A cuffless continuous blood pressure signal reconstruction method and system

Abstract

The present disclosure provides a cuffless continuous blood pressure signal reconstruction method and system, relating to the technical field of blood pressure monitoring, comprising: acquiring synchronous PPG signals and ECG signals; using a trained attention time sequence network, extracting PPG feature vectors and ECG feature vectors from the PPG signals and ECG signals respectively, and performing multi-modal fusion on the PPG feature vectors and ECG feature vectors, and based on the fused feature vectors, reconstructing a cuffless continuous blood pressure signal; wherein the feature vector extraction increases an information screening machine and a time sequence learning machine between the encoder and the decoder, which are respectively used for learning a channel probability distribution vector and a time sequence weight vector, guiding the decoder to decode high-level semantic features and the time sequence relationship of PPG and ECG before and after; the present disclosure proposes an attention time sequence network based on multi-modal fusion, and realizes the reconstruction of continuous blood pressure signals by analyzing the photoplethysmogram signal and the electrocardiogram signal.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of blood pressure monitoring, in particular to a cuffless continuous blood pressure signal reconstruction method and system. BACKGROUND

[0002] Hypertension is an important risk factor for stroke, cardiovascular and cerebrovascular diseases, and chronic kidney disease; due to the 'white coat' effect and behavioral factors, the measurement value of discrete blood pressure (BP) deviates from the true value in the range of -24mmHg to 33mmHg, and most medical associations recommend using cuffless continuous BP measurement technology.

[0003] According to the photoplethysmogram (PPG) and electrocardiogram (ECG), the pulse transit time (PTT), pulse arrival time (PAT), and pulse wave velocity (PWV) are calculated, which is a common cuffless BP detection method; the essence of these methods is to mathematically and physically model (MPM) the propagation process of blood between two given points to achieve the estimation of instantaneous BP, which is affected by the distance between the measured points, the thickness of the blood vessel wall, and the density of the blood. Although the pre-ejection period study can assist PAT, the pre-ejection period will suddenly change due to external factors and aging.

[0004] Machine learning (ML) based on feature engineering is applied in the estimation of instantaneous BP value; PPG and BP have a nonlinear relationship, so it is a common method to input morphological, time domain, frequency domain, and nonlinear domain features into a nonlinear regressor; but the nonlinear regressor of ML has high parameter adjustment difficulty and is difficult to overcome the specificity difference of the subjects, and the nonlinear analysis ability for long-range signals needs to be further improved.

[0005] Deep learning (DL) based on data-driven has a more powerful nonlinear modeling capability, and analyzing single-modal signals for BP estimation has the advantage of convenience, but single-modal signals are greatly disturbed by physiological and environmental factors; DL needs diversified data as a training set and the mapping relationship between BP signals and other physiological signals is not clear, and there is a risk of information loss in estimating BP using single-modal signals.

[0006] In order to solve the defects of single-modal signals in BP estimation, researchers use different modal signals as the source domain of DL or use different neural networks to analyze single-modal signals in multiple modalities; although the technology based on multi-modal fusion can reduce the error of BP estimation, the above research simply linearly adds, subtracts or cascades different features, and fails to effectively mine and utilize the nonlinear relationship of multi-modalities.

[0007] In summary, MPM and ML are affected by the variability of various physiological parameters, and it is difficult to model long-range BP signals and analyze patient-specificities; although multi-modal fusion and DL technology has shown better performance than other methods, the above methods fail to reconstruct continuous BP signals, and when processing multi-modal data, the game problem of insufficient complementary learning of individual information and difficult to eliminate redundancy of common information is faced. SUMMARY

[0008] To solve the above problems, the present disclosure provides a cuffless continuous blood pressure signal reconstruction method and system, which proposes an attention time sequence network (AT-Net) based on multi-modal fusion, and realizes the reconstruction of continuous BP signals by analyzing photoplethysmography (PPG) and electrocardiogram (ECG) signals.

[0009] According to some embodiments, the present disclosure adopts the following technical solutions:

[0010] A cuffless continuous blood pressure signal reconstruction method, comprising:

[0011] synchronously acquiring PPG signals and ECG signals, and performing standardization processing thereon;

[0012] extracting PPG feature vectors and ECG feature vectors from the PPG signals and the ECG signals respectively by using the trained attention time sequence network, performing multi-modal fusion on the PPG feature vectors and the ECG feature vectors, and reconstructing a cuffless continuous blood pressure signal based on the fused feature vectors;

[0013] wherein the feature vector extraction is to add an information screening machine and a time sequence learning machine between an encoder and a decoder, which are respectively used to learn a channel probability distribution vector and a time sequence weight vector, and guide the decoder to decode high-level semantic features and decode the time sequence relationship of PPG and ECG.

