Blood pressure prediction apparatus and device

Abstract

The application discloses a blood pressure prediction device and equipment, comprising: an information acquisition module: for acquiring one-dimensional physiological signals and personalized information; a signal processing module: for cutting one-dimensional physiological signals into segments and preprocessing; a feature extraction module: for inputting one-dimensional physiological signal segments and personalized information into a blood pressure value prediction model, the blood pressure value prediction model comprising a feature extraction model and a feature fusion model, and being used for feature extraction of one-dimensional physiological signal segments and personalized information and output of high-dimensional features; a blood pressure prediction module: for feature fusion of two high-dimensional features through the feature fusion model, and output of predicted blood pressure values. The blood pressure prediction device provided by the application can reduce prediction errors, remove noise in collected signals, screen out signals capable of blood pressure value prediction, improve the accuracy of blood pressure prediction, and quickly predict the blood pressure values of prediction objects.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of blood pressure measurement, in particular to a blood pressure prediction device and equipment. BACKGROUND

[0002] Blood pressure is an important physiological parameter for cardiovascular disease diagnosis, and is respectively referred to as arterial blood pressure, capillary pressure and venous blood pressure in different blood vessels. The blood pressure commonly referred to by people refers to arterial blood pressure, which is the starting point of systemic circulation, the pressure of blood ejected from the left ventricle into the aortic vessel wall, referred to as aortic pressure. The maximum value of blood pressure is referred to as systolic blood pressure (SBP), and the normal range is 100 to 120 mmHg; the minimum value of blood pressure is referred to as diastolic blood pressure (DBP), and the normal range is 60 to 80 mmHg. Central venous blood pressure refers to the blood pressure of blood flowing back into the right atrium and thoracic great vein, which is the terminal point of systemic circulation. And the arteries and veins are connected together through the microcirculation in the body. After the systemic circulation blood is ejected from the left ventricle, it finally reaches the terminal microartery of the artery through various functional types of arterial vessels. In general, central venous pressure can represent the function of the right heart, while arterial blood pressure is the superposition of left heart function and peripheral vascular resistance. The research object of the present subject is arterial blood pressure.

[0003] The World Health Organization's 2020 statistical report shows that 17.9 million people died of cardiovascular and cerebrovascular diseases worldwide, accounting for 43.7% of all non-communicable disease deaths. Cardiovascular disease, which ranks first in mortality, is hypertension. There are many causes of hypertension, such as family inheritance and long-term high mental stress, excessive obesity and diabetes, which can all cause hypertension. Hypertension patients often feel sudden numbness of the limbs to the fingers and toes, and limited activity. The real harm of hypertension is the damage to important organs such as the heart, brain and kidneys, which can cause cerebral vascular rupture, myocardial infarction and renal failure. For example, stroke is a malignant disease with high mortality. Even if the patient survives, most of them will face the risk of paralysis. The higher the blood pressure, the higher the incidence of stroke. In addition, hypertensive patients often have a certain degree of pathological arteriosclerosis, such as cerebral arteriosclerosis. When cerebral arteriosclerosis reaches a certain degree, if there is a short or excessive excitement in life, such as anger, sudden accidents, intense exercise, etc., the blood pressure will suddenly rise, leading to stroke. Therefore, hypertension seriously threatens human life and health and quality of life.

[0004] The above background knowledge is only used to help understand the inventive scheme and technical content of the present application, and it does not belong to the prior art of the present patent application. Without clear evidence that the above content has been disclosed before the filing date of the present patent application, the above background should not be used to evaluate the creativity and novelty of the present application. SUMMARY

[0005] The present application aims to solve the technical problems of poor real-time performance and low accuracy of blood pressure prediction devices, and provides a blood pressure prediction device and equipment.

[0006] The technical problems of the present application are solved by the following technical solutions:

[0007] A blood pressure prediction device comprises:

[0008] An information acquisition module is configured to acquire one-dimensional physiological signals and personalized information of a prediction object, wherein the one-dimensional physiological signals include electrocardiogram signals and pulse wave signals;

[0009] A signal processing module is configured to cut the one-dimensional physiological signals into one-dimensional physiological signal segments and pre-process the one-dimensional physiological signal segments;

[0010] A feature extraction module is configured to input the one-dimensional physiological signal segments and the personalized information into a constructed blood pressure value prediction model, wherein the blood pressure value prediction model comprises a feature extraction model and a feature fusion model, and is configured to extract features from the pre-processed one-dimensional physiological signal segments and output one-dimensional physiological signal high-dimensional features, and extract features from the personalized information and output personalized high-dimensional features;

[0011] A blood pressure prediction module is configured to fuse the one-dimensional physiological signal high-dimensional features and the personalized high-dimensional features through the feature fusion model, and output predicted blood pressure values.

[0012] In some embodiments, the personalized information includes at least one of gender, age, height, and weight.

[0013] In some embodiments, the pre-processing specifically includes:

[0014] Data denoising, at least one of low-pass filtering, wavelet threshold denoising, and band-pass filter denoising is used;

[0015] Quality assessment, to assess whether the denoised one-dimensional physiological signal segments can be used for blood pressure prediction, and the assessment indicators include signal energy, peak-to-peak interval, peak factor, and kurtosis factor in the one-dimensional physiological signal segments;

[0016] Peak-valley detection, using first-order difference to perform signal transformation on the pulse wave signal, finding wave troughs, calibrating and positioning, so as to detect the wave peaks and wave troughs in the signal.

[0017] In some embodiments, wavelet threshold denoising is used for the pulse wave signal, and a band-pass filter is used for the electrocardiogram signal.

[0018] In some embodiments, the peak-valley detection specifically comprises:

[0019] Obtaining the pulse wave signal;

[0020] Calculating the first-order difference of the pulse wave signal, performing signal transformation, and the method of the signal transformation is:

[0021]

[0022] where y k is the input one-dimensional physiological signal, Ay k is the result of the first-order difference, w is the weighted time window, and the transformed signal is a slope and signal;

[0023] Setting an initial threshold, traversing the slope and signal from the beginning, finding the first signal point with a signal value greater than the threshold and setting a judgment window, finding the wave trough of the slope and signal waveform, and updating the threshold;

[0024] Calibrating the detected wave trough;

[0025] Positioning the index of the wave peak according to the found wave trough index to obtain the wave peak of the slope and signal, and finally outputting the wave peak and wave trough of the signal.

[0026] In some embodiments, the feature extraction model extracts features from the preprocessed one-dimensional physiological signal segment by building a convolutional neural network, and extracts features from the personalized information by building a machine learning model.

[0027] In some embodiments, the feature extraction from the preprocessed one-dimensional physiological signal segment by building a convolutional neural network specifically comprises: dividing the processed one-dimensional physiological signal into a training set, a validation set, and a test set, performing neural network training, and obtaining the one-dimensional physiological signal high-dimensional features corresponding to the preliminary one-dimensional physiological signal through training of a regressor.

[0028] The feature extraction from the personalized information by building a machine learning model specifically comprises: dividing the processed personalized information into a training set, a validation set, and a test set, performing machine learning training, and obtaining the personalized high-dimensional features corresponding to the preliminary personalized information through training of a regressor.

