号采集装置对测试对象进行实时的PPG信号采集及数据采样处理，得到实时PPG信号数据；设备还可以通过与自身连接的PPG信号采集装置对测试对象进行实时的PPG信号采集及数据采样处理，得到实时PPG信号数据。
例如，获取到的实时PPG信号数据长度为10秒，与标定PPG信号数据长度一致。
步骤3，根据标定PPG信号数据和实时PPG信号数据，进行卷积神经网络CNN模型的输入数据准备处理，生成CNN输入四维张量；
具体包括：步骤31，按照预设的采样频率，对实时PPG信号数据进行信号数据采样处理生成实时PPG数据序列，对标定PPG信号数据进行信号数据采样处理生成标定PPG数据序列；
其中，实时PPG数据序列的数据长度L 1与标定PPG数据序列的数据长度L 2相同；
此处，采样频率存储在设备的本地存储介质中；
例如，实时、标定PPG信号数据均为10秒，采样频率阈值为125赫兹(hertz,Hz)，则实时、标定PPG数据序列的数据长度：L 1＝L 2＝10*125＝1250，实时PPG数据序列具体为实时PPG数据序列[1250]，标定PPG数据序列具体为标定PPG数据序列[1250]；
步骤32，按照预设的片段长度，对实时PPG数据序列，按顺序进行数据 片段划分处理，得到实时片段总数个实时PPG一维张量；再将实时片段总数个实时PPG一维张量融合到一个二维张量中，生成实时PPG二维张量；
其中，实时片段总数＝int(L 1/片段长度)；int()为取整函数；实时PPG一维张量的形状为1×W 1；W 1为实时PPG一维张量的1维参数，且W 1＝片段长度；实时PPG二维张量的形状为B 1×W 2；B 1为实时PPG二维张量的2维参数，且B 1＝实时片段总数；W 2为实时PPG二维张量的1维参数，且W 2＝W 1；
此处，片段长度存储在设备的本地存储介质中；
这里，片段长度的值由CNN模型的输入数据长度决定；片段划分方式采用顺序片段划分方式，则相邻片段之间的数据没有重合；
例如，片段长度为250，对实时PPG数据序列[1250]，按顺序进行数据片段划分处理，则B 1＝实时片段总数＝int(1250/250)＝5，W 2＝W 1＝250，实时PPG二维张量的形状具体为5×250，这里将之表示为实时PPG二维张量[5,250]；
步骤33，按照片段长度，对标定PPG数据序列，按顺序进行数据片段划分处理，得到标定片段总数个标定PPG一维张量；再将标定片段总数个标定PPG一维张量融合到一个二维张量中，生成标定PPG二维张量；
其中，标定片段总数＝int(L 2/片段长度)；标定PPG一维张量的形状为1×W 3；W 3为标定PPG一维张量的1维参数，且W 3＝片段长度；标定PPG二维张量的形状为B 2×W 4；B 2为标定PPG二维张量的2维参数，且B 2＝标定片段总数；W 4为标定PPG二维张量的1维参数，且W 4＝W 3；
例如，片段长度为250，对标定PPG数据序列[1250]，按顺序进行数据片段划分处理，则B 2＝标定片段总数＝int(1250/250)＝5，W 4＝W 3＝250，标定PPG二维张量的形状具体为5×250，这里将之表示为标定PPG二维张量[5,250]；
步骤34，对实时PPG二维张量和标定PPG二维张量，按照实时数据在前、标定数据在后的顺序，进行二维张量融合处理，生成融合二维张量；
其中，融合二维张量的形状为B 3×W 5；B 3为融合二维张量的2维参数，且B 3＝B 2＝B 1；W 5为融合二维张量的1维参数，且W 5＝W 2+W 4；融合二维张量包括B 3 个融合一维张量；融合一维张量由对应的实时PPG一维张量和标定PPG一维张量，按实时PPG一维张量在前、标定PPG一维张量在后的顺序，拼接而成；
例如，实时PPG二维张量[5,250]的数据具体为{(Z 1,1,…Z 1,250),(Z 2,1,…Z 2,250),(Z 3,1,…Z 3,250),(Z 4,1,…Z 4,250),(Z 5,1,…Z 5,250)}，其中(Z 1,1,…Z 1,250)、(Z 2,1,…Z 2,250)、(Z 3,1,…Z 3,250)、(Z 4,1,…Z 4,250)、(Z 5,1,…Z 5,250)依次为实时PPG二维张量中第1、第2、第3、第4、第5实时PPG一维张量的数据；
标定PPG二维张量[5,250]的数据具体为{(D 1,1,…D 1,250),(D 2,1,…D 2,250),(D 3,1,…D 3,250),(D 4,1,…D 4,250),(D 5,1,…D 5,250)}，其中(D 1,1,…D 1,250)、(D 2,1,…D 2,250)、(D 3,1,…D 3,250)、(D 4,1,…D 4,250)、(D 5,1,…D 5,250)依次为标定PPG二维张量中第1、第2、第3、第4、第5标定PPG一维张量的数据；
则，对实时PPG二维张量[5,250]和标定PPG二维张量[5,250]进行二维张量融合处理之后，B 3＝B 2＝B 1＝5，W 5＝W 2+W 4＝250+250＝500，融合二维张量的形状为5×500，这里将之表示为融合二维张量[5，500]；
在融合二维张量[5，500]中，包括5个融合一维张量，每个融合一维张量内包括W 5＝W 2+W 4＝250+250＝500个数据，其中前250个数据是实时PPG二维张量[5,250]中对应的实时PPG一维张量的数据，后250个数据是标定PPG二维张量[5,250]中对应的标定PPG一维张量的数据；
即，融合二维张量的数据格式应是：
{(Z 1,1,…Z 1,250,D 1,1,…D 1,250),(Z 2,1,…Z 2,250,D 2,1,…D 2,250),(Z 3,1,…Z 3,250,D 3,1,…D 3,250),(Z 4,1,…Z 4,250,D 4,1,…D 4,250),(Z 5,1,…Z 5,250,D 5,1,…D 5,250)}，
其中，
(Z 1,1,…Z 1,250,D 1,1,…D 1,250),(Z 2,1,…Z 2,250,D 2,1,…D 2,250),(Z 3,1,…Z 3,250,D 3,1,…D 3, 250),(Z 4,1,…Z 4,250,D 4,1,…D 4,250),(Z 5,1,…Z 5,250,D 5,1,…D 5,250)依次为融合二维张量中第1、第2、第3、第4、第5融合一维张量的数据；
步骤35，按CNN模型的四维张量输入数据格式，将融合二维张量的形状从二维张量形状升维到四维张量形状，生成CNN输入四维张量；
其中，CNN输入四维张量的形状为B 4×H 1×W 6×C 1；B 4为CNN输入四维张量的4维参数，且B 4＝B 3；H 1为CNN输入四维张量的3维参数，且H 1＝2；W 6为CNN输入四维张量的2维参数，且W 6＝W 5/2；C 1为CNN输入四维张量的1维参数，且C 1＝1。
这里，将融合二维张量的形状从二维张量形状升维到四维张量形状，其过程只对张量形状重新设置，没有破坏张量内的实际数据顺序。
