Physiological semantic consistent cuff-free blood pressure model comparative learning pre-training method

Abstract

The invention provides a physiological semantic consistent cuff-free blood pressure model comparative learning pre-training method, and the method comprises the steps: firstly carrying out the unsupervised training of a model through employing a label-free PPG signal, so as to improve the generalization capability of the model; the method comprises the steps that firstly, a PPG signal with a label is adopted for conducting supervised training on a model to train the prediction performance of the model, finally, a small number of PPG signals with the label of a target object are adopted for conducting personalized training on the model to enable the model to have the performance of predicting blood pressure data of the target object, and in the unsupervised training process, the prediction performance of the model is improved. Data enhancement is realized based on blood pressure key physiological knowledge to ensure the quality and practicability of enhanced data, and the diversity of model extraction features is improved through random feature masking and feature resampling. Compared with an existing blood pressure prediction method, the demanded quantity of label data in the training process is reduced, and meanwhile, data enhancement and feature reconstruction are carried out based on the blood pressure key physiological information so as to improve the generalization ability of a blood pressure prediction model.

Description

Technical Field

[0001] The present invention relates to the fields of medical health, physiological measurement, deep learning, data enhancement, and more specifically to a comparative learning pre-training method for a cuffless blood pressure model with consistent physiological semantics. Background Art

[0002] According to statistics from the World Health Organization, about 1.65 million people die from cardiovascular diseases each year, among which hypertension is one of the most important factors. Therefore, accurate and continuous blood pressure prediction is crucial for the early detection, prevention and treatment of hypertension and related cardiovascular diseases. Traditional blood pressure measurement technology is invasive or based on inflatable cuffs. Invasive blood pressure measurement is invasive and may even cause infection, and the extremely high cost also limits its promotion and use. Blood pressure measurement devices based on inflatable cuffs have a poor user experience due to cuff expansion, making them unsuitable for chronic disease management and community health screening.

[0003] In the past few decades, scholars have tried to develop a cuff-free, non-invasive, continuous blood pressure prediction technology based on physiological signals. Among them, photoplethysmography (PPG) is a simple and economical optical method to detect changes in blood volume in the peripheral circulatory system. In recent years, researchers have applied deep learning technology to predict blood pressure due to its advantages such as large model capacity and good robustness to noisy signals. Existing blood pressure prediction methods based on deep models require a large amount of data, but repeated cuff inflation is very time-consuming and uncomfortable, so it is very laborious to obtain blood pressure labels for PPG signals. On the other hand, currently popular wearable devices such as Huawei bracelets, Xiaomi Watch, and Apple Watch have a large amount of unlabeled PPG signal data. These wearable devices have a wider range of daily activities and demographic data, and their potential remains to be developed. Contrastive learning is a promising self-supervised paradigm that can use unlabeled data for representation learning. Contrastive learning has made significant progress in the field of computer vision. However, compared with methods for computer vision tasks, time series enhancement methods must consider the semantic uncertainty that arises during data enhancement. Therefore, there are still many problems to be solved when directly applying contrastive learning to the task of blood pressure prediction.

[0004] In summary, in the current research on blood pressure prediction models, model training is overly dependent on the feature distribution of specific data sets and the blood pressure prediction model trained on the basis of limited labeled data has insufficient generalization ability, which limits the model's ability to be applied to new data or different patient groups. In addition, although contrastive learning can improve the generalization ability of the model, existing contrastive learning faces challenges in semantic consistency when enhancing physiological time series data. Traditional data enhancement techniques are not suitable for time series data, which will destroy the time dependency in physiological signals and ignore the inherent complexity of physiological signals. Although the current time series data enhancement methods mainly focus on the time domain and frequency domain of the signal, they will also destroy the time dependency in physiological signals. Therefore, it is crucial to improve the generalization ability of the model and ensure the semantic consistency between the physiological time series data after data enhancement and the original data.

[0005] It should be noted that this background technology is only used to introduce the relevant information of the present invention to help understand the technical solution of the present invention, but it does not mean that the relevant information is necessarily the prior art. If there is no evidence that the relevant information has been disclosed before the application date of the present invention, the relevant information shall not be regarded as the prior art. Summary of the invention

[0006] Therefore, the purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a cuffless blood pressure model comparative learning pre-training method with consistent physiological semantics. The purpose of the present invention is achieved through the following technical solutions.

