Blood pressure measuring method based on wrist single-path PPG signal

Abstract

The invention belongs to the technical field of human health monitoring, and discloses a blood pressure measurement method based on a wrist single-path PPG signal, and the method comprises the steps: collecting a first original wrist signal in a T1 time period and a second original wrist signal in a T2 time period corresponding to two different positions in a wrist part of a user through a wearable device; calculating the distance between the two acquisition positions; performing compensation processing on the first original wrist signal and the second original wrist signal by adopting a generative compensation deep neural network to obtain a complete first standard wrist signal and a complete second standard wrist signal within a time period T = T1 + T2; feature extraction is conducted on the first standard wrist signal and the second standard wrist signal, blood flow dynamic features, blood vessel hardness features and heart index features are obtained, and the reference blood pressure of the user is measured through a hybrid intelligent network model.

Description

Technical Field

[0001] The present invention belongs to the technical field of human health monitoring, and in particular relates to a blood pressure measurement method based on a single-channel PPG signal at the wrist. Background Art

[0002] Blood pressure, as an important indicator of physical health, often represents a person's cardiovascular health status. Traditional blood pressure measurement methods usually rely on cuff-type sphygmomanometers to measure the pressure of blood in the arteries by pressurizing and releasing pressure. Although this method is accurate, it is complicated to operate and requires specialized equipment and technicians during the measurement process, making it unsuitable for daily self-monitoring and continuous monitoring. With the advancement of technology and the increasing demand for health monitoring, monitoring technology based on photoplethysmography (PPG) signals has become a commonly used physiological signal detection method in wearable devices due to its non-invasive, real-time and low-cost characteristics.

[0003] Currently, most blood pressure measurement methods based on PPG signals rely on multi-channel PPG signals or need to be combined with other physiological signals (such as ECG signals) to implement. The changes in pulse wave velocity are calculated by multi-signal peak differences to measure blood pressure, but these methods increase the complexity and cost of the equipment. As for the blood pressure measurement method using a single-channel PPG signal, since the single-channel PPG signal cannot directly provide relevant characteristics of pulse wave velocity, the accuracy of its measurement results is lower than that of the multi-signal fusion method. Summary of the invention

[0004] The present invention provides a blood pressure measurement method based on a single-channel PPG signal at the wrist, which solves the problem of low measurement accuracy of the existing single-channel PPG signal, and overcomes technical problems such as equipment complexity and increased cost caused by other blood pressure measurement methods based on PPG signals due to reliance on multiple signals or additional sensors.

[0005] The present invention can be achieved through the following technical solutions:

[0006] A blood pressure measurement method based on a single-channel PPG signal at the wrist, comprising the following steps:

[0007] Step 1: Use a wearable device to collect a first original wrist signal in a T1 time period and a second original wrist signal in a T2 time period corresponding to two different positions on the user's wrist, and calculate the distance between the two collection positions;

[0008] Step 2: Using a generative compensation deep neural network to perform compensation processing on the first original wrist signal and the second original wrist signal, to obtain a complete first standard wrist signal and a complete second standard wrist signal within a time period of T=T1+T2;

[0009] Step 3: extracting features of the first standard wrist signal and the second standard wrist signal respectively to obtain blood flow dynamics features, blood vessel hardness features, and heart index features, and using a hybrid intelligent network model to measure the user's baseline blood pressure;

[0010] Among them, the vascular hardness feature is characterized by pulse wave velocity PWV, and the pulse wave velocity PWV is calculated based on the average delay time of the peak of the first standard wrist signal and the second standard wrist signal and the distance between the two acquisition positions. The hybrid intelligent network model takes the LightGBM model as the model body, and uses the particle swarm optimization algorithm Particle Swarm Optimization (PSO) and the simulated annealing algorithm Simulated Annealing (SA) to optimize the parameters of the LightGBM model for construction.

[0011] Furthermore, for M users, N groups of PPG signal data within T time periods are collected respectively, and a total of M*N groups of PPG signal data are obtained to form a data set.

[0012] During training, for each set of PPG signal data, the PPG signal data in the T1 time period is used as the input of the generator in the generative compensation deep neural network for generative training, and then its output and the PPG signal data in the T2 time period are input into the discriminator in the generative compensation deep neural network for authenticity judgment, and the training is continued alternately until the generator and the discriminator reach a balance and the training is completed;

[0013] The trained generative compensation deep neural network is used to generate a first original wrist signal in the T1 time period corresponding to a first generated PPG signal in the T2 time period, and a second original wrist signal in the T2 time period corresponding to a second generated PPG signal in the T1 time period.

