A method for estimating invasive blood pressure waveform based on photoelectric volume pulse wave

Abstract

The application provides a method for invasive blood pressure waveform estimation based on photoelectric plethysmogram, comprising: collecting photoelectric plethysmogram signals of a monitored human body, and collecting continuous blood pressure waveforms as reference blood pressure signals; performing data preprocessing to obtain training data; adopting a BEFS algorithm to screen information parameters with discriminative significance as prior information of a deep learning model; building a deep learning model for continuous estimation and monitoring of sleeveless blood pressure waveforms, including a deep feature extraction network, a prior information fusion module and a knowledge distillation training architecture; dividing the training data into a training set, a validation set and a test set, and training and validating the model; collecting user PPG data, inputting the data into the deep learning model and outputting the predicted blood pressure waveform of the user. The application optimizes deep features and integrates morphological features of original PPG to formulate a KD-Informer, provides continuous blood pressure waveform prediction and can better reflect continuous changes of blood pressure in an arterial blood vessel.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical detection, in particular to a method for estimating invasive blood pressure waveform based on photoelectric plethysmogram. BACKGROUND

[0002] Long-term monitoring of blood pressure parameters helps to improve the prognosis effect, thereby reducing the mortality rate of cardiovascular diseases CVD. Therefore, a device capable of non-invasive monitoring of blood pressure for a long time is very important for the diagnosis and treatment of CVD patients. However, the traditional cuff blood pressure measurement device is large in size, inconvenient to carry, and not suitable for long-term monitoring. In order to realize portable, comfortable and long-term blood pressure monitoring, many studies have proposed sleeveless blood pressure measurement technology and devices, which mostly use optical principles to measure photoplethysmography (PPG) signals, from which physiological parameters such as blood oxygen, heart rate, blood pressure, etc. can be derived. This method can obtain continuous and dynamic blood pressure parameters, effectively making up for the inconvenience and intermittent measurement of traditional cuff blood pressure measurement. Many existing methods combine electrocardiogram (ECG) to extract pulse transient time (PTT), i.e. the time interval of pulse wave propagation in blood vessels, and then estimate blood pressure based on pulse wave velocity (PWV) theory. However, so far, there are few mature sleeveless blood pressure monitoring devices based on PWV theory. At present, it has not been fully proven that PTT can accurately estimate blood pressure, and frequent use of a cuff blood pressure meter is required for calibration. Finally, ECG measurement requires electrodes or hand contact to form a closed loop path, which is poor in ease of use, and also makes it difficult to achieve all-weather continuous blood pressure monitoring.

[0003] The above-mentioned method of estimating blood pressure based on PWV theory has certain problems. In comparison, it may be easier to learn deeper physiological relevance by extracting morphological parameters from PPG waveform to establish a blood pressure model. Many studies have also improved signal preprocessing methods, thereby effectively suppressing the interference of baseline drift and abnormal noise commonly found in physiological signals. Deep learning methods are used to directly extract deep features from the original waveform, thereby avoiding the shortcomings of artificially defined features or shallow learning algorithms in fitting blood pressure. In addition, considering that there is more useful information describing cardiovascular function in the original waveform, artificially defined morphological parameters may miss these features, and sequence learning methods have more potential for accurate blood pressure estimation through cross-correlation operations on complete signals. SUMMARY

[0004] The purpose of the present application is to provide a method for invasive blood pressure waveform estimation based on photoplethysmogram, by simultaneously optimizing deep features and integrating original PPG morphological features, a new Informer architecture with a knowledge distillation (KD) strategy (KD-Informer) is developed for blood pressure waveform estimation, which provides continuous blood pressure waveform prediction and better reflects the continuous changes of arterial blood pressure, and has higher prediction value for cardiovascular event prediction and target organ damage.

[0005] To achieve the above purpose, the present application provides the following scheme:

[0006] A method for invasive blood pressure waveform estimation based on photoplethysmogram, comprising the following steps:

[0007] S1, collecting photoplethysmogram signals of the monitored human body, and collecting continuous blood pressure waveform as reference blood pressure signal;

[0008] S2, data preprocessing is performed on the photoplethysmogram signals and reference blood pressure signal fragments obtained in step S1 to obtain training data;

[0009] S3, a large number of PPG morphological parameters are calculated by the training data obtained by data preprocessing in step S2, and then the BEFS algorithm is used to screen information parameters with discriminative significance as prior information of the deep learning model;