[0029] In some embodiments, the training set is used to train the blood pressure value prediction model to obtain the blood pressure value prediction model; the validation set is used to verify the trained blood pressure value prediction model, adjust the model parameters until the average error and standard deviation of the model meet the preset conditions; and the test set is used to test the blood pressure value prediction model that meets the preset conditions.

[0030] In some embodiments, the signal processing module for one-dimensional physiological signal processing further comprises:

[0031] cutting the one-dimensional physiological signal for 10 seconds before the cutting, obtaining a one-dimensional physiological signal segment, and performing quality assessment;

[0032] determining whether the signal quality is qualified according to the quality assessment result, if not, continuing to wait for 10 seconds to cut the next segment;

[0033] if the quality assessment meets the requirements, the one-dimensional physiological signal segment is used for input into the blood pressure value prediction model;

[0034] performing the next cutting, cutting another 10-second signal from the last cutting position, then performing quality assessment again, and determining whether the signal quality is qualified according to the quality assessment result, until the whole one-dimensional physiological signal is completely detected; inputting the 10-second one-dimensional physiological signal, and outputting the quality assessment result and the corresponding blood pressure value.

[0035] The present application also provides a blood pressure prediction device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements the following steps when executing the computer program:

[0036] S1: obtaining one-dimensional physiological signal and personalized information of a prediction object; the one-dimensional physiological signal comprises electrocardiogram signal and pulse wave signal;

[0037] S2: cutting the one-dimensional physiological signal into one-dimensional physiological signal segments, and pre-processing the one-dimensional physiological signal segments;

[0038] S3: constructing a blood pressure value prediction model, the blood pressure value prediction model comprising a feature extraction model and a feature fusion model; inputting the one-dimensional physiological signal segments and personalized information into the blood pressure value prediction model, the feature extraction model performing feature extraction on the pre-processed one-dimensional physiological signal segments and outputting one-dimensional physiological signal high-dimensional features, and simultaneously performing feature extraction on the personalized information and outputting personalized high-dimensional features;

[0039] S4: the feature fusion model performing feature fusion on the one-dimensional physiological signal high-dimensional features and the personalized high-dimensional features, and outputting predicted blood pressure value.

[0040] The beneficial effects of the present application compared with the prior art include:

[0041] The blood pressure prediction device provided by the application can reduce prediction error, and the signal processing module is used for pre-processing one-dimensional physiological signals, removing noise in the one-dimensional physiological signals, and screening signals capable of blood pressure value prediction, so as to improve the accuracy of blood pressure prediction; meanwhile, the blood pressure value prediction model is fully utilized to extract one-dimensional physiological signal high-dimensional features and personalized high-dimensional features through the feature extraction module, and the blood pressure prediction module is used for fusing the two, so that the blood pressure value of the prediction object can be quickly predicted, thereby improving the real-time performance of blood pressure prediction.

[0042] In addition, in some embodiments, the following beneficial effects are also achieved:

[0043] The signal processing module of the application detects the wave peak and valley of the signal based on the slope and function, the slope and function can enhance the rising slope of the waveform and suppress the falling slope, so that the signal wave peak and valley can be detected in near real time, and the prediction real-time performance and accuracy of the blood pressure prediction device are further improved. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 Fig. 1 is a structural schematic diagram of the blood pressure prediction device in the embodiment of the application;

[0045] Figure 2 Fig. 3 is a flowchart of the signal processing of the one-dimensional physiological signal by the signal processing module in the embodiment of the application;

[0046] Figure 3 Fig. 4 is a flowchart of the pre-processing of the one-dimensional physiological signal by the signal processing module in the embodiment of the application;

[0047] Figure 4 Fig. 5 is a flowchart of the peak and valley detection in the embodiment of the application;

[0048] Figure 5 Fig. 6 is a flowchart of the use method of the blood pressure prediction device in the embodiment of the application;

[0049] Figure 6 Fig. 7 is a model overall flowchart of the blood pressure prediction device in the embodiment of the application;

[0050] Figure 7 Fig. 8 is a blood pressure prediction flowchart of the blood pressure prediction device in the embodiment of the application;

[0051] Figure 8a Fig. 9 is a PPG signal schematic diagram in the embodiment of the application;

[0052] Figure 8b Fig. 10 is an SSF signal schematic diagram in the embodiment of the application;

[0053] Figure 9is a flow chart of extracting high-dimensional features of one-dimensional physiological signals in the embodiment of the present application;

[0054] Figure 10 is a flow chart of extracting high-dimensional features of personalized information in the embodiment of the present application;

[0055] Figure 11 is a flow chart of training a blood pressure value prediction model in the embodiment of the present application;

[0056] Figure 12 is a flow chart of quality assessment in the embodiment of the present application;

[0057] Figure 13a is a comparison diagram of predicted systolic pressure 1 and true systolic pressure 1 in the embodiment of the present application;

[0058] Figure 13b is a comparison diagram of predicted diastolic pressure 1 and true diastolic pressure 1 in the embodiment of the present application;

[0059] Figure 13c is a comparison diagram of predicted systolic pressure 2 and true systolic pressure 2 in the embodiment of the present application;

[0060] Figure 13d is a comparison diagram of predicted diastolic pressure 2 and true diastolic pressure 2 in the embodiment of the present application;

[0061] Figure 14a is an ECG signal diagram in the embodiment of the present application;

[0062] Figure 14b is a PPG signal diagram in the embodiment of the present application;

[0063] Figure 15 is a high-dimensional feature fusion diagram of one-dimensional physiological signals and personalized information in the embodiment of the present application. DETAILED DESCRIPTION

[0064] The present application will be further described below with reference to the drawings and in conjunction with preferred embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0065] It should be noted that the left, right, up, down, top, bottom and other orientation terms in the present embodiment are only relative concepts or are referenced to the normal use state of the product, and should not be considered as limiting.

[0066] Deep learning is a subfield of artificial intelligence (AI) that is rapidly developing in the medical field. Deep learning systems can accept multiple data types, such as images or time series, which will be processed through multiple processing layers based on neural networks to learn the representation of data step by step and ultimately provide output. In the healthcare field, deep learning technology can have potential uses for diagnosis, treatment, population health management, and management and regulation.

[0067] The embodiment of the present application provides a blood pressure prediction device for predicting blood pressure of a human body through electrocardio and pulse wave signals based on one-dimensional physiological signals, which combines deep learning, and the main content comprises: acquiring electrocardio and pulse wave signals throughout the night, and adding personalized information (including but not limited to gender, age, height, and weight) of a patient, wherein the one-dimensional physiological signals are preprocessed.

[0068] The specific preprocessing steps include quality assessment, data denoising, peak-valley detection, and finally cutting into 10-second signal segments in a time alignment manner; a multi-modal learning model capable of data regression is constructed, the model is composed of two branches, a convolutional neural network branch based on a large convolution kernel is used to extract one-dimensional physiological signal information, and a multi-layer perceptron is used to process personalized information, and finally feature fusion is performed in a decision layer to output corresponding blood pressure values.