[0014] Further, the first original wrist signal and the second original wrist signal are denoted as PQ1 and PQ2 respectively, the first generated PPG signal and the second generated PPG signal are denoted as PS1 and PS2 respectively,

[0015] The following formula is used to synthesize the first standard wrist signal PH1 and the second standard wrist signal PH2 by splicing:

[0016] PH1[kT]={PQ1, PS1}={ph11, ph12, ph13,…, ph12n}, kT=1, 2, 3,…, 2n;

[0017] PH2[kT]={PS2, PQ2}={ph21, ph22, ph23,..., ph22n}, kT=1, 2, 3,..., 2n.

[0018] Furthermore, a photoelectric sensor and an acceleration sensor are arranged inside the wearable device, wherein the photoelectric sensor is used to collect the PPG signal of the wrist, and the acceleration sensor is used to collect the acceleration signal of the wrist.

[0019] The photoelectric sensor is used to collect the first original wrist signal at the collection position I, and then the wearable device is moved to the collection position II to collect the second original wrist signal. At the same time, the acceleration sensor is used to collect the acceleration signal during the movement to calculate the distance between the collection position I and the collection position II.

[0020] Furthermore, when collecting the first original wrist signal and the second original wrist signal, the corresponding acceleration signals are collected by using the acceleration sensor at the same sampling frequency, which are respectively recorded as A1 and A2, as shown below. They correspond to the first original wrist signal and the second original wrist signal one by one:

[0021] A1[k]={a11,a12,a13,…,a1n},k=1,2,3,…,n

[0022] A2[k]={a21,a22,a23,…,a2n},k=1,2,3,…,n

[0023] The mean, variance, smoothness, signal entropy, skewness and kurtosis of the acceleration signals A1 and A2 are calculated respectively and used as the input of the machine learning classification model to classify these acceleration signals, find out the abnormal clusters and their corresponding PPG signals in the first original wrist signal and the second original wrist signal, and eliminate them, thereby realizing denoising of the first original wrist signal and the second original wrist signal.

[0024] Furthermore, the wearable device is used to collect the first PPG signal and the second PPG signal in the T time period corresponding to two different positions on the wrists of multiple users, and the corresponding blood flow dynamics characteristics, vascular hardness characteristics, and cardiac index characteristics are extracted, and the corresponding systolic and diastolic blood pressures are measured to construct a data set;

[0025] The hybrid intelligent network model is trained with the blood flow dynamics characteristics, vascular hardness characteristics, and cardiac index characteristics in the dataset as input and the corresponding systolic and diastolic blood pressure in the dataset as output. The trained hybrid intelligent network model is then used to measure the baseline blood pressure of the user to be tested.

[0026] The beneficial technical effects of the present invention are as follows:

[0027] 1. Simplify equipment complexity and enhance convenience:

[0028] By moving the wearable device along the wrist and changing the collection position of the PPG signal, single-channel PPG signal collection at two collection positions can be achieved using only one photoelectric sensor in the wearable device itself, thereby realizing blood pressure measurement, avoiding the need for multi-channel signals and additional sensors, such as two photoelectric sensors or ECG signal measuring instruments, simplifying the design and manufacturing of the device, and reducing production costs and equipment complexity.

[0029] 2. Improve measurement accuracy and signal stability:

[0030] Taking into account movement interference, the acceleration sensor in the wearable device is used to synchronously collect acceleration signals in the T1 time period and the T2 time period. The data classification model is used to eliminate the influence of the user's motion state on the PPG signal, thereby enhancing the stability and reliability of the PPG signal and ensuring that accurate blood pressure monitoring results can still be provided during the user's daily activities.

[0031] At the same time, a generative compensation deep neural network is used to compensate the first original wrist signal and the second original wrist signal to obtain the first generated wrist PPG signal PS1 within the T2 time period and the second generated wrist PPG signal PS2 within the T1 time period, and the denoised signal PQ1 of the first original wrist signal and the first generated wrist signal PS1, the second generated wrist signal PS2 and the denoised signal PQ2 of the second original wrist signal are respectively combined to obtain the first synthetic wrist signal PH1 and the second synthetic wrist signal PH2 within the total time period T=T1+T2, that is, a group of PPG signals of different wrist positions corresponding to each other in time are obtained, and then combined with the acceleration signal and movement time of the acceleration sensor, the distance between the two sampling positions can be calculated, thereby making up for the shortcoming of the existing technology that the single-channel PPG signal cannot extract the pulse wave conduction velocity characteristics, and effectively improving the accuracy of blood pressure measurement and signal quality.