[0010] S4, a deep learning model for continuous estimation and monitoring of sleeveless blood pressure waveform is established, a new Informer architecture with KD strategy (KD-Informer) is obtained, the KD-Informer uses Informer as a backbone network for learning shared high-level features, and uses ResNet structure with compression and excitation modules for fusion of prior information, based on the response of the KD strategy, the knowledge learned by the teacher Informer is transferred to the lightweight student Informer;

[0011] S5, the training data obtained by data preprocessing in step S2 is divided into training set, validation set and test set according to a certain proportion, wherein the training set and the validation set are used to train and select parameters for the deep learning model established in step S4, and the test set is used to test the generalization ability of the deep learning model obtained after training and parameter selection;

[0012] S6, collecting the photoplethysmogram signal data of the user to be measured, inputting the deep learning model obtained in step S5, and outputting the predicted blood pressure waveform of the user.

[0013] Further, in the step S2, the data preprocessing includes resampling, filtering and noise reduction, signal segmentation, time alignment and prior feature extraction operations.

[0014] Further, in the step S3, the BEFS algorithm is used to screen information parameters with discriminative significance as prior information of the deep learning model, specifically including:

[0015] The Wrapper and Embedded methods are used to filter out feature subsets with the top k positions respectively, for estimating systolic pressure and diastolic pressure, and the combined set is taken as a baseline set.

[0016] Then, in the backward elimination stage of the feature parameters, the least important features are sequentially eliminated, the remaining feature set is input to the regressor to fit the blood pressure value, and the final feature set is considered according to the optimized regression result as the prior information of the deep learning model.

[0017] Further, in the step S5, the training set, the validation set and the test set are divided in the ratio of 7:1.5:1.5.

[0018] According to the specific embodiments provided by the present application, the following technical effects are disclosed: the method for estimating invasive blood pressure waveform based on photoelectric plethysmogram provided by the present application fully utilizes the advantage of the Transformer model in time dependence analysis, applies the Transformer as the network backbone, combines the morphological features extracted from the PPG signal, and uses the ResNet structure with a squeeze-and-excitation module (SE-ResNet) to fuse it with the deep features; the introduced prior information can promote the model to utilize more discriminative features in the learning process, thereby achieving higher prediction accuracy; at the same time, in order to reduce memory consumption and realize lightweight calculation, the present application adopts a response-based KD strategy to transfer the knowledge learned by the teacher model to the lightweight student model parameters; in addition, by modifying the ChiMerge algorithm, the backward elimination feature selection (BEFS) and the feature fusion method, not only a sparse subset of useful features can be obtained, but also the sequence features captured by the neural network in the learning process can be effectively combined. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0020] Figure 1 is a flow chart of the method for estimating invasive blood pressure waveform based on photoplethysmogram according to the present application;

[0021] Figure 2 is a specific experimental design flow chart of the method for estimating invasive blood pressure waveform based on photoplethysmogram according to the present application;

[0022] Figure 3 is a direct display diagram of the morphological features designed in the data processing according to the present application;

[0023] Figure 4 is the overall network framework of the deep learning model according to the present application;

[0024] Figure 5 is an input signal schematic diagram according to the present application;

[0025] Figure 6 is an output result schematic diagram according to the present application. DETAILED DESCRIPTION

[0026] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0027] The purpose of the present application is to provide a method for estimating invasive blood pressure waveform based on photoplethysmogram, which formulates a new Informer architecture with a knowledge distillation (KD) strategy (KD-Informer) for blood pressure waveform estimation by simultaneously optimizing deep features and integrating morphological features of original PPG, provides continuous blood pressure waveform prediction, and can better reflect the continuous changes of blood pressure in the arterial blood vessels, has higher prediction value for cardiovascular event prediction and target organ damage.

[0028] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0029] As Figure 1As shown, the method for estimating invasive blood pressure waveform based on photoplethysmogram provided by the application comprises the following steps:

[0030] S1, collect the photoplethysmogram signal of the monitored human body, and collect the continuous blood pressure waveform as a reference blood pressure signal; the method of the application uses the MIMIC (Multiple Parameter Intelligent Monitoring of Intensive Care) data set, which contains more than 20000 physiological data recorded from the patient monitor in the medical, surgical and intensive care rooms of Boston Beth Israel Hospital, each record usually contains 24 to 48 hours of continuous data;

[0031] S2, data preprocessing is performed on the photoplethysmogram signal and the reference blood pressure signal segment obtained in step S1 to obtain training data; the data preprocessing mainly includes resampling, filtering and denoising, signal segmentation, time alignment and prior feature extraction operations;

