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

Abstract

The invention 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, and the method comprises the steps: respectively projecting an obtained photoelectric volume pulse signal and an electrocardiosignal to a high-resolution space to obtain a first output, and carrying out the multi-stage decomposition of the first output to obtain a blood pressure feature matrix; global dimension reduction compression and local high-dimensional mapping are carried out on the blood pressure feature vectors, and the result is weighted on the blood pressure feature vectors to obtain a multi-modal blood pressure feature set; decoding the multi-modal blood pressure features, learning a nonlinear coupling relation of the multi-modal blood pressure features, and obtaining a blood pressure estimation value; defining a target function, and optimizing parameters of the blood pressure estimation model. According to the method, the low-calculation-cost reasoning mode of the tree algorithm is combined with the multi-modal feature extraction capacity of the multi-scale and multi-resolution neural network, the light weight of the algorithm is achieved on the premise of high-precision continuous blood pressure estimation, and the calculation cost meets the requirement of wearable equipment.

Description

Technical Field

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

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

[0003] Hypertension is one of the main causes of cardiovascular and cerebrovascular diseases. Clinical cuff-type electronic blood pressure monitors are affected by external, behavioral and psychological factors, leading to missed diagnosis of hidden hypertension and misdiagnosis of pseudohypertension. Mainstream medical associations recommend continuous, dynamic and long-term monitoring of BP based on wearable devices.

[0004] Wearable devices require that the computational complexity of the BP estimation algorithm be lower than a certain threshold and ensure low errors. Existing BP estimation algorithms are mainly based on empirical or hypothetical mathematical and 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 large errors in BP estimation and dynamic tracking of a wide range of subjects. DL methods can construct arbitrary function mapping relationships by automatically analyzing the distribution structure of big data, but the huge computational complexity of DL exceeds the computing power of wearable devices. Under the premise of high-precision continuous BP estimation, it is a significant task to compress the computational cost of the algorithm within the computing range of wearable devices and realize hardware deployment. Summary of the invention

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

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

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

[0008] Acquire a photoelectric volumetric pulse signal and an electrocardiogram signal, project the photoelectric volumetric pulse signal and the electrocardiogram signal to a high-resolution space, respectively, obtain a first output, perform multi-level decomposition on the first output using a multi-scale convolution calculation, and obtain a blood pressure feature matrix;

[0009] Perform global dimensionality 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 global dimensionality reduction compression and local high-dimensional mapping to the blood pressure feature vector to obtain a multimodal blood pressure feature set;

[0010] Decoding the multimodal blood pressure features, learning the nonlinear coupling relationship of the multimodal blood pressure features, and obtaining the blood pressure estimation value;

[0011] The objective function is defined and the parameters of the blood pressure estimation model with multi-scale neural networks as binary tree nodes are optimized.

[0012] As an optional implementation, global dimensionality 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, and a 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 high weight and to suppress redundant information in the feature vector with low weight, thereby achieving global dimensionality reduction compression of the blood pressure feature vector.

[0013] As an optional implementation, 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 accumulated channel by channel and multiplied element by element to obtain channel feature vectors and space feature vectors, multiple groups of convolutions are used to map the channel feature vectors and space feature vectors to high-dimensional space respectively to obtain high-dimensional channel vectors and high-dimensional space vectors, and an attention mechanism is used to perform weighted fusion of the high-dimensional channel vectors and high-dimensional space vectors to achieve local high-dimensional mapping of the blood pressure feature vectors.

[0014] As an optional implementation, a multi-head decoder is used to decode the multi-modal blood pressure features, and the multi-head decoder is composed of a head set and a backbone network.

[0015] As an optional implementation, the head set projects the feature vectors of the photoplethysmography signal and the electrocardiogram signal to the target domain respectively. and The backbone network is and Unified mapping to BP domain

[0016] As an optional implementation, the root network of the binary tree neural network uses separable convolution and channel fusion to project the photoelectric volumetric pulse signal and the electrocardiogram signal into the high-resolution space respectively, and the sub-network of the binary tree neural network uses a multi-scale mechanism to decompose the high-resolution space into multiple low-resolution sub-spaces.