[0032] 3. Application of intelligent optimization algorithm and comprehensive feature extraction:

[0033] The LightGBM model is combined with particle swarm optimization PSO and simulated annealing SA algorithm to construct a hybrid intelligent network model, which can efficiently optimize the parameters of the hybrid intelligent network model, improve the robustness and generalization ability of the measurement model, better adapt to the individual differences of different users, and provide personalized blood pressure measurement results. At the same time, it extracts a variety of blood characteristic parameters including blood flow dynamics characteristics, vascular hardness characteristics, and cardiac index characteristics to comprehensively reflect the user's cardiovascular health status, provide rich input information for blood pressure measurement, enhance the accuracy and reliability of the measurement model, and provide a scientific basis for the user's health management. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is an overall flow chart of an embodiment of the present invention;

[0035] Figure 2 is a schematic diagram of a process in which a user moves a blood pressure measurement device on the wrist in an embodiment of the present invention;

[0036] Figure 3 is a schematic diagram of a signal synthesized from a first original wrist signal and a first generated wrist signal in an embodiment of the present invention;

[0037] Figure 4 is a signal schematic diagram of a first denoised wrist signal and a second denoised wrist signal in an embodiment of the present invention;

[0038] Figure 5 It is a comparison diagram of the measured blood pressure of the hybrid intelligent network model in the embodiment of the present invention, the measured blood pressure after the adaptive genetic algorithm parameters are optimized, and the real reference blood pressure. DETAILED DESCRIPTION

[0039] The specific implementation of the present invention is described in detail below with reference to the accompanying drawings and preferred embodiments.

[0040] Considering the limitations of traditional blood pressure measurement methods, especially the blood pressure measurement method based on a single PPG signal from the wrist, which often faces the problem of low measurement accuracy, while the blood pressure measurement method based on multiple signals or additional sensors increases the complexity and cost of the equipment, such as Figure 1 As shown, the present invention proposes a blood pressure measurement method based on a single-channel PPG signal on the wrist, which uses a wearable device to collect a single-channel PPG signal on the wrist and an acceleration signal, and improves the accuracy and stability of blood pressure measurement by improving the signal acquisition process and optimizing the signal processing algorithm and model, thereby providing users with a convenient and efficient blood pressure monitoring solution. Users can monitor their blood pressure anytime and anywhere through wearable devices such as smart watches or fitness bracelets, without the need for cumbersome operations and specialized technicians. It is suitable for daily self-monitoring and continuous monitoring, providing a convenient health management tool and a scientific basis for users' health management.

[0041] The details are as follows:

[0042] Step 1: Use a wearable device to collect a first original wrist signal in a T1 time period and a second original wrist signal in a T2 time period corresponding to two different positions on the user's wrist, and calculate the distance between the two collection positions.

[0043] Considering the complexity of simplifying the acquisition equipment, we can use a photoelectric sensor and an acceleration sensor of the wearable device itself to complete the acquisition of PPG signals and acceleration signals twice, which can be used for subsequent blood pressure calculation and distance calculation. This does not require additional photoelectric sensors, which helps to reduce costs and reduce the size of wearable devices.

[0044] First, the photoelectric sensor is used to collect the first original wrist signal at the collection position I, and then the wearable device is moved to the collection position II to collect the second original wrist signal. At the same time, the acceleration sensor is used to collect the acceleration signal during the movement to calculate the distance between the collection position I and the collection position II. Figure 2 Describes the process of a user moving a wearable device on the wrist. When moving the wearable device, the wrist on which the device is worn should be kept as still as possible, and only the wearable device should be moved.

[0045] The user puts the wearable device on the wrist and moves it. After the movement is completed, the movement time is recorded as Td. At the same time, the acceleration sensor in the wearable device is used to collect the movement state of the user's device to obtain three-axis acceleration signals Ax, Ay, and Az. The three-axis acceleration signals in the Td time period are integrated over time Td to obtain the three-axis movement speeds Vx, Vy, and Vz in the xyz directions. The movement speeds in the xyz directions are integrated over time Td again to obtain the distances Dx, Dy, and Dz in the xyz directions. The maximum value of the movement distances in the three directions is taken as the wrist direction, thereby determining the movement distance of the device in the wrist direction, that is, the distance Ddif between the collection position I and the collection position II.

[0046] The collected first original wrist signal and the second original wrist signal are respectively recorded as PQ1 and PQ2, which are expressed as follows:

[0047] PQ1[k]={p11,p12,p13,…,p1n},k=1,2,3,…,n

[0048] PQ2[k]={p21,p22,p23,…,p2n},k=1,2,3,…,n

[0049] Preferably, 64 Hz is used as the sampling frequency and the pulse wave data acquired within 30 seconds is used as the processing unit, and in this case n=1920.

[0050] In addition, human blood pressure will remain relatively stable in a resting environment, and the measurement results will be relatively accurate. If a photoelectric sensor is used to obtain two sets of PPG signal data by changing the collection position for subsequent blood pressure testing, considering that there may be errors in the definition of movement and stillness of wearable devices, we can denoise the PPG data through acceleration data. Therefore, when collecting the first original wrist signal and the second original wrist signal, we also collect the corresponding acceleration signals A1 and A2 at the same collection frequency, which are expressed as follows:

[0051] A1[k]={a11,a12,a13,…,a1n},k=1,2,3,…,n

[0052] A2[k]={a21,a22,a23,…,a2n},k=1,2,3,…,n

[0053] Preferably, 64 Hz is used as the sampling frequency and the pulse wave data acquired within 30 seconds is used as the processing unit, and in this case n=1920.

