CN116369877A - A non-invasive blood pressure estimation method based on photoplethysmography - Google Patents

Abstract

The invention provides a non-invasive blood pressure estimation method based on photoplethysmography. The method includes: collecting and monitoring photoplethysmography signals and invasive continuous blood pressure waveform signals of the human body, and performing initialization processing on them, and performing initialization processing on the Feature selection of the data, screening of information parameters with discriminative significance, and feature fusion. According to the data after feature fusion, a two-stage model of non-invasive blood pressure estimation based on photoplethysmography is built. The two-stage model of non-invasive blood pressure estimation based on photoplethysmography is trained, and the user's photoplethysmography signal is collected and input into the two-stage model of non-invasive blood pressure estimation based on photoplethysmography for non-invasive blood pressure estimation. The non-invasive blood pressure estimation method based on photoplethysmography provided by the present invention can realize non-invasive blood pressure estimation through photoplethysmography, improve the estimation accuracy, and can provide reference for blood pressure disease detection.

Description

A non-invasive blood pressure estimation method based on photoplethysmography

technical field

The invention relates to the technical field of artificial intelligence, pulse wave and non-invasive blood pressure monitoring, in particular to a non-invasive blood pressure estimation method based on photoplethysmography.

Background technique

Statistical reports show that the prevalence of hypertension in the general population of the United States rose to 45.4% in 2017, and it is estimated that the number of deaths related to hypertension accounted for 19.2% of all deaths. Long-term continuous monitoring of blood pressure (BP) is necessary to predict risk and make appropriate therapeutic intervention, which can effectively improve prognosis and reduce mortality. The gold standard for obtaining continuous blood pressure is invasive arterial cannulation. However, this method requires inserting a thin tube into the blood vessel to bring the pressure sensor into direct contact with the blood, which is not only painful but also increases the risk of infection. Therefore, continuous noninvasive blood pressure monitoring is more suitable for health management beyond intensive care. Current cuff-based auscultation and oscillometric techniques require inflation and occlusion of vessels during testing, which is inconvenient for long-term monitoring.

Additionally, several studies have proposed cuffless measurement methods for portable and comfortable blood pressure monitoring. These methods use various sensors to acquire pulse waveforms, such as photoplethysmography (PPG), ultrasound, and radar. According to the pulse wave velocity (PWV) theory, there are even studies combining multimodal physiological signals, such as electrocardiogram (ECG) or ballistocardiogram (BCG). Among them, the blood pressure prediction method based on pulse transit time (pulse transient times, PTT) is the most common technique. However, most algorithms require combining PPG with signals such as ECG or BCG monitored elsewhere, requiring additional sensors or sticky electrode placement, which hinders long-term blood pressure measurement for users. Additionally, current non-invasive blood pressure monitoring technologies have some challenges that must be addressed. First, existing techniques do not consider individual differences in optical signals. For example, a weak pulse may not be present in the elderly or those with cardiovascular disease. These special populations are often accompanied by symptoms such as high blood pressure. It can lead to incorrect blood pressure predictions if modeled as a normal population.

Furthermore, current techniques and inventive approaches are still limited by dataset quality, experimental design, and algorithmic bottlenecks. Finally, most techniques are restricted to a single dataset for training and evaluation, which introduces a high risk of overfitting given the diversity of the population. Therefore, it is necessary to design a non-invasive blood pressure estimation method based on photoplethysmography.

Contents of the invention

The purpose of the present invention is to provide a non-invasive blood pressure estimation method based on photoplethysmography, which can realize non-invasive blood pressure estimation through photoplethysmography, improve the estimation accuracy, and provide reference for blood pressure disease detection.

To achieve the above object, the present invention provides the following scheme:

A method for estimating noninvasive blood pressure based on photoplethysmography, comprising the steps of:

Step 1: Collect and monitor the photoplethysmography signal of the human body and the invasive continuous blood pressure waveform signal;

Step 2: Initialize the acquired photoplethysmography signal and invasive continuous blood pressure waveform signal;

Step 3: Perform feature selection on the initialized data, screen information parameters with discriminative significance, and perform feature fusion;

Step 4: Build a two-stage model of non-invasive blood pressure estimation based on photoplethysmography based on the data after feature fusion;

Step 5: According to the initialized and processed data, train the two-stage model of non-invasive blood pressure estimation based on photoplethysmography;

Step 6: Collect the user's photoplethysmography signal and input it into the two-stage model of non-invasive blood pressure estimation based on photoplethysmography for non-invasive blood pressure estimation;

Step 7: Proving the robustness of the various morphological features selected by the PMFL provided by the embodiments of the present invention through experiments;

Step 8: Quantitative verification and evaluation of the non-invasive blood pressure estimation method based on photoplethysmography provided by the embodiment of the present invention.

Optionally, in step 1, collect and monitor the photoplethysmography signal of the human body and the invasive continuous blood pressure waveform signal, specifically:

Collect and monitor the photoplethysmographic pulse wave signal of the human body, that is, the PPG signal, and the invasive continuous blood pressure waveform signal, wherein the invasive continuous blood pressure waveform signal is used as a reference signal.

