Blood pressure estimation method and system using multi-scale neural network as binary tree node

Abstract

The application belongs to the technical field of blood pressure measurement, and discloses a blood pressure estimation method and system with a multi-scale neural network as a binary tree node, which comprises the following steps: projecting an obtained photoelectric volume pulse signal and an electrocardio signal to a high-resolution space respectively to obtain a first output; performing multi-stage decomposition on the first output to obtain a blood pressure feature matrix; performing global dimension reduction compression and local high-dimensional mapping on the blood pressure feature vector, and weighting the results to the blood pressure feature vector to obtain a multi-modal blood pressure feature set; performing decoding operation on the multi-modal blood pressure feature, learning the nonlinear coupling relationship of the multi-modal blood pressure feature, and obtaining a blood pressure estimation value; and defining a target function to optimize blood pressure estimation model parameters. The application combines the low computational cost reasoning mode of the tree algorithm with the multi-modal feature extraction capability of the multi-scale and multi-resolution neural network, realizes the light weight of the algorithm under the premise of high-precision continuous blood pressure estimation, and the calculation cost meets the requirements of wearable devices.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of blood pressure measurement, in particular to a blood pressure estimation method and system taking a multi-scale neural network as a binary tree node. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] Hypertension is one of the main causes of cardiovascular and cerebrovascular diseases. The cuff-type electronic sphygmomanometer in clinical practice is affected by external factors, behavioral factors and psychological factors, resulting in high concealment of missed diagnosis of hypertension and misdiagnosis of false hypertension. The mainstream medical association recommends continuous, dynamic and long-term monitoring of BP based on wearable devices.

[0004] The wearable device requires the computational complexity of the BP estimation algorithm to be below a certain threshold and to ensure a low error. Existing BP estimation algorithms are mainly based on empirical or hypothetical mathematical physical modeling (MPM), feature engineering-based machine learning (ML) and end-to-end deep learning (DL). MPM and ML methods usually have lower computational complexity, but have larger errors in BP estimation and dynamic tracking of a wide range of subjects. The DL method can construct an arbitrary function mapping relationship by automatically analyzing the distribution structure of big data, but the huge computational complexity of DL exceeds the computing capacity of the wearable device. Under the premise of high-precision continuous BP estimation, it is an important work to compress the computational cost of the algorithm within the computing range of the wearable device and realize hardware deployment. SUMMARY

[0005] To solve the above problems, the present application proposes a blood pressure estimation method and system taking a multi-scale neural network as a binary tree node, which combines the low computational cost reasoning method of tree algorithm with the multi-modal feature extraction capability of multi-scale and multi-resolution neural network, realizes the lightweight optimization of the algorithm under the premise of high-precision continuous BP estimation, and the computational cost meets the requirements of the wearable device.

[0006] To achieve the above purpose, the present application adopts the following technical solutions:

[0007] In a first aspect, the present application provides a blood pressure estimation method taking a multi-scale neural network as a binary tree node, comprising the following steps:

[0008] Obtain the photoplethysmographic signal and the electrocardiogram signal, project the photoplethysmographic signal and the electrocardiogram signal into a high-resolution space respectively to obtain a first output, and perform multi-level decomposition on the first output by using multi-scale convolution calculation to obtain a blood pressure feature matrix;

[0009] The global dimension reduction compression and the local high-dimensional mapping are performed on each group of blood pressure feature vectors in the blood pressure feature matrix, and the results of the global dimension reduction compression and the local high-dimensional mapping are weighted on the blood pressure feature vectors to obtain a multi-modal blood pressure feature set;

[0010] The multi-modal blood pressure feature is subjected to a decoding operation, and a nonlinear coupling relationship of the multi-modal blood pressure feature is learned to obtain a blood pressure estimation value.

[0011] A target function is defined, and a blood pressure estimation model parameter taking a multi-scale neural network as a binary tree node is optimized.

[0012] As an optional implementation, the global dimension reduction compression is performed on each group of blood pressure feature vectors in the blood pressure feature matrix, specifically, two different compressions are performed on each group of blood pressure feature vectors in the blood pressure feature matrix, the two different compression results are vertically spliced to obtain a two-dimensional matrix, convolution calculation is performed on the two-dimensional matrix to obtain a feature vector, an attention mechanism is used to focus on important information related to blood pressure in the feature vector with a high weight and suppress redundant information in the feature vector with a low weight, and the global dimension reduction compression of the blood pressure feature vector is realized.

[0013] As an optional implementation, the local high-dimensional mapping is performed on each group of blood pressure feature vectors in the blood pressure feature matrix, specifically, the blood pressure feature vectors are subjected to channel-by-channel accumulation and element-by-element multiplication operations to obtain channel feature vectors and spatial feature vectors, a plurality of groups of convolutions are used to map the channel feature vectors and the spatial feature vectors to high-dimensional spaces respectively to obtain high-dimensional channel vectors and high-dimensional spatial vectors, and an attention mechanism is used to weight and fuse the high-dimensional channel vectors and the high-dimensional spatial vectors to realize the local high-dimensional mapping of the blood pressure feature vectors.