[0054] Step 2: Use a generative compensation deep neural network to compensate the first original wrist signal and the second original wrist signal to obtain a complete first standard wrist signal and a complete second standard wrist signal within the time period T=T1+T2.

[0055] Since blood pressure remains relatively stable in a resting environment, the present invention uses a generative compensation deep neural network to generate compensation for the PPG signal to obtain a group of PPG signals at different wrist positions that correspond to each other in time.

[0056] S2.1. Filter the first original wrist signal P1 and the second original wrist signal P2 through a bandpass filter to remove noise and interference components in the signals, and obtain a first denoised wrist signal PQ1′ and a second denoised wrist signal PQ2′. Preferably, use a third-order Butterworth filter with a frequency range of 0.1 Hz to 5 Hz to filter the PPG signal to filter out the noise part in the signal and retain the information of the blood characteristic and heart characteristic frequency bands.

[0057] In addition to the noise caused by the above-mentioned acquisition frequency, obtaining two sets of PPG signal data by moving the wearable device will also produce noise interference in a non-resting state. Therefore, we can judge the motion interference through the three-axis accelerometer signals Ax, Ay, and Az within time T, and further eliminate the possibility that the first original wrist signal PQ1 within the T1 time period and the second original wrist signal PQ2 within the T2 time period are affected by motion interference.

[0058] First, extract multiple acceleration features, including but not limited to the mean, variance, smoothness, signal entropy, skewness, kurtosis, etc. These features can effectively reflect the changes in the acceleration signal and the user's motion status within the T time period. For the acceleration signal A, the feature calculation formula is as follows:

[0059] A[k]={a1,a2,a3,…,an},n=1920,k=1,2,3,…,n

[0060] Mean:

[0061] variance:

[0062] Waveform Smoothness:

[0063] Information Entropy:

[0064] Skewness:

[0065] Kurtosis:

[0066] The above formula is used to calculate the characteristics of multiple groups of acceleration signals A1 and A2 respectively to construct a data set.

[0067] Then, a motion interference discrimination model is established based on the extracted multiple features.

[0068] A machine learning algorithm such as a K-nearest neighbor algorithm, a support vector machine, a decision tree, a random forest or a neural network can be used as the model body to classify wrist signals affected by motion interference and wrist signals that are not interfered with. The present invention selects to use a support vector machine as a motion interference discrimination model, and uses each data in the data set as the input of a machine learning classification model such as a support vector machine to classify these acceleration signals, distinguish abnormal clusters from normal clusters, and complete the training.

[0069] Under the guidance of the trained motion interference discrimination model, the first synthetic wrist signal PH1 and the second synthetic wrist signal PH2 are classified and judged to find out the abnormal clusters and their corresponding PPG signals in the first synthetic wrist signal and the second synthetic wrist signal, and eliminate them to obtain the first denoised wrist signal PQ1 and the second denoised wrist signal PQ2.

[0070] Figure 3 The first denoised wrist signal and the second denoised wrist signal from which motion interference and noise effects have been removed in the embodiment are shown.

[0071] S2.2. Use a generative compensation deep neural network to generate and compensate the first denoised wrist PPG signal PQ1 and the second denoised wrist signal PQ2, respectively, to obtain the first generated wrist PPG signal PS1 within the time period T2, the second generated wrist PPG signal PS2 within the time period T1, and combine the first denoised wrist signal PQ1 and the first generated wrist signal PS1, the second generated wrist signal PS2 and the second denoised wrist signal PQ2, respectively, to obtain the first synthesized wrist signal PH1 and the second synthesized wrist signal PH2 within the total time period T=T1+T2.

[0072] Specifically, by using deep learning technology and deep neural network models, combined with the characteristic that blood pressure will remain relatively stable in a resting environment, a single-channel PPG signal is compensated and generated within a specific time period to obtain a set of PPG signals with different positions but corresponding times. The deep neural network model may adopt but is not limited to: generative adversarial networks, variational autoencoders, recursive neural networks, long short-term memory networks, autoregressive models, etc. The present invention selects to use a generative compensatory deep neural network as the model structure. The model structure improves the accuracy and consistency of single-channel PPG signal generation by introducing a spatiotemporal self-attention mechanism and a dynamic compensation strategy.