Optionally, in step 2, the acquired photoplethysmography signal and invasive continuous blood pressure waveform signal are initialized, specifically:

Resampling, filtering, noise reduction, signal segmentation, phase matching and normalization processing are carried out on the obtained photoplethysmography signal and invasive continuous blood pressure waveform signal, and the prior feature extraction is performed on the normalized signal, from The morphological features were extracted from the normalized PPG signal, and the normalized signal and the extracted morphological features were used as training data for training the two-stage model of non-invasive blood pressure estimation based on photoplethysmography.

Optionally, in step 3, feature selection is performed on the initialized data, information parameters with discriminative significance are screened, and feature fusion is performed, specifically:

Obtain the normalized signal and the extracted morphological features, and use the PPG morphological feature learning algorithm (PPG morphological feature learning, PMFL) to screen the morphological features. First, use the base learner to perform a set of relatively important features Sorting and filtering, followed by infiltration using recursive feature elimination to find the best feature combination and determine the final optimized feature set, based on the final optimized feature set, the depth acquired by the two-stage model based on photoplethysmographic NIBP estimation The features are fused with the morphological features in the final optimized feature set, and the proportion of different feature sets is determined by setting feature weights:

F _f =ε·F _m +(1-ε)·F _d (1)

In the formula, F _f represents fusion features, F _m and F _d represent morphological features and depth features, respectively.

Optionally, in step 4, a two-stage model of noninvasive blood pressure estimation based on photoplethysmography is built according to the data after feature fusion, specifically:

According to the normalized signal and its corresponding fusion features, a two-stage model of non-invasive blood pressure estimation based on photoplethysmography is built. The cuff blood pressure group classification model and the cuffless blood pressure estimation pipeline based on automatic machine learning are composed of the SEM-ResNet model and the Auto-Regressor model. The SEM-ResNet model is used to classify the acquired user PPG signal For different blood pressure groups, the Auto-Regressor model is used to automatically establish a fine-grained regressor for the PPG signal of each blood pressure group for non-invasive blood pressure estimation.

Optionally, build a SEM-ResNet model, specifically:

Build the SEM-ResNet model, where the SEM-ResNet model includes ResNet, multiple sub-networks and neural networks, where SEM-ResNet takes ResNet as the backbone network, and integrates the residual of the squeeze-and-excitation (SE) module Differential connections are used to learn shared low-level features, and multiple sub-networks use convolution kernels of different scales to jointly learn high-level specific signal features, and the neural network embeds the extracted fusion features to provide prior information.

Optionally, build an Auto-Regressor model, specifically:

Build an Auto-Regressor model, automatically build a fine-grained regressor for the PPG signal of each blood pressure group obtained by the SEM-ResNet model, and obtain accurate non-invasive blood pressure estimation. Among them, a stacked integrated learning optimization algorithm is constructed through the AutoML pipeline , the Auto-Regressor model is defined through the H2O AutoML framework. First, a meta-level regressor is trained to find the best combination of the basic level regressors. If the prediction errors of the basic learners are all low, and the constructed models are If there is a significant difference, it is judged that the model performs well, and the AutoML pipeline enters the base regression stage, and random searches are performed in different model sets, which can generate a variety of base regressors, and produce influential when paired with the stacking method. Ensemble, where meta-level regressors are set in the meta-level regression stage to train using the k-fold cross-validation predictions of the base regressors, and to test on datasets outside the fold.

Optionally, in step 5, according to the initialized and processed data, the two-stage model of non-invasive blood pressure estimation based on photoplethysmography is trained, specifically:

Divide the initialized data into training set, verification set and test set according to the ratio of 7:1.5:1.5, train and select parameters for the two-stage model of non-invasive blood pressure estimation through the training set and verification set, and test the saved data through the test set For the generalization ability of the optimal system model, in the process of training the network, the parameters are updated through the Adam optimizer, where the learning rate is 0.001, the weight decay is 0.999, and the momentum is 0.8.

Optionally, in step 6, the user’s photoplethysmography signal is collected and input into the two-stage model of noninvasive blood pressure estimation based on photoplethysmography to perform noninvasive blood pressure estimation, specifically:

The user's PPG signal is collected, and the signal waveform is sent as input to the non-invasive blood pressure estimation two-stage model to obtain the user's corresponding blood pressure value to realize non-invasive blood pressure estimation.

According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: The non-invasive blood pressure estimation method based on photoplethysmography provided by the present invention, the method includes collecting and monitoring the photoplethysmography signal of the human body and the invasive continuous blood pressure waveform Signal, initialize the acquired photoplethysmography signal and invasive continuous blood pressure waveform signal, perform feature selection on the initialized data, screen the information parameters with identification significance, and perform feature fusion, according to the data after feature fusion Build a two-stage model of non-invasive blood pressure estimation based on photoplethysmography, train the two-stage model of non-invasive blood pressure estimation based on photoplethysmography according to the initialized data, collect user photoplethysmography signals, and input them into the photoelectric plethysmography In the two-stage model of non-invasive blood pressure estimation of volume pulse wave, non-invasive blood pressure estimation is performed; this method integrates SE module and SEM-ResNet model, and can effectively obtain cross-scale features from PPG signals with multi-information fusion. In addition, the deep network first The empirical knowledge is provided by morphological features and fused with deep features to help the model obtain more discriminative features in the learning process and achieve higher classification accuracy; this method deploys an automatic machine learning ( automated machine learning, AutoML) pipeline to obtain the best BP prediction model without a lot of expert experience; this method designs a feature selection algorithm called PPG morphological feature learning (PPGmorphological feature learning, PMFL), through the visualization The SHapley Additive Explanation (SHAP) value quantifies the degree of contribution of a subset of features to demonstrate the robustness of the multiple morphological features selected by PMFL.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required in the embodiments. Obviously, the accompanying drawings in the following description are only some of the present invention. Embodiments, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without paying creative labor.