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

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

[0019] The multimodal information transformation module is configured to: perform global dimensionality 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 dimensionality reduction compression and local high-dimensional mapping to the blood pressure feature vector to obtain a multimodal blood pressure feature set;

[0020] The multi-head decoder module is configured to: perform a 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] The model training module is configured to: define an objective function and optimize the parameters of a blood pressure estimation model with a multi-scale neural network as a binary tree node.

[0022] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in the first aspect is performed.

[0023] In a fourth aspect, the present invention provides a computer-readable storage medium for storing computer instructions, wherein when the computer instructions are executed by a processor, the method described in the first aspect is performed.

[0024] In a fifth aspect, the present invention provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the method described in the first aspect is implemented.

[0025] Compared with the prior art, the present invention 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, using a binary tree neural network (Bi-Tree-Net) as a framework, and using a multi-modal information transformation module (MITM) and a multi-head decoder (MHDE) as an aid to achieve lightweight end-to-end blood pressure estimation. The binary tree neural network combines the tree algorithm of machine learning with the multi-scale and multi-resolution technology of deep learning, and learns the morphology of low-amplitude waves such as reflected waves, descending isthmuses and dicrotic waves, as well as the gradient change trends of systole and diastole in a low-computational-cost manner, thereby achieving lightweight algorithms under the premise of high-precision continuous blood pressure estimation, and the computational cost meets the requirements of wearable devices.

[0027] The multimodal information transformation module performs global averaging and high-dimensional mapping on the vectors encoded by the binary tree neural network to reduce the timing redundancy between the continuous transformation trend of blood pressure and the low-amplitude wave morphology, as well as the sample redundancy between the photoplethysmogram signal and the electrocardiogram signal.

[0028] The multi-head decoder module decodes the heterogeneous blood pressure information between the photoplethysmogram signal and the electrocardiogram signal in parallel to enhance the binary tree neural network's ability to characterize the peaks and troughs of the photoplethysmogram signal and the QRS segment and TP segment of the electrocardiogram signal.

[0029] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0031] Figure 1 A flow chart of a blood pressure estimation method using a multi-scale neural network as a binary tree node provided in Example 1 of the present invention;

[0032] Figure 2 A schematic diagram of the structure of a binary tree neural network provided in Example 1 of the present invention;

[0033] Figure 3 A schematic diagram of the structure of a multimodal information conversion module provided in Example 1 of the present invention;

[0034] Figure 4 This is an experimental flow chart of a blood pressure estimation method using a multi-scale neural network as a binary tree node provided in Example 1 of the present invention;

[0035] Figure 5 Bland-Altman graph and Correlation graph of the blood pressure estimation method using a multi-scale neural network as a binary tree node provided in Example 1 of the present invention;

[0036] Figure 6 The MAE and RMSE results of the blood pressure estimation method using a multi-scale neural network as a binary tree node provided in Example 1 of the present invention in N-type, P-type and S-type subjects;

[0037] Figure 7 A comparison chart of the actual value and the predicted value of the blood pressure estimation method using a multi-scale neural network as a binary tree node in the P category 12334 subjects provided by Example 1 of the present invention;

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

[0039] Fig. 9 The characteristic thermal diagrams of PPG and ECG before and after MITM weighting of the present invention. DETAILED DESCRIPTION

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

[0041] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0042] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "include" and "have" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0043] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0044] Example 1

[0045] like Figure 1 As shown, this embodiment provides a blood pressure estimation method using a multi-scale neural network as a binary tree node, comprising the following steps:

[0046] S1 acquires a photoelectric volumetric pulse signal and an electrocardiogram signal, projects the photoelectric volumetric pulse signal and the electrocardiogram signal to a high-resolution space, respectively, obtains a first output, and uses a multi-scale convolution calculation to perform multi-level decomposition on the first output to obtain a blood pressure feature matrix;

[0047] 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, and weights the results of global dimensionality reduction compression and local high-dimensional mapping to the blood pressure feature vector to obtain a multi-modal blood pressure feature set;

[0048] 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 the blood pressure estimation value;

[0049] S4 defines the objective function and optimizes the parameters of the blood pressure estimation model with the multi-scale neural network as the binary tree nodes.