[0073] By performing multi-scale time series modeling on the PPG signal, the dynamic changes and correlations of the signal at different time points are captured, and by utilizing the conditional generative adversarial network cGAN, specific conditional information such as time period is introduced into the generation process to achieve accurate compensation generation of the PPG signal within a specific time period, thereby enhancing the consistency and smoothness of the generated signal and improving the accuracy and reliability of the generated signal. Among them, the generator in the cGAN network can identify and generate a set of missing parts in the PPG signal through conditional information, capture the characteristic structure in the original PPG signal through a deep learning algorithm, measure the most likely output state of the missing part, and finally obtain a set of PPG signal data that correspond to each other in time.

[0074] The training and use of the generative compensatory deep neural network are as follows:

[0075] For M users, N groups of PPG signal data in T time periods are collected respectively, and a total of M*N groups of PPG signal data are obtained to form a data set. It also includes preprocessing of PPG signal data such as noise removal and normalization processing, generating training data sets and test data sets containing complete PPG signals and corresponding condition information such as T1 time period and T2 time period. The data set comes from 68 volunteers, of which 36 are hypertensive patients and 32 are healthy volunteers. The data is collected at the wrist, and the total data collection time is about 500 hours;

[0076] Next, the generator and discriminator network structures are designed. The generator generates the missing PPG signal part through conditional information. The discriminator is used to judge the authenticity of the signal, initialize the network parameters, set the optimization algorithm such as Adam optimizer and hyperparameters such as learning rate. As a preference, when training the cGAN model, the following hyperparameters can be set: the optimizer of the generator and the discriminator uses the Adam optimizer, the learning rate is 0.0002, the first-order momentum parameter beta1 is 0.5, and the second-order momentum parameter beta2 is 0.9. The loss function of the generator uses the mean square error, and the loss function of the discriminator uses the cross entropy loss. During the training process, the batch size is set to 64, the number of training rounds is 200, and the noise vector dimension is 100. In order to prevent overfitting, the weight decay can be set to 0.001, and the regularization method with a Dropout rate of 0.3 is used. In addition, the learning rate decays by half every 50 epochs to dynamically adjust the learning speed of the model. These hyperparameter settings can effectively improve the generation capability and stability of the cGAN model, ensuring that the generated PPG signals meet the expected goals in terms of quality and consistency.

[0077] During the training process, for each set of PPG signal data, the PPG signal data in the T1 time period is used as the input of the generator in the generative compensation deep neural network for generative training, and then its output and the PPG signal data in the T2 time period are input into the discriminator in the generative compensation deep neural network for authenticity judgment, and the training is continued alternately until the generator and the discriminator reach a balance and the training is completed;

[0078] That is, input conditional information such as PPG signal data and noise vector in the T1 time period to the generator, the generator outputs the generated PPG signal, and calculates the error between the generated signal and the real signal such as the PPG signal data in the T2 time period, such as the mean square error, and updates the generator parameters through the back propagation algorithm; at the same time, input the real PPG signal and the generated PPG signal to the discriminator, the discriminator outputs the judgment on the authenticity of the signal, calculates the loss of the discriminator such as the cross entropy loss, and updates the discriminator parameters through the back propagation algorithm, so as to alternately train the generator and the discriminator until the generator and the discriminator reach a balance, at which time the error between the signal generated by the generator and the real signal is minimized;

[0079] Finally, the trained Generative Compensation Deep Neural Network is used to generate a first original wrist signal in time period T1 corresponding to a first generated PPG signal in time period T2, and a second original wrist signal in time period T2 corresponding to a second generated PPG signal in time period T1.

[0080] Evaluate the trained cGAN model. The model effect is as follows Figure 4 As shown, it can be seen that the quality of the generated signal is consistent with the original signal. Then, the first denoised wrist PPG signal PQ1 and the second denoised wrist PPG signal PQ2 are used to generate compensation through the trained cGAN-based generative compensation deep neural network to obtain the first generated wrist PPG signal PS1 in the T2 time period and the second generated wrist PPG signal PS2 in the T1 time period.

[0081] S2.3, when performing signal synthesis, the first denoised wrist signal PQ1 and the first generated wrist signal PS1, the second generated wrist signal PS2 and the second denoised wrist signal PQ2 are respectively combined to obtain the first synthesized wrist signal PH1 and the second synthesized wrist signal PH2 within a total time period T, and the specific synthesis rules are as follows:

[0082] PH1[kT]={PQ1, PS1}={ph11, ph12, ph13,…, ph12n}, kT=1,2,3,…,2n

[0083] PH2[kT]={PS2, PQ2}={ph21,ph22,ph23,…,ph22n},kT=1,2,3,…,2n

[0084] Figure 4 The first original wrist signal and the first generated wrist signal in the embodiment are shown. Through the trained generative compensatory deep neural network, the generated wrist signal in the second half can include the blood pressure related features of the original wrist signal in the first half.