[0069] The embodiment of the present application can quickly and accurately predict the current systolic blood pressure (SBP) and diastolic blood pressure (DBP) of a human body under the condition of providing one-dimensional physiological signals and personalized information of the human body, and compared with a model using only single-channel one-dimensional physiological signals for prediction, the blood pressure value prediction model has higher accuracy.

[0070] The following is a summary of the embodiment of the present application:

[0071] As shown in Figure 1 The embodiment of the present application provides a blood pressure prediction device, which comprises: an information acquisition module, a signal processing module, a feature extraction module, and a blood pressure prediction module.

[0072] The information acquisition module in the blood pressure prediction device is used to acquire one-dimensional physiological signals and personalized information of a prediction object; the one-dimensional physiological signals comprise electrocardio signals and pulse wave signals.

[0073] The personalized information comprises at least one of gender, age, height, and weight.

[0074] The signal processing module in the blood pressure prediction device is used to cut the one-dimensional physiological signals into one-dimensional physiological signal segments, and pre-process the one-dimensional physiological signal segments.

[0075] Specifically, the signal processing module processes the one-dimensional physiological signals in the following steps:

[0076] cutting a one-dimensional physiological signal for 10 seconds to obtain a one-dimensional physiological signal segment and performing quality assessment;

[0077] determining whether the signal quality is qualified according to the quality assessment result, if not, continuing to wait for 10 seconds to cut the next segment;

[0078] if the quality assessment meets the requirements, the one-dimensional physiological signal segment is used for input into a blood pressure value prediction model;

[0079] performing the next cutting, cutting another 10-second signal from the last cutting position, then performing quality assessment again, and determining whether the signal quality is qualified according to the quality assessment result, until the whole one-dimensional physiological signal is completely detected; inputting the 10-second one-dimensional physiological signal, and outputting the quality assessment result and the corresponding blood pressure value.

[0080] Specifically, the signal processing module pre-processes the one-dimensional physiological signal segment as shown in Figure 3 , specifically including:

[0081] data denoising, at least one of low-pass filtering, wavelet threshold denoising and band-pass filter is used for denoising;

[0082] Further, in the data denoising, the wavelet threshold denoising is used for the pulse wave signal, and the band-pass filter is used for the electrocardiogram signal.

[0083] quality assessment, assessing whether the one-dimensional physiological signal segment after denoising can be used for blood pressure prediction, and the assessment indexes include signal energy, peak-to-peak interval, peak factor and kurtosis factor in the one-dimensional physiological signal segment;

[0084] peak and valley detection, using first-order difference to perform signal transformation, find wave trough, calibration and positioning on the pulse wave signal, so as to detect the wave peak and wave trough in the signal.

[0085] Further, the peak and valley detection as shown in Figure 4 , specifically including:

[0086] obtaining the pulse wave signal;

[0087] calculating the first-order difference of the pulse wave signal to perform signal transformation, and the method of signal transformation is:

[0088]

[0089] wherein y k is the input one-dimensional physiological signal, Δy k is the result of first-order difference, w is a weighted time window, and the transformed signal is a slope signal and a signal;

[0090] Set an initial threshold, traverse the slope and signal from the beginning, find the first signal point whose signal value is greater than the threshold and set the judgment window, find the trough of the slope and signal waveform, and update the threshold value at the same time;

[0091] Calibrate the detected troughs;

[0092] According to the found trough index, the peak index is located, the peaks of the slope and signal are obtained, and finally the peaks and troughs of the signal are output.

[0093] The feature extraction module in the blood pressure prediction device is used to input the one-dimensional physiological signal segment and the personalized information into the constructed blood pressure value prediction model, the blood pressure value prediction model includes a feature extraction model and a feature fusion model, and is used for feature extraction on the preprocessed one-dimensional physiological signal segment and output as one-dimensional physiological signal high-dimensional features, and feature extraction on the personalized information and output as personalized high-dimensional features;

[0094] Specifically, the feature extraction model extracts features from the preprocessed one-dimensional physiological signal segment by building a convolutional neural network, and extracts features from the personalized information by building a machine learning model.

[0095] The specific steps of extracting features from the preprocessed one-dimensional physiological signal segment are: dividing the processed one-dimensional physiological signal into a training set, a validation set and a test set, training a neural network, and obtaining preliminary one-dimensional physiological signal high-dimensional features corresponding to the one-dimensional physiological signal through training of a regressor.

[0096] The specific steps of extracting features from the personalized information are: dividing the processed personalized information into a training set, a validation set and a test set, training a machine learning model, and obtaining preliminary personalized high-dimensional features corresponding to the personalized information through training of a regressor.

[0097] Further, the training set is used to train the blood pressure value prediction model to obtain the blood pressure value prediction model; the validation set is used to verify the blood pressure value prediction model obtained by training, adjust the model parameters until the average error and standard deviation of the model meet the preset conditions; and the test set is used to test the blood pressure value prediction model meeting the preset conditions.

[0098] The blood pressure prediction module is used to perform feature fusion on the one-dimensional physiological signal high-dimensional features and the personalized high-dimensional features through the feature fusion model, and output the predicted blood pressure value.

[0099] The embodiment of the present application also provides a blood pressure prediction device, which includes a processor, a memory and a computer program stored in the memory and configured to be executed by the processor, and the blood pressure prediction device uses the method as shown in Figure 5 When the processor executes the computer program, the following steps are implemented.

[0100] S1: obtaining one-dimensional physiological signals and personalized information of a prediction object;

[0101] Specifically, the one-dimensional physiological signals include electrocardiogram signals and pulse wave signals; and the personalized information includes at least one of gender, age, height and weight.

[0102] S2: cutting the one-dimensional physiological signals into one-dimensional physiological signal segments, and preprocessing the one-dimensional physiological signal segments;

[0103] S3: constructing a blood pressure value prediction model, the blood pressure value prediction model including a feature extraction model and a feature fusion model; inputting the one-dimensional physiological signal segments and the personalized information into the blood pressure value prediction model, the feature extraction model performing feature extraction on the preprocessed one-dimensional physiological signal segments and outputting one-dimensional physiological signal high-dimensional features, and simultaneously performing feature extraction on the personalized information and outputting personalized high-dimensional features;

[0104] Specifically, the feature extraction model performs feature extraction on the preprocessed one-dimensional physiological signal segments by building a convolutional neural network, and performs feature extraction on the personalized information by building a machine learning model.

[0105] Further, the feature extraction on the preprocessed one-dimensional physiological signal segments by building a convolutional neural network specifically includes: dividing the processed one-dimensional physiological signals into a training set, a validation set and a test set, performing neural network training, and obtaining one-dimensional physiological signal high-dimensional features corresponding to the one-dimensional physiological signals by training a regressor.

[0106] The feature extraction on the personalized information by building a machine learning model specifically includes: dividing the processed personalized information into a training set, a validation set and a test set, performing machine learning training, and obtaining personalized high-dimensional features corresponding to the personalized information by training a regressor.

[0107] Further, the training set is used to train the one-dimensional regression model to obtain the blood pressure value prediction model; the validation set is used to verify the blood pressure value prediction model obtained by training, to adjust the model parameters until the average error and the standard deviation of the model meet preset conditions; and the test set is used to test the blood pressure value prediction model meeting the preset conditions.