[0032] S3, a large number of PPG morphological parameters are calculated by the training data obtained by the data preprocessing in step S2, and then the BEFS algorithm is used to screen information parameters with discriminative significance as prior information of the deep learning model; the BEFS algorithm is divided into baseline set generation and backward elimination stage of feature parameters, Wrapper and Embedded methods are used to filter out the top k feature subsets respectively, which are used to estimate systolic pressure and diastolic pressure, and the combined set is used as the baseline set; then, in the backward elimination stage of feature parameters, the least important features are eliminated in turn, and the remaining feature set is input into the regressor to fit the blood pressure value, and the number of the final feature set is considered according to the optimized regression result; in addition, this step also considers the method design of feature fusion of prior information and deep information extracted by neural network;

[0033] S4, a deep learning model for continuous estimation and monitoring of cuffless blood pressure waveform is built, as shown in Figure 2 As shown, a new KD-Informer architecture with KD strategy is obtained, KD-Informer uses Informer as the backbone network for learning shared high-level features, uses ResNet structure with compression and excitation modules for fusion of prior information, and based on the response of the KD strategy, the knowledge learned by the teacher Informer is transferred to the lightweight student Informer; wherein, Informer is the main framework, blood pressure waveform prediction mainly focuses on the mapping relationship between different modal physiological sequences, therefore the sparse probability multi-head self-attention (MHSA) structure of the mask in the original network is deleted to avoid sequence information leakage;

[0034] S5, the training data obtained by data preprocessing in step S2 is divided into training set, validation set and test set according to a certain proportion, wherein the training set and the validation set are used to train and select parameters of the deep learning model established in step S4, and the test set is used to test the generalization ability of the deep learning model after training and parameter selection; the network parameters are updated by continuously sending the training data to the system model in batches and training the network, and after repeating the above operation N times, the obtained parameters are regarded as optimal parameters;

[0035] S6, collecting the to-be-measured photoplethysmography (PPG) signal data of the user, inputting the deep learning model obtained in step S5, and outputting the predicted blood pressure waveform of the user.

[0036] Wherein, predicting the blood pressure waveform based on the PPG signal can be summarized as a real-time long sequence prediction problem. Therefore, the model needs to learn and obtain the mapping relationship between the two sequences through a fixed window length time lag window. In the method of the application, for the input data PPG sequence (i.e. PPG waveform) of the time window t, the output is the corresponding blood pressure sequence (i.e. blood pressure waveform). Unlike the single variable prediction problem, the feature dimension of the model is not limited to the case of a single variable. The derivative of the PPG sequence, i.e. the velocity sequence of the PPG, the acceleration sequence of the PPG, can be selected. The recommended model accepts input and output, which is given by formula (1).

[0037] Y=F(X;M;θ) (1)

[0038] Wherein, F(·) is a model function, represents the introduction of prior knowledge, which is an optional parameter, such as the morphological features of the PPG waveform; and theta represents the hyperparameters of the deep learning model.

[0039] Figure 2 The experimental design flowchart of the method of the application is illustrated. The PPG and blood pressure signals collected in the database are of different lengths and are disturbed by outliers and baseline drift, so the signals need to be preprocessed, as shown in the block in Figure 3 The signal is uniformly resampled to 125Hz in the method of the application, which helps to align the time of the sequence. In addition, since wavelet transform is suitable for analyzing non-stationary signals, it has better time-frequency positioning in signal mutation, compression reconstruction and signal denoising. Therefore, wavelet transform is selected with sym4 wavelet as the basis wavelet to denoise the signal, and the noise PPG signal is independently decomposed in two stages according to the soft threshold function. Then, the filtered signal is segmented using a fixed size of 8 seconds and a sliding step of 3 seconds. Finally, the maximum and minimum normalization is adopted in the method of the application to ensure that the model can converge quickly. The MIMIC data set consists of 80454 8-second segments, with an average systolic pressure of 140.21 ± 28.97 mmHg and a diastolic pressure of 68.62 ± 15.60 mmHg.

[0040] Finally, morphological features are extracted from the normalized PPG signal to provide prior information for the model decision. In total, 75 interpretable features are extracted from the PPG signal, including time domain, frequency domain and nonlinear features, such as the time parameter signal skewness (Equation 2), the dimensionless indicator edge factor (Equation 3), and the area parameter K value (Equation 4). Most of the features are shown in Table 1. Figure 4

[0041]

[0042]

[0043]

[0044] where p max , p min and p mean represent the maximum, minimum and mean of the PPG amplitude within a cycle, respectively.