[0085] Step 3: Extract features of the first standard wrist signal and the second standard wrist signal respectively to obtain blood flow dynamics features, blood vessel hardness features, and heart index features, and use a hybrid intelligent network model to measure the user's baseline blood pressure.

[0086] S3.1 When extracting features from the first standard wrist signal S1 and the second standard wrist signal S2, calculation and analysis of multiple key blood pressure-related feature parameters are involved.

[0087] Preferably, the blood flow dynamics characteristics include pulse rate and pulse waveform characteristics. The pulse rate is usually calculated by identifying and analyzing the period of the pulse waveform, which reflects the stability and rhythmicity of the heart rate. The pulse waveform characteristics include the pulse wave rise time, the pulse wave fall time, the pulse waveform area and the pulse wave amplitude. These characteristics are extracted by:

[0088] The peaks and troughs of the first standard wrist signal S1 and the second standard wrist signal S2 are calculated to obtain the peak sets SP1 and SP2 and the trough sets SV1 and SV2. The time difference between the peak sets SP1 and SP2 is counted, the time interval between every two adjacent peaks is calculated, and the average value Δt of all time intervals is obtained. Finally, the pulse rate is calculated by multiplying the inverse of the average time interval by 60.

[0089] The pulse wave rise time P_rise is the average time difference from the trough to the peak of the pulse wave, the pulse wave fall time P_fall is the average time difference from the peak to the trough of the pulse wave, the pulse waveform area P_area is the average value of the area enclosed by the pulse wave in one cycle, and the pulse wave amplitude P_amplitude is the average value of the difference between the peak value and the trough value of the pulse wave in the same cycle.

[0090] By extracting and analyzing these features, the dynamic characteristics of blood flow can be more comprehensively reflected, thereby improving the accuracy and stability of blood pressure measurement.

[0091] Preferably, the vascular hardness feature is mainly calculated by pulse wave velocity PWV. Compared with the traditional single-channel measurement method, the present invention can calculate the pulse wave velocity feature strongly correlated with blood pressure by the peak time difference between the first standard wrist signal and the second standard wrist signal.

[0092] The pulse wave velocity PWV estimates the speed of blood propagation in the arteries by measuring the average time delay Δt and the distance Ddif between the peak values ​​of the first standard wrist signal S1 and the second standard wrist signal S2 corresponding to the two acquisition positions. It is a key indicator for measuring arterial hardness and reflects the health of blood vessels and the risk of cardiovascular disease.

[0093] Preferably, the cardiac index features include the average adjacent pulse intervals IBI, the proportion of adjacent pulse peak intervals with a difference of more than 20 milliseconds pNN20, and the proportion of adjacent pulse peak intervals with a difference of more than 50 milliseconds pNN50. These indicators reflect important information such as the activity of the cardiac autonomic nervous system, cardiac workload, and myocardial function. Among them, IBI represents the average time interval between adjacent pulse wave peaks, while pNN20 and pNN50 features are important indicators of pulse variability, which are used to evaluate cardiac health and the function of the autonomic nervous system. The specific calculation is as follows:

[0094] Calculate the difference sequence of adjacent pulse intervals and count the number of times in the difference sequence that exceeds 20 milliseconds, i.e., NN20. Divide NN20 by the total number of pulse intervals N to obtain the pNN20 feature, i.e., pNN20=NN20 / N, which represents the proportion of the number of times the adjacent pulse interval exceeds 20 milliseconds in the total number of times. Similarly, count the number of times in the difference sequence that exceeds 50 milliseconds, i.e., NN50. Divide NN50 by the total number of pulse intervals N to obtain the pNN50 feature, i.e., pNN5=NN50 / N, which represents the proportion of the number of times the adjacent pulse interval exceeds 50 milliseconds in the total number of times. The pNN20 and pNN50 features can reflect the variability of the heart, and further reflect the heart's pumping capacity and blood flow pressure.

[0095] S3.2 After feature extraction is completed, machine learning models such as linear regression, support vector regression, decision tree regression, random forest regression, LightGBM, neural network regression, etc. can be used to regress and calculate the user's blood pressure through the extracted blood flow dynamics features, vascular hardness features and cardiac index features.

[0096] In the present invention, LightGBM is selected as the main body of the model, and the particle swarm optimization algorithm PSO is combined with the simulated annealing algorithm SA to optimize the parameters of the LightGBM model to complete the construction of the hybrid intelligent network model. This method can effectively find the optimal feature combination in the feature space to improve the accuracy of baseline blood pressure measurement. Of course, the particle swarm optimization algorithm PSO or the simulated annealing algorithm SA can also be used alone to optimize the parameters of the LightGBM model, as long as the recognition accuracy can meet the actual application requirements.