例如，将融合二维张量[5，500]从二维张量形状升维到四维张量形状，B 4＝B 3＝5，W 6＝W 5/2＝5/2＝250，得到的CNN输入四维张量的形状为5×2×250×1，这里将之表示为CNN输入四维张量[5,2,250,1]。
步骤4，利用CNN模型，对CNN输入四维张量进行多层卷积池化计算，生成CNN输出四维张量；
具体包括：将CNN输入四维张量做为第一输入四维张量，再将第一输入四维张量送入CNN模型的第一层卷积网络层，进行第一层卷积池化计算，生成第一输出四维张量；接着将第一输出四维张量做为第二输入四维张量，再将第二输入四维张量送入CNN模型的第二层卷积网络层，进行第二层卷积池化计算，生成第二输出四维张量；直至最后，将倒数第二个输出四维张量做为最后一个输入四维张量，再将最后一个输入四维张量送入CNN模型的最后一层卷积网络层，进行最后一层卷积池化计算，生成CNN输出四维张量；
其中，CNN模型包括多层卷积网络层；卷积网络层包括卷积层和池化层；CNN输出四维张量的形状为B 5×H 2×W 7×C 2；B 5为CNN输出四维张量的4维参数，且B 5＝B 4；H 2为CNN输出四维张量的3维参数；W 7为CNN输出四维张量的2维参数；C 2为CNN输出四维张量的1维参数。
其中，将第一输入四维张量送入CNN模型的第一层卷积网络层，进行第一层卷积池化计算，生成第一输出四维张量，具体为将第一输入四维张量， 送入第一层卷积网络层的第一卷积层，进行第一卷积计算，生成第一卷积四维张量；将第一卷积四维张量，送入第一层卷积网络层的第一池化层，进行第一池化计算，生成第一输出四维张量。
例如，CNN模型的网络结构如图2a为本发明实施例一提供的卷积神经网络的结构示意图所示，总层数为四层，则，
将CNN输入四维张量做为第一输入四维张量；再将第一输入四维张量送入CNN模型的第一层卷积网络层的第一卷积层进行第一卷积计算，生成第一卷积四维张量；再将第一卷积四维张量，送入第一层卷积网络层的第一池化层进行第一池化计算，生成第一输出四维张量；
将第一输出四维张量做为第二输入四维张量；再将第二输入四维张量送入CNN模型的第二层卷积网络层的第二卷积层进行第二卷积计算，生成第二卷积四维张量；再将第二卷积四维张量，送入第二层卷积网络层的第二池化层进行第二池化计算，生成第二输出四维张量；
将第二输出四维张量做为第三输入四维张量；再将第三输入四维张量送入CNN模型的第三层卷积网络层的第三卷积层进行第三卷积计算，生成第三卷积四维张量；再将第三卷积四维张量，送入第三层卷积网络层的第三池化层进行第三池化计算，生成第三输出四维张量；
将第三输出四维张量做为第四输入四维张量；再将第四输入四维张量送入CNN模型的第四层卷积网络层的第四卷积层进行第四卷积计算，生成第四卷积四维张量；再将第四卷积四维张量，送入第四层卷积网络层的第四池化层进行第四池化计算，生成第四输出四维张量；这里的第四输出四维张量就是最后输出的CNN输出四维张量。
这里由前文可知，在CNN模型中，每经过一层卷积层或池化层，输入数据的形状都会发生变化，但依然保持4维张量形式，其中4维参数(片段总数)不会发生变化；3维、2维参数的变化与每一个卷积层的卷积核大小以及滑动步长的设定有关，也与池化层的池化窗口大小和滑动步长有关；1维参数 的变化与卷积层中选定的输出空间维数(卷积核的个数)有关。在实际应用中，网络中层数的设定，以及每一层各种参数的设定都要根据经验和实验结果进行不停的修正。
步骤5，根据CNN输出四维张量，进行人工神经网络ANN模型的输入数据准备处理，生成ANN输入二维张量；
具体包括：步骤51，按ANN模型的二维张量输入数据格式，将CNN输出四维张量的形状从四维张量形状降维到二维张量形状，生成ANN输入二维张量；其中，ANN输入二维张量的形状为B 6×W' 1；B 6为ANN输入二维张量的2维参数，且B 6＝B 5；W' 1为ANN输入二维张量的1维参数，且W' 1＝H 2*W 7*C 2；ANN输入二维张量包括B 6个ANN输入一维张量；ANN输入一维张量的形状为1×W' 2；W' 2为ANN输入一维张量的1维参数,且W' 2＝W' 1；
这里，将CNN输出四维张量的形状从四维张量形状降维到二维张量形状实际就是将CNN输出四维张量的形状从四维降到二维，其过程只对张量形状重新设置，没有破坏张量内的实际数据顺序；
例如，CNN输出四维张量的形状为5×2×20×64，则B 6＝B 5＝5，W 8＝H 2*W 7*C 2＝2*20*64＝2560，ANN输入二维张量的形状应为5×2560，这里将之表示为ANN输入二维张量[5,2560]；对应的，ANN输入二维张量包括5个ANN输入一维张量，ANN输入一维张量的形状应为1×2560，这里将之表示为ANN输入一维张量[2560]；
步骤52，在每个ANN输入一维张量的末端，增加标定舒张压数据和标定收缩压数据；ANN输入一维张量的形状变为1×W' 3；W' 3为ANN输入一维张量的1维参数，且W' 3＝W' 2+2＝W' 1+2＝H 2*W 7*C 2+2；ANN输入二维张量的形状变为B 6×W 8；W 8为ANN输入二维张量新的1维参数，且W 8＝W' 3＝H 2*W 7*C 2+2。
这里，在每个ANN输入一维张量的末端，添加两个数据：标定舒张压数据和标定收缩压数据，以此来提高ANN模型的计算精度；完成数据添加后，ANN输入一维张量的形状从1×W' 2变为1×W' 3；对应的ANN输入二维张量的形 状从B 6×W' 1变为B 6×W 8；
例如，未做数据添加前的ANN输入二维张量的形状为5×2560，未做数据添加前的ANN输入一维张量的形状为1×2560；完成了数据添加之后的ANN输入一维张量的形状变为1×2562，对应的ANN输入二维张量的形状变为5×2562，这里将之表示为ANN输入二维张量[5,2562]。
步骤6，利用ANN模型，对ANN输入二维张量进行回归计算，生成ANN输出二维张量；
具体包括：将ANN输入二维张量做为第一输入二维张量，再将第一输入二维张量送入ANN模型的第一层全连接层，进行第一层全连接计算，生成第一输出二维张量；接着将第一输出二维张量做为第二输入二维张量，再将第二输入二维张量送入ANN模型的第二层全连接层，进行第二层全连接计算，生成第二输出二维张量；直至最后，将倒数第二个输出二维张量做为最后一个输入二维张量，再将最后一个输入二维张量送入ANN模型的最后一层全连接层，进行最后一层全连接计算，生成ANN输出二维张量；
其中，ANN模型包括多层全连接层；ANN输出二维张量的形状为B 7×W 9；B 7为ANN输出二维张量的2维参数，且B 7＝B 6；W 9为ANN输出二维张量的1维参数，且W 9＝2；ANN输出二维张量包括B 7个ANN输出一维张量；ANN输出一维张量包括舒张压相对数据和收缩压相对数据。
例如，ANN模型的网络结构如图2b为本发明实施例一提供的人工神经网络的结构示意图所示,ANN模型包括四层全连接层，则