[0014] According to some embodiments, the present disclosure adopts the following technical solutions:

[0015] A cuffless continuous blood pressure signal reconstruction system, comprising an acquisition module and a reconstruction module:

[0016] The acquisition module is configured to synchronously acquire PPG signals and ECG signals, and perform standardization processing thereon;

[0017] The reconstruction module is configured to extract PPG feature vectors and ECG feature vectors from the PPG signals and the ECG signals respectively by using the trained attention time sequence network, perform multi-modal fusion on the PPG feature vectors and the ECG feature vectors, and reconstruct a cuffless continuous blood pressure signal based on the fused feature vectors;

[0018] Among them, the feature vector extraction is to add an information screening machine and a time sequence learning machine between the encoder and the decoder, which are respectively used for learning a channel probability distribution vector and a time sequence weight vector, and guiding the decoder to decode high-level semantic features and decode the time sequence relationship before and after PPG and ECG.

[0019] According to some embodiments, the present disclosure adopts the technical scheme as follows:

[0020] A computer program product comprising a computer program which, when executed by a processor, implements the method for reconstructing a cuffless continuous blood pressure signal.

[0021] According to some embodiments, the present disclosure adopts the technical scheme as follows:

[0022] A non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the method for reconstructing a cuffless continuous blood pressure signal.

[0023] According to some embodiments, the present disclosure adopts the technical scheme as follows:

[0024] An electronic device comprising a processor, a memory and a computer program; wherein the processor is connected with the memory, and the computer program is stored in the memory; when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device implements the method for reconstructing a cuffless continuous blood pressure signal.

[0025] Compared with the prior art, the present disclosure has the beneficial effects that:

[0026] Combined with DL and multi-modal fusion technology, the present application designs an attention time sequence network (AT-Net) with information screening and time sequence feature mining capability, and balances the redundancy and complementarity of PPG and ECG in multi-modal fusion, so as to realize high-precision continuous BP signal reconstruction, specifically:

[0027] (1) Multi-modal fusion: using multi-modal fusion method to learn the autocorrelation information and cross-correlation information of PPG and ECG respectively.

[0028] (2) Information screening machine: integrating multiple parallel attention mechanisms to focus on individual information and common information of features at different scales in the ventricular systole and diastole process to eliminate the redundancy of the two.

[0029] (3) Time sequence learning machine: based on the memory function of GRU to capture the dynamic transformation trend of PPG and ECG to improve the time sequence feature representation ability of AT-Net to the ventricular ejection period. BRIEF DESCRIPTION OF DRAWINGS

[0030] The accompanying drawings, which are incorporated in and constitute a part of this specification, are included to provide a further understanding of the disclosure, illustrate preferred embodiments of the disclosure, and to explain the principles of the disclosure.

[0031] Figure 1 Method flow chart of an embodiment of the disclosure.

[0032] Figure 2 Network structure diagram of an embodiment of the disclosure.

[0033] Figure 3 Flow chart of an information screening machine of an embodiment of the disclosure.

[0034] Figure 4 Flow chart of a memory unit in a timing learning machine of an embodiment of the disclosure.

[0035] Figure 5 Example diagram of a reconstructed continuous arterial BP waveform of an embodiment of the disclosure. DETAILED DESCRIPTION

[0036] The disclosure will be further described below with reference to the drawings and embodiments.

[0037] It should be noted that the following detailed description is illustrative only, and is intended to provide further description of the disclosure. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the disclosure belongs.

[0038] It should be noted that the terms used herein are only intended to describe specific embodiments, and are not intended to limit exemplary embodiments according to the disclosure. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should be further understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of a feature, step, operation, device, component, and / or combination thereof.

[0039] Embodiment 1

[0040] In an embodiment of the disclosure, a cuffless continuous blood pressure signal reconstruction method is provided, as shown in Figure 1 comprises:

[0041] Step 1, obtain synchronized PPG signals and ECG signals, and perform standardization processing thereon;

[0042] Step 2, use the trained attention timing network to extract PPG feature vectors and ECG feature vectors from the PPG signals and ECG signals, respectively, and perform multi-modal fusion on the PPG feature vectors and ECG feature vectors, and based on the fused feature vectors, reconstruct a cuffless continuous blood pressure signal.