FIG. 1 is a schematic flow chart of a noninvasive blood pressure estimation method based on photoplethysmography according to an embodiment of the present invention;

Fig. 2 is a schematic flow chart of the experimental design of the non-invasive blood pressure estimation method based on photoplethysmography according to the embodiment of the present invention;

Figure 3 is a schematic diagram of the statistical distribution of the data set;

Figure 4 is a schematic diagram of visual display of morphological features;

Fig. 5 is an overall network frame diagram;

Fig. 6 is a display diagram of input signal and output result;

Fig. 7 is a schematic diagram of the pseudocode of the two-stage framework algorithm;

Fig. 8 is a schematic diagram of the pseudocode of the PPG morphological feature learning algorithm;

Fig. 9 is a schematic diagram of the pseudocode of the stacking integration optimization algorithm.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

The purpose of the present invention is to provide a non-invasive blood pressure estimation method based on photoplethysmography, which can realize non-invasive blood pressure estimation through photoplethysmography, improve the estimation accuracy, and provide reference for blood pressure disease detection.

In order to make the above objects, features and advantages of the present invention more comprehensible, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

As shown in Figure 1, the non-invasive blood pressure estimation method based on photoplethysmography provided by the embodiment of the present invention includes the following steps:

Step 1: Collect and monitor the photoplethysmography signal of the human body and the invasive continuous blood pressure waveform signal;

Step 2: Initialize the acquired photoplethysmography signal and invasive continuous blood pressure waveform signal;

Step 3: Perform feature selection on the initialized data, screen information parameters with discriminative significance, and perform feature fusion;

Step 4: Build a two-stage model of non-invasive blood pressure estimation based on photoplethysmography based on the data after feature fusion;

Step 5: According to the initialized and processed data, train the two-stage model of non-invasive blood pressure estimation based on photoplethysmography;

Step 6: Collect the user's photoplethysmography signal and input it into the two-stage model of non-invasive blood pressure estimation based on photoplethysmography to perform non-invasive blood pressure estimation.

Step 7: Proving the robustness of the various morphological features selected by the PMFL provided by the embodiments of the present invention through experiments;

Step 8: Quantitative verification and evaluation of the non-invasive blood pressure estimation method based on photoplethysmography provided by the embodiment of the present invention.

In step 1, collect and monitor the photoplethysmography signal of the human body and the invasive continuous blood pressure waveform signal, specifically:

Collect and monitor the photoplethysmographic pulse wave signal of the human body, that is, the PPG signal, and the invasive continuous blood pressure waveform signal, wherein the invasive continuous blood pressure waveform signal is used as a reference signal;

The inventive method adopts MIMIC (multi-parameter intelligent monitoring of intensive care) data set, has included more than 20,000 physiological data recorded from the patient monitor of Boston Beth Israel Hospital's Department of Internal Medicine, Surgery and Intensive Care Unit, and each record usually includes 24 Up to 48 hours of continuous data;

X^t表示PPG信号的输入维度为d_x，一共有N条信号，每条信号有s_x个样本点，而输出是相应的血压预测值(即在时间窗口t内对应的收缩压和舒张压的数值)，Y^t代表具有s_y个样本点的血压序列(即血压波形)，这里计算其相应的血压数值，得到时间窗口t内的包含M个峰值/>

和谷值/>

的稀疏集合/>

The method of the present invention uses a private data set as an independent database to verify the proposed model. This private database contains more than 60,000 pieces of physiological data recorded from the patient monitor in the intensive care unit, and each record usually contains 30 seconds of continuous data. ; The prediction of blood pressure waveform based on PPG signal can be generalized as a real-time long sequence prediction problem. Therefore, the model needs to learn and obtain the mapping relationship between two PPGs and blood pressure values through a time-delay window with a fixed window length. In the method of the present invention, the input data PPG sequence (i.e. PPG waveform) and the corresponding blood pressure waveform for the time window t are expressed as

X ^t indicates that the input dimension of the PPG signal is d _x , there are a total of N signals, each signal has s _x sample points, and the output is the corresponding predicted value of blood pressure (that is, the corresponding systolic and diastolic blood pressure in the time window t value), Y ^t represents the blood pressure sequence (that is, the blood pressure waveform) with s _y sample points, and the corresponding blood pressure value is calculated here to obtain the time window t containing M peaks />

and valley />

sparse collection of />

The input signal dimension of this model is not limited to the single variable case, and the derivative of the PPG sequence can be selected, that is, the velocity sequence of the PPG, and the acceleration sequence of the PPG. The proposed model accepts an input and an output, given by the following formula:

Y _s = F(X; P; θ) (1)

代表引入的先验知识，即通过特征选择算法筛选的PPG波形的形态学特征，为可选参数，θ则代表深度学习模型的超参数。In the formula, F(·) is the model function,

Represents the introduced prior knowledge, that is, the morphological features of the PPG waveform screened by the feature selection algorithm, which is an optional parameter, and θ represents the hyperparameter of the deep learning model.