将ANN输入二维张量做为第一输入二维张量，再将第一输入二维张量送入ANN模型的第一层全连接层，进行第一层全连接计算，生成第一输出二维张量；
接着将第一输出二维张量做为第二输入二维张量，再将第二输入二维张量送入ANN模型的第二层全连接层，进行第二层全连接计算，生成第二输出二维张量；
接着将第二输出二维张量做为第三输入二维张量，再将第三输入二维张量送入ANN模型的第三层全连接层，进行第三层全连接计算，生成第三输出二维张量；
最后将第三输出二维张量做为第四输入二维张量，再将第四输入二维张量送入ANN模型的第四层全连接层，进行第四层全连接计算，生成第四输出二维张量；这里的第四输出二维张量就是最后输出的ANN输出二维张量；
这里由前文可知，ANN模型由全连接层组成，全连接层的每一个结点都与上一层的所有结点相连，用来把前边提取到的特征综合起来，每层全连接层可以设置该层的结点个数以及激活函数(ReLU较多，也可以改成其他)。本实施例ANN模型的最后一层全连接层的节点数为2，对应的最终的输出二维张量的形状具体为B 7×2。
例如，ANN模型包括四层全连接层，ANN输入二维张量形状为5×2560，则B 7＝B 6＝5，输出二维张量的形状为5×2；这里将之表示为输出二维张量[5,2]；
ANN输出二维张量[5,2]内的数据应为
{(R db1,R sb1),(R db2,R sb2),(R db3,R sb3),(R db4,R sb4),(R db5,R sb5)}，
其中，(R db1,R sb1),(R db2,R sb2),(R db3,R sb3),(R db4,R sb4),(R db5,R sb5)依次为ANN输出二维张量中第1、第2、第3、第4、第5个ANN输出一维张量的数据(R sbi为收缩压相对数据、R dbi为舒张压相对数据，i的取值从1到5)。
步骤7，根据标定舒张压数据、标定收缩压数据和ANN输出二维张量，进行血压数据计算，生成血压二维张量；
具体包括：当预设的相对关系信息为差值关系时，使用标定舒张压数据，对ANN输出二维张量中的所有舒张压相对数据，进行舒张压增值处理；并使用标定收缩压数据，对ANN输出二维张量中的所有收缩压相对数据，进行收缩压增值处理；再将完成增值处理的ANN输出二维张量做为血压二维张量；
其中，血压二维张量的形状为B 8×W 10；B 8为血压二维张量的2维参数，且B 8＝B 7；W 10为血压二维张量的1维参数，且W 10＝2；血压二维张量包括B 8个血压 一维张量；血压一维张量包括舒张压数据和收缩压数据；舒张压数据为对应的舒张压相对数据与标定舒张压数据相加的和；收缩压数据为对应的收缩压相对数据与标定收缩压数据相加的和。
此处，相对关系信息至少包括差值相对关系(差值关系)。
这里，在实施例一中，相对关系信息具体为差值关系时，意味着将ANN输出二维张量中的舒张压数相对据加上标定舒张压数据，就可以得到舒张压数据；将收缩压相对数据加上标定收缩压数据，就可以得到收缩压数据。
例如，
ANN输出二维张量[5,2]为{(7,23),(6,22),(7,21),(11,20),(9,18)}，
标定舒张压数据为74mmHg、标定收缩压数据为113mmHg，B 8＝B 7＝5，W 10＝2，血压二维张量的形状为5×2(这里将之表示为血压二维张量[5,2])；
则，血压二维张量[5,2]的数据应为：
{(7+74,23+113),(6+74,22+113),(7+74,21+113),(11+74,20+113),(9+74,18+113)}＝{(81,136),(80,135),(81,134),(85,133),(83,131)}。
步骤8，当预设的预测类型信息为第一类型时，根据血压二维张量，进行均值血压数据计算，生成舒张压预测数据和收缩压预测数据；当预测类型信息为第二类型时，对血压二维张量，进行血压数据提取处理，生成舒张压预测数据序列和收缩压预测数据序列；
具体包括：当预测类型信息为第一类型时，计算血压二维张量中的所有舒张压数据的平均值，生成舒张压预测数据；计算血压二维张量中的所有收缩压数据的平均值，生成收缩压预测数据；当预测类型信息为第二类型时，依次提取血压二维张量中的舒张压数据，组成舒张压预测数据序列；提取血压二维张量中的收缩压数据，组成收缩压预测数据序列。
此处，本发明实施例支持两种类型的预测数据输出：第一类型，输出一对均值血压预测数据(舒张压预测数据和收缩压预测数据)；第二类型，按片段提取血压数据，组成动态的血压数据序列(舒张压预测数据序列和收缩压 预测数据序列)。预测类型信息存储在设备的本地存储介质中，本发明实施例通过读取预测类型信息的内容，来选择具体采用哪种类型进行预测数据输出。
例如，预测类型信息为第一类型，血压二维张量[5,2]的数据为{(81,136),(80,135),(81,134),(85,133),(83,131)}，则:
舒张压预测数据＝(81+80+81+85+83)/5＝82(mmHg)；
收缩压预测数据＝(136+135+134+133+131)/5≈134(mmHg)。
又例如，预测类型信息为第二类型，血压二维张量[5,2]的数据为{(81,136),(80,135),(81,134),(85,133),(83,131)}，则:
舒张压预测数据序列[5]的数据内容具体为(81,80,81,85,83)；
收缩压预测数据序列[5]的数据内容具体为(136,135,134,133,131)。
图3为本发明实施例二提供的一种融合标定光体积描计信号数据的血压预测装置的模块结构图，该装置可以为前述实施例所描述的终端设备或者服务器，也可以为能够使得前述终端设备或者服务器实现本发明实施例提供的方法的装置，例如该装置可以是前述终端设备或者服务器的装置或芯片系统。如图3所示，该装置包括：
获取模块31用于获取标定光体积描计PPG信号数据和对应的标定舒张压数据与标定收缩压数据；获取实时PPG信号数据；其中，实时PPG信号数据的时间长度与标定PPG信号数据的时间长度相同；
人工智能计算模块32用于根据标定PPG信号数据和实时PPG信号数据，进行卷积神经网络CNN模型的输入数据准备处理，生成CNN输入四维张量；接着利用CNN模型，对CNN输入四维张量进行多层卷积池化计算，生成CNN输出四维张量；再根据CNN输出四维张量，进行人工神经网络ANN模型的输入数据准备处理，生成ANN输入二维张量；接着利用ANN模型，对ANN输入二维张量进行回归计算，生成ANN输出二维张量；
血压预测模块33用于根据标定舒张压数据、标定收缩压数据和ANN输出二维张量，进行血压数据计算，生成血压二维张量；当预设的预测类型信息 为第一类型时，根据血压二维张量，进行均值血压数据计算，生成舒张压预测数据和收缩压预测数据；当预测类型信息为第二类型时，对血压二维张量，进行血压数据提取处理，生成舒张压预测数据序列和收缩压预测数据序列。
本发明实施例提供的一种融合标定光体积描计信号数据的血压预测装置，可以执行上述方法实施例中的方法步骤，其实现原理和技术效果类似，在此不再赘述。