[0050] Multi-scale mechanism and multi-resolution technology can effectively obtain the local morphology of low-amplitude waves such as reflected waves, descending isthmus and dicrotic waves, as well as 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 the complex original problem into multiple sub-problems, and achieve efficient reasoning by synchronously processing multiple sub-problems. In order to obtain the local morphology and global rhythm of blood pressure (BP) from photoplethysmography (PPG) and electrocardiogram (ECG) signals in a low computational cost manner, the present invention uses a binary tree algorithm as the basic architecture and multi-scale and multi-resolution neural networks as the nodes of the tree to design a binary tree neural network (Bi-Tree-Net), whose structure is as follows: Figure 2 As shown in the figure, Bi-Tree-Net decomposes the BP estimation task into multi-scale and multi-resolution neural network nodes for parallel processing. Multiple nodes use different receptive fields to perform nonlinear transformation to extract multi-modal blood pressure features.

[0051] Assume that is an arbitrary set of real numbers, and the PPG and ECG inputs of 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 input to output is the process of mapping high resolution to low resolution. In Bi-Tree-Net, the present disclosure sets the process of decomposing the root network into sub-networks according to the nature of the task as the paradigm of mapping high resolution to low resolution.

[0052] Specifically, the root network of Bi-Tree-Net projects PPG and ECG into high-resolution space respectively, and the sub-network uses a multi-scale mechanism to decompose the high-resolution space into multiple low-resolution sub-spaces. The Bi-Tree-Net consists of a root network and sub-networks, where the depth Included subnetworks. For any input After mapping the root network, output tree out1 The resolution is f. After mapping the sub-network, we get The group resolution is Output.

[0053] The root network uses separable convolution and channel fusion to transform X∈(X ppg ,X ecg ) is projected to the high-resolution space as shown in formula (1).

[0054]

[0055] Where k1 and k2 are the scales of the convolution kernel (k1≠k2). The dimensionality change of this process is

[0056] The sub-network uses multi-scale convolution calculations to calculate the tree out1 Decompose the image level by level to achieve mapping from high resolution to multiple groups of low resolution. The decomposition process of any subnetwork in is shown in formula (2-3).

[0057]

[0058] k s1 and k s2 is the scale of the convolution kernel (k s1 ≠k s2 ), For Depth The resolution change of this decomposition process is Finally, after 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-by-channel convolution, and * represents convolution multiplication.

[0059] The BP feature vectors encoded at different scales and resolutions contain the local morphological changes of the opening and closing of the aorta and mitral valves and the global gradient changes of the ascending and descending branches. Different BP feature vectors have a complementary relationship. However, the global gradient changes contain local morphological changes and any two sub-networks in Bi-Tree-Net have the same source domain, which leads to There is information redundancy between the global trend and local morphology of BP and the sample redundancy between PPG and ECG. 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 multimodal information transformation module (MITM). MITM performs global dimensionality reduction compression and local high-dimensional mapping on BP feature vectors to achieve feature selection of different BP vectors. The structure of MITM is as follows: Figure 3 shown.

[0060] Global dimensionality reduction compression: First, for any set of BP feature vectors out of Bi-Tree-Net iConduct F gap (·) and F gmp (·) Two different compressions. As shown in formula (4-5),

[0061]

[0062] Then, z gap and z gmp Vertical splicing to obtain the two-dimensional matrix z gap-gmp , and use convolution calculation to achieve z gap and z gmp Point-to-point fusion is performed to obtain the feature vector z. As shown in formula (6-7),

[0063]

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

[0065]

[0066] Local high-dimensional mapping: First, out i Perform channel-by-channel accumulation Γ add and the element-wise product Γ mul Operation, get the channel feature vector and spatial eigenvectors As shown in formula (9-10),