[0108] S4: the feature fusion model performs feature fusion on the one-dimensional physiological signal high-dimensional features and the personalized high-dimensional features, and outputs predicted blood pressure values.

[0109] The blood pressure prediction device will be further described below by specific embodiments with reference to the accompanying drawings:

[0110] The blood pressure value prediction model used in the embodiment of the present application is obtained by deep learning method according to the existing one-dimensional physiological signals (including electrocardiogram signal ECG and pulse wave signal PPG) and personalized information (gender, age), and the blood pressure value obtained after the corresponding patient wears PSG (sleep monitoring) equipment for monitoring all night is used as a label. The one-dimensional physiological signal is input into a one-dimensional convolutional neural network, which is called a “feature extraction network”; the personalized information is input into a multi-layer perceptron, which is also a kind of neural network, and is often attributed to a machine learning model due to its relatively simple structure. The two networks extract the features of the two inputs respectively, then perform fusion, and finally output the predicted blood pressure value.

[0111] As shown in Figure 1 The blood pressure prediction device provided by the embodiment of the present application comprises an information acquisition module, a signal processing module, a feature extraction module and a blood pressure prediction module. The information acquisition module acquires one-dimensional physiological signals (including electrocardiogram signals and pulse wave signals) and personalized information of a prediction object. The information acquisition module can use PSG (sleep monitoring) equipment. The signal processing module cuts the one-dimensional physiological signals acquired by the information acquisition module into one-dimensional physiological signal segments, and pre-processes the one-dimensional physiological signal segments. The feature extraction module extracts features from the one-dimensional physiological signal segments pre-processed by the signal processing module, and outputs one-dimensional physiological signal high-dimensional features. The personalized information acquired by the information acquisition module is extracted for features and output as personalized high-dimensional features. Finally, the blood pressure prediction module performs feature fusion on the one-dimensional physiological signal high-dimensional features and the personalized high-dimensional features extracted by the feature extraction module, and outputs the predicted blood pressure value.

[0112] As shown in Figure 6 The blood pressure value prediction model is divided into two branches. The physiological signal branch (including electrocardiogram and pulse wave signal) uses SEResNet (a kind of deep learning model) to extract features. The personalized branch (including gender, age, etc.) uses a multi-layer perceptron to extract features. The multi-layer perceptron is composed of two fully connected layers, which are used to fit a one-dimensional linear equation y=ax+b.

[0113] Compared with the existing measurement method using cuff pressurization, the embodiment of the present application uses deep learning method combined with one-dimensional physiological signals of different modalities and personalized information of the measured person, which is convenient for the patient to understand the disease in time and treat early, is convenient and convenient, and the measured person does not need to wear various cuff devices, is easy to operate, and the recognition errors of SBP and DBP are-0.15 and-0.13 mmHg respectively.

[0114] The one-dimensional physiological signal (ECG, PPG data) of the subject in the embodiment of the present application is obtained within a period of time, and the acquisition method is to use overnight PSG, and the sampling rate of the sampling device includes but is not limited to 125 Hz. The blood pressure true value can be the value measured based on the cuff method within 5 minutes.

[0115] The step of obtaining the personalized information modeling of the measured person in the embodiment of the present application includes: selecting the screened features, including but not limited to gender, age, height, weight and the like; and pre-processing the selected personalized information.

[0116] As shown in Figure 7 The blood pressure prediction device provided by the embodiment of the present application performs the steps of blood pressure prediction as follows:

[0117] S101: obtaining one-dimensional physiological signals of a measured person within a period of time through an information acquisition module, wherein the one-dimensional physiological signals include electrocardiogram signals and pulse wave signals, and the personalized information (including but not limited to gender, age, height, weight) of the measured person, and taking the one-dimensional physiological signals and the personalized information as input data; and obtaining blood pressure true value data obtained by the measured person receiving PSG device detection at the same time.

[0118] S102: pre-processing the one-dimensional physiological signals through a signal processing module, including quality evaluation, data denoising, peak-valley detection;

[0119] The purpose of the quality evaluation in the embodiment of the present application is to evaluate whether a one-dimensional physiological signal within a period of time can be used for blood pressure prediction, and the main indicators are signal energy, peak-to-peak interval, peak factor and kurtosis factor within the one-dimensional physiological signal segment;

[0120] The data denoising in the embodiment of the present application respectively adopts wavelet threshold denoising and band-pass filter denoising. For the PPG signal, wavelet threshold denoising is used, the Morlet wavelet is used, the decomposition is 7 levels, the threshold selection rule is sureshrink (a threshold rule of wavelet threshold denoising), all low-frequency components are set to zero, and high-frequency components are set to zero using the threshold. The ECG signal is filtered using a band-pass filter with a passband of 1 Hz to 35 Hz.

[0121] In specific implementation, the signal filtering method of the one-dimensional physiological signal (including the ECG signal and the PPG signal) pre-processing includes any one of the following: low-pass filtering, band-pass filtering, wavelet threshold filtering and a method based on a neural network.

[0122] The principle of the wavelet threshold denoising is as follows:

[0123] There are useful signals and noises in the original signal, and their frequencies are different. Therefore, the original signal is decomposed by using wavelet transform to obtain a set of wavelet coefficients. Then a certain threshold is defined, and the coefficients greater than the threshold are regarded as noises and directly set to zero. Finally, the original signal is reconstructed by using wavelet to complete the filtering operation.

[0124] The principle of band-pass filtering is as follows:

[0125] A frequency range is set, and signals exceeding the range become 0, that is, are filtered out.

[0126] The peak-valley detection in the embodiment of the application is to transform, search, calibrate and locate the PPG signal by using first-order difference, so as to detect the wave peaks and wave troughs in the signal.

[0127] Specifically, the peak-valley detection is to detect the wave peaks and wave troughs of the signal by using a method based on a slope sum function (SSF). The purpose of the SSF function is to enhance the rising slope of the waveform and suppress the falling slope. At time i, the window-weighted z i is defined as follows:

[0128]

[0129] y k is the original signal, k represents the kth sampling point, and Δy k =y k +1-y k represents the difference between adjacent signals; Δu k is the result of Δy k , as shown in the formula, Δy k is kept unchanged when it is greater than 0, and the result is set to zero when it is less than 0; and w is the window size in the SSF function.

[0130] The original PPG signal is transformed into an SSF signal by using the above formula in the embodiment of the application, wherein the PPG signal is as shown in Figure 8a , the SSF signal is as shown in Figure 8b ; then an initial threshold is set, the SSF signal is traversed and the threshold is continuously updated; then the detected PPG peaks and valleys are calibrated; and finally the peak-valley values in a certain signal segment are obtained.

[0131] The first step of the SSF function is to transform the PPG signal into an SSF signal, so that the signal becomes simpler. Therefore, the wave peak and valley detection is easier. The specific PPG signal is as shown in Figure 8a , and the transformed SSF signal is as shown in Figure 8b . Figure 8a the lower end point in Figure 8bThe lower end point in the signal is corresponding to the lower end point of the blood pressure signal.

[0132] In a specific implementation, before the preprocessing operation on the one-dimensional physiological signal, a "cut and detect" method is used to cut the one-dimensional physiological signal into one-dimensional physiological signal segments, and then a blood pressure value prediction model is used to predict the blood pressure of the whole one-dimensional physiological signal of the subject.