需要说明的是，应理解以上装置的各个模块的划分仅仅是一种逻辑功能的划分，实际实现时可以全部或部分集成到一个物理实体上，也可以物理上分开。且这些模块可以全部以软件通过处理元件调用的形式实现；也可以全部以硬件的形式实现；还可以部分模块通过处理元件调用软件的形式实现，部分模块通过硬件的形式实现。例如，获取模块可以为单独设立的处理元件，也可以集成在上述装置的某一个芯片中实现，此外，也可以以程序代码的形式存储于上述装置的存储器中，由上述装置的某一个处理元件调用并执行以上确定模块的功能。其它模块的实现与之类似。此外这些模块全部或部分可以集成在一起，也可以独立实现。这里所描述的处理元件可以是一种集成电路，具有信号的处理能力。在实现过程中，上述方法的各步骤或以上各个模块可以通过处理器元件中的硬件的集成逻辑电路或者软件形式的指令完成。
例如，以上这些模块可以是被配置成实施以上方法的一个或多个集成电路，例如：一个或多个特定集成电路(Application Specific Integrated Circuit，ASIC)，或，一个或多个数字信号处理器(Digital Signal Processor，DSP)，或，一个或者多个现场可编程门阵列(Field Programmable Gate Array，FPGA)等。再如，当以上某个模块通过处理元件调度程序代码的形式实现时，该处理元件可以是通用处理器，例如中央处理器(Central Processing Unit，CPU)或其它可以调用程序代码的处理器。再如，这些模块可以集成在一起，以片上系统(System-on-a-chip，SOC)的形式实现。
在上述实施例中，可以全部或部分地通过软件、硬件、固件或者其任意 组合来实现。当使用软件实现时，可以全部或部分地以计算机程序产品的形式实现。该计算机程序产品包括一个或多个计算机指令。在计算机上加载和执行该计算机程序指令时，全部或部分地产生按照本发明实施例所描述的流程或功能。上述计算机可以是通用计算机、专用计算机、计算机网络、或者其他可编程装置。上述计算机指令可以存储在计算机可读存储介质中，或者从一个计算机可读存储介质向另一个计算机可读存储介质传输，例如，上述计算机指令可以从一个网站站点、计算机、服务器或数据中心通过有线(例如同轴电缆、光纤、数字用户线路((Digital Subscriber Line，DSL))或无线(例如红外、无线、蓝牙、微波等)方式向另一个网站站点、计算机、服务器或数据中心进行传输。上述计算机可读存储介质可以是计算机能够存取的任何可用介质或者是包含一个或多个可用介质集成的服务器、数据中心等数据存储设备。上述可用介质可以是磁性介质，(例如，软盘、硬盘、磁带)、光介质(例如，DVD)、或者半导体介质(例如固态硬盘(solid state disk，SSD))等。
图4为本发明实施例三提供的一种电子设备的结构示意图。该电子设备可以为前述的终端设备或者服务器，也可以为与前述终端设备或者服务器连接的实现本发明实施例方法的终端设备或服务器。如图4所示，该电子设备可以包括：处理器41(例如CPU)、存储器42、收发器43；收发器43耦合至处理器41，处理器41控制收发器43的收发动作。存储器42中可以存储各种指令，以用于完成各种处理功能以及实现本发明上述实施例中提供的方法和处理过程。优选的，本发明实施例涉及的电子设备还包括：电源44、系统总线45以及通信端口46。系统总线45用于实现元件之间的通信连接。上述通信端口46用于电子设备与其他外设之间进行连接通信。
在图4中提到的系统总线可以是外设部件互连标准(Peripheral Component Interconnect，PCI)总线或扩展工业标准结构(Extended Industry Standard Architecture，EISA)总线等。该系统总线可以分为地址总线、数据总线、控制总线等。为便于表示，图中仅用一条粗线表示，但并 不表示仅有一根总线或一种类型的总线。通信接口用于实现数据库访问装置与其他设备(例如客户端、读写库和只读库)之间的通信。存储器可能包含随机存取存储器(Random Access Memory，RAM)，也可能还包括非易失性存储器(Non-Volatile Memory)，例如至少一个磁盘存储器。
上述的处理器可以是通用处理器，包括中央处理器CPU、网络处理器(Network Processor，NP)、图形处理器(Graphics Processing Unit，GPU)等；还可以是数字信号处理器DSP、专用集成电路ASIC、现场可编程门阵列FPGA或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。
需要说明的是，本发明实施例还提供一种计算机可读存储介质，该存储介质中存储有指令，当其在计算机上运行时，使得计算机执行上述实施例中提供的方法和处理过程。
本发明实施例还提供一种运行指令的芯片，该芯片用于执行上述实施例中提供的方法和处理过程。
本发明实施例还提供一种程序产品，该程序产品包括计算机程序，该计算机程序存储在存储介质中，至少一个处理器可以从上述存储介质读取上述计算机程序，上述至少一个处理器执行上述实施例中提供的方法和处理过程。
本发明实施例提供一种融合标定光体积描计信号数据的血压预测方法、装置、电子设备、计算机程序产品及计算机可读存储介质，首先获取测试对象的标定PPG信号数据和对应的标定血压数据(标定舒张压数据和标定收缩压数据)；再使用训练成熟的用于预测相对血压数据的CNN+ANN人工智能血压预测网络，对测试对象的实时PPG信号数据和标定PPG信号数据的融合数据，进行血压预测运算得到相对血压数据(包括舒张压相对数据和收缩压相对数据)；再根据相对关系信息，对标定血压数据和相对血压数据，进行绝对血压数据计算得到最终的预测血压数据。由此，提高了人工智能血压预测网络的预测准确度。
专业人员应该还可以进一步意识到，结合本文中所公开的实施例描述的 各示例的单元及算法步骤，能够以电子硬件、计算机软件或者二者的结合来实现，为了清楚地说明硬件和软件的可互换性，在上述说明中已经按照功能一般性地描述了各示例的组成及步骤。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本发明的范围。
结合本文中所公开的实施例描述的方法或算法的步骤可以用硬件、处理器执行的软件模块，或者二者的结合来实施。软件模块可以置于随机存储器(RAM)、内存、只读存储器(ROM)、电可编程ROM、电可擦除可编程ROM、寄存器、硬盘、可移动磁盘、CD-ROM、或技术领域内所公知的任意其它形式的存储介质中。
以上所述的具体实施方式，对本发明的目的、技术方案和有益效果进行了进一步详细说明，所应理解的是，以上所述仅为本发明的具体实施方式而已，并不用于限定本发明的保护范围，凡在本发明的精神和原则之内，所做的任何修改、等同替换、改进等，均应包含在本发明的保护范围之内。