[0067]

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

[0069]

[0070] Finally, the attention mechanism F is used attention2 (·) Perform weighted fusion on CH and SP to obtain the local feature vector of BP As shown in formula (13),

[0071]

[0072] After completing global dimensionality reduction compression and local high-dimensional mapping, and Weighted outi Get the multimodal BP feature vector set i As shown in formula (14),

[0073]

[0074] Out tree Repeat the above operation for each set of BP feature vectors in to obtain the multimodal BP feature set: represents the sum of elements, ⊙ represents the multiplication of elements, Matrix multiplication.

[0075] Simple linear addition or cascade operation on the multimodal BP feature set of MITM will cause dimensionality disaster, multiple comparisons and loss of nonlinear relationships. The present disclosure proposes a multi-head decoder (MHDE) consisting of a head set and a backbone network to enhance the feature representation capabilities of the peaks and troughs of PPG and the QRS and TP segments of ECG and realize the learning of nonlinear coupling relationships in multimodal BP features.

[0076] Assume that the head set of MHDE is head i Set i The decoding process is as shown in formula (15-17), where 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 steps, the Set of any input source MBPF Projected to the target domain Assume that the head set projects the feature vectors of PPG and ECG to the target domain respectively and Backbone Network bn is to and Unified mapping to BP domain The process is as shown in formula (18).

[0079]

[0080] like and The error between them is loss ppg , and The error between them 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 |loss ppg |+|loss ecg |The minimum value is reached.

[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 As shown. The present disclosure obtains various types of subjects from the Medical Information Mart for Intensive Care (MIMIC) database. After preprocessing, the present disclosure 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 advanced research.

[0090] MIMIC contains synchronized PPG, ECG and BP signals, all with a sampling rate of 125Hz. In order to facilitate the training and testing of BP estimation tasks, the signals of MIMIC were sorted and stored in the University of California, Irvine (UCI) machine learning database. The data disclosed in this paper comes from the UCI machine learning database, and the number of each record is fully recorded.

[0091] The present invention divides the subjects into three categories according to the Office blood pressure threshold: Normal (N), Prehypertension (P) and Stage 1 Hypertension (S). 100 subjects are randomly selected from each category, and the duration of all records is greater than or equal to 8 minutes. The present invention uses a sliding window with a window length of 1 second and a step length of 1 second to segment the records, and 167432 segments are obtained for PPG, ECG and BP respectively. MAP, DBP and SBP are defined as the average, minimum and maximum values ​​of each BP segment, respectively. Max-Min Normalization is performed on the PPG segment and the ECG segment respectively.

[0092] The present disclosure divides the subjects into training set (80%), validation set (10%) and test set (10%) according to the inter-patient paradigm, and all experiments adopt 10-fold cross validation. The training set consists of three types of subjects, N, P and S, instead of training for a specific type of subject. The experimental environment is 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. The present disclosure designs a 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 α, β and γ 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 evaluation indicators for BP estimation results, and uses parameter quantity (Par) and computational complexity (FLOPs) as lightweight evaluation indicators. The calculation process of MAE and RMSE is shown in formulas (28) and (29). In addition, the present disclosure also uses the standards of the British Hypertension Society (BHS) and the Association for the Advancement of Medical Instrumentation (AAMI) to evaluate the clinical application value of the algorithm.

[0099]

[0100] The Bland-Altman diagram and Correlation diagram of the present disclosure are as follows: Figure 5 As shown. The Bland-Altman plot shows 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 BP estimated by the present disclosure has a high consistency with the true BP.

[0101] Figure 5 (d) Correlation plot representing 300 subjects. The Pearson Correlation Coefficient (PCC) between the estimated MAP, SBP, and DBP and the true values ​​were 0.8027, 0.9464, and 0.5363, respectively, and the p-values ​​were all less than 10. -5 This shows that the BP estimated by the present invention has a high correlation with the true BP, and the results are statistically significant. The ME of SBP is the lowest and the PCC is the highest, which shows that the BP focal loss function of the present invention has a good effect.