[0045] In addition, the pulse rate variability (PRV) is used to describe the variation pattern within the sequence window, such as the approximate entropy (ApEn) to describe the degree of sequence disorder. For the PPG sequence x, the embedded sequence W = {w(i), i = 1, 2,..., N + m - 1} is obtained using a window cut-off length of m = 2, a similarity tolerance of r = 0.1 x SD, and the ApEn is defined as.

[0046]

[0047] where represents the average similarity rate of all length m sub-segments of the entire sequence under the similarity criterion r, and the calculation formula is as follows:

[0048]

[0049] where w(j) satisfies max(|w(i)-w(j)|)≤r) / (N-m+1).

[0050] ​In addition, the method of the present application discretizes the calculated continuous features by supervised ChiMerge binning, which has the following advantages: 1) Discrete features have better stability for abnormal data; 2) Sparse vector inner product multiplication operation is relatively efficient; 3) Discrete features help to simplify the model and reduce the risk of overfitting. ChiMerge relies on the chi-square test to combine adjacent intervals with low chi-square values (indicating that they have similar class distributions) until the stopping criterion is met. The calculation method is shown in formula (7), A ij represents the number of instances in the ith interval, the jth class, respectively, and E ij represents the expected frequency of A ij .

[0051]

[0052] The set of continuous features can be represented as where is called the s-dimensional feature set collection at time t, represents the average blood pressure value of the tth feature set. The method of the present application is further improved, first, the blood pressure value is equally divided; then the value of the corresponding class is calculated, and the smallest interval is combined, so as to realize feature discretization.

[0053] A large number of PPG morphological parameters are calculated by the training data obtained in the data preprocessing stage of step S2, however, too many feature sets may contain redundant information, and feature selection and fusion operations are still needed to make the deep neural network better learn the morphological related prior information. In addition, this step also considers the design of feature fusion method between prior information and deep information extracted by neural network. The BEFS algorithm is designed in the method of the present application to screen features with discriminant significance as prior information of deep model.

[0054] After obtaining the optimized feature set, the method of the present application considers designing a feature fusion strategy to combine the deep features of the KD-Informer encoder with the traditional morphological features. The classical fusion method is to determine the proportion of different feature sets by setting feature weights, and the calculation method is as follows:

[0055] F f = ε·F m + (1-ε)·F d (8)

[0056] where F f represents the fused feature, F m and F dThese represent morphological features and depth features, respectively. However, while this linear fusion method is simple to deploy, the hyperparameters are difficult to automatically adjust according to the data distribution, requiring extensive experimentation to select a suitable scaling factor. Therefore, this invention designs a novel feature fusion method that can adaptively adjust the feature scaling factor. This method updates the feature weights by training a multilayer perceptron, and the fusion operation is defined as follows.

[0057]

[0058] in, b is the bias value. The nonlinear spatial mapping of fully connected layers can better balance prior knowledge and information from deep features.

[0059] Considering the high computational complexity and large storage requirements of deep learning models, in step S4, the method of this invention employs a response-based KD strategy to transfer the knowledge learned by the teacher Informer to a lightweight student Informer. The main idea of ​​this method is to train an efficient student model under the guidance of the teacher model to achieve considerable accuracy. The implicit knowledge learned from the pre-trained teacher model can help the student model mimic the behavior of the teacher model; the specific extraction process is as follows... Figure 4 As shown. The soft odds (logits) of the teacher model contain useful information about a specific class and are used for softmax regression tasks. The temperature parameter T can control the importance of each label in a classification task; for example, higher temperatures are more important for a specific class z. i A weaker probability distribution p is generated. i (As shown in Equation 10). However, the output of the blood pressure prediction task is a time series, not a probability distribution or single-class value as in a classification task. In regression tasks, there are no such soft odds (logits); therefore, the method of this invention does not use the temperature parameter, but instead uses the predicted output of the teacher model as a soft label to guide the student's training. The KD-Informer loss can be defined as Equation (11). The first term is the distillation loss, which measures the distance between the teacher's soft prediction and the student's prediction; the second term is the student loss, which measures the distance between the student's prediction and the reference blood pressure sequence. Hyperparameters are used to adjust the degree of contribution between the two loss functions.

[0060]

[0061]

[0062] Step S5 divides the data after step S2 preprocessing into training set, validation set and test set according to the ratio of 7:1.5:1.5, wherein the training set and the validation set train and select parameters for the system model established in step 4, and the test set is used for final inspection of the generalization ability of the optimal system model saved in step S5. The data division adopts an inter-individual segmentation method to ensure that each group of data does not contain data of the same patient, thereby avoiding information leakage. In the process of training the network, the Adam optimizer is used to update the parameters, the learning rate is 0.001, the weight decay is 0.999, and the momentum is 0.8.