Claims (12)

  1. 一种融合标定光体积描计信号数据的血压预测方法，其特征在于，所述方法包括：
    获取标定光体积描计PPG信号数据和对应的标定舒张压数据与标定收缩压数据；
    获取实时PPG信号数据；所述实时PPG信号数据的时间长度与所述标定PPG信号数据的时间长度相同；
    根据所述标定PPG信号数据和所述实时PPG信号数据，进行卷积神经网络CNN模型的输入数据准备处理，生成CNN输入四维张量；
    利用所述CNN模型，对所述CNN输入四维张量进行多层卷积池化计算，生成CNN输出四维张量；
    根据所述CNN输出四维张量，进行人工神经网络ANN模型的输入数据准备处理，生成ANN输入二维张量；
    利用所述ANN模型，对所述ANN输入二维张量进行回归计算，生成ANN输出二维张量；
    根据所述标定舒张压数据、所述标定收缩压数据和所述ANN输出二维张量，进行血压数据计算，生成血压二维张量；
    当预设的预测类型信息为第一类型时，根据所述血压二维张量，进行均值血压数据计算，生成舒张压预测数据和收缩压预测数据；当所述预测类型信息为第二类型时，对所述血压二维张量，进行血压数据提取处理，生成舒张压预测数据序列和收缩压预测数据序列。
  2. 根据权利要求1所述的融合标定光体积描计信号数据的血压预测方法，其特征在于，所述根据所述标定PPG信号数据和所述实时PPG信号数据，进行卷积神经网络CNN模型的输入数据准备处理，生成CNN输入四维张量，具体包括：
    按照预设的采样频率，对所述实时PPG信号数据进行信号数据采样处理 生成实时PPG数据序列，对所述标定PPG信号数据进行信号数据采样处理生成标定PPG数据序列；所述实时PPG数据序列的数据长度L 1与所述标定PPG数据序列的数据长度L 2相同；
    按照预设的片段长度，对所述实时PPG数据序列，按顺序进行数据片段划分处理，得到实时片段总数个实时PPG一维张量；再将所述实时片段总数个所述实时PPG一维张量融合到一个二维张量中，生成实时PPG二维张量；
    其中，所述实时片段总数＝int(L 1/片段长度)；所述int()为取整函数；所述实时PPG一维张量的形状为1×W 1；所述W 1为实时PPG一维张量的1维参数，且W 1＝片段长度；所述实时PPG二维张量的形状为B 1×W 2；所述B 1为所述实时PPG二维张量的2维参数，且B 1＝实时片段总数；所述W 2为所述实时PPG二维张量的1维参数，且W 2＝W 1；
    按照所述片段长度，对所述标定PPG数据序列，按顺序进行数据片段划分处理，得到标定片段总数个标定PPG一维张量；再将所述标定片段总数个所述标定PPG一维张量融合到一个二维张量中，生成标定PPG二维张量；
    其中，所述标定片段总数＝int(L 2/片段长度)；所述标定PPG一维张量的形状为1×W 3；所述W 3为标定PPG一维张量的1维参数，且W 3＝片段长度；所述标定PPG二维张量的形状为B 2×W 4；所述B 2为所述标定PPG二维张量的2维参数，且B 2＝标定片段总数；所述W 4为所述标定PPG二维张量的1维参数，且W 4＝W 3；
    对所述实时PPG二维张量和所述标定PPG二维张量，按照实时数据在前、标定数据在后的顺序，进行二维张量融合处理，生成融合二维张量；
    其中，所述融合二维张量的形状为B 3×W 5；所述B 3为所述融合二维张量的2维参数，且B 3＝B 2＝B 1；所述W 5为所述融合二维张量的1维参数，且W 5＝W 2+W 4；所述融合二维张量包括所述B 3个融合一维张量；所述融合一维张量由对应的所述实时PPG一维张量和所述标定PPG一维张量，按所述实时PPG一维张量在前、所述标定PPG一维张量在后的顺序，拼接而成；
    按所述CNN模型的四维张量输入数据格式，将所述融合二维张量的形状从二维张量形状升维到四维张量形状，生成所述CNN输入四维张量；
    其中，所述CNN输入四维张量的形状为B 4×H 1×W 6×C 1；所述B 4为所述CNN输入四维张量的4维参数，且B 4＝B 3；所述H 1为所述CNN输入四维张量的3维参数，且H 1＝2；所述W 6为所述CNN输入四维张量的2维参数，且W 6＝W 5/2；所述C 1为所述CNN输入四维张量的1维参数，且C 1＝1。
  3. 根据权利要求2所述的融合标定光体积描计信号数据的血压预测方法，其特征在于，所述利用所述CNN模型，对所述CNN输入四维张量进行多层卷积池化计算，生成CNN输出四维张量，具体包括：
    将所述CNN输入四维张量做为第一输入四维张量，再将所述第一输入四维张量送入所述CNN模型的第一层卷积网络层，进行第一层卷积池化计算，生成第一输出四维张量；接着将所述第一输出四维张量做为第二输入四维张量，再将所述第二输入四维张量送入所述CNN模型的第二层卷积网络层，进行第二层卷积池化计算，生成第二输出四维张量；直至最后，将倒数第二个输出四维张量做为最后一个输入四维张量，再将所述最后一个输入四维张量送入所述CNN模型的最后一层卷积网络层，进行最后一层卷积池化计算，生成所述CNN输出四维张量；
    其中，所述CNN模型包括多层所述卷积网络层；所述卷积网络层包括卷积层和池化层；所述CNN输出四维张量的形状为B 5×H 2×W 7×C 2；所述B 5为所述CNN输出四维张量的4维参数，且B 5＝B 4；所述H 2为所述CNN输出四维张量的3维参数；所述W 7为所述CNN输出四维张量的2维参数；所述C 2为所述CNN输出四维张量的1维参数。
  4. 根据权利要求3所述的融合标定光体积描计信号数据的血压预测方法，其特征在于，所述将所述第一输入四维张量送入所述CNN模型的第一层卷积网络层，进行第一层卷积池化计算，生成第一输出四维张量，具体包括：
    将所述第一输入四维张量，送入所述第一层卷积网络层的第一卷积层， 进行第一卷积计算，生成第一卷积四维张量；将所述第一卷积四维张量，送入所述第一层卷积网络层的第一池化层，进行第一池化计算，生成所述第一输出四维张量。
  5. 根据权利要求3所述的融合标定光体积描计信号数据的血压预测方法，其特征在于，所述根据所述CNN输出四维张量，进行人工神经网络ANN模型的输入数据准备处理，生成ANN输入二维张量，具体包括：
    按所述ANN模型的二维张量输入数据格式，将所述CNN输出四维张量的形状从四维张量形状降维到二维张量形状，生成所述ANN输入二维张量；
    其中，所述ANN输入二维张量的形状为B 6×W' 1；所述B 6为所述ANN输入二维张量的2维参数，且B 6＝B 5；所述W' 1为所述ANN输入二维张量的1维参数，且W' 1＝H 2*W 7*C 2；所述ANN输入二维张量包括所述B 6个ANN输入一维张量；所述ANN输入一维张量的形状为1×W' 2；所述W' 2为所述ANN输入一维张量的1维参数,且W' 2＝W' 1；