[0133] More specifically, as shown in Figure 2 The one-dimensional physiological signal processing method of the signal processing module includes the following steps:

[0134] A1: Obtain a one-dimensional physiological signal in a period of time, cut the first 10 seconds of the one-dimensional physiological signal, obtain a one-dimensional physiological signal segment, and perform quality assessment.

[0135] A2: Determine whether the signal quality is qualified according to the quality assessment result, if the signal after quality assessment does not meet the detection requirement, continue to wait for 10 seconds to cut the next segment, that is, input the next one-dimensional physiological signal segment.

[0136] A3: If the quality assessment meets the requirement, input the cut one-dimensional physiological signal segment and the obtained personalized information into the blood pressure prediction model.

[0137] A4: Perform the next cutting, cut another 10-second signal from the cutting position, and then repeat A2 until the whole one-dimensional physiological signal is detected completely.

[0138] The one-dimensional physiological signal processing method in the embodiment of the application can also be applied to other one-dimensional physiological signal regression prediction tasks, including but not limited to heart rate regression of electrocardiogram signals, QRS wave (QRS wave is a waveform of an electrocardiogram) prediction, etc.

[0139] In a specific implementation, the input one-dimensional physiological signal needs to be cut into a 10-second one-dimensional physiological signal segment, the input is a pulse wave, an electrocardiogram signal and personalized information, and the label is a blood pressure true value.

[0140] The preprocessing of the one-dimensional physiological signal in the embodiment of the application mainly has two functions, one is to ensure that the signal quality is qualified, and the other is to align the physiological characteristics of the signal.

[0141] Firstly, only the signal with qualified quality can carry useful information, otherwise it is impossible to perform blood pressure prediction.

[0142] Secondly, the feature alignment considers the physiological significance of blood pressure prediction.

[0143] Blood pressure is generated by the pressure wave of the heart pump, and the pressure wave is called "pulse wave signal" PPG when it propagates to the position of the pulse. The electrocardiogram ECG is detected by a special device when the voltage changes when the heart pumps. Therefore, blood pressure is strongly related to PPG and ECG. According to previous studies, blood pressure is more related to PPG, so ECG can be used as an auxiliary signal. Therefore, in the preprocessing of the embodiment of the application, only the peaks and troughs of PPG are detected, and the signal slicing is performed based on this, so that the deep learning model can better learn the corresponding features.

[0144] S103: input the data obtained by preprocessing into the blood pressure value prediction model through the feature extraction module, the blood pressure value prediction model comprising a feature extraction model and a feature fusion model, the first branch of the feature extraction model using a convolutional neural network model to extract features of one-dimensional physiological signals of the subject within a certain period of time, and the second branch of the feature extraction model using a machine learning model to obtain high-dimensional features of personalized information.

[0145] Figure 9 The blood pressure prediction device provided by the embodiment of the application extracts one-dimensional physiological signal features through a feature extraction model, and the flowchart comprises the following steps:

[0146] B1: obtaining one-dimensional physiological signal data;

[0147] B2: building a convolutional neural network to extract features;

[0148] B3: outputting high-dimensional features of one-dimensional physiological signals.

[0149] The feature extraction model of the first branch uses the convolutional layer, the pooling layer, the activation layer and the batch normalization layer in the neural network to extract features of one-dimensional physiological signals (i.e. PPG and ECG signals), and outputs feature vectors at corresponding positions.

[0150] Figure 10 The blood pressure prediction device provided by the embodiment of the application extracts personalized information features through a feature extraction model, and the flowchart comprises the following steps:

[0151] C1: obtaining personalized information data;

[0152] C2: building a machine learning model to extract features;

[0153] C3: outputting high-dimensional features of personalized information;

[0154] The feature extraction model of the second branch is characterized by using a full connection layer and an activation layer in a neural network to extract personalized information of the subject, wherein the personalized information is personalized discrete information.

[0155] In a specific implementation, the convolutional neural network includes but is not limited to a CNN, a ResNet, a VGG, and an AlexNet.

[0156] The ECG (electrocardiogram) and PPG (pulse wave) signals are divided into training data, validation data, and test data as dual-channel data information.

[0157] Specifically, the signals are cut into 10-second segments, and the preprocessed one-dimensional physiological signals are put into a trained convolutional neural network to obtain high-dimensional features. The personalized information is extracted by a multi-layer perceptron, which is a machine learning model.

[0158] In a specific implementation, the personalized information mainly includes but is not limited to gender, age, height, and weight information. The machine learning model includes but is not limited to a multi-layer perceptron, a support vector machine (SVM), and a random forest.

[0159] The training of the blood pressure value prediction model of the embodiment is as shown in Figure 11 and includes the following steps:

[0160] S201, obtaining patient information in a certain period of time, including one-dimensional physiological signals and personalized information;

[0161] S202, dividing a training set, a validation set, and a test set;

[0162] S203, training a blood pressure value prediction model;

[0163] S204, calibrating the trained blood pressure value prediction model using validation data until the error meets the preset requirements.

[0164] Specifically, the processed one-dimensional physiological signals are divided into a training set, a validation set, and a test set, and a neural network is trained to obtain one-dimensional physiological signal high-dimensional features corresponding to the one-dimensional physiological signals by training a regressor.

[0165] Meanwhile, the processed personalized information samples are divided into a training set, a validation set, and a test set, and a machine learning training is performed to obtain personalized high-dimensional features corresponding to the personalized information by training a regressor.

[0166] Then, the one-dimensional regression model is trained using the training set to obtain a blood pressure value prediction model.

[0167] Then, the blood pressure value prediction model trained is verified by using the above verification set, and the model parameters are adjusted until the average error and standard deviation of the model meet the preset conditions.

[0168] Finally, the blood pressure value prediction model meeting the preset conditions is tested by using the above test set.

[0169] In specific embodiments, the selection of the cutting position is set according to different tasks, because the sampling frequency of the one-dimensional physiological signal in the embodiment is 125 Hz, and 10 seconds is 1250 sampling points, so the length of time is a relatively reasonable length for quality evaluation and subsequent peak-valley detection. Considering that the quality evaluation and peak-valley detection method can be almost real-time, the length of the signal segment can be lengthened or shortened. In this way, the input signal can meet the quality requirements while also playing a key role in subsequent model training.

[0170] S104: The blood pressure prediction module uses a feature fusion model to fuse the one-dimensional physiological signal high-dimensional features and the personalized high-dimensional features, and finally uses a multilayer perception model to obtain the predicted blood pressure value.

[0171] Specifically, the multilayer perception model is composed of two fully connected layers, which is used to map the fused features to a specific target space, and is used to fit the coefficients a and b in y=ax+b.

[0172] Specifically, the blood pressure value level fusion is performed in the feature fusion layer.

[0173] In specific implementation, the feature fusion mode includes but is not limited to: direct addition, direct multiplication, Cartesian product, and neural network fusion based mode.