[0007] According to a first aspect of the present invention, a method for constructing a blood pressure prediction model is proposed, wherein the blood pressure prediction model is used to predict the corresponding systolic and diastolic pressures of a target object based on its photoplethysmogram. The method comprises: step S1, obtaining a first training set comprising a plurality of unlabeled photoplethysmograms, a second training set comprising a plurality of labeled photoplethysmograms, and a third training set comprising a plurality of labeled photoplethysmograms of the target object, wherein the label of each photoplethysmogram is its corresponding systolic and diastolic pressures; step S2, constructing an initial model, wherein the initial model comprises a feature encoding module, a feature reconstruction module, a projection module, and a prediction module, wherein the feature encoding module is used to encode the photoplethysmogram into a feature representation, and the feature reconstruction module is used to resample the feature representation to obtain a new feature representation of the photoplethysmogram. The projection module is used to convert the new feature representation of the photoplethysmogram into its corresponding low-dimensional feature vector, and the prediction module is used to predict the systolic and diastolic pressures corresponding to the photoplethysmogram based on the low-dimensional feature vector of the photoplethysmogram; step S3, based on the key physiological information of blood pressure, the photoplethysmogram in the first training set is data enhanced to obtain multiple positive sample pairs, each positive sample pair includes the original photoplethysmogram and the photoplethysmogram after data enhancement, and the initial model is trained by contrastive learning based on all the positive sample pairs, and the parameters of the feature encoding module, the feature reconstruction module, and the projection module are adjusted based on the preset first loss function; step S4, based on the second training set, the model trained in step S3 is supervised trained and the model parameters are adjusted based on the preset second loss function to obtain an initial blood pressure prediction model; step S5,.

[0008] According to the first aspect of the present invention, the feature encoding module is based on the ConvNeXt network architecture and consists of four consecutive convolutional neural networks, wherein the number of convolutional blocks of the four consecutive convolutional neural networks is configured as (2, 2, 6, 2) in sequence, and the number of channels of the four consecutive convolutional neural networks is configured as (64, 128, 256, 512) in sequence.

[0009] According to the first aspect of the present invention, the feature reconstruction module is configured to: sequentially use random feature masking and feature resampling to reconstruct the feature representation encoded by the encoding module to obtain a new feature representation, and dynamically adjust the weight of the feature resampling based on the self-attention mechanism.

[0010] According to the first aspect of the present invention, the projection module is a fully connected layer including a preset number of neurons.

[0011] According to a first aspect of the present invention, the prediction module consists of two fully connected layers and uses a rectified linear unit as an activation function.

[0012] According to a first aspect of the present invention, data enhancement is performed on each photoelectric volumetric pulse wave in a first training set based on key physiological information of blood pressure in the following manner: all key signal sampling points in the photoelectric volumetric pulse wave containing key physiological information of blood pressure are obtained, wherein the key signal sampling points include signal sampling points corresponding to the start of all systolic pulses, signal sampling points corresponding to the maximum peak value during the ejection period of all systolic blood, signal sampling points corresponding to all transitions from systolic to diastolic blood, signal sampling points corresponding to the lowest pressure in all arterial systems, and all morphological feature points of the accelerated photoelectric volumetric pulse wave corresponding to the photoelectric volumetric pulse wave signal, and signal sampling points corresponding to the diastolic peak, wherein the accelerated photoelectric volumetric pulse wave is the second-order derivative of the photoelectric volumetric pulse wave. ; Based on the acquired signal sampling points, the photoplethysmogram is data enhanced by any of the following methods: Method 1, which is a data enhancement operation of adding noise to the amplitude of other signal sampling points in the photoplethysmogram except the key signal sampling point and its adjacent signal sampling points, wherein the noise is a value randomly generated by a pre-configured normal distribution probability density function; Method 2, which is a data enhancement operation of adjusting the amplitude corresponding to the signal sampling points in a part of the sampling interval of the photoplethysmogram except the key signal sampling point and its adjacent signal sampling points to 0 and keeping the amplitude of other signal sampling points unchanged; Method 3, which is a data enhancement operation of rolling the photoplethysmogram to multiple signal sampling points, wherein the number of signal sampling points of the rolling displacement is determined by the following method:

[0013]

[0014] Wherein, T represents the number of signal sampling points of rolling displacement, M represents the total number of signal sampling points corresponding to the beginning of the systolic pulse, and O t represents the index of the signal sampling point corresponding to the start of the t-th systolic pulse, O t-1 Represents the index of the signal sampling point corresponding to the beginning of the t-1th cardiac systolic pulse; Method 4, which is to divide the photoelectric volumetric pulse wave into multiple sub-waves, and perform data enhancement operations on all the sub-waves in random arrangement, wherein each sub-wave is a waveform between the signal sampling points corresponding to the beginning of two adjacent cardiac systolic pulses; Method 5, which is to perform data enhancement on the photoelectric volumetric pulse wave by sequentially using the operations in Method 1 and Method 4.