[0102] The results of the algorithm disclosed in this disclosure in subjects of category N, category P and category S are as follows Figure 6As shown. In the three types of subjects, the distribution intervals of MAE and RMSE are [3.13, 4.84] mmHg and [4.57, 6.66] mmHg respectively. This shows that the algorithm disclosed in the present invention can maintain accurate estimation effects when facing diverse subjects in clinical environments, and is less affected by subject specificity.

[0103] Compared to the N-type subjects, the algorithm's errors in the P and S categories 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 P-type subjects had the highest error, and the predicted values ​​were usually greater than the true values. This is because the P-type is between the N-type and the S-type, and some subjects have SBP in the P-type and DBP in the N-type.

[0104] The present disclosure adopted the BHS standard and the AAMI standard to evaluate the performance of the algorithm. The results showed that the performance of the present disclosure in 9 BP estimation tasks reached Level A of the BHS standard, and passed the AAMI test in three categories of subjects: N, P, and S.

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

[0106] (1) Comparison of errors: Mohebbian et al., Panwa et al., and Qiu et al. used intra-patient paradigm and sample overlap preprocessing methods, and their test processes had information leakage problems. The present invention is an inter-patient paradigm and non-overlapping sample division method, and the test results are more in line with clinical standards. Hajj et al.'s MAE for DBP was 2.60 mmHg, but the MAE for SBP was 4.61 mmHg, and the standard deviation (STD) of the error was 7.81 mmHg. The STD of the present invention was 3.54 mmHg, which shows that the stability of the present invention is better than that of Hajj et al. Ali et al. tested 40 subjects who did not meet the AAMI standard. Liang et al. obtained the best results, but the proportion of hypertensive subjects in their research data was small, and the results of the algorithm in non-normal subjects were not shown. Shoeibi et al. obtained 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 computational costs: Panwa et al.'s Par is 0.23M lower than that of the present invention, but its FLOPs is 31.56M higher than that of the present invention. FLOPs is a key indicator reflecting reasoning efficiency, which means that the present invention takes less time to complete a BP estimation. Rong et al. used a 523×396 time-frequency graph as input and used two-dimensional CNN and Bi-LSTM as multi-path feature encoders, which resulted in higher computational costs than existing studies. Qiu et al. used Res Net 25 as the basis and increased the depth of the model to improve the accuracy of BP estimation, which made its FLOPs exceed 267 times that of the present invention. Compared with multi-path and deep large neural networks, the present invention has lower FLOPs than existing advanced studies and faster reasoning speed in BP estimation tasks.

[0108] This paper 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 means that the larger the amplitude range of the real BP, the more difficult it is to estimate.

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

[0110] Figure 7 The figure shows the comparison between the actual value and the predicted value of subject No. 12334 of category P. The standard deviation (STD) of the error of the mean pressure (MAP) of this subject is 3.03 mmHg, the STD of 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 MAP greater than SBP, which indicates that in the same subject, the STD of BP is the main factor leading to positive deviation.

[0111] Figure 8 The MAE distributions of MAP, SBP, and DBP are shown. Figure 8 (a) shows that there is no significant difference in MAE between 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 is quite different from that of the other two. The MAE distribution of SBP and DBP is the most different, with the PCC and STD of their MAE being 0.25 and 4.42 mmHg, respectively.

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

[0113] The present disclosure analyzes the PCC of MAP, SBP and DBP and calculates the p-value. 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, and the p-value is less than 10 -5 At the same time, the mean, amplitude range and dispersion of SBP are greater than those of MAP and DBP. The above situation is the key factor that makes the estimation of SBP more difficult than that of MAP and DBP, and the estimated value of SBP is usually greater than the true value.

[0114] The present disclosure explores the influence of ECG on BP estimation results through comparative experiments. The results show that the introduction of ECG has a better improvement effect on Class S and can effectively reduce the estimation error of SBP.