[0174] The training of the blood pressure value prediction model in the embodiment includes the following steps:

[0175] According to the time sequence, the original one-dimensional physiological signal is cut into 10-second slices, and the corresponding blood pressure values on these slices are labeled according to the blood pressure values generated by PSG monitoring, and the blood pressure values include systolic and diastolic pressure values.

[0176] The personalized information of the measured person is extracted, and the one-dimensional physiological signal information is added as the input of the blood pressure prediction neural network.

[0177] The blood pressure value prediction model is constructed, which includes a feature extraction model and a feature fusion model. Two different feature extraction branches are used to extract high-dimensional features of the one-dimensional physiological signal and the personalized information, and finally the blood pressure value level fusion is performed in the feature fusion model.

[0178] The blood pressure value prediction model built is trained by using a 10-second one-dimensional physiological signal segment.

[0179] In the embodiment of the present application, the reason for quality assessment of data input is that the one-dimensional physiological signal collected by PSG or other wearable devices has noise that can and cannot be removed due to various reasons, so quality assessment of data input is needed to monitor the generated blood pressure value and label the corresponding blood pressure value on the slice; the one-dimensional physiological signal segment is evaluated, and only when the requirement is met, the signal segment can be input into the blood pressure value prediction model for prediction.

[0180] The advantage of the embodiment of the present application lies in the way of processing one-dimensional physiological signals by using a signal processing module, which continuously inputs signals into the model (except for blank signals generated when the device is connected and impulse signals generated when the device is disconnected, generally, signal segments will not be removed), the core lies in that the peak-valley detection method of the embodiment can detect signal peaks and valleys in near real time, and the neural network method calculation is also very fast, so the overall model can meet the requirement of near real-time prediction of human blood pressure.

[0181] As shown in Figure 12 , quality assessment is performed on a certain signal segment input by a subject, including:

[0182] Step S41: acquiring one-dimensional physiological signal data in a certain period of time;

[0183] Step S42: slicing the acquired one-dimensional physiological signal according to a length of 10 seconds;

[0184] Step S43: calculating signal energy and peak factor in the signal segment, and detecting the maximum value in the signal segment.

[0185] In a specific embodiment, the signal segment needs to be filtered and denoised first, and the amplitude range of the processed one-dimensional physiological signal is generally 0 to 3 millivolts.

[0186] In a specific embodiment, the calculation formula of signal energy is:

[0187]

[0188] Where x(t) is the energy signal, T represents the calculation interval, which is 10 seconds in the embodiment, t is the time, and d is the differential symbol.

[0189] The calculation formula of the peak factor is:

[0190]

[0191] Where Xp is the detected peak amplitude, Xrms refers to the root mean square value of the signal segment, and the ratio of the two is used to measure the proportion of the peak amplitude in the whole signal.

[0192] The calculation formula of the kurtosis factor is as follows:

[0193]

[0194] Where μ and σ are the mean and standard deviation of the one-dimensional physiological signal segment, respectively.

[0195] The kurtosis factor is used to measure the steepness of the waveform. Too steep indicates that the waveform rises very quickly, and it is likely that it is not the physiological waveform of most people.

[0196] Step S44: output whether the signal segment is qualified.

[0197] The peak-valley detection in the embodiment of the present application is as shown in Figure 4 The peak-valley detection of the certain signal segment input by the subject includes:

[0198] Step S231: acquire one-dimensional physiological signal data within a certain period of time;

[0199] Step S232: calculate the first-order difference of the input signal to perform signal transformation;

[0200] In a specific embodiment, at time i, the window-weighted z i The definition is as follows: the method of signal transformation is:

[0201]

[0202] Where y k is the input one-dimensional physiological signal, Δy k is the result of the first-order difference, w is the weighting time window, and the transformed signal is called the SSF signal.

[0203] Step S233: set an initial threshold, traverse the signal from the beginning, find the point just greater than the threshold and set a judgment window, find the wave valley, and update the threshold.

[0204] In a specific embodiment, the size of the judgment window is related to the sampling rate, and the proportion is 0.125 to 0.128.

[0205] In a specific embodiment, the threshold is updated to 0.6 times the maximum value in the judgment window.

[0206] Step S234: calibrate the detected wave valley.

[0207] ​In a specific embodiment, the human heart pumps blood at certain intervals, which fall within a certain range. While these intervals vary from person to person, they do not exceed this range. Therefore, if the detected trough interval exceeds this range, it can be assumed that one of the two detections has failed, and a decision must be made. This ensures the model's effectiveness in feature extraction of one-dimensional physiological signals and the final blood pressure regression prediction.

[0208] Step S235: Locate the index of the peak based on the found trough index, and output the final result.

[0209] The peak and valley detection method consists of four steps: transformation, wave finding, calibration, and positioning. In some embodiments, the calibration step can be omitted because it has little impact on accuracy, and omitting it can further improve the detection speed.

[0210] In a specific embodiment, once the trough index is determined, the peak must be located between adjacent troughs. Therefore, it is only necessary to find the index corresponding to the maximum value between adjacent troughs, which is the peak index. Ultimately, this method can output the peaks and troughs of the signal at once and achieve real-time output.

[0211] This embodiment uses mean error (ME), mean absolute error (MAE), standard deviation (STD), and Pearson correlation coefficient (r) to measure model accuracy.

[0212] Existing methods using single-channel pulse wave (PPG) for prediction, referencing a method proposed in a paper in November 2021, provide ME, MAE, and STD values ​​of -0.23, 5.16, and 8.50 for systolic blood pressure, respectively, without providing r values; and ME, MAE, and STD values ​​of 0.59, 2.89, and 4.78 for diastolic blood pressure, respectively, without providing r values.

[0213] like Figure 13a-13d As shown, in this embodiment of the invention, the predicted blood pressure value is compared with the reference value. The r-value measures the correlation between the predicted value and the true value. It can be understood as the degree of fit between the predicted value and the true value. Figure 13a The mean error (ME) was 0.16, the mean absolute error (MAE) was 5.99, the standard deviation (STD) was 6.90, and the correlation coefficient (r) was 0.8993. Figure 13b The mean error (ME) was -0.51, the mean absolute error (MAE) was 3.33, the standard deviation (STD) was 4.10, and the correlation coefficient (r) was 0.8653. Figure 13c The mean error (ME) was -0.17, the mean absolute error (MAE) was 3.98, the standard deviation (STD) was 4.62, and the correlation coefficient (r) was 0.9564. Figure 13dThe mean error (ME) is -0.24, the mean absolute error (MAE) is 2.33, the standard deviation (STD) is 2.95, and the correlation coefficient (r) is 0.9340. Line l1 represents the ideal line (predicted value equals actual value), and line l2 represents the actual line fitted from the predicted value. The closer the actual line is to the ideal line, the better. Although... Figure 13a The overall error is better, but the reason is that its error is more balanced between positive and negative, so they are all "neutralized" (for example, if there are two predicted values, one with an error of 10 and the other with an error of -10, then the overall error is 0). Figure 13c The error in the value is relatively large, but it is very close to the ideal straight line. Obviously, the r value can "more comprehensively" evaluate the effect of the model.

[0214] The international standard AAMI requires ME±STD to be between 5±8, and the results obtained from the embodiments of the present invention fully meet the requirements.