In step 2, the acquired photoplethysmography signal and invasive continuous blood pressure waveform signal are initialized, specifically:

The concrete experimental design flow chart of the inventive method is as shown in Figure 2, and the PPG and blood pressure signal length that collects in the database are different, and are subjected to the interference of outlier and baseline drift, therefore need to carry out preliminary processing to signal, as shown in Figure 5 As shown, resampling, filtering, noise reduction, signal segmentation, phase matching and normalization processing are carried out on the obtained photoplethysmography signal and invasive continuous blood pressure waveform signal, specifically: the method of the present invention uniformly resamples the signal to 125Hz, Contribute to the time alignment of sequence, in addition, because wavelet transform is suitable for analyzing non-stationary signal, has better time-frequency positioning in signal mutation, compression reconstruction and signal denoising, therefore, the present invention selects sym4 wavelet as the wavelet of basic wavelet Transformation is used to denoise the signal, and the noisy PPG signal is decomposed independently in two stages according to the soft threshold function. Then, the filtered signal is segmented using a time-delay window with a fixed size of 8 seconds and a sliding step of 3 seconds. Finally, , the method of the present invention adopts the maximum and minimum normalization to ensure that the model can quickly converge. After the preliminary data processing, the statistical distribution of the data set is as shown in Figure 3;

Finally, a priori feature extraction is performed on the normalized signal to provide prior information for model decision-making. The normalized signal and the extracted morphological features are used as training data for The two-stage model of non-invasive blood pressure estimation of waves is trained, and the present invention extracts 75 interpretable features from the PPG signal respectively, including time domain, frequency domain and nonlinear features, such as the time parameter signal skewness is formula (2), The non-dimensional index edge factor is formula (3), and the area parameter K value is formula (4), among which, most of the features are shown in Figure 4;

In the formula, p _max , p _min and p _mean respectively represent the maximum value, minimum value and average value of the PPG amplitude within the period;

设置一个比例因子τ来划分时间序列，从而形成一组构建的粗粒度的序列

In addition, the method of the present invention additionally uses pulse rate variability (pulse rate variability, PRV) to describe the change pattern in the sequence window, such as the sample entropy (SampEn) describing the degree of sequence disorder, in particular, multi-scale entropy (MSE) Focusing on quantifying the nonlinear dynamic characteristics of complex systems at multiple scales, the calculation process of MSE is shown in detail here, for the PPG sequence

Set a scale factor τ to divide the time series, thus forming a set of coarse-grained series constructed

Then, SampEn can be derived for coarse-grained sequences. First, construct a continuous sequence vector set with an embedding dimension of m

G _m (j)={g(j+k)}, 0≤k≤m-1, 1≤j≤S _x -m (6)

Secondly, the distance between the m-dimensional coarse-grained vector sequences G _m (i) and G _m (j) is defined as the maximum value of the difference between the corresponding elements of the two vectors, and the calculation formula is:

d[g(i), g(j)]=max(|g(i+k)-g(j+k)|), 0≤k≤m-1 (7)

并找出所有的i的平均值B^m(r)为：Then, set the similarity tolerance r, calculate the number d[g(i), g(j)]≤r of each i(1≤j≤S _x -m) value, and calculate the template matching number B _i and the total distance ratio of numbers

And find the average B ^m (r) of all i as:

然后计算出所有的i的均值A^m(r)为：Increase the vector dimension to m+1, reconstruct m+1 dimension vectors G _m+1 (i) and G _m+1 (j), where, 1≤j≤S _x -m, j≠i, calculate G _m+ The number of distances between ₁ (i) and G _m+1 (j) less than the similarity tolerance value r, that is, the number of template matches A _i , the ratio of the number of template matches A _i to the total distance is

Then calculate the mean A ^m (r) of all i as:

The SampEn of the coarse-grained PPG sequence can be calculated as:

In the method of the present invention, the embedding dimension can be selected as m=2, similar to the threshold r=0.15×SDNN, where SDNN represents the standard deviation of the PPG peak interval.

In step 3, feature selection is performed on the initialized data, information parameters with discriminative significance are screened, and feature fusion is performed, specifically:

Through the training data obtained in the preliminary data processing stage in step 2, a large number of PPG morphological parameters are calculated. However, too many feature sets may contain redundant information, and feature selection and fusion operations are still required to make the deep neural network Better learn morphological-related prior information. Therefore, the normalized signal and the extracted morphological features are obtained, and the morphological features are screened by the PMFL algorithm. The PMFL algorithm is specifically PPG morphological feature learning (PPGmorphological feature learning). learning, PMFL) feature selection algorithm, which is used to screen discriminative features as the prior information of the deep model. This method can also quantify the contribution of the feature subset through the visualized SHAP value to prove the variety of PMFL selection. The robustness of morphological features. Specifically, the PMFL algorithm first uses a base learner to sort and filter a set of relatively important features, and then uses recursive feature elimination (RFE) to arrange them to find an optimal Specifically, the PMFL algorithm is divided into a baseline set generation and a backward elimination stage of feature parameters. The filtering method screens a subset of features with top-k bits for estimating systolic and diastolic blood pressure, and takes the filtered feature set as the baseline set. Then, in the backward elimination stage of feature parameters, the least important features are eliminated by the method of REF in turn, and the remaining feature sets are input into the base regressor to fit the blood pressure value, and the final features are considered according to the optimized regression results The number of sets to determine the final optimized feature set,

According to the final optimized feature set, the method of the present invention considers and designs a feature fusion strategy, combines the depth features acquired by the SEM-ResNet model with the morphological features in the final optimized feature set, and determines the proportion of different feature sets by setting feature weights :

F _f =ε·F _m +(1-ε)·F _d (11)

In the formula, F _f represents fusion features, F _m and F _d represent morphological features and depth features, respectively.