[0115] Comparative experiments have shown that using ECG to assist PPG can reduce the error of BP estimation, and it is more effective for patients with more severe hypertension and higher SBP. In addition, the error of PPG in BP estimation is lower than that of ECG, because the morphology of PPG and BP is similar and PPG reflects the changes in blood flow, that is, the correlation between PPG and BP is higher.

[0116] The present disclosure further explores the effect of ECG on BP through masking technology. After masking the QRS segment representing ventricular depolarization and the TP segment representing ventricular repolarization. The results show that in the N and P classes, masking the QRS segment leads to a higher increase in error, which indicates that the QRS segment has a major effect on the N and P classes. In the S class, masking the TP segment leads to a higher increase in error, which indicates that the TP segment has a greater effect on the S class than the QRS segment.

[0117] The results of the masking experiment show that as BP gradually increases, the effect of the TP segment gradually increases. This is because the QRS segment represents the contraction of the ventricles and the flow of blood to the arteries, and the cardiac dynamics information it contains can reflect the BP of normal or mildly hypertensive subjects. However, the TP segment of hypertensive patients is prolonged, and the P-wave terminal force is strongly correlated with hypertension, which makes the TP segment more able to reflect the BP fluctuation information of subjects with severe hypertension.

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

[0119] The distribution of PPG features and ECG features before and after MITM weighting is as follows: Fig. 9 As shown. By comparison, it can be found that in the feature map without MITM weighting, most of the features appear in blue with low usefulness. This shows that the extracted features have little effect in the BP estimation task. In the feature map after MITM weighting, most of the features of PPG appear in yellow-green with high usefulness, and the important features of ECG appear in bright yellow-green. This shows that after MITM weighting, the information in PPG and ECG that is highly relevant to BP is given special attention.

[0120] Bi-Tree-Net based on distributed thinking 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 BP estimation tasks, and can be directly deployed on Jetson Nano, mobile devices and other IoT devices. Based on the characteristics of BP, this paper analyzes the influence of its amplitude range and discreteness on the prediction results, studies the reasons for the high difficulty of SBP estimation, and compares the relationship between different bands of ECG signals and BP. The research results are helpful to improve the measurement accuracy and dynamic tracking ability of clinical wearable and portable electronic blood pressure monitors.

[0121] Example 2

[0122] This embodiment provides a blood pressure estimation system using a multi-scale neural network as a binary tree node, including:

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

[0124] The multimodal information transformation module is configured to: perform global dimensionality 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 dimensionality reduction compression and local high-dimensional mapping to the blood pressure feature vector to obtain a multimodal blood pressure feature set;

[0125] The multi-head decoder module is configured to: perform a 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 an objective function and optimize the parameters of a blood pressure estimation model with a multi-scale neural network as a binary tree node.

[0127] It should be noted that the above modules correspond to the steps described in Example 1, and the examples and application scenarios implemented by the above modules and the corresponding steps are the same, but are not limited to the contents disclosed in the above Example 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 further embodiments, there is also provided:

[0129] An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in Embodiment 1 is performed. For the sake of brevity, it will not be described in detail here.

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

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

[0132] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the method described in Example 1 is completed.

[0133] The method in Example 1 can be directly embodied as a hardware processor, or a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it is not described in detail here.

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

[0135] The present invention 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, such as instructions included in a program module, which are executed in a device on a real or virtual processor of the target to perform the process / method as described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform specific tasks or implement specific abstract data types. In various embodiments, the functions of program modules can be combined or divided between program modules as needed. Machine executable instructions for program modules can be executed in local or distributed devices. In distributed devices, program modules can be located in local and remote storage media.

[0136] The computer program code for implementing the method of the present invention can be written in one or more programming languages. These computer program codes can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the computer or other programmable data processing device, causes the function / operation specified in the flow chart and / or block diagram to be implemented. The program code can be executed completely on a computer, partially on a computer, as an independent software package, partially on a computer and partially on a remote computer or completely on a remote computer or server.