[0215] Therefore, the embodiments of the present invention have smaller errors, higher accuracy, and more comprehensive evaluation indicators.

[0216] In some embodiments, it is assumed that a PSG device is used for signal acquisition, and only ECG and PPG signals are used (e.g., Figure 14a and Figure 14b As shown), the individual information of the test subject (gender, age, assumed to be 0 and 52, where 0 represents male and 52 represents 52 years old) is determined at the same time.

[0217] The acquired ECG and PPG signals are input. First, this embodiment of the invention pre-captures and buffers a 10-second signal segment. This signal segment undergoes quality assessment (primarily evaluating for blank signals, abnormal impulse signals, and other signals that clearly lack one-dimensional physiological signal characteristics). After quality assessment, the method filters the signal. Then, peak and valley detection is performed on the filtered signal using the SSF method. Finally, the method processes the signal segment (ECG and PPG signals, with a shape format of (1, 2, fs) after this series of preprocessing operations. 10), where 1 represents one signal segment, 2 represents two channels, fs represents the sampling rate, and 10 represents 10 seconds. 10 represents the length of the signal segment. Together with the subject's personalized information (gender and age, in shape format (1, 2), where 1 represents one sample and 2 represents two features, namely gender and age), the data can be input into the model to obtain the subject's blood pressure value within 10 seconds (in shape format (1, 2), where 1 represents one sample and 2 represents two outputs, namely systolic and diastolic blood pressure values, assumed to be 115 mmHg and 71 mmHg).

[0218] Since ECG, PPG and personalized information are used, which belong to multiple modalities, the neural network model in the embodiment of the application belongs to a "multi-modal model", and the blood pressure prediction process learned is called "multi-modal learning". "Feature fusion" is divided into two parts, one part is the fusion of one-dimensional physiological signals, mainly the splicing of ECG and PPG in the channel, and the "one-dimensional physiological signal feature" obtained through the "feature extraction sub-network" is obtained; the second part is to splice and fuse the "personalized information feature" extracted by the machine learning model and the "one-dimensional physiological signal feature" in the previous step of model decision. This process is "feature fusion".

[0219] In the feature fusion model, a shared space of multi-modal joint representation is constructed, V phy represents the high-dimensional feature of SEResNet and LKA extracted one-dimensional physiological signal, and the high-dimensional feature of personalized information is represented by Therefore, the multi-modal feature is defined as f pp The expression is as follows:

[0220]

[0221] Where Concat(·) represents the concatenation of multiple input vectors, V phy represents the high-dimensional feature of one-dimensional physiological signal, V per represents the high-dimensional feature of personalized information.

[0222] The fused multi-modal feature has a large amount of redundant information, so a "shared layer" is used to extract the feature. As shown in Figure 15 The first convolutional layer is used to smooth the fused feature. Then, two convolutional layers are used to generate a deep representation. The final output is used for the shared characteristics of the downstream task.

[0223] This embodiment uses the Concat operation for feature fusion, which splices the high-dimensional feature of personalized information to the back of the one-dimensional physiological signal high-dimensional feature vector.

[0224] For example: the one-dimensional physiological signal feature vector is [0, 1, 2, 2], and the personalized information high-dimensional feature vector is [3, 4], and the spliced vector is [0, 1, 2, 2, 3, 4].

[0225] Feature after feature fusion f pp Continue to use some convolutional layers to extract features, and then input them into a multi-layer perceptron, respectively, for predicting systolic pressure and diastolic pressure each one, and finally set the output unit to 1, indicating that only one value is output.

[0226] Experimental example:

[0227] The peak-valley detection method is essentially to find the maximum and minimum values in the window as the wave peak and wave trough of the whole signal through the sliding window.

[0228] In the traditional peak-valley detection method, a one-dimensional physiological signal is given, the length of which is 3206.1 million sampling points (about 71 hours), the detection time of the traditional method is 286.94 seconds (about 4.78 minutes), and the number of detected peak and valley points is 327492.

[0229] And the detection time of the SSF method proposed in the embodiment of the application is 26 seconds, and the number of detected peak and valley points is 339940.

[0230] Comparative results show that the peak and valley points detected by the traditional method have a large number of false detection and missed detection, while the peak and valley points detected by the SSF method are basically real.

[0231] Compared with the traditional method, the SSF method improves the accuracy by 10%, and the speed is about 11 times that of the traditional method.

[0232] Through calculation, it is known that the sampling speed of 3206.1 million samples is 27 seconds, so a piece of f s The detection time of a 10-sample signal (the input signal length of the SSF method in the embodiment of the application, f s is 1ms, indicating that the sampling speed of the SSF method in the embodiment of the application is extremely fast.

[0233] The traditional blood pressure prediction regression method is based on a mathematical formula, the content of which is BP = a ln (PTT) + b, where a and b are constant parameters to be determined, which are determined according to the measured population, and PTT is pulse transit time, i.e. pulse wave transmission time, which has the physiological meaning of the time taken by the blood flow pumped by the heart to the measured PPG position. It has the following disadvantages:

[0234] First, the existing blood pressure prediction regression method needs frequent calibration. With the addition of new measured data or due to the change of time, the constant parameters need to be measured again, that is, the so-called "personalized calibration", which is very inconvenient.

[0235] Second, although the existing blood pressure prediction regression method has mathematical theory support, it is found that the theory has some assumptions that are unacceptable in modern times, so the blood pressure value calculated according to the existing blood pressure prediction regression method cannot be trusted.

[0236] Third, PTT is calculated based on ECG and PPG. Mainly by detecting the peaks of ECG and PPG respectively (the peak of ECG is the highest peak, that is, the peak of R wave), and time alignment is performed, and then the difference between the peaks of ECG and PPG is calculated, which is PTT.

[0237] However, ECG is much more complex than PPG signal (ECG has QRS complex, u wave, etc.), and the R wave peak of some subjects is not obvious, which will lead to detection failure and further PTT calculation error, so the time cannot be aligned, and finally the blood pressure value calculation error result will be caused.

[0238] In addition, since this method needs to store the previous peak detection results, and in order to calculate PTT, if the peak detection error occurs in the middle of PPG and ECG, the subsequent results will also be calculated error, in this case, the detection results need to be manually set to zero and start again from the failed place.

[0239] Fourth, the calculation of PTT needs to detect the peaks of ECG and PPG. Among them, PPG is relatively simple, and even using the traditional sliding window method can basically detect accurately. However, ECG is much more complex, and the detection accuracy is relatively low, but PTT requires high accuracy of ECG and PPG peak detection, which causes a contradiction. Therefore, in order to achieve the required accuracy, a more precise sliding window method must be used, and the sliding step is very small, but the result needs to be calibrated after detection, which actually exchanges speed for accuracy, so the existing blood pressure prediction method has poor real-time performance.

[0240] The neural network model used in the embodiment of the present application has a very fast calculation speed, so the peak-valley detection in the preprocessing operation is the largest time-consuming. As demonstrated before, a 125Hz 10 signal segment needs only 1ms for detection, and the peak and valley values can be obtained at the same time, so the method in the embodiment of the present application nearly meets the real-time requirement.