[0014] As an optional implementation, a multi-head decoder is used to perform the decoding operation on the multi-modal blood pressure feature, and the multi-head decoder is composed of a head set and a trunk network.

[0015] As an optional implementation, the head set projects the feature vectors of the photoplethysmography signal and the electrocardiogram signal to a target domain and The trunk network uniformly maps and to a BP domain

[0016] As an optional implementation, the root network of the binary tree neural network projects the photoplethysmography signal and the electrocardiogram signal to a high-resolution space by using a split convolution and channel fusion, and the subnetwork of the binary tree neural network decomposes the high-resolution space into a plurality of low-resolution subspaces by using a multi-scale mechanism.

[0017] In a second aspect, the present application provides a blood pressure estimation system with a multi-scale neural network as a binary tree node, comprising:

[0018] A binary tree neural network module is configured to: acquire a photoplethysmogram signal and an electrocardiogram signal, project the photoplethysmogram signal and the electrocardiogram signal into a high-resolution space respectively, obtain a first output, perform multi-level decomposition on the first output by using multi-scale convolution calculation, and obtain a blood pressure feature matrix;

[0019] A multi-modal information transformation module is configured to: perform global dimension reduction compression and local high-dimensional mapping on each group of blood pressure feature vectors in the blood pressure feature matrix, and weight the results of the global dimension reduction compression and the local high-dimensional mapping to the blood pressure feature vectors to obtain a multi-modal blood pressure feature set;

[0020] A multi-head decoder module is configured to: perform decoding operation on the multi-modal blood pressure features, learn the nonlinear coupling relationship of the multi-modal blood pressure features, and obtain a blood pressure estimation value.

[0021] A model training module is configured to: define a target function, and optimize the blood pressure estimation model parameter with the multi-scale neural network as the binary tree node.

[0022] In a third aspect, the present application provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and running on the processor, when the computer instructions are run by the processor, the method of the first aspect is completed.

[0023] In a fourth aspect, the present application provides a computer readable storage medium for storing computer instructions, when the computer instructions are executed by a processor, the method of the first aspect is completed.

[0024] In a fifth aspect, the present application provides a computer program product comprising a computer program, when the computer program is executed by a processor, the method of the first aspect is completed.

[0025] Compared with the prior art, the present application has the following beneficial effects:

[0026] The present disclosure proposes a blood pressure estimation method and system with a multi-scale neural network as a binary tree node, taking a binary tree neural network (Bi-Tree-Net) as a framework, and taking a multi-modal information transformation module (MITM) and a multi-head decoder (MHDE) as an auxiliary to realize a lightweight end-to-end blood pressure estimation. The binary tree neural network combines the tree algorithm of machine learning and the multi-scale and multi-resolution technology of deep learning to learn the morphology of low-amplitude waves such as reflected waves, descending sinus, and heavy beats, and the gradient change trend of systole and diastole in a low-computing-cost manner. Under the premise of high-precision continuous blood pressure estimation, the algorithm is lightweight, and the computing cost meets the requirements of wearable devices.

[0027] The multi-modal information transformation module globally averages and high-dimensionally maps the vectors encoded by the binary tree neural network to reduce the temporal redundancy between the continuous transformation trend and low-amplitude waveform morphology of blood pressure and the sample redundancy between the photoplethysmogram and the electrocardiogram.

[0028] The multi-head decoder module parallel decodes the heterogeneous blood pressure information between the photoplethysmogram and the electrocardiogram to enhance the feature representation capability of the binary tree neural network to the peaks and troughs of the photoplethysmogram and the QRS segment and TP segment of the electrocardiogram.

[0029] Advantages of the additional aspects of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0030] The accompanying drawings, which constitute a part of this specification, are included to provide a further understanding of the present application, and are incorporated by reference in this specification. The explanations of the illustrative embodiments of the present application and its drawings serve to explain the present application. They should not be considered as unnecessarily limiting the present application.

[0031] Figure 1 A flowchart of a blood pressure estimation method taking a multi-scale neural network as a binary tree node provided for Embodiment 1 of the present application;

[0032] Figure 2 A structure diagram of a binary tree neural network provided for Embodiment 1 of the present application;

[0033] Figure 3 A structure diagram of a multi-modal information transformation module provided for Embodiment 1 of the present application;

[0034] Figure 4 An experimental flowchart of a blood pressure estimation method taking a multi-scale neural network as a binary tree node provided for Embodiment 1 of the present application;

[0035] Figure 5 A Bland-Altman graph and a Correlation graph of a blood pressure estimation method taking a multi-scale neural network as a binary tree node provided for Embodiment 1 of the present application;

[0036] Figure 6 MAE and RMSE results of a blood pressure estimation method taking a multi-scale neural network as a binary tree node provided for Embodiment 1 of the present application in N-class, P-class and S-class subjects;

[0037] Figure 7 A comparison graph of true values and predicted values of a blood pressure estimation method taking a multi-scale neural network as a binary tree node provided for Embodiment 1 of the present application in a 12334th subject of P-class;

[0038] Figure 8 MAE distribution of MAP, SBP and DBP of the present application;

[0039] Figure 9 Feature heat map of PPG and ECG before and after MITM weighting of the present application. DETAILED DESCRIPTION

[0040] The present application will be further described below in conjunction with the accompanying drawings and embodiments.