[0015] According to the first aspect of the present invention, the preset first loss function is configured as:

[0016]

[0017] in, represents the contrastive learning loss, z i 、z jrepresents the original photoplethysmogram in a positive sample pair and the corresponding feature vector of the photoplethysmogram after data enhancement, sim(·) is the function for calculating the similarity of two feature vectors, N is the number of positive sample pairs processed simultaneously in one training process, exp(·) represents the exponential function, τ represents the hyperparameter, z k It represents the kth eigenvector among the 2N eigenvectors corresponding to the N positive sample pairs processed simultaneously in a training process. f(k,i) is the indicator function. When k≠i, the value of f(k,i) is 1, and when k=i, the value of f(k,i) is 0.

[0018] According to the first aspect of the present invention, the preset second loss function is configured as:

[0019]

[0020] Among them, l MAE represents the mean absolute error loss, n represents the number of photoplethysmograms in the third training set, represents the predicted systolic and diastolic pressures corresponding to the i-th photoplethysmogram in the third training set, Represents the true systolic pressure and true diastolic pressure corresponding to the i-th photoplethysmogram in the third training set.

[0021] According to the second aspect of the present invention, a blood pressure prediction method is proposed, which is used to predict the corresponding systolic and diastolic blood pressure of a target object based on its photoplethysmogram. The method comprises: obtaining the photoplethysmogram of the target object; using a target blood pressure prediction model constructed by any method described in the first aspect of the present invention to predict the systolic and diastolic blood pressure of the target object based on the photoplethysmogram of the target object.

[0022] According to a third aspect of the present invention, a blood pressure prediction system is proposed, which includes: a feature encoding module for encoding a photoplethysmogram into a feature representation; a feature reconstruction module for resampling the feature representation to obtain a new feature representation of the photoplethysmogram; a projection module for converting the new feature representation of the photoplethysmogram into its corresponding low-dimensional feature vector; and a prediction module for predicting the systolic and diastolic pressures corresponding to the photoplethysmogram based on the low-dimensional feature vector of the photoplethysmogram.

[0023] According to a fourth aspect of the present invention, an electronic device is proposed, comprising: one or more processors; and a memory, wherein the memory is used to store executable instructions; the one or more processors are configured to implement the steps of any one of the methods described in the first aspect of the present invention by executing the executable instructions.

[0024] Compared with the prior art, the advantages of the present invention are:

[0025] The present invention proposes a blood pressure prediction model construction method and a blood pressure prediction method by combining contrastive learning and data enhancement technology. The method first uses unlabeled PPG signals to perform unsupervised training on the model to improve the generalization ability of the model, and then uses labeled PPG signals to perform supervised training on the model to train the prediction performance of the model. Finally, a small amount of labeled PPG signals of target objects are used to perform personalized training on the model so that the model has the performance of predicting the blood pressure data of the target object. In the unsupervised training process, data enhancement is implemented based on key physiological knowledge of blood pressure to ensure the quality and practicality of the enhanced data, and random feature masking and feature resampling are used to improve the diversity of model extracted features. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The embodiments of the present invention are further described below with reference to the accompanying drawings, in which:

[0027] Figure 1 A step diagram of a method for constructing a blood pressure prediction model according to an embodiment of the present invention;

[0028] Figure 2 A schematic diagram of key signal sampling points according to an embodiment of the present invention;

[0029] Figure 3 A data enhancement comparison diagram according to an embodiment of the present invention;

[0030] Figure 4 A method for constructing a blood pressure prediction model according to an embodiment of the present invention;

[0031] Figure 5 A comparative experimental result diagram according to an embodiment of the present invention;

[0032] Figure 6 Graph showing ablation experiment results according to an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail by specific embodiments below. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0034] As mentioned in the background technology section, in the current research on blood pressure prediction models, model training is overly dependent on the feature distribution of a specific data set and the blood pressure prediction model trained on limited labeled data has insufficient generalization ability, which limits the model's ability to be applied to new data or different patient groups. In addition, although contrastive learning can improve the generalization ability of the model, existing contrastive learning faces challenges in semantic consistency when enhancing physiological time series data. Traditional data enhancement techniques are not suitable for time series data, which will destroy the time dependency in physiological signals and ignore the inherent complexity of physiological signals. Current time series data enhancement methods mainly focus on the time domain and frequency domain of signals, which will also destroy the time dependency in physiological signals.