    在每个所述ANN输入一维张量的末端，增加所述标定舒张压数据和所述标定收缩压数据；所述ANN输入一维张量的形状变为1×W' 3；所述W' 3为所述ANN输入一维张量的1维参数，且W' 3＝W' 2+2＝W' 1+2＝H 2*W 7*C 2+2；所述ANN输入二维张量的形状变为B 6×W 8；所述W 8为所述ANN输入二维张量新的1维参数，且W 8＝W' 3＝H 2*W 7*C 2+2。
  6. 根据权利要求5所述的融合标定光体积描计信号数据的血压预测方法，其特征在于，所述利用所述ANN模型，对所述ANN输入二维张量进行回归计算，生成ANN输出二维张量，具体包括：
    将所述ANN输入二维张量做为第一输入二维张量，再将所述第一输入二维张量送入所述ANN模型的第一层全连接层，进行第一层全连接计算，生成第一输出二维张量；接着将所述第一输出二维张量做为第二输入二维张量，再将所述第二输入二维张量送入所述ANN模型的第二层全连接层，进行第二层全连接计算，生成第二输出二维张量；直至最后，将倒数第二个输出二维 张量做为最后一个输入二维张量，再将所述最后一个输入二维张量送入所述ANN模型的最后一层全连接层，进行最后一层全连接计算，生成所述ANN输出二维张量；
    其中，所述ANN模型包括多层所述全连接层；所述ANN输出二维张量的形状为B 7×W 9；所述B 7为所述ANN输出二维张量的2维参数，且B 7＝B 6；所述W 9为所述ANN输出二维张量的1维参数，且W 9＝2；所述ANN输出二维张量包括所述B 7个ANN输出一维张量；所述ANN输出一维张量包括舒张压相对数据和收缩压相对数据。
  7. 根据权利要求6所述的融合标定光体积描计信号数据的血压预测方法，其特征在于，所述根据所述标定舒张压数据、所述标定收缩压数据和所述ANN输出二维张量，进行血压数据计算，生成血压二维张量，具体包括：
    当预设的相对关系信息为差值关系时，使用所述标定舒张压数据，对所述ANN输出二维张量中的所有所述舒张压相对数据，进行舒张压增值处理；并使用所述标定收缩压数据，对所述ANN输出二维张量中的所有所述收缩压相对数据，进行收缩压增值处理；再将完成增值处理的所述ANN输出二维张量做为所述血压二维张量；
    其中，所述血压二维张量的形状为B 8×W 10；所述B 8为所述血压二维张量的2维参数，且B 8＝B 7；所述W 10为所述血压二维张量的1维参数，且W 10＝2；所述血压二维张量包括所述B 8个血压一维张量；所述血压一维张量包括舒张压数据和收缩压数据；所述舒张压数据为对应的所述舒张压相对数据与所述标定舒张压数据相加的和；所述收缩压数据为对应的所述收缩压相对数据与所述标定收缩压数据相加的和。
  8. 根据权利要求7所述的融合标定光体积描计信号数据的血压预测方法，其特征在于，所述当预设的预测类型信息为第一类型时，根据所述血压二维张量，进行均值血压数据计算，生成舒张压预测数据和收缩压预测数据；当所述预测类型信息为第二类型时，对所述血压二维张量，进行血压数据提取 处理，生成舒张压预测数据序列和收缩压预测数据序列，具体包括：
    当所述预测类型信息为所述第一类型时，计算所述血压二维张量中的所有所述舒张压数据的平均值，生成所述舒张压预测数据；计算所述血压二维张量中的所有所述收缩压数据的平均值，生成所述收缩压预测数据；
    当所述预测类型信息为所述第二类型时，依次提取所述血压二维张量中的所述舒张压数据，组成所述舒张压预测数据序列；提取所述血压二维张量中的所述收缩压数据，组成所述收缩压预测数据序列。
  9. 一种融合标定光体积描计信号数据的血压预测装置，其特征在于，所述装置包括：
    获取模块用于获取标定光体积描计PPG信号数据和对应的标定舒张压数据与标定收缩压数据；获取实时PPG信号数据；其中，所述实时PPG信号数据的时间长度与所述标定PPG信号数据的时间长度相同；
    人工智能计算模块用于根据所述标定PPG信号数据和所述实时PPG信号数据，进行卷积神经网络CNN模型的输入数据准备处理，生成CNN输入四维张量；接着利用所述CNN模型，对所述CNN输入四维张量进行多层卷积池化计算，生成CNN输出四维张量；再根据所述CNN输出四维张量，进行人工神经网络ANN模型的输入数据准备处理，生成ANN输入二维张量；接着利用所述ANN模型，对所述ANN输入二维张量进行回归计算，生成ANN输出二维张量；
    血压预测模块用于根据所述标定舒张压数据、所述标定收缩压数据和所述ANN输出二维张量，进行血压数据计算，生成血压二维张量；当预设的预测类型信息为第一类型时，根据所述血压二维张量，进行均值血压数据计算，生成舒张压预测数据和收缩压预测数据；当所述预测类型信息为第二类型时，对所述血压二维张量，进行血压数据提取处理，生成舒张压预测数据序列和收缩压预测数据序列。
  10. 一种电子设备，其特征在于，包括：存储器、处理器和收发器；
    所述处理器用于与所述存储器耦合，读取并执行所述存储器中的指令，以实现权利要求1-8任一项所述的方法步骤；
    所述收发器与所述处理器耦合，由所述处理器控制所述收发器进行消息收发。
  11. 一种计算机程序产品，其特征在于，所述计算机程序产品包括计算机程序代码，当所述计算机程序代码被计算机执行时，使得所述计算机执行权利要求1-8任一项所述的方法。
  12. 一种计算机可读存储介质，其特征在于，所述计算机存储介质存储有计算机指令，当所述计算机指令被计算机执行时，使得所述计算机执行权利要求1-8任一项所述的方法的指令。
PCT/CN2021/088023 2020-10-12 2021-04-19 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 WO2022077889A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
US18/248,755 US20230371831A1 (en) 2020-10-12 2021-04-19 Method and apparatus for predicting blood pressure by fusing calibrated photoplethysmographic signal data
EP21878943.6A EP4226856A1 (en) 2020-10-12 2021-04-19 Method and apparatus for predicting blood pressure by fusing calibrated photoplethysmographic signal data