In step 4, a two-stage model of noninvasive blood pressure estimation based on photoplethysmography is built according to the data after feature fusion, specifically:

According to the normalized signal and its corresponding fusion features, a two-stage model of non-invasive blood pressure estimation based on photoplethysmography, that is, the SMART-BP model, is built. The two-stage model of non-invasive blood pressure estimation includes two stages, respectively. Coarse-grained classification phase (CCP) and fine-grained regression phase (fine-grained regression phase, FRP), corresponding to the coarse-grained classification phase (CCP) is based on deep learning without The cuff blood pressure group classification model, that is, the SEM-ResNet model, corresponds to the fine-grained regression phase (FRP), which is a cuff-free blood pressure estimation pipeline model based on automatic machine learning, that is, the Auto-Regressor model ;

For the coarse-grained classification phase (CCP), the method of the present invention uses a deep residual network (residual network, ResNet) that combines a squeeze-and-excitation (SE) module and a multi-scale kernel (SEM-ResNet), which efficiently obtains cross-scale features from multi-information fused PPG signals. In addition, the prior knowledge of deep networks is provided by morphological features and fused with deep features, thus helping the model acquire More discriminant features to achieve higher classification accuracy;

Correspondingly, the SEM-ResNet model is used to divide the obtained user PPG signal into different blood pressure groups through the neural network, for example, hypotension, normal blood pressure and high blood pressure;

In the fine-grained regression phase (FRP), the method of the present invention deploys an automatic machine learning (automated machine learning, AutoML) pipeline for each BP interval to obtain the best BP prediction model without requiring a large number of expert experience;

Correspondingly, the Auto-Regressor model is used to automatically establish a fine-grained regressor for the PPG signal of each blood pressure group for non-invasive blood pressure estimation.

Build the SEM-ResNet model, specifically:

Build the SEM-ResNet model, which can automatically learn the dense morphological features in the PPG waveform, and obtain higher classification accuracy than the shallow model. The method of the present invention deploys multi-scale feature fusion learning to achieve cross-scale information, thereby further Mining discriminative features. Specifically, the SEM-ResNet model includes ResNet, multiple sub-networks and neural networks. Among them, SEM-ResNet uses ResNet as the backbone network, and the residual connection of the fusion SE module is used to learn shared low-level features. The sub-networks use convolution kernels of different scales to jointly learn specific signal features of high-level scales, and the neural network embeds the extracted fusion features to provide prior information;

定义为：The overall framework is shown in Figure 6, the overall structure of SEM-ResNet and the schematic forms of ResNet and CBR-3-1. Conv-3-1 means a one-dimensional convolution operation with a kernel size of 3 and a span size of 1; Batch Norm is an abbreviation for batch normalization; ReLU means a rectified linear unit activation function; MaxPool and AvgPool mean maximum pooling and averaging Pooling, the shared feature map obtained from the single-scale CNN layer, and sent to sub-networks of different scales, for the input fusion feature F _f , the output of the corresponding branch

defined as:

表示由融合特征F_f得到的分支特征，Conv_b和/>

代表单尺度CNN层和特定尺度的子网络，θ_b和/>

分别是网络参数，计算Softmax损失进行单分支模型训练，并计算出每个类别的后验概率为：In the formula,

Indicates the branch feature obtained by the fusion feature F _f , Conv _b and />

represent single-scale CNN layers and scale-specific sub-networks, θ _b and />

They are the network parameters, calculate the Softmax loss for single-branch model training, and calculate the posterior probability of each category as:

是模型为输入的PPG信号分配标签y_i∈{0,1,...,C}的概率，例如，将血压分类为低血压、正常血压和高血压类别可以表示为y_i∈{0,1,2}，θ_k是类别y_i的参数，因此，对于所有可观察的实例，使用交叉熵损失函数，为：In the formula,

is the probability that the model assigns a label y _i ∈ {0,1,...,C} to the input PPG signal, for example, classifying blood pressure into hypotension, normotension and hypertension categories can be expressed as y _i ∈ {0, 1, 2}, θ _k is the parameter of category y _i , so, for all observable instances, using the cross-entropy loss function, is:

In the formula, I{ } refers to the index function;

The present invention uses the SE module to learn the feature channel, which can integrate the time information learned by the one-dimensional convolution operation. The SE module can learn to use global information, selectively emphasize useful features, and suppress other features. Among them, SE-ResNet The structure is shown in Figure 6. For any feature map input F={f ₁ ,f ₁ ,..,f _m }, where f _m ∈ R ^t , the statistic z∈R ^m is compressed by its time dimension t F And resulting, the cth element of z is computed by the squeeze operation:

After aggregation, an excitation operation consisting of two linear layers is performed to generate channel modulation weights. The present invention allows multiple channels to be activated, so a simple gating mechanism is adopted, that is, the Sigmoid activation method is:

s=Excitation(z,θ)=σ(σ(θ ^T z)) (16)

where σ denotes the Sigmoid function, θ is the parameter, and the final output of this module is obtained by rescaling F with the activation s as:

Scale(·)为通道间的乘法；in,

Scale( ) is the multiplication between channels;

The objective function for global fusion of multi-scale features for backbone extraction is defined as:

表示通过串联分支特征的融合特征图，通过联合优化多个分支的损失来训练整个SEM-ResNet模型，可以学习到跨时空尺度的潜在互补信息，在本发明方法中，最终的损失函数如下所示：in,

Indicates that by concatenating the fusion feature maps of branch features and jointly optimizing the losses of multiple branches to train the entire SEM-ResNet model, potential complementary information across temporal and spatial scales can be learned. In the method of the present invention, the final loss function is as follows :

where γ _j is the weight coefficient corresponding to each branch.