[0041] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. 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 this application belongs.

[0042] It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting, as the scope of the application will be limited only by the appended claims. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. Unless otherwise required by context, singular terms shall include pluralities and vice versa. Plural elements can be separated by a hyphenated form and can contain hyphenated forms.

[0043] The embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0044] Embodiment 1

[0045] As shown in the figure, the present embodiment provides a blood pressure estimation method taking a multi-scale neural network as a binary tree node, comprising the following steps: Figure 1 S1 obtains a photoplethysmography signal and an electrocardiogram signal, projects the photoplethysmography signal and the electrocardiogram signal into a high-resolution space respectively, obtains a first output, performs multi-level decomposition on the first output by using a multi-scale convolution calculation, and obtains a blood pressure feature matrix;

[0046] S2 performs global dimensionality reduction compression and local high-dimensional mapping on each group of blood pressure feature vectors in the blood pressure feature matrix, weights the results of the global dimensionality reduction compression and the local high-dimensional mapping to the blood pressure feature vectors, and obtains a multi-modal blood pressure feature set;

[0047] S3 performs decoding operation on the multi-modal blood pressure features, learns the nonlinear coupling relationship of the multi-modal blood pressure features, and obtains a blood pressure estimation value;

[0048]

[0049] ​S4 defines the objective function, and optimizes the blood pressure estimation model parameter taking the multi-scale neural network as the binary tree node.

[0050] The multi-scale mechanism and multi-resolution technology can effectively obtain the local morphology of low-amplitude waves such as reflected waves, descending isthmus, and double-pulsed waves, and the global gradient change trend in systole and diastole. However, the implementation of multi-scale and multi-resolution usually relies on multi-path and deep neural networks, which leads to a significant increase in the computational cost of the algorithm. The essence of the tree algorithm of machine learning (ML) is to decompose a complex original problem into multiple sub-problems, and to realize efficient reasoning by simultaneously processing multiple sub-problems. In order to obtain the local morphology and global rhythm of blood pressure (BP) from photoplethysmography (PPG) and electrocardiogram (ECG) in a low computational cost manner, the present disclosure designs a binary tree neural network (Bi-Tree-Net) with a binary tree algorithm as the basic architecture and a multi-scale and multi-resolution neural network as the node of the tree, as shown in the structure of Figure 2 Bi-Tree-Net decomposes the BP estimation task on the nodes of the multi-scale and multi-resolution neural network for parallel processing. Multiple nodes use different receptive fields for non-linear transformation to extract multi-modal blood pressure features.

[0051] Assume, is an arbitrary set of real numbers, and the PPG and ECG input into Bi-Tree-Net are and The output BP value is {MAP, DBP, SBP}. The BP estimation task is equivalent to that is, the essence of the input to the output is a high-resolution mapping to a low-resolution process. In Bi-Tree-Net, the present disclosure sets the process of decomposing the root network into sub-networks as the norm of high-resolution mapping to low-resolution according to the nature of the task.

[0052] Specifically, the root network of Bi-Tree-Net projects PPG and ECG into a high-resolution space, and the sub-networks use a multi-scale mechanism to decompose the high-resolution space into multiple low-resolution subspaces. If the depth of Bi-Tree-Net is then Bi-Tree-Net is composed of one root network and sub-networks, where the depth contains sub-networks. For any input after mapping by the root network, the output tree out1 has a resolution of f. After mapping by level sub-networks, the output is obtained a set of outputs with a resolution of .

[0053] Root network utilizes separated convolution and channel fusion to project X e (X ppg , ecg ) to high resolution space as formula (1) shows,

[0054]

[0055] where k1 and k2 are the scale of convolution kernel (k1≠k2). The dimension change of this process is

[0056] Sub-networks employ multi-scale convolution computation to decompose tree out1 in a hierarchical manner to achieve high resolution to multiple groups of low resolution mapping. The depth of any sub-network decomposition process is shown in formula (2-3),

[0057]

[0058] k s1 and k s2 are the scale of convolution kernel (k s1 ≠k s2 ), is any output of depth . The resolution change of this decomposition process is Finally, after level nonlinear mapping and decomposition, the output of Bi-Tree-Net is li is the length of the i-th BP feature vector. represents channel-wise convolution, and * represents convolution multiplication.