[0035] Before specifically introducing the embodiments of the present invention, some of the terms used therein are explained as follows:

[0036] Photoplethysmography (PPG): Photoplethysmography is a non-invasive biometric technology that obtains pulse wave signals by measuring changes in blood volume in skin blood vessels. PPG technology uses LED light sources and detectors to capture skin color changes caused by the cardiac cycle through transmission or reflection, thereby recording the pulsation state of blood vessels and measuring pulse waves.

[0037] In order to better understand the present invention, the inventive concept of the present invention is briefly introduced first. In the scheme of the present invention, the model is firstly trained unsupervisedly using unlabeled PPG signals to improve the generalization ability of the model, and then the model is supervised trained using labeled PPG signals to train the prediction performance of the model, and finally a small number of labeled PPG signals of the target object are used to perform personalized training on the model so that the model has the performance of predicting the blood pressure data of the target object, wherein, in the unsupervised training process, data enhancement is implemented based on key physiological knowledge of blood pressure to ensure the quality and practicality of the enhanced data, and random feature masking and feature resampling are used to improve the diversity of model extracted features.

[0038] In summary, the present invention proposes a method for constructing a blood pressure prediction model, wherein the blood pressure prediction model is used to predict the corresponding systolic and diastolic blood pressure of a target object according to the photoplethysmogram of the target object. The steps of the method are as follows: Figure 1As shown, it includes: step S1, obtaining a first training set including multiple unlabeled photoplethysmograms, a second training set including multiple labeled photoplethysmograms, and a third training set including multiple labeled photoplethysmograms of target objects, wherein the label of each photoplethysmogram is its corresponding systolic pressure and diastolic pressure; step S2, building an initial model, the initial model includes a feature encoding module, a feature reconstruction module, a projection module, and a prediction module, the feature encoding module is used to encode the photoplethysmogram into a feature representation, the feature reconstruction module is used to resample the feature representation to obtain a new feature representation of the photoplethysmogram, the projection module is used to convert the new feature representation of the photoplethysmogram into its corresponding low-dimensional feature vector, and the prediction module is used to convert the low-dimensional feature vector of the photoplethysmogram based on the feature representation of the photoplethysmogram. Predict the systolic and diastolic pressures corresponding to the photoplethysmogram; step S3, based on the key physiological information of blood pressure, perform data enhancement on the photoplethysmogram in the first training set to obtain multiple positive sample pairs, each positive sample pair includes the original photoplethysmogram and the photoplethysmogram after data enhancement, train the initial model based on all the positive sample pairs by contrastive learning, and adjust the parameters of the feature encoding module, the feature reconstruction module, and the projection module based on the preset first loss function; step S4, based on the second training set, perform supervised training on the model trained in step S3, and adjust the model parameters based on the preset second loss function to obtain an initial blood pressure prediction model; step S5, based on the third training set, perform personalized supervised training on the initial blood pressure prediction model, and adjust the model parameters based on the preset second loss function to obtain a target blood pressure prediction model.

[0039] In order to better understand the present invention, each step is described in detail below in conjunction with specific embodiments.

[0040] According to one embodiment of the present invention, in step S1, the photoplethysmograms in the first training set and the second training set are photoplethysmograms of non-target objects. The present invention uses unlabeled photoplethysmograms to train the initial model, which can effectively mine the value of a large amount of unlabeled PPG data collected by current wearable devices, thereby fully releasing the potential of these data.

[0041] According to one embodiment of the present invention, the feature encoding module is configured to be composed of four consecutive convolutional neural networks based on the ConvNeXt network architecture, wherein the number of convolution blocks of the four consecutive convolutional neural networks is configured to be (2, 2, 6, 2) in sequence, and the number of channels of the four consecutive convolutional neural networks is configured to be (64, 128, 256, 512) in sequence.

[0042] ConvNeXt is a variant of Convolutional Neural Network (CNN), which improves and optimizes the traditional CNN architecture. ConvNext improves the performance and efficiency of image processing and pattern recognition tasks. The key features of ConvNeXt include: Convolutional Block: It is the basic building block of ConvNeXt. Each block consists of depth-separable convolution, layer normalization (LayerNorm), GELU activation function and point-by-point convolution. These components work together on input features to achieve efficient feature extraction and transformation; Number of neural network channels: ConvNeXt carefully configures the number of channels at different stages to adapt to feature representation and