[0063] In step S6, the PPG signal of the user is collected (as shown in Figure 5 The signal waveform is input into the deep learning model, and the blood pressure signal of the user can be automatically output, as shown in Figure 6 The non-invasive estimation of the invasive blood pressure waveform based on the photoplethysmogram is realized.

[0064] The method for estimating the invasive blood pressure waveform based on the photoplethysmogram provided by the application has the following advantages: (1) the input 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 speed, the step of acquiring electrocardiogram signals is omitted, and excessive derivative calculation is also unnecessary. Therefore, the device can be conveniently integrated into a bracelet or other equipment, and does not require a blood pressure measuring device such as a cuff, thereby freeing the user from the constraints of the cuff and making the device more portable; (2) the blood pressure estimation method of the application can realize continuous blood pressure waveform estimation, long-term blood pressure monitoring, and blood pressure measurement in daily life, and does not cause trauma or discomfort to the human body during measurement; (3) the application applies a Transformer as the network backbone to extract a large amount of information from the PPG signal, combines morphological features extracted from the PPG signal, and introduces these priori information to promote the model to utilize more discriminative features in the learning process. The input signal contains more information, so that the measured blood pressure result is more stable and has higher prediction accuracy; (4) the application adopts a knowledge distillation strategy to transfer the knowledge learned by the teacher model to the parameters of the lightweight student model, and embeds the student model into the actual measurement device, which can greatly reduce the cost of the device; (5) the method of the application designs a detailed feature selection strategy, which can not only obtain a sparse subset of useful features, but also effectively combine the sequence features captured by the neural network in the learning and training process through the modified ChiMerge algorithm, backward elimination feature selection (BEFS) and feature fusion method.

[0065] The principles and implementation manners of the present application are described by using specific examples in the present application, and the above examples are only used to help understand the method of the present application and its core idea; meanwhile, for the general technical personnel in the art, the specific implementation manners and application ranges will be changed according to the idea of the present application. In conclusion, the content of the present specification should not be understood as the limitation of the present application.

Claims

1. A method for invasive blood pressure waveform estimation based on photoplethysmography (PPG), characterized in that, Includes the following steps: S1 collects and monitors the photoplethysmography (PPG) signal of the human body, and simultaneously collects continuous blood pressure waveforms as reference blood pressure signals. S2, perform data preprocessing on the photoplethysmography pulse wave signal and reference blood pressure signal segment obtained in step S1 to obtain training data; S3, using the training data obtained through the data preprocessing in step S2, calculates a large number of PPG morphological parameters. Then, the BEFS algorithm is used to select informational parameters with discriminative significance as prior information for the deep learning model, specifically including: The Wrapper and Embedded methods were used to filter out the top k feature subsets, which were then used to estimate systolic and diastolic blood pressure. The merged set was then used as the baseline set. Then, in the backward elimination stage of the feature parameters, the features with the lowest importance are eliminated in turn, and the remaining feature set is input into the regressor to fit the blood pressure value. The final feature set is considered based on the optimized regression result as prior information of the deep learning model. S4. A deep learning model for continuous estimation and monitoring of cuffless blood pressure waveforms is constructed, resulting in a novel Informer architecture, KD-Informer, with a KD strategy. KD-Informer uses an Informer as the backbone network for learning shared high-level features and employs a ResNet structure with compression and activation modules for fusing prior information. Based on the response-based KD strategy, the knowledge learned by the teacher Informer is transferred to the lightweight student Informer. A new feature fusion method adaptively adjusts the feature scaling factor. This method updates the feature weights by training a multilayer perceptron. The fusion operation is defined as follows: ； in, b is the bias value; the nonlinear spatial mapping of the fully connected layer can balance prior knowledge and information from deep features; S5. The training data obtained from the data preprocessing in step S2 is divided into training set, validation set and test set according to a certain ratio. The training set and validation set are used to train and select parameters for the deep learning model established in step S4. The test set is used to test the generalization ability of the deep learning model obtained after training and parameter selection. S6: Collect the user's photoplethysmography (PPG) signal data, input it into the deep learning model obtained in step S5, and output the user's predicted blood pressure waveform.

2. The method for invasive blood pressure waveform estimation based on photoplethysmography (PPG) according to claim 1, characterized in that, In step S2, the data preprocessing includes resampling, filtering and noise reduction, signal segmentation, time alignment, and prior feature extraction.

3. The method for invasive blood pressure waveform estimation based on photoplethysmography (PPG) according to claim 1, characterized in that, In step S5, the training set, validation set, and test set are divided in a ratio of 7:1.5:1.5.

Images