[0059] According to different scales and resolution encoded BP feature vectors contain the local morphological changes of aortic and mitral valve opening and closing and the global gradient changes of ascending and descending branches, and different BP feature vectors have complementary relationship. However, the global gradient change contains local morphological changes and any two sub-networks in Bi-Tree-Net have the same source domain, which leads to information redundancy problem. In order to overcome the temporal redundancy between the global trend and local morphology of BP and the sample redundancy between PPG and ECG, the present disclosure proposes a multi-modal information transformation module (MITM). MITM performs global dimension reduction compression and local high-dimensional mapping on BP feature vectors to achieve feature selection of different BP vectors, and the structure of MITM is shown in Figure 3 .

[0060] Global dimension reduction compression: first, any group of BP feature vectors out iF gap (·) and F gmp (·) two different compressions. As shown in equations (4-5),

[0061]

[0062] Then, z gap and z gmp are concatenated along the vertical direction to obtain a two-dimensional matrix z gap-gmp and point-to-point fusion of z gap and z gmp is achieved by convolution calculation to obtain the feature vector z. As shown in equations (6-7),

[0063]

[0064] Finally, the attention mechanism F attention1 (·) is used to focus on important information related to BP in z with high weight and suppress redundant information in z with low weight to obtain the global feature vector of BP As shown in equation (8), where σ(·) represents the Sigmiod function and δ(·) represents the Relu function.

[0065]

[0066] Local high-dimensional mapping: First, out i is accumulated channel by channel to obtain Γ add and element-by-element multiplication of Γ mul operation to obtain the channel feature vector and the spatial feature vector As shown in equations (9-10),

[0067]

[0068] Then, t groups of convolutions are used to map Ch and Sp to high-dimensional spaces respectively to obtain high-dimensional channel vector CH and spatial vector SP. As shown in equations (11-12),

[0069]

[0070] Finally, the attention mechanism F attention2 (·) is used to weight and fuse CH and SP to obtain the local feature vector of BP As shown in equation (13),

[0071]

[0072] After completing the global dimension reduction compression and local high-dimensional mapping, out and are weighted to outi get multimodal BP feature vector set i . As formula (14),

[0073]

[0074] Out tree for each group of BP feature vectors, the above operation is repeated to get the multimodal BP feature set represent the sum of elements, and ⊙ represents the multiplication of elements, matrix multiplication.

[0075] Simple linear addition or concatenation operation on the multimodal BP feature set of MITM will cause dimension disaster, multiple comparisons and loss of nonlinear relationship. The present disclosure proposes a multi-head decoder (MHDE) composed of a head set and a trunk network to enhance the feature representation ability of the wave crest and trough of PPG and the QRS segment and TP segment of ECG and realize the learning of nonlinear coupling relationship in multimodal BP features.

[0076] Suppose the head set of MHDE is head i The decoding process of set i is as formula (15-17), wherein is deconvolution, is the weight matrix, k inv is the size of the deconvolution kernel, and θ(*) is the ELU activation function.

[0077]

[0078] After repeating the above operation, set MBPF of any input source is projected to the target domain Suppose the head set projects the feature vectors of PPG and ECG to the target domains and respectively, and the trunk network W bn is to map and to the BP domain The process is as formula (18).

[0079]

[0080] If the error between and is loss ppg , and is loss ecg, then the target of the backbone network is formula (19).

[0081]

[0082] The backbone network performs nonlinear transformation on the feature vectors of PPG and ECG, making the loss ppg |+|loss ecg | Reached minimum value.

[0083] First, using the weight matrix right Each eigenvalue in is weighted, as shown in formula (20-21),

[0084]

[0085] Then, using the weight matrix BP st Perform regression calculation to obtain the estimated value BP out , as shown in formula (22-23),

[0086]

[0087] Among them, v and y are the dimensions of the output respectively.

[0088] Experimental verification:

[0089] The experimental process is as follows Figure 4 This paper obtains data from the Medical Information Mart for Intensive Care (MIMIC) database for various types of subjects. After preprocessing, this paper uses Bi-Tree-Net, MITM, and MHDE to perform feature encoding, selection, and decoding of PPG and ECG, and compares the results with international standards and leading research.

[0090] MIMIC includes synchronized PPG, ECG, and BP signals, all sampled at 125Hz. To facilitate training and testing of BP estimation tasks, the MIMIC signals have been collated and stored in the University of California, Irvine (UCI) Machine Learning Repository. The data disclosed herein is sourced from the UCI Machine Learning Repository, and each record is fully documented.

[0091] This paper divides the subjects into three categories according to the Office blood pressure threshold: Normal (N), Prehypertension (P), and Stage 1 Hypertension (S). 100 subjects were randomly selected from each category, and the duration of all records was greater than or equal to 8 minutes. This paper uses a sliding window with a window length of 1 second and a step size of 1 second to segment the records. 167,432 segments are obtained for PPG, ECG, and BP, respectively. MAP, DBP, and SBP are defined as the mean, minimum, and maximum values ​​of each BP segment, respectively. Max-Min Normalization is performed on the PPG and ECG segments.