[0043] Among them, the extraction of the feature vector is to add an information screening machine and a timing learning machine between the encoder and the decoder, which are respectively used to learn a channel probability distribution vector and a timing weight vector, so as to guide the decoder to decode high-level semantic features and decode the timing relationship between PPG and ECG.

[0044] As an embodiment, a sleeveless continuous blood pressure signal reconstruction method of the present disclosure is based on multi-modal fusion and proposes an attention timing network (AT-Net). The reconstruction of continuous BP signals is realized by analyzing photoplethysmography (PPG) and electrocardiogram (ECG) signals. The specific implementation process is as follows:

[0045] Step S1: Signal acquisition

[0046] The training data used in this embodiment is derived from the Multiparameter Intelligent Monitoring in Intensive Care (MIMIC) dataset, and the sampling rate of all signals is 125 Hz. In this embodiment, data with a recording duration greater than or equal to 8 minutes and no obvious noise is selected from the sorted dataset, and a total of 1400 subjects are obtained.

[0047] In clinical practice, some patients may have very small and very high BP values. The BP range of this embodiment is set to 50mmHg to 200mmHg. In addition, the synchronous PPG and ECG are subjected to Z-score standardization and are segmented into 1-second segments. All data are divided into training set: validation set: test set according to the ratio of 8:1:1. According to the Office blood pressure threshold, 1400 subjects are divided into 4 groups, and the division standard and subject proportion are shown in Table 1. The mean arterial pressure (MAP), diastolic blood pressure (DBP) and systolic blood pressure (SBP) are the average value, minimum value and maximum value of the BP segment, respectively.

[0048] Table 1. Division standard and subject proportion table

[0049]

[0050] The above data is used for network training. For the person to be detected in actual application, the PPG signal and ECG signal of the person to be detected are synchronously collected and standardized, and then input into the trained network for sleeveless continuous blood pressure signal reconstruction.

[0051] Step S2: Reconstructing continuous blood pressure signal using attention timing network

[0052] The attention time sequence network (AT-Net) based on multi-modal fusion in the embodiment adopts the framework of U-Net, and the overall structure of the network is as shown in Figure 2 The encoder, the information screening machine, the time sequence learning machine, the decoder, and the multi-modal fusion module are sequentially connected.

[0053] The encoder is composed of one-dimensional convolutions of three scales of 1x7, 1x5, and 1x3, and is used to extract the implicit information related to blood pressure in the PPG signal and the ECG signal; the multi-scale mechanism is used to analyze the amplitudes and gradients of different wave bands in the multi-modal signals to learn the complementarity between blood flow and cardiac electrical activity.

[0054] The information screening machine is based on the attention mechanism, and performs two excitation operations on the output of the encoder to learn the importance of each feature channel in the information mining process, and obtains a channel probability distribution vector for guiding the decoding of high-level semantic features by the decoder.

[0055] The time sequence learning machine is constructed based on 128 memory units of GRU, and captures the time transformation trend of long-time signals from the output of the encoder to obtain a time sequence weight vector for guiding the decoding of the time sequence relationship between PPG and ECG by the decoder.

[0056] The decoder is composed of deconvolutions of three scales of 1x3, 1x5, and 1x7, and is used to use the channel probability distribution vector and the time sequence weight vector to perform weighted decoding on the implicit information extracted by the encoder to obtain the feature vector related to BP in the PPG signal and the ECG signal.

[0057] It can be seen that, unlike the U-Net, the output of the c-th layer convolution in the embodiment is converted into a channel probability distribution vector S2 and a time sequence weight vector G after two independent encodings (i.e., the information screening machine and the time sequence learning machine), and the output of the 6-c-th deconvolution is weighted using the vectors S2 and G and is used as the input of the 6-c+1-th deconvolution.