[0097] Among them, the particle swarm optimization algorithm imitates the optimization process of group behavior. Each particle updates its own position by memorizing its own best position and the best position of the entire group, thereby iterating towards the global optimal solution. The simulated annealing algorithm imitates the physical process of metal annealing, and jumps out of the local optimal solution by accepting the probability of a worse solution, so that it is more likely to find the global optimal solution. PSO and SA are combined, and the global search capability of PSO is used to quickly explore the feature space, while the local optimization capability of SA is used to finely adjust the parameters to achieve a more accurate optimization effect. This combination is particularly important in the baseline blood pressure measurement of the present invention, because the baseline blood pressure is affected by a variety of physiological and environmental factors, and accurate and comprehensive feature extraction and optimization are required. PSO can help quickly find potential advantage combinations among a large number of possible feature combinations, while SA can perform more refined optimization in the feature space and select the most measurable feature subset.

[0098] In addition to feature selection, optimizing the hyperparameters of the hybrid intelligent network model is also crucial to improving the performance and generalization ability of the measurement model. As a global optimization algorithm, PSO can quickly explore potential optimization directions in the hyperparameter space. Each particle represents a hyperparameter combination. It updates the parameter combination by memorizing its own optimal position and the optimal position of the entire group, and gradually iterates towards the direction of the global optimal solution. This global search capability enables PSO to quickly find the parameter combination of the global optimal solution in a variety of hyperparameter combinations, while SA performs local search in the candidate hyperparameter set determined by PSO. SA simulates the process of metal annealing and jumps out of the local optimal solution by accepting a certain probability of differential solution, so as to be more likely to find a parameter configuration close to the global optimal solution. This local optimization capability enables SA to further adjust parameters in the initial optimization direction determined by PSO to improve the performance and generalization ability of the model.

[0099] The wearable device is used to collect the first PPG signal and the second PPG signal in the T time period corresponding to two different positions on the wrists of multiple users, and the corresponding blood flow dynamics characteristics, vascular hardness characteristics, and cardiac index characteristics are extracted, and the corresponding systolic and diastolic blood pressures are measured to construct a data set;

[0100] The hybrid intelligent network model is trained with the blood flow dynamics characteristics, vascular hardness characteristics, and cardiac index characteristics in the data set as input, and the corresponding systolic and diastolic blood pressures in the data set as output. The trained hybrid intelligent network model is then used to measure the baseline blood pressure of the user to be tested, such as regressing the systolic and diastolic blood pressure data within a 30-second time period.

[0101] Preferably, when the user's real blood pressure data is obtained, the parameters of the hybrid intelligent network model of the present invention are further optimized according to the user's real blood pressure data to obtain a more accurate blood pressure measurement. The specific method is as follows:

[0102] After obtaining the real blood pressure data detected by the user using an electronic sphygmomanometer or a cuff sphygmomanometer, the blood pressure data output by the hybrid intelligent network model is compared for error estimation, and the specific error values ​​of systolic and diastolic pressures are obtained by calculating the root mean square error. Based on the error value and the existing adaptive parameter optimization algorithm, the back propagation algorithm of parameter gradient descent is used to iteratively optimize the blood pressure regression model multiple times, so that the output of the model after parameter optimization gradually approaches the real blood pressure data. When the specific error value cannot be reduced, the adaptive parameter optimization algorithm ends, and the accuracy of the optimized model will be greatly improved.

[0103] Figure 5It is a comparison chart of the blood pressure measured by the hybrid intelligent network model, the blood pressure measured after adaptive genetic algorithm parameter optimization and the real reference blood pressure. Compared with the real reference systolic and diastolic blood pressure, the hybrid intelligent network model can measure accurate blood pressure values, and the model after adaptive genetic parameter optimization can further improve the accuracy of blood pressure measurement. Compared with the model without parameter optimization, the error between the blood pressure measured by the model after parameter optimization and the real blood pressure is further reduced.

[0104] In summary, compared with the existing single-channel PPG signal method for measuring blood pressure, the present invention uses a generative compensation deep neural network to generate a group of PPG signals at different wrist positions that correspond to each other in time, thereby making up for the shortcoming that the single-channel PPG signal in the existing technology cannot extract the pulse wave conduction velocity characteristics. At the same time, combined with the motion characteristic parameters of the three-axis acceleration signal, it effectively eliminates the influence of the user's motion state on the signal, and improves the stability and accuracy of the blood pressure measurement results. At the same time, since the present invention only uses a single-channel wrist PPG signal, it greatly simplifies the complexity of the device and improves the convenience of the device and the user's experience.

[0105] Although specific embodiments of the present invention are described above, those skilled in the art should understand that these are merely examples and that various changes or modifications may be made to these embodiments without departing from the principles and essence of the present invention. Therefore, the scope of protection of the present invention is limited by the appended claims.