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011085575.XA CN112336325B (zh) 2020-10-12 2020-10-12 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
CN202011085575.X 2020-10-12

Publications (1)

Publication Number Publication Date
WO2022077889A1 true WO2022077889A1 (zh) 2022-04-21

Family

ID=74360449

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/088023 WO2022077889A1 (zh) 2020-10-12 2021-04-19 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

Country Status (4)

Country Link
US (1) US20230371831A1 (zh)
EP (1) EP4226856A1 (zh)
CN (1) CN112336325B (zh)
WO (1) WO2022077889A1 (zh)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
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
US20230293117A1 (en) * 2022-03-17 2023-09-21 Microlife Corporation Method for estimating blood pressures using photoplethysmography signal analysis and system using the same
CN116919363A (zh) * 2022-04-12 2023-10-24 乐普（北京）医疗器械股份有限公司 一种基于大数据特征的个性化血压预测方法和装置

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP2544125A1 (en) * 2011-07-04 2013-01-09 Sabirmedical, S.L. Methods and systems for non-invasive measurement of blood pressure
CN104665768A (zh) * 2013-10-03 2015-06-03 塔塔咨询服务有限公司 生理参数的监测
CN109452935A (zh) * 2017-09-06 2019-03-12 塔塔咨询服务有限公司 使用统计后处理从血管容积图估计血压的无创方法和系统
US20200093386A1 (en) * 2018-09-21 2020-03-26 Imec Vzw Method of generating a model for heart rate estimation from a photoplethysmography signal and a method and a device for heart rate estimation
CN111063420A (zh) * 2019-12-18 2020-04-24 清华大学 基于神经网络识别心理压力的方法及装置
CN111248883A (zh) * 2020-02-21 2020-06-09 乐普（北京）医疗器械股份有限公司 一种血压预测的方法和装置
CN111291727A (zh) * 2020-03-17 2020-06-16 乐普（北京）医疗器械股份有限公司 一种光体积变化描记图法信号质量检测方法和装置
CN111358453A (zh) * 2020-03-17 2020-07-03 乐普（北京）医疗器械股份有限公司 一种血压分类预测方法和装置
CN111358451A (zh) * 2020-03-17 2020-07-03 乐普（北京）医疗器械股份有限公司 一种血压预测方法和装置
CN112336325A (zh) * 2020-10-12 2021-02-09 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