Build the Auto-Regressor model, specifically:

Build an Auto-Regressor model, automatically build a fine-grained regressor for the PPG signal of each blood pressure group obtained by the SEM-ResNet model, and obtain accurate non-invasive blood pressure estimation;

Machine learning models are less dependent on data and do not rely on hardware operations. Selecting appropriate features through domain knowledge can also greatly reduce the complexity of the algorithm. However, model building, training, and hyperparameter optimization require a lot of time spent by experienced experts. AutoML, as an easily deployable system, is generally understood as a pipeline that can perform scientific tasks with minimal human intervention. AutoML can optimize algorithms to minimize specific loss functions associated with specific content domains. AutoML algorithms have been successfully implemented using techniques such as meta-learning, reinforcement learning, genetic programming, and superposition ensemble algorithms. AutoML performed well compared to machine learning models manually tuned by data science experts;

In order to automate the pipeline selection, the method of the present invention uses the H2O AutoML framework to define the model. The framework has the following advantages: 1) It outperforms other AutoML frameworks; 2) It has a highly scalable fully automated algorithm that can automatically train a large number of candidate models. Available algorithms are tested in a fixed order with hyperparameters defined by experts or selected by random grid search, and the best performing configurations are aggregated to form an ensemble. The method of the present invention constructs a stacking (Stacking) integrated learning optimization algorithm through the AutoML pipeline. First, a meta-level regressor is trained to find the best combination of base-level regressors. If the prediction errors of the base learners are all low, and there is a significant difference between the constructed models, then the stacked model will perform well. The AutoML pipeline then enters a base-level regression phase, where random searches across different ensembles of models can yield diverse base regressors and, when paired with stacking methods, an influential ensemble. In addition, the method of the present invention sets the meta-level regressor in the meta-level regression stage to use the k-fold cross-validation prediction value of the base regressor for training, and then tests on the data set outside the fold.

In step 5, according to the initialized and processed data, the two-stage model of non-invasive blood pressure estimation based on photoplethysmography is trained, specifically:

Divide the initialized data into training set, verification set and test set according to the ratio of 7:1.5:1.5, train and select parameters for the two-stage model of non-invasive blood pressure estimation through the training set and verification set, and test the saved data through the test set The generalization ability of the optimal system model, in the process of training the network, the parameters are updated through the Adam optimizer, and the data segmentation adopts the method of individual segmentation to ensure that each group of data does not contain the data of the same patient, thereby avoiding information leakage , where the learning rate is 0.001, the weight decay is 0.999, and the momentum is 0.8.

In step 6, the user’s photoplethysmography signal is collected and input into the two-stage model of noninvasive blood pressure estimation based on photoplethysmography to perform noninvasive blood pressure estimation, specifically:

As shown in Figure 6, the user's PPG signal is collected, and the signal waveform is sent as an input into the two-stage model of non-invasive blood pressure estimation to obtain the corresponding blood pressure value of the user to realize non-invasive blood pressure estimation.

In step 7, the robustness of various morphological features selected by the PMFL provided by the embodiment of the present invention is proved through experiments. In order to compare the impact of the proposed PMFL algorithm on BP prediction performance, the patent of the present invention constructs a supervised feature weighting algorithm Relief, a weakly supervised feature selection algorithm (WSF) based on spectral analysis, and a maximum correlation and minimum redundancy based on mutual information. Redundancy algorithm (MRMR), and unsupervised multi-cluster feature selection (MCFS). The patent of the present invention shows the BP prediction errors of various feature selection methods in Table 1, and calculates the robustness index of the corresponding methods. The predictive performance of the subset of features selected by the PMFL algorithm outperforms other methods. In addition, the PMFL algorithm has the lowest ANHI and CD values, indicating a strong consistency of feature weights and rankings. In contrast, PCC has the highest value, implying that the subset of features has the highest correlation and the strongest robustness.

Table 1 Performance comparison with reference feature selection algorithms

Note, *, these are dimensionless quantities; ANHI, mean normalized Hamming index; CD, Canberra distance; PCC, Pearson correlation coefficient.

In Step 8, the predictive performance of the proposed two-stage model for photoplethysmography-based NIBP estimation was experimentally verified. In order to further verify the effectiveness of the proposed method, the patent of the present invention established a variety of structural models for comparison, including: 1) a single-stage AutoML algorithm for direct regression; 2) an end-to-end deep learning algorithm that directly sends PPG signals to into the model and automatically predict the corresponding BP value, including DenseNet, Transformer and SEM-ResNet models; 3) AutoML's two-stage method (AutoML-AutoML) is used in both CCP and FRP. Table 2 shows the comparative results of the experiments. The main quantitative indicators include mean error (ME), mean standard deviation (SDE), and mean absolute difference (MAE). It can be observed that the proposed SMART-BP outperforms other works in most evaluation metrics. For DBP prediction, all models meet the AAMI criteria. For SBP prediction, most models pass the AAMI standard except DenseNet.