[0058] Suppose that the output of the c-th layer convolution in the encoder is U c ={U c1 ,U c2 ,U c3 ,…,U cn}, the channel probability distribution vector S2 obtained after the learning of the information screening machine is S2={s 21 ,s 22 ,s 23 ,…,s 2n}, and the time sequence weight vector G obtained after the learning of the time sequence learning machine is G={g1,g2,g3,…,g n}. If the output of the 6-c-th deconvolution is U cT ={UcT1 ,U cT2 ,U cT3 ,…,U cTn}, then the input U of the 6-c+1 layer deconvolution ct =(S2⊕G)⊙U cT ={U ct1 ,U ct2 ,U ct3 ,…,U ctn}, where U cti =(s 2i +g i )×U cTi The feature vector after weighting by vectors S2 and G can reduce the redundancy of the encoder's multimodal information and preserve the temporal nature of PPG and ECG.

[0059] Using an encoder and a decoder for feature extraction is an existing technology. This embodiment is based on this technology and adds an information screening machine and a timing learning machine to learn the channel probability distribution vector and the timing weight vector respectively, guiding the decoder to decode high-level semantic features and the timing relationship between PPG and ECG, so as to solve the problem of redundancy and complementarity between PPG and ECG.

[0060] The information screening machine, time series learning machine, and multimodal fusion module are explained in detail below.

[0061] 1. Information screening machine

[0062] The implicit information extracted from different bands of PPG and ECG after multi-scale encoding contains information redundancy. Indiscriminately adding this information to the decoder will lead to a decline in model performance. To eliminate the redundancy of individual and common information at different scales during ventricular contraction and relaxation, this embodiment proposes an information filtering machine based on the attention mechanism. Its principle is shown in Algorithm 1 in Table 2:

[0063]

[0064]

[0065] Assume that is an arbitrary set of real numbers, and the input signal of AT-Net is X undergoes a nonlinear transformation in the encoder After that, we get multi-dimensional features

[0066] Information filtering machines, such as Figure 3 As shown, first, the compression operation F sq (·) will U cConvert to a vector with global representation characteristics To obtain the channel importance under the wide field of view; then, use the weight vector And the Relu function to perform the first excitation operation on Z Get the feature vector To increase the sparsity of AT-Net; second, in order to learn the importance of each feature channel in the information mining process and convert it to a weight vector, use the weight vector The second excitation operation is performed on the obtained feature vector S1 And pass it through the Softmax activation function to convert S1 into a channel probability distribution vector with a probability sum of 1 Where m, n, H and L are positive integers.

[0067] This embodiment uses the channel probability distribution vector S2 as a weight to guide the decoder to decode high-level semantic features.

[0068] 2, Time sequence learning machine

[0069] PPG, ECG and BP are all time sequence signals. When using neural networks to mine deep abstract features, preserving the time sequence relationship of the signals helps to decode the blood flow and cardiac electrical activity during the ventricular ejection period. Convolutional neural networks have strong spatial information learning ability, but their time sequence information learning ability is weak. This embodiment designs a time sequence learning machine based on GRU to mine the time transformation relationship of PPG and ECG. Based on a GRU with 128 memory cells, the long-time signal time transformation trend is captured from the output of the encoder to obtain a time sequence weight vector, which is used to guide the decoder to decode the time sequence relationship of PPG and ECG before and after. Its principle is shown in Algorithm 2 in Table 3:

[0070]

[0071]

[0072] Suppose, given any time [τ1,τ n ], the input signal After c times of nonlinear transformation in the encoder The feature vector extracted is

[0073] For a certain memory cell, as shown in Figure 4 The specific processing flow is as follows:

[0074] The input at the current time The corresponding feature is U tt The hidden state at the previous time is h t-1, first, the weight matrix W r For the matrix [h t-1 ,U tt ] Perform multi-dimensional linear transformation To learn h t-1 with U tt The temporal relationship between them. The weight matrix W z For the matrix [h t-1 ,U tt ] Perform linear transformations in high-dimensional spaces To obtain U tt Secondly, Sigmoid transforms the temporal relationship and long-term memory of PPG and ECG into weight values ​​r through nonlinear transformation. t and weight z t , r t With h t-1 Perform the Hadamard product to get h t-1 The remaining information. Then, the weight W has a t *h t-1 with U tt Perform nonlinear weighting and obtain candidate hidden states after a secondary nonlinear transformation using the Tanh function Finally, after selective forgetting (1-z t )*h t-1 and selective memory After that, the hidden state h at the next moment is obtained t and the output g of the current memory unit t , τ1, τ t and τ n is a positive real number, and a, b, c, d, and t are positive integers.

[0075] Using the output G of the memory unit t The timing weight vector G composed of guides the decoder to perform deep decoding on the temporal relationship between PPG and ECG.