[0092] This study used a cross-patient model to divide the subjects into a training set (80%), a validation set (10%), and a test set (10%). All experiments employed 10-fold cross-validation. The training set consisted of three categories of subjects: N, P, and S, rather than training on specific subject types. The experimental environment was Python 3.9, Keras 2.7.0, and TensorFlow 2.7.0.

[0093] To determine the important hyperparameters of Bi-Tree-Net, this paper explores the following ranges: depth The range of is {3,4,5}, the range of resolution f is {64,128,256}, and finally determined The performance of f=128 is the best. The sizes of convolution kernels in Bi-Tree-Net are set according to the default rules of DL: k1=7, k2=1, k s1 =3 and k s2 = 5. t in MITM is set to 8 according to the lightweight rule. The output dimensions of MHDE are set to: v = 128, y = 125.

[0094] SBP is the difficulty of BP estimation task. This paper designs BP focal loss function to improve the sensitivity of the algorithm to SBP, such as formula (24-27),

[0095] loss = α × loss map +β×loss sbp +γ×loss dbp (twenty four)

[0096]

[0097] BP truei is the true value of BP. Regarding the selection of α, β and γ, the present disclosure explores the following combinations: {0.3, 0.4, 0.3}, {0.25, 0.5, 0.25}, {0.2, 0.6, 0.2}. Finally, it is determined that when a, b and g are set to {0.25, 0.5, 0.25} respectively, the loss is the smallest.

[0098] The present disclosure uses mean absolute error (MAE) and root mean square error (RMSE) as the evaluation index of BP estimation results, and uses parameter quantity (Par) and calculation quantity (FLOPs) as the lightweight evaluation index. The calculation process of MAE and RMSE is shown in formulas (28) and (29). In addition, the present disclosure also uses the standard of British Hypertension Society (BHS) and the standard of Association for the Advancement of Medical Instrumentation (AAMI) to evaluate the clinical application value of the algorithm.

[0099]

[0100] The Bland-Altman graph and the Correlation graph of the present disclosure are shown in Figure 5 It is concluded from the Bland-Altman graph that the mean error (ME) of MAP is -0.73 mmHg, and the 95% consistency limit is [-10.64, 10.07] mmHg. The ME of SBP is -0.03 mmHg, and the 95% consistency limit is [-11.33, 11.40] mmHg. The ME of DBP is -0.93 mmHg, and the 95% consistency limit is [-9.83, 7.97] mmHg. This shows that the estimated BP of the present disclosure has high consistency with the true BP.

[0101] Figure 5 (d) represents the Correlation graph of 300 subjects. The Pearson Correlation Coefficient (PCC) of the estimated MAP, SBP and DBP with the true value is 0.8027, 0.9464 and 0.5363 respectively, and the p-value is less than 10 -5 This shows that the estimated BP of the present disclosure has high correlation with the true BP, and the result has statistical significance. The ME of SBP is the lowest and the PCC is the highest, which shows that the BP focal loss function of the present disclosure has good effect.

[0102] The results of the algorithm of the present disclosure in N, P and S subjects are shown in Figure 6The distribution intervals of MAE and RMSE are [3.13, 4.84] mmHg and [4.57, 6.66] mmHg, respectively. This shows that the algorithm of the present disclosure can maintain accurate estimation effect when facing diversified subjects in the clinical environment, and is less affected by subject specificity.

[0103] Compared with N-type subjects, the error of the algorithm in P-type and S-type subjects increased slightly. The MAE and RMSE of MAP increased by 0.29 mmHg and 0.48 mmHg, respectively. The MAE and RMSE of SBP increased by 0.78 mmHg and 1.04 mmHg, respectively. The MAE and RMSE of DBP increased by 0.23 mmHg and 0.28 mmHg, respectively. The error of P-type subjects is the highest, and the predicted value is usually greater than the true value. This is because P-type is between N-type and S-type, and some subjects have SBP belonging to P-type and DBP belonging to N-type.

[0104] The present disclosure uses BHS standards and AAMI standards to evaluate the performance of the algorithm, and the results show that the performance of the present disclosure in the nine BP estimation tasks has reached the A level of the BHS standard, and has passed the AAMI test in N, P and S three types of subjects.

[0105] Comparison of the present disclosure with the results of advanced research:

[0106] (1) Error comparison: Mohebbian et al., Panwa et al. and Qiu et al. use patient-in and sample overlap preprocessing methods, and the test process has information leakage problem. The present disclosure is patient-out and non-overlapping sample division method, and the test result is more in line with the clinical standard. The MAE of DBP of Hajj et al. is 2.60 mmHg, but the MAE of SBP is 4.61 mmHg, and the standard deviation (STD) of the error is 7.81 mmHg. The STD of the present disclosure is 3.54 mmHg, which shows that the stability of the present disclosure is better than that of Hajj et al. Ali et al. tested 40 subjects, which did not meet the AAMI standard. Liang et al. achieved the best result, but the proportion of hypertensive subjects in the research data was small, and the result of the algorithm in non-normal subjects was not shown. Shoeibi et al. achieved suboptimal results, but they selected noise-free BP signals and did not consider subjects with SBP < 90 mmHg or DBP < 60 mmHg.