Table 2 Performance comparison of different frameworks on private datasets (mmHg).

The present invention proposes a two-stage model for non-invasive blood pressure estimation, which is separately modeled for the potential physiological differences of hypotension, normotension and hypertension groups, named SMART-BP (SeM-resnet and Auto-Regressor based on aTwo-stage framework for Blood Pressure estimation). This approach allows each blood pressure group to be modeled individually for highly accurate blood pressure predictions. SMART-BP is significantly different from existing technologies. It first uses deep learning algorithms to roughly classify the original PPG signal into three BP categories, and then establishes an automatic optimization pipeline of machine learning regressors for each interval to accurately Predict fine-grained BP values;

The input of the present invention is only composed of pulse waves, so the acquisition circuit only needs to acquire PPG signals. Compared with the traditional method based on pulse wave propagation velocity, the step of collecting ECG signals is omitted, and there is no need for too much derivation In this way, it can be easily integrated into devices such as wristbands, without the need for blood pressure measurement equipment such as cuffs, and it is freed from the constraints of cuffs, making the device more portable.

The blood pressure estimation algorithm of the present invention can realize continuous blood pressure estimation and long-term monitoring of blood pressure, and can be used for blood pressure measurement in daily life without causing trauma and discomfort to the human body.

The present invention uses deep learning and automatic machine learning pipelines as the network backbone to extract a large amount of information in PPG signals, combined with morphological features extracted from PPG signals, and introducing these prior information can promote the model to use more discrimination in the learning process The input signal contains richer information, which makes the measured blood pressure result more stable and achieves higher prediction accuracy.

In order to improve the prediction accuracy of blood pressure, the present invention adopts a two-stage framework to perform high-precision automatic modeling for specific groups of people, so as to realize the prediction accuracy meeting medical standards.

The method of the present invention designs a feature selection strategy based on the recursive feature elimination method, which not only can obtain a sparse subset of useful features to improve the fine-grained blood pressure estimation accuracy in the AutoML modeling stage, but also in the process of learning and training, the sequence features captured by the neural network can also be effectively combined.

The algorithm pseudo code used in the present invention is shown in Fig. 7, Fig. 8 and Fig. 9 .

The non-invasive blood pressure estimation method based on photoplethysmography provided by the present invention, the method includes collecting and monitoring photoplethysmogram signals and invasive continuous blood pressure waveform signals of the human body, and analyzing the obtained photoplethysmography signals and invasive continuous blood pressure waveform signals Carry out initialization processing, perform feature selection on the data after initialization processing, screen information parameters with discriminative significance, and perform feature fusion, build a two-stage model of noninvasive blood pressure estimation based on photoplethysmography based on the data after feature fusion, Based on the completed data, the two-stage model of non-invasive blood pressure estimation based on photoplethysmography is trained, and the user’s photoplethysmography signal is collected and input into the two-stage model of non-invasive blood pressure estimation based on photoplethysmography to perform non-invasive blood pressure estimation; This method integrates the SE module and the SEM-ResNet model, and can effectively obtain cross-scale features from the multi-information fusion PPG signal. In addition, the prior knowledge of the deep network is provided by the morphological features and fused with the deep features, thus helping The model obtains more discriminative features during the learning process and achieves higher classification accuracy; this method deploys an automated machine learning (AutoML) pipeline for each BP interval to obtain the best BP prediction model, while Extensive expert experience is not required; the method devises a feature selection algorithm called PPG morphological feature learning (PMFL) to quantify feature subsets by visualizing SHapley Additive explanation (SHAP) values degree of contribution to demonstrate the robustness of PMFL selection for multiple morphological features.

In this paper, specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; meanwhile, for those of ordinary skill in the art, according to the present invention Thoughts, there will be changes in specific implementation methods and application ranges. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims (9)

1. A noninvasive blood pressure estimation method based on photoplethysmography is characterized by comprising the following steps:

step 1: collecting and monitoring photoelectric volume pulse wave signals and invasive continuous blood pressure waveform signals of a human body;

step 2: initializing the obtained photoelectric volume pulse wave signal and the invasive continuous blood pressure waveform signal;

step 3: feature selection is carried out on the initialized data, information parameters with identification significance are screened, and feature fusion is carried out;

step 4: constructing a noninvasive blood pressure estimation double-stage model based on photoelectric volume pulse waves according to the data after feature fusion;

step 5: training a noninvasive blood pressure estimation double-stage model based on the photoelectric volume pulse wave according to the initialized data;

step 6: the method comprises the steps of collecting a photoelectric volume pulse wave signal of a user, inputting the photoelectric volume pulse wave signal into a noninvasive blood pressure estimation double-stage model based on the photoelectric volume pulse wave, and carrying out noninvasive blood pressure estimation.

2. The method for noninvasive blood pressure estimation based on photoplethysmogram according to claim 1, wherein in step 1, the photoplethysmogram signal and the invasive continuous blood pressure waveform signal of the monitored human body are collected, specifically:

the method comprises the steps of collecting a photoelectric volume pulse wave signal of a monitored human body, namely a PPG signal and an invasive continuous blood pressure waveform signal, wherein the invasive continuous blood pressure waveform signal is used as a reference signal.