Claims

1. A blood pressure measurement method based on a single-channel PPG signal from the wrist, characterized in that The following steps are involved: Step 1: Use a wearable device to collect a first original wrist signal in a T1 time period and a second original wrist signal in a T2 time period corresponding to two different positions on the user's wrist, and calculate the distance between the two collection positions; Step 2: Using a generative compensation deep neural network to perform compensation processing on the first original wrist signal and the second original wrist signal, to obtain a complete first standard wrist signal and a complete second standard wrist signal within a time period of T=T1+T2; Step 3: extracting features of the first standard wrist signal and the second standard wrist signal respectively to obtain blood flow dynamics features, blood vessel hardness features, and heart index features, and using a hybrid intelligent network model to measure the user's baseline blood pressure; Among them, the vascular hardness feature is characterized by pulse wave velocity PWV, and the pulse wave velocity PWV is calculated based on the average delay time of the peak of the first standard wrist signal and the second standard wrist signal and the distance between the two acquisition positions. The hybrid intelligent network model takes the LightGBM model as the model body, and uses the particle swarm optimization algorithm PSO combined with the simulated annealing algorithm Simulated Annealing SA to optimize the parameters of the LightGBM model for construction.

2. The blood pressure measurement method based on a single-channel PPG signal at the wrist according to claim 1, characterized in that: For M users, N groups of PPG signal data are collected in T time periods, and a total of M*N groups of PPG signal data are obtained to form a data set. During training, for each set of PPG signal data, the PPG signal data in the T1 time period is used as the input of the generator in the generative compensation deep neural network for generative training, and then its output and the PPG signal data in the T2 time period are input into the discriminator in the generative compensation deep neural network for authenticity judgment, and the training is continued alternately until the generator and the discriminator reach a balance and the training is completed; The trained generative compensation deep neural network is used to generate a first original wrist signal in the T1 time period corresponding to a first generated PPG signal in the T2 time period, and a second original wrist signal in the T2 time period corresponding to a second generated PPG signal in the T1 time period.

3. The blood pressure measurement method based on a single-channel PPG signal at the wrist according to claim 2, characterized in that: The first original wrist signal and the second original wrist signal are respectively PQ1 and PQ2, the first generated PPG signal and the second generated PPG signal are respectively PS1 and PS2, The following formula is used to synthesize the first standard wrist signal PH1 and the second standard wrist signal PH2 by splicing: PH1[kT]={PQ1, PS1}={ph11, ph12, ph13,…, ph12n}, kT=1, 2, 3,…, 2n; PH2[kT]={PS2, PQ2}={ph21, ph22, ph23,..., ph22n}, kT=1, 2, 3,..., 2n.

4. The blood pressure measurement method based on a single-channel PPG signal at the wrist according to claim 1, characterized in that: The wearable device is provided with a photoelectric sensor and an acceleration sensor inside, wherein the photoelectric sensor is used to collect the PPG signal of the wrist, and the acceleration sensor is used to collect the acceleration signal of the wrist. The photoelectric sensor is used to collect the first original wrist signal at the collection position I, and then the wearable device is moved to the collection position II to collect the second original wrist signal. At the same time, the acceleration sensor is used to collect the acceleration signal during the movement to calculate the distance between the collection position I and the collection position II.

5. The blood pressure measurement method based on a single-channel PPG signal at the wrist according to claim 4, characterized in that: When collecting the first original wrist signal and the second original wrist signal, the corresponding acceleration signals are collected by using the acceleration sensor at the same sampling frequency, which are respectively recorded as A1 and A2, as shown below. They correspond to the first original wrist signal and the second original wrist signal one by one. A1[k]={a11,a12,a13,…,a1n},k=1,2,3,…,n A2[k]={a21,a22,a23,…,a2n},k=1,2,3,…,n The mean, variance, smoothness, entropy, skewness and kurtosis of the acceleration signals A1 and A2 are calculated respectively and used as the input of the machine learning classification model to classify these acceleration signals, find out the abnormal clusters and their corresponding PPG signals in the first original wrist signal and the second original wrist signal, and eliminate them, thereby realizing denoising of the first original wrist signal and the second original wrist signal.

6. The blood pressure measurement method based on a single-channel PPG signal at the wrist according to claim 1, characterized in that: The wearable device is used to collect the first PPG signal and the second PPG signal in the T time period corresponding to two different positions on the wrists of multiple users, and the corresponding blood flow dynamics characteristics, vascular hardness characteristics, and cardiac index characteristics are extracted, and the corresponding systolic and diastolic blood pressures are measured to construct a data set; The hybrid intelligent network model is trained with the blood flow dynamics characteristics, vascular hardness characteristics, and cardiac index characteristics in the dataset as input and the corresponding systolic and diastolic blood pressure in the dataset as output. The trained hybrid intelligent network model is then used to measure the baseline blood pressure of the user to be tested.

Images