[0107] (2) Comparison of the cost of calculation: Par of Panwa et al. is 0.23M lower than the present disclosure, but FLOPs is 31.56M higher than the present disclosure. FLOPs is a key indicator reflecting the inference efficiency, which shows that the present disclosure takes less time to complete a BP estimation. Rong et al. takes the time-frequency graph of 523x396 as input, and takes two-dimensional CNN and Bi-LSTM as multi-path feature encoder, which leads to higher calculation cost than existing research. Qiu et al. is based on Res Net 25, and increases the depth of the model to improve the accuracy of BP estimation, which makes its FLOPs exceed 267 times of the present disclosure. Compared with the large neural network with multi-path and deep layers, the FLOPs of the present disclosure is lower than the existing advanced research, and the inference speed in the BP estimation task is faster.

[0108] The present disclosure analyzes the relationship between the range of MAE and the amplitude range of BP signal, and the results show that the amplitude range of BP signal is positively correlated with MAE. This shows that the larger the amplitude range of the real BP, the more difficult the estimation.

[0109] The present disclosure analyzes the influence of the dispersion of BP signal on the estimation results, and the results show that the increase of the dispersion of BP signal will lead to the increase of the estimation error, and the main source of error increase is positive deviation.

[0110] Figure 7 The comparison chart of the true value and the predicted value of the subject of P type No. 12334. The standard deviation (STD) of the error of the mean arterial pressure (MAP) of this subject is 3.03 mmHg, and the STD of the DBP is 1.46 mmHg. The proportion of positive deviation of MAP is 78.75%, and the proportion of positive deviation of DBP is 39.38%. In terms of mean, SBP is greater than MAP, but in terms of STD, MAP is greater than SBP. The proportion of positive deviation is that MAP is greater than SBP, which shows that in the same subject, the STD of BP is the main factor leading to positive deviation.

[0111] Figure 8 The MAE distribution of MAP, SBP and DBP is shown. Figure 8 (a) shows that there is no obvious difference in the MAE of MAP and DBP, and the PCC and STD of MAE are 0.66 and 2.62 mmHg respectively. Figure 8 (b) and (c) show that the MAE distribution of SBP has a large difference with the MAE distribution of the other two. The difference between SBP and DBP is the largest, and the PCC and STD of the MAE of the two are 0.25 and 4.42 mmHg respectively.

[0112] Figure 8Region 1 of (b) represents the segment where the MAE of SBP exceeds 15 mmHg and the difference between MAP and the MAE of SBP is greater than 10 mmHg, wherein 76.01% of the segments belong to the case where the predicted value of SBP is greater than the true value. Figure 8 Region 2 of (c) represents the segment where the MAE of SBP exceeds 15 mmHg and the difference between DBP and the MAE of SBP is greater than 10 mmHg, wherein 80.63% of the segments belong to the case where the predicted value of SBP is greater than the true value.

[0113] The PCCs of MAP, SBP and DBP are analyzed and the p-values are calculated in the present disclosure. The PCC of MAP and DBP is 0.7782, the PCC of MAP and SBP is 0.5611, and the PCC of DBP and SBP is 0.4001, all of which are less than 10 -5 -4. The mean, range of amplitude and dispersion of SBP are all greater than those of MAP and DBP. The above-mentioned cases are the key factors leading to the difficulty of estimating SBP being higher than that of MAP and DBP, and the estimated value of SBP is usually greater than the true value.

[0114] The influence of ECG on the estimation result of BP is explored by comparative experiments in the present disclosure, and the results show that the introduction of ECG has a better improvement effect on S class, and can effectively reduce the estimation error of SBP.

[0115] The comparative experiments prove that using ECG to assist PPG can reduce the error of BP estimation, and the effect is better for patients with more serious hypertension and higher SBP. In addition, the error of PPG in the BP estimation task is lower than that of ECG, because PPG is similar to BP in shape and PPG reflects the change of blood flow, that is, the correlation between PPG and BP is higher.

[0116] The effect of ECG on BP is further explored by the masking technology in the present disclosure. The mask represents the QRS segment of ventricular depolarization and the TP segment representing ventricular repolarization. The results show that in N class and P class, the masked QRS segment leads to a higher increase in error, which shows that QRS segment has a major effect on N class and P class. In S class, the masked TP segment leads to a higher increase in error, which shows that the effect of TP segment on S class is greater than that of QRS segment.

[0117] The results of the mask experiment show that the role of the TP segment gradually increases as the BP gradually increases. This is because the QRS segment represents ventricular contraction and blood flow to the arteries, and the cardiac dynamics information contained therein can reflect the BP of subjects with normal or mild hypertension. However, the TP segment of hypertensive patients is prolonged, and the P-wave terminal force has a strong correlation with hypertension, which makes the TP segment more capable of reflecting the BP fluctuation information of subjects with severe hypertension.