[0076] 3. Multimodal Fusion

[0077] The mutual assistance of other physiological signals and PPG can improve the accuracy of BP estimation. Among various physiological signals, ECG can accurately reflect the changes in skin potential during blood flow. Algorithms based on PTT, PAT and PWV all use ECG as an auxiliary signal, which shows that there are features in ECG that are closely related to BP.

[0078] This embodiment uses an encoder, an information filter, a time series learning machine, and a decoder to perform feature extraction, which actually maps the source domains of PPG and ECG to their respective feature domains. Specifically,

[0079] The synchronous PPG and ECG are taken as inputs of the AT-Net, and continuous BP signal end-to-end modeling is realized according to different source domains. If an arbitrary source domain is given , the PPG segment is , and the ECG segment is An encoder is used to map the source domain to a high-dimensional space , so as to obtain two groups of high-dimensional feature vectors and Then, a decoder is used to project the high-dimensional space to the feature domain and the feature domain , so as to obtain two groups of final feature vectors and

[0080] After obtaining the two groups of final feature vectors, the PPG feature vector y ppg and the ECG feature vector y ecg have corresponding complementary relationships, but also have a certain degree of information redundancy. In order to realize the information coupling of the two groups of feature vectors, the PPG feature vector y ppg and the ECG feature vector y ecg are modeled again, so as to uniformly map the feature domain and to the BP target domain .

[0081] It is assumed that the BP (i.e. the real blood pressure signal) in True is BP True1 = {BP True2 , BP True3 , BP Truen , …, BP ppg}, the modeled BP (i.e. the output signal obtained by taking PPG as input) in PPg is BP PPG1 = {BP PPG2 , BP PPG3 , BP PPGn , …, BP ecg}, the modeled BP (i.e. the output signal obtained by taking ECG as input) in ECG is BP ECG1 = {BP ECG2 , BP ECG3 , BP ECGn , …, BP PPGi}, and any (BP ECGi + BPError relationship between BP Truei and BP PPGi is shown in Table 4:

[0082] Table 4 Error situation of BP PPGi , BP ECGi and BP Truei

[0083]

[0084]

[0085] Clinical data is unlabeled data, so it is impossible to select the estimated value with the minimum error from BP PPGi and BP ECGi as the clinical prediction result. If BP PPGi and BP ECGi are simply added and averaged, there is a 2 / 3 probability that the estimated value will deviate from the true value.

[0086] In order to avoid the phenomenon of deviation , the embodiment fuses BP and BP by relearning to obtain the transition domain BP , and then maps BP to BP . The process is shown in formulas (1-2):

[0087]

[0088]

[0089] Specifically, assuming that the error between the feature domain BP and the target domain BP is γ ppg , and the error between the feature domain BP and the target domain BP is γ ecg , as shown in formulas (3-4):

[0090]

[0091]

[0092] The embodiment dynamically adjusts BP PPG and BP ECG , uses nonlinear transformation means to make | γ ppg + γ ecg | reach the minimum value, and the process of nonlinear transformation is shown in formulas (5-8):

[0093]

[0094]

[0095]

[0096] BP PRE =U{BP PRE1 ,BP PRE2 ,BP PRE3 ,…,BP PREn} (8)

[0097] First, using the weight matrix and bias BP 22G With BP ECG Each BP value in is weighted to achieve This process is based on the back propagation of the loss function to achieve BP PPGi With BP ECGi Dynamic adjustment of , that is, using low weights to suppress estimates with large errors. Then, using the weight matrix and bias right BP vector BP in pre Perform regression calculation to obtain BP True Closest BP PRE , as the reconstructed cuff-free continuous blood pressure signal; among them, and are all 1×n vectors.

[0098] In order to evaluate the effect of this method, the mean absolute error (MAE) and root mean square error (RMSE) are used as evaluation indicators. The calculation process of MAE and RMSE is shown in formulas (9) and (10).

[0099]

[0100]

[0101] This example also uses the British Hypertension Society (BHS) standard and the Association for the Advancement of Medical Instrumentation (AAMI) standard to evaluate the clinical application value of this method.

[0102] This embodiment reconstructs BP signals based on continuous PPG and ECG. In order to observe the long-range tracking capability of the algorithm, a record is randomly selected from each type of subject as a visualization sample. The reconstruction effect of BP signal is shown in Figure 2. Figure 5The results show that the AT-Net of the embodiment can accurately reconstruct the BP signal according to the PPG and the ECG.