[0137] In the context of the present invention, computer program codes or related data may be carried by any appropriate carrier to enable a device, apparatus or processor to perform the various processes and operations described above. Examples of carriers include signals, computer readable media, and the like. Examples of signals may include electrical, optical, radio, acoustic or other forms of propagation signals, such as carrier waves, infrared signals, and the like.

[0138] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0139] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A blood pressure estimation method using a multi-scale neural network as a binary tree node, characterized in that: The following steps are involved: Acquire a photoelectric volumetric pulse signal and an electrocardiogram signal, project the photoelectric volumetric pulse signal and the electrocardiogram signal to a high-resolution space, respectively, obtain a first output, perform multi-level decomposition on the first output using a multi-scale convolution calculation, and obtain a blood pressure feature matrix; Perform global dimensionality 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 global dimensionality reduction compression and local high-dimensional mapping to the blood pressure feature vector to obtain a multimodal blood pressure feature set; Decoding the multimodal blood pressure features, learning the nonlinear coupling relationship of the multimodal blood pressure features, and obtaining the blood pressure estimation value; The objective function is defined and the parameters of the blood pressure estimation model with multi-scale neural networks as binary tree nodes are optimized.

2. The blood pressure estimation method using a multi-scale neural network as a binary tree node as claimed in claim 1, characterized in that: Global dimensionality 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, and the two-dimensional matrix is ​​convolved to obtain a feature vector. The attention mechanism is used to focus on important information related to blood pressure in the feature vector with high weight and suppress redundant information in the feature vector with low weight, so as to achieve global dimensionality reduction compression of the blood pressure feature vector.

3. The blood pressure estimation method using a multi-scale neural network as a binary tree node as claimed in claim 1, characterized in that: A 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 accumulated channel by channel and multiplied element by element to obtain channel feature vectors and spatial feature vectors. Multiple groups of convolutions are used to map the channel feature vectors and spatial feature vectors to high-dimensional spaces respectively to obtain high-dimensional channel vectors and high-dimensional space vectors. The attention mechanism is used to perform weighted fusion of the high-dimensional channel vectors and high-dimensional space vectors to achieve local high-dimensional mapping of the blood pressure feature vectors.

4. The blood pressure estimation method using a multi-scale neural network as a binary tree node as claimed in claim 1, characterized in that: A multi-head decoder is used to decode the multi-modal blood pressure features, and the multi-head decoder is composed of a head set and a backbone network.

5. The blood pressure estimation method using a multi-scale neural network as a binary tree node as claimed in claim 4, characterized in that: The head set projects the feature vectors of the photoplethysmography signal and the electrocardiogram signal to the target domain respectively. and The backbone network is and Unified mapping to BP domain 6. The blood pressure estimation method using a multi-scale neural network as a binary tree node as claimed in claim 1, characterized in that: The root network of the binary tree neural network uses separable convolution and channel fusion to project the photoelectric volumetric pulse signal and the electrocardiogram signal into the high-resolution space respectively, and the sub-network of the binary tree neural network uses a multi-scale mechanism to decompose the high-resolution space into multiple low-resolution subspaces.

7. A blood pressure estimation system using a multi-scale neural network as a binary tree node, characterized in that: include: The binary tree neural network module is configured to: obtain a photoelectric volumetric pulse signal and an electrocardiogram signal, project the photoelectric volumetric pulse signal and the electrocardiogram signal to a high-resolution space, respectively, obtain a first output, and perform multi-level decomposition on the first output using a multi-scale convolution calculation to obtain a blood pressure feature matrix; The multimodal information transformation module is configured to: perform global dimensionality 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 dimensionality reduction compression and local high-dimensional mapping to the blood pressure feature vector to obtain a multimodal blood pressure feature set; The multi-head decoder module is configured to: perform a 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; The model training module is configured to: define an objective function and optimize the parameters of a blood pressure estimation model with a multi-scale neural network as a binary tree node.

8. An electronic device, characterized in that: The method comprises a memory and a processor and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method according to any one of claims 1 to 6 is completed.

9. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the method described in any one of claims 1 to 6.

10. A computer program product, characterized in that The invention comprises a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 6.

Images