[0118] The disclosure verifies the effect of MITM in BP fitting through ablation experiments, and the results show that MITM can reduce the MAE and RMSE of all subjects, and the improvement effect of S class and SBP is more obvious. This shows that MITM can improve the generalization of BP estimation, and has high sensitivity to the case with large amplitude range and high dispersion.

[0119] The distribution of PPG features and ECG features before and after MITM weighting is as shown in Figure 9 It can be found by comparison that in the feature map without MITM weighting, most of the features are blue, which means that the extracted features have lower usefulness in the BP estimation task. In the feature map after MITM weighting, most of the PPG features are yellow-green, which means that the important features of ECG are bright yellow-green. This shows that after MITM weighting, the information related to BP in PPG and ECG is given more attention.

[0120] Bi-Tree-Net based on the distributed idea realizes lightweight end-to-end BP estimation and maintains high generalization. Bi-Tree-Net has the advantages of low error, strong stability and low computational cost in the BP estimation task, and can be directly deployed on Jetson Nano, mobile devices and other Internet of Things devices. Based on the characteristic analysis of BP, the disclosure analyzes the influence of the amplitude range and dispersion on the prediction result, studies the reason why SBP estimation is difficult, and compares the relationship between different wave segments of electrocardiogram and BP. The research results are helpful to improve the measurement accuracy and dynamic tracking ability of clinical wearable and portable electronic sphygmomanometers.

[0121] Embodiment 2

[0122] The embodiment provides a blood pressure estimation system taking a multi-scale neural network as a binary tree node, comprising:

[0123] The binary tree neural network module is configured to: acquire a photoplethysmogram and an electrocardiogram, project the photoplethysmogram and the electrocardiogram into a high-resolution space respectively to obtain a first output, perform multi-level decomposition on the first output by using multi-scale convolution calculation to obtain a blood pressure feature matrix;

[0124] The multi-modal information transformation module is configured to perform global dimension reduction compression and local high-dimensional mapping on each group of blood pressure feature vectors in the blood pressure feature matrix, and to obtain a multi-modal blood pressure feature set by weighting the results of the global dimension reduction compression and the local high-dimensional mapping to the blood pressure feature vectors.

[0125] The multi-head decoder module is configured to perform decoding operation on the multi-modal blood pressure features, learn the nonlinear coupling relationship of the multi-modal blood pressure features, and obtain a blood pressure estimation value.

[0126] The model training module is configured to define a target function, and optimize the blood pressure estimation model parameters with the multi-scale neural network as the binary tree nodes.

[0127] It should be noted that the above modules correspond to the steps described in Embodiment 1, and the above modules have the same examples and application scenarios as the corresponding steps, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above modules as part of the system can be executed in a computer system such as a set of computer executable instructions.

[0128] In more embodiments, there are also provided:

[0129] An electronic device includes a memory and a processor, and computer instructions stored on the memory and running on the processor, when the computer instructions are run by the processor, the method described in Embodiment 1 is completed. For brevity, it will not be repeated here.

[0130] It should be understood that in the embodiments, the processor can be a central processing unit CPU, and the processor can also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, ready-to-program gate arrays FPGA or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0131] The memory can include read-only memory and random access memory, and provide instructions and data to the processor, and a part of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0132] A computer readable storage medium for storing computer instructions, when the computer instructions are executed by the processor, the method described in Embodiment 1 is completed.

[0133] The method in embodiment 1 can be directly embodied as being completed by a hardware processor or being completed by a combination of hardware and software modules in the processor. The software modules can be located in a storage medium mature in the art, such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, an electrically erasable programmable memory, a register, or the like. The storage medium is located in a memory, and a processor reads information in the memory and completes the steps of the above method in combination with hardware. To avoid repetition, no longer detailed description is made herein.

[0134] A computer program product includes a computer program, which, when executed by a processor, implements the method described in embodiment 1.

[0135] The present application also provides at least one computer program product tangibly stored on a non-transitory computer readable storage medium. The computer program product includes computer executable instructions, for example, instructions embodied in program modules, executed by devices at the destination, real or virtual processors, to perform processes / methods as described above. Generally, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform particular tasks or implement particular abstract data types. In various embodiments, the functions of the program modules can be combined or divided as desired in various embodiments. Machine executable instructions for program modules can be executed within a local or distributed device. In a distributed device, program modules can be located in local and remote storage media.

[0136] Computer program code for carrying out operations of the present application can be written in one or more programming languages. These computer program codes can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus, so that the program codes, when executed by the computer or other programmable data processing apparatus, cause the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program codes can be executed entirely on a computer, partially on a computer, as a standalone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0137] In the context of the present application, computer program code or related data can be carried by any appropriate carrier to enable a device, apparatus or processor to perform various processes and operations described above. Examples of the carrier include a signal, a computer readable medium, and the like. Examples of the signal can include an electrical, optical, radio, sound or other forms of propagated signals, such as a carrier wave, an infrared signal, and the like.