[0103] In the ventricular ejection period, the rapid ejection velocity causes the slope and amplitude of the rising branch to rapidly increase in a short time, and the slow ejection velocity causes the slopes of the front branch and the rear branch of the falling branch to slowly decrease. The method of the embodiment has good fitting effects on the rapid increase and slow decrease of the amplitude and the slope. Especially for the dicrotic wave (RW) in the BP signal, which has small jitter and is easy to form an inflection point, the method of the embodiment has good fitting effect and can accurately reconstruct the inflection point formed by the RW. The reconstruction effect of the RW is as shown in FIG. 6. Figure 5 As shown in FIG. 6, the method of the embodiment can better fit the RW, and can accurately reconstruct the inflection point formed by the RW.

[0104] The embodiment collects part of the existing technologies about BP estimation between 2019 and 2024, and compares the results of the embodiment, and the comparison results are as shown in Table 5. The MAE and RMSE of the mean arterial pressure (MAP) in the embodiment are 1.24 mmHg and 2.19 mmHg respectively, the MAE and RMSE of the systolic blood pressure (SBP) are 2.94 mmHg and 4.49 mmHg respectively, and the MAE and RMSE of the diastolic blood pressure (DBP) are 1.65 mmHg and 2.77 mmHg respectively. Through comparison, it is found that part of the indexes are only second to the scheme proposed by Hajj and Zhang. However, the scheme proposed by Hajj only selects 500 groups of waveforms, and the scheme proposed by Zhang only tests 25 healthy subjects.

[0105] In contrast, the embodiment uses 1400 subjects, and the BP interval is [50mmHg, 200mmHg], that is, the test data of the embodiment is more diversified, and the test result has higher reliability. In addition, the task of the embodiment is not simply the detection of SBP and DBP, but the reconstruction of continuous BP signal, that is, the embodiment is more in line with the requirements of continuous, long-range and dynamic monitoring.

[0106] Table 5 Method Comparison

[0107]

[0108]

[0109] In summary, the embodiment adopts a multi-modal fusion strategy and designs a new AT-Net to realize the reconstruction of continuous BP signals; the encoder of the AT-Net extracts the implicit information related to BP in PPG and ECG, the information screening machine solves the information redundancy problem in the feature extraction process, the time sequence learning machine captures the time transformation trend of the long-time signal, and the decoder accurately decodes the semantic vector related to BP in PPG and ECG; compared with existing algorithms, the embodiment realizes the task conversion from estimating discrete BP values to reconstructing continuous BP signals, and has lower error and stronger reliability.

[0110] Embodiment 2

[0111] In an embodiment of the present disclosure, a cuffless continuous blood pressure signal reconstruction system is provided, comprising an acquisition module and a reconstruction module:

[0112] The acquisition module is configured to acquire synchronized PPG signals and ECG signals and perform standardization processing thereon.

[0113] The reconstruction module is configured to extract PPG feature vectors and ECG feature vectors from the PPG signals and the ECG signals respectively by using a trained attention time sequence network, perform multi-modal fusion on the PPG feature vectors and the ECG feature vectors, and reconstruct a cuffless continuous blood pressure signal based on the fused feature vectors.

[0114] Among them, the feature vectors are extracted by adding an information screening machine and a time sequence learning machine between the encoder and the decoder, which are respectively used to learn a channel probability distribution vector and a time sequence weight vector to guide the decoder to decode high-level semantic features and decode the time sequence relationship before and after PPG and ECG.

[0115] Embodiment 3

[0116] In an embodiment of the present disclosure, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the cuffless continuous blood pressure signal reconstruction method.

[0117] Embodiment 4

[0118] In an embodiment of the present disclosure, a non-transitory computer readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement the cuffless continuous blood pressure signal reconstruction method.

[0119] Embodiment 5

[0120] An electronic device is provided in one embodiment of the present disclosure, comprising: a processor, a memory and a computer program; wherein the processor is connected with the memory, and the computer program is stored in the memory; when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the method for realizing the continuous blood pressure signal reconstruction without cuff.

[0121] The present disclosure is described with reference to the flowcharts and / or block diagrams of the method, device (system) and computer program product according to the embodiments of the present disclosure. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device that implements the function specified in the flowcharts and / or block diagrams. Figure 1 The function specified in one flow or multiple flows and / or blocks. Figure 1 The function specified in one block or multiple blocks.