[0241] The implementation principle of the embodiment of the present application is to learn the information carried by ECG, PPG and personalized data through a deep learning model. A blood pressure value prediction model is designed for input (ECG, PPG and personalized information). By fusing the high-dimensional features of ECG, PPG and the high-dimensional features of personalized information, the blood pressure value prediction model can better learn.

[0242] The above is further detailed description of the present application in combination with specific preferred embodiments, and cannot be deemed as limitation of the specific implementation of the present application to these descriptions. For those skilled in the art to which the present application belongs, without departing from the concept of the present application, a number of equivalent substitutions or obvious variations can be made, and the performance or use is the same, which should be deemed as falling within the protection scope of the present application.

Claims

1. A blood pressure prediction device, characterized in that, include: Information acquisition module: used to acquire one-dimensional physiological signals and personalized information of the prediction object; The one-dimensional physiological signal includes electrocardiogram signal and pulse wave signal; Signal processing module: used to cut the one-dimensional physiological signal into one-dimensional physiological signal segments and to preprocess the one-dimensional physiological signal segments; Feature extraction module: used to input the one-dimensional physiological signal fragment and personalized information into the constructed blood pressure numerical prediction model. The blood pressure numerical prediction model includes a feature extraction model and a feature fusion model. It is used to extract features from the preprocessed one-dimensional physiological signal fragment and output it as a one-dimensional physiological signal high-dimensional feature. At the same time, it extracts features from the personalized information and outputs it as a personalized high-dimensional feature. Blood pressure prediction module: used to perform feature fusion on the one-dimensional physiological signal high-dimensional features and the personalized high-dimensional features through the feature fusion model, and output the predicted blood pressure value; The preprocessing specifically includes: Data denoising employs at least one of low-pass filtering, wavelet thresholding, and band-pass filtering. Quality assessment evaluates whether the denoised one-dimensional physiological signal segment can be used for blood pressure prediction. The assessment indicators include signal energy, peak-to-peak interval, peak factor, and kurtosis factor within the one-dimensional physiological signal segment. Peak and valley detection uses first-order difference to transform the pulse wave signal, find the valley, calibrate and locate it, thereby detecting the peaks and valleys in the signal. The peak-valley detection specifically includes: Acquire the pulse wave signal; Calculate the first-order difference of the pulse wave signal and perform signal transformation. The signal transformation method is as follows: Where y k It is the input one-dimensional physiological signal, Δy k It is the result of the first-order difference, where w is the weighted time window, and the transformed signal is the slope and the signal. Set an initial threshold, traverse the slope and signal from the beginning, find the first signal point whose signal value is greater than the threshold and set a judgment window, find the trough of the slope and signal waveform, and update the threshold at the same time; The detected troughs are calibrated; Based on the found trough index, the peak index is located, the slope and the peak of the signal are obtained, and finally the peak and trough of the signal are output.

2. The blood pressure prediction device according to claim 1, characterized in that: The personalized information includes at least one of gender, age, height, and weight.

3. The blood pressure prediction device according to claim 1, characterized in that: Wavelet thresholding was used to denoise the pulse wave signal; a bandpass filter was used to denoise the electrocardiogram signal.

4. The blood pressure prediction device according to claim 1, characterized in that: The feature extraction model extracts features from preprocessed one-dimensional physiological signal segments by building a convolutional neural network, and extracts features from personalized information by building a machine learning model.

5. The blood pressure prediction device according to claim 4, characterized in that: The specific steps of extracting features from the preprocessed one-dimensional physiological signal fragments by building a convolutional neural network are as follows: the processed one-dimensional physiological signal is divided into a training set, a validation set and a test set, the neural network is trained, and the high-dimensional features of the one-dimensional physiological signal corresponding to the one-dimensional physiological signal are obtained by training a regressor. The step of extracting features from the personalized information by building a machine learning model specifically involves dividing the processed personalized information into a training set, a validation set, and a test set, performing machine learning training, and obtaining the initial personalized high-dimensional features corresponding to the personalized information through training a regressor.

6. The blood pressure prediction device according to claim 5, characterized in that: The training set is used to train the blood pressure numerical prediction model to obtain the blood pressure numerical prediction model; the validation set is used to validate the trained blood pressure numerical prediction model and adjust the model parameters until the model's mean error and standard deviation meet the preset conditions; the test set is used to test the blood pressure numerical prediction model that meets the preset conditions.

7. The blood pressure prediction device according to claim 1, characterized in that: The signal processing module further includes the following for processing one-dimensional physiological signals: The one-dimensional physiological signal was cut 10 seconds before the procedure to obtain a one-dimensional physiological signal fragment, and its quality was assessed. The signal quality is determined based on the quality assessment results. If it is not qualified, wait 10 seconds to cut the next segment. If the quality assessment meets the requirements, the one-dimensional physiological signal fragment is used as input into the blood pressure numerical prediction model; The next cut is performed, starting from the previous cut position to cut out another 10-second signal. Then, a quality assessment is performed, and the signal quality is judged to be qualified based on the quality assessment results, until the entire one-dimensional physiological signal is completely detected. Input a 10-second one-dimensional physiological signal and output the quality assessment results and the corresponding blood pressure value.

8. A blood pressure prediction device, characterized in that, Includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, performs the following steps: S1: Obtain one-dimensional physiological signals and personalized information of the prediction object; the one-dimensional physiological signals include electrocardiogram signals and pulse wave signals; S2: Cut the one-dimensional physiological signal into one-dimensional physiological signal segments and preprocess the one-dimensional physiological signal segments; S3: Construct a blood pressure numerical prediction model, which includes a feature extraction model and a feature fusion model; input the one-dimensional physiological signal segment and personalized information into the blood pressure numerical prediction model; the feature extraction model extracts features from the preprocessed one-dimensional physiological signal segment and outputs them as one-dimensional physiological signal high-dimensional features, and at the same time extracts features from the personalized information and outputs them as personalized high-dimensional features. S4: The feature fusion model performs feature fusion on the one-dimensional physiological signal high-dimensional features and the personalized high-dimensional features, and outputs the predicted blood pressure value; The preprocessing specifically includes: Data denoising employs at least one of low-pass filtering, wavelet thresholding, and band-pass filtering. Quality assessment evaluates whether the denoised one-dimensional physiological signal segment can be used for blood pressure prediction. The assessment indicators include signal energy, peak-to-peak interval, peak factor, and kurtosis factor within the one-dimensional physiological signal segment. Peak and valley detection uses first-order difference to transform the pulse wave signal, find the valley, calibrate and locate it, thereby detecting the peaks and valleys in the signal. The peak-valley detection specifically includes: Acquire the pulse wave signal; Calculate the first-order difference of the pulse wave signal and perform signal transformation. The signal transformation method is as follows: Where y k It is the input one-dimensional physiological signal, Δy k It is the result of the first-order difference, where w is the weighted time window, and the transformed signal is the slope and the signal. Set an initial threshold, traverse the slope and signal from the beginning, find the first signal point whose signal value is greater than the threshold and set a judgment window, find the trough of the slope and signal waveform, and update the threshold at the same time; The detected troughs are calibrated; Based on the found trough index, the peak index is located, the slope and the peak of the signal are obtained, and finally the peak and trough of the signal are output.

Images