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103598876B (zh) * 2013-11-22 2016-08-17 哈尔滨工业大学深圳研究生院 数据处理方法及系统
CN105943005B (zh) * 2016-06-01 2019-08-06 合肥芯福传感器技术有限公司 基于光电绿光脉搏与心电图混合的无创血压检测装置
KR102042700B1 (ko) * 2017-11-17 2019-11-08 가천대학교 산학협력단 딥러닝 기반의 혈압 예측 시스템 및 방법
US11229404B2 (en) * 2017-11-28 2022-01-25 Stmicroelectronics S.R.L. Processing of electrophysiological signals
CN109363652B (zh) * 2018-09-29 2021-05-07 天津惊帆科技有限公司 基于深度学习的ppg信号重构方法及设备
CN111358455B (zh) * 2020-03-17 2022-07-29 乐普（北京）医疗器械股份有限公司 一种多数据源的血压预测方法和装置
CN111358452B (zh) * 2020-03-17 2022-07-29 乐普（北京）医疗器械股份有限公司 一种基于同步信号进行血压预测的方法和装置
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

Patent Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP2544125A1 (en) * 2011-07-04 2013-01-09 Sabirmedical, S.L. Methods and systems for non-invasive measurement of blood pressure
CN104665768A (zh) * 2013-10-03 2015-06-03 塔塔咨询服务有限公司 生理参数的监测
CN109452935A (zh) * 2017-09-06 2019-03-12 塔塔咨询服务有限公司 使用统计后处理从血管容积图估计血压的无创方法和系统
US20200093386A1 (en) * 2018-09-21 2020-03-26 Imec Vzw Method of generating a model for heart rate estimation from a photoplethysmography signal and a method and a device for heart rate estimation
CN111063420A (zh) * 2019-12-18 2020-04-24 清华大学 基于神经网络识别心理压力的方法及装置
CN111248883A (zh) * 2020-02-21 2020-06-09 乐普（北京）医疗器械股份有限公司 一种血压预测的方法和装置
CN111291727A (zh) * 2020-03-17 2020-06-16 乐普（北京）医疗器械股份有限公司 一种光体积变化描记图法信号质量检测方法和装置
CN111358453A (zh) * 2020-03-17 2020-07-03 乐普（北京）医疗器械股份有限公司 一种血压分类预测方法和装置
CN111358451A (zh) * 2020-03-17 2020-07-03 乐普（北京）医疗器械股份有限公司 一种血压预测方法和装置
CN112336325A (zh) * 2020-10-12 2021-02-09 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

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
HAIXIA FAN, XIAOHUI CHEN: "A Continuous Non-invasive Blood Pressure Prediction Method Based on Improved SVR Learning", JOURNAL OF SYSTEM SIMULATION, GAI-KAN BIANJIBU , BEIJING, CN, vol. 32, no. 9, 30 September 2020 (2020-09-30), CN , XP055923421, ISSN: 1004-731X *
SHIMAZAKI SHOTA; KAWANAKA HARUKI; ISHIKAWA HIROKI; INOUE KOICHI; OGURI KOJI: "Cuffless Blood Pressure Estimation from only the Waveform of Photoplethysmography using CNN", 2019 41ST ANNUAL INTERNATIONAL CONFERENCE OF THE IEEE ENGINEERING IN MEDICINE AND BIOLOGY SOCIETY (EMBC), 23 July 2019 (2019-07-23), pages 5042 - 5045, XP033624618, DOI: 10.1109/EMBC.2019.8856706 *

Also Published As

Publication number Publication date
CN112336325A (zh) 2021-02-09
US20230371831A1 (en) 2023-11-23
EP4226856A1 (en) 2023-08-16
CN112336325B (zh) 2024-02-23

Similar Documents

Publication Publication Date Title
WO2022077889A1 (zh) 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
CN111292853B (zh) 基于多参数的心血管疾病风险预测网络模型及其构建方法
CN107977709B (zh) 预测血管树血管路径上的血流特征的深度学习模型和系统
JP7304901B2 (ja) 不整脈検出方法、装置、電子装置およびコンピュータ記憶媒体
CN109620205B (zh) 心电数据分类方法、装置、计算机设备和存储介质
TWI742408B (zh) 一種圖像處理方法、電子設備及儲存介質
WO2022077888A1 (zh) 一种光体积描计信号的散点图分类方法和装置
CN109243620B (zh) 基于血药浓度监测的药效优化方法及装置
EP4108164A1 (en) Blood pressure prediction method and device
CN112270212B (zh) 基于多导联心电信号生成心搏标签数据序列的方法和装置
EP3769312A1 (en) Systems and methods for personalized medication therapy management
CN114863225B (zh) 图像处理模型训练方法、生成方法、装置、设备及介质
CN111358451B (zh) 一种血压预测方法和装置
CN114611879A (zh) 一种基于多任务学习的临床风险预测系统
Vijayaganth et al. Smart sensor based prognostication of cardiac disease prediction using machine learning techniques
Raihan et al. Chronic Renal Disease Prediction using Clinical Data and Different Machine Learning Techniques
Alavi et al. Sequentially-reduced representation of artificial neural network to determine cardiovascular intrinsic frequencies
WO2023197425A1 (zh) 一种基于大数据特征的个性化血压预测方法和装置
Akdemïr et al. Prediction of cardiac end-systolic and end-diastolic diameters in m-mode values using adaptive neural fuzzy inference system
Liu et al. PSC-Net: Integration of Convolutional Neural Networks and Transformers for Physiological Signal Classification
CN116189906B (zh) 脓毒症实时预警的神经网络模型、方法和电子设备
US11551082B2 (en) Parallel sequence reductions with recursive neural networks
CN114822849B (zh) 基于数字孪生的数据监测方法、装置、设备和存储介质
US20230148969A1 (en) Sequentially-reduced artificial intelligence methodology for instantaneous determination of waveform intrinsic frequencies
Alavi et al. A machine learning approach for computation of cardiovascular intrinsic frequencies

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: 21878943

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2021878943

Country of ref document: EP

Effective date: 20230512