3. The method for non-invasive blood pressure estimation based on photoplethysmography according to claim 2, wherein in step 2, the acquired photoplethysmography signal and the invasive continuous blood pressure waveform signal are initialized, specifically:

resampling, filtering and denoising, signal segmentation, phase matching and normalization are carried out on the obtained photoelectric volume pulse wave signals and the invasive continuous blood pressure waveform signals, priori feature extraction is carried out on the signals after normalization, morphological features are extracted from the normalized PPG signals, and the signals after normalization and the morphological features obtained through extraction are used as training data for training a noninvasive blood pressure estimation dual-stage model based on the photoelectric volume pulse waves.

4. The method for non-invasive blood pressure estimation based on photoplethysmogram according to claim 3, wherein in step 3, feature selection is performed on the initialized data, information parameters with identification significance are screened, and feature fusion is performed, specifically:

the method comprises the steps of obtaining normalized signals and extracted morphological characteristics, screening the morphological characteristics through a PPG morphological characteristic learning algorithm, firstly sequencing and filtering a group of relatively important characteristics by using a base learner, then iterating by using a recursive characteristic elimination method, finding out the optimal characteristic combination, determining a final optimized characteristic set, fusing depth characteristics obtained by a noninvasive blood pressure estimation dual-stage model based on photoelectric volume pulse waves with morphological characteristics in the final optimized characteristic set according to the final optimized characteristic set, and determining the proportion of different characteristic sets by setting characteristic weights:

F _f ＝ε·F _m +(1-ε)·F _d (1)

in the method, in the process of the invention,F _f represents fusion features, F _m And F _d Respectively representing morphological features and depth features.

5. The method for noninvasive blood pressure estimation based on photoplethysmogram according to claim 4, wherein in step 4, a noninvasive blood pressure estimation dual-stage model based on photoplethysmogram is built according to the feature fused data, specifically:

according to the normalized signals and the fusion characteristics corresponding to the signals, a non-invasive blood pressure estimation double-stage model based on a photoelectric volume pulse wave is built, wherein the non-invasive blood pressure estimation double-stage model, namely a SMART-BP model, consists of a non-cuff blood pressure group type classification model based on deep learning and a non-cuff blood pressure estimation pipeline based on automatic machine learning, namely an SEM-ResNet model and an Auto-regress model, wherein the SEM-ResNet model is used for dividing acquired user PPG signals into different blood pressure groups through a neural network, and the Auto-regress model is used for automatically building a fine-granularity Regressor for the PPG signals of each blood pressure group to perform non-invasive blood pressure estimation.

6. The non-invasive blood pressure estimation method based on photoplethysmography according to claim 5, wherein constructing an SEM-ResNet model specifically comprises:

and establishing an SEM-ResNet model, wherein the SEM-ResNet model comprises a ResNet, a plurality of sub-networks and a neural network, the SEM-ResNet takes the ResNet as a backbone network, residual connection of the fusion extrusion-excitation module is used for learning shared low-level features, the plurality of sub-networks cooperatively learn high-level specific signal features by using convolution kernels with different scales, and the neural network is embedded into the extracted fusion features and is used for providing prior information.

7. The non-invasive blood pressure estimation method based on photoplethysmography according to claim 6, wherein the Auto-regress model is built specifically as follows:

an Auto-regress model is built, a fine-granularity Regressor is automatically built for PPG signals of each blood pressure group obtained by an SEM-ResNet model, accurate noninvasive blood pressure estimation is obtained, a stacked integrated learning optimization algorithm is built through an AutoML pipeline, the Auto-Regressor model is defined through an H2OAutoML frame, firstly, a metalevel Regressor is trained and used for finding the optimal combination of the base level Regressor, if the prediction errors of the base learner are very low and the constructed models have significant differences, the model is judged to perform well, the AutoML pipeline enters a base regression stage, random search is carried out in different model sets, diversified base regressors can be generated, and an influencing set is generated when the Regressor is matched with a stacking method, wherein the metalevel Regressor is trained by using k-fold cross-validation predicted values of the base Regressor in the metalevel regression stage, and a test is carried out on a data set outside fold.

8. The method for non-invasive blood pressure estimation based on photoplethysmogram according to claim 7, wherein in step 5, the non-invasive blood pressure estimation dual-stage model based on photoplethysmogram is trained based on the initialized data, specifically:

dividing initialized data into a training set, a verification set and a test set according to the proportion of 7:1.5:1.5, training and parameter selection are carried out on the noninvasive blood pressure estimation double-stage model through the training set and the verification set, the generalization capability of the stored optimal system model is checked through the test set, and parameter updating is carried out through an Adam optimizer in the process of training a network, wherein the learning rate is 0.001, the weight attenuation is 0.999, and the momentum is 0.8.

9. The method for non-invasive blood pressure estimation based on photoplethysmogram according to claim 8, wherein in step 6, the photoplethysmogram signal of the user is collected and input into a non-invasive blood pressure estimation two-stage model based on photoplethysmogram to perform non-invasive blood pressure estimation, specifically:

and collecting PPG signals of the user, and sending the signal waveforms into a noninvasive blood pressure estimation dual-stage model as input to obtain blood pressure values corresponding to the user, so as to realize noninvasive blood pressure estimation.

Images