[0138] Those skilled in the art can understand that the units and algorithm steps of each example described in combination with the present embodiment can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software manner depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0139] Although the specific embodiments of the present application are described above in combination with the drawings, it is not a limitation on the scope of protection of the present application. Those skilled in the art should understand that various modifications or variations made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the scope of protection of the present application.

Claims

1. A blood pressure estimation method using a multi-scale neural network as a binary tree node, characterized by, The method comprises the following steps: obtaining a photoplethysmogram signal and an electrocardiogram signal, projecting the photoplethysmogram signal and the electrocardiogram signal into a high-resolution space respectively to obtain a first output, performing multi-level decomposition on the first output by using multi-scale convolution calculation to obtain a blood pressure feature matrix; performing global dimension reduction compression and local high-dimensional mapping on each group of blood pressure feature vectors in the blood pressure feature matrix, and weighting the results of the global dimension reduction compression and the local high-dimensional mapping to the blood pressure feature vectors to obtain a multi-modal blood pressure feature set; performing decoding operation on the multi-modal blood pressure feature, learning the nonlinear coupling relationship of the multi-modal blood pressure feature, and obtaining a blood pressure estimation value; defining a target function, and optimizing the blood pressure estimation model parameter with the multi-scale neural network as a binary tree node; performing global dimension reduction compression on each group of blood pressure feature vectors in the blood pressure feature matrix, specifically, performing two different compressions on each group of blood pressure feature vectors in the blood pressure feature matrix, longitudinally splicing the two different compression results to obtain a two-dimensional matrix, performing convolution calculation on the two-dimensional matrix to obtain a feature vector, using an attention mechanism to focus on important information related to blood pressure in the feature vector with a high weight and to suppress redundant information in the feature vector with a low weight, and realizing the global dimension reduction compression on the blood pressure feature vector; performing local high-dimensional mapping on each group of blood pressure feature vectors in the blood pressure feature matrix, specifically, performing channel-by-channel accumulation and element-by-element multiplication operation on the blood pressure feature vectors to obtain channel feature vectors and spatial feature vectors, respectively mapping the channel feature vectors and the spatial feature vectors to a high-dimensional space by using multiple groups of convolution to obtain high-dimensional channel vectors and high-dimensional space vectors, and using an attention mechanism to weight and fuse the high-dimensional channel vectors and the high-dimensional space vectors to realize the local high-dimensional mapping of the blood pressure feature vector; the root network of the binary tree neural network projects the photoplethysmogram signal and the electrocardiogram signal into a high-resolution space by using a split convolution and channel fusion, and the subnetwork of the binary tree neural network decomposes the high-resolution space into multiple low-resolution subspaces by using a multi-scale mechanism. 2.The blood pressure estimation method with a multi-scale neural network as a binary tree node of claim 1, wherein, The multi-modal blood pressure feature is decoded by using a multi-head decoder, and the multi-head decoder is composed of a head set and a backbone network. 3.The blood pressure estimation method with a multi-scale neural network as a binary tree node of claim 2, wherein, The head set projects feature vectors of the photoelectric volume pulse signal and the electrocardio signal to a target domain respectively and The backbone network is to map the feature vectors of the photoelectric volume pulse signal and the electrocardio signal to the BP domain and .​ 4. A blood pressure estimation system with multiscale neural networks as nodes of a binary tree, characterized in that, The blood pressure estimation method with the multi-scale neural network as a binary tree node according to any one of claims 1-3 comprises: a binary tree neural network module configured to obtain a photoplethysmogram signal and an electrocardiogram signal, project the photoplethysmogram signal and the electrocardiogram signal into a high-resolution space respectively to obtain a first output, perform multi-level decomposition on the first output by using multi-scale convolution calculation to obtain a blood pressure feature matrix; a multi-modal information transformation module configured to perform global dimension reduction compression and local high-dimensional mapping on each group of blood pressure feature vectors in the blood pressure feature matrix, and weight the results of the global dimension reduction compression and the local high-dimensional mapping to the blood pressure feature vectors to obtain a multi-modal blood pressure feature set; a multi-head decoder module configured to perform decoding operation on the multi-modal blood pressure feature, learn the nonlinear coupling relationship of the multi-modal blood pressure feature, and obtain a blood pressure estimation value; The model training module is configured to define a target function and optimize parameters of a blood pressure estimation model with a multi-scale neural network as a binary tree node.

5. An electronic device, comprising: A computer program product, comprising a memory and a processor, and computer instructions stored on the memory and run on the processor, when the computer instructions are run by the processor, the method of any one of claims 1-3 is completed.

6. A computer-readable storage medium, characterized in that, A computer program product for storing computer instructions, when the computer instructions are executed by a processor, the method of any one of claims 1-3 is completed.

7. A computer program product, characterised in that, A computer program product for storing computer instructions, when the computer instructions are executed by a processor, the method of any one of claims 1-3 is completed.

Images