[0122] These computer program instructions can also be loaded into a computer or other programmable data processing device to cause a series of operation steps to be executed on the computer or other programmable data processing device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable data processing device provide a process for implementing the function specified in the flowcharts and / or block diagrams. Figure 1 The function specified in one flow or multiple flows and / or blocks. Figure 1 The function specified in one block or multiple blocks.

[0123] Although the specific embodiments of the present disclosure are described above with reference to the accompanying drawings, the present disclosure is not limited to the above embodiments, and various modifications or changes can be made by those skilled in the art without departing from the technical solutions of the present disclosure.

Claims

1. A cuffless continuous blood pressure signal reconstruction method, characterized in that, The method comprises the following steps: acquiring synchronous PPG signals and ECG signals and performing standardization processing on the signals; extracting PPG feature vectors and ECG feature vectors from the PPG signals and the ECG signals respectively by using a trained attention time sequence network, performing multi-modal fusion on the PPG feature vectors and the ECG feature vectors, and reconstructing a cuffless continuous blood pressure signal based on the fused feature vectors; wherein the feature vector extraction is performed by adding an information screening machine and a time sequence learning machine between an encoder and a decoder, which are respectively used for learning a channel probability distribution vector and a time sequence weight vector to guide the decoder to decode high-level semantic features and the time sequence relationship between PPG and ECG; the information screening machine learns the importance of each feature channel in the information mining process by performing twice excitation operation on the output of the encoder based on an attention mechanism, and obtains the channel probability distribution vector; the time sequence learning machine captures the time transformation trend of a long-time signal from the output of the encoder based on a GRU, and obtains the time sequence weight vector.

2. The cuffless continuous blood pressure signal reconstruction method of claim 1, wherein, The attention time sequence network comprises an encoder, an information screening machine, a time sequence learning machine, a decoder and a multi-modal fusion module connected in sequence.

3. The cuffless continuous blood pressure signal reconstruction method of claim 2, wherein, The encoder is composed of multi-scale one-dimensional convolution and is used for extracting implicit information related to blood pressure in the PPG signals and the ECG signals; the decoder is composed of multi-scale one-dimensional deconvolution and is used for weighting decoding the implicit information extracted by the encoder by using the channel probability distribution vector and the time sequence weight vector, so as to obtain feature vectors related to BP in the PPG signals and the ECG signals.

4. The cuffless continuous blood pressure signal reconstruction method of claim 1, wherein, The multi-modal fusion is performed by learning the minimum error between the feature domain and the target domain, using the learned weight matrix and bias to map the PPG feature vectors and the ECG feature vectors from their respective feature domains to the target domain of blood pressure, and obtaining the fused feature vectors.

5. A cuffless continuous blood pressure signal reconstruction system, characterized by, The method comprises an acquisition module and a reconstruction module: the acquisition module is configured to acquire synchronous PPG signals and ECG signals and perform standardization processing on the signals; the reconstruction module is configured to extract PPG feature vectors and ECG feature vectors from the PPG signals and the ECG signals respectively by using a trained attention time sequence network, perform multi-modal fusion on the PPG feature vectors and the ECG feature vectors, and reconstruct a cuffless continuous blood pressure signal based on the fused feature vectors; wherein the feature vector extraction is performed by adding an information screening machine and a time sequence learning machine between an encoder and a decoder, which are respectively used for learning a channel probability distribution vector and a time sequence weight vector to guide the decoder to decode high-level semantic features and the time sequence relationship between PPG and ECG; the information screening machine learns the importance of each feature channel in the information mining process by performing twice excitation operation on the output of the encoder based on an attention mechanism, and obtains the channel probability distribution vector; the time sequence learning machine captures the time transformation trend of a long-time signal from the output of the encoder based on a GRU, and obtains the time sequence weight vector.

6. A computer program product comprising a computer program, characterized in that, The computer program is executed by a processor to implement the method for reconstructing a cuffless continuous blood pressure signal according to any one of claims 1-4.

7. A non-transitory computer-readable storage medium, comprising: The non-transitory computer readable storage medium is configured to store computer instructions, and the computer instructions are configured to be executed by a processor to implement the method of claim 1-4.

8. An electronic device, comprising: The method comprises: A processor, a memory and a computer program, wherein the processor is connected with the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to implement the method of claim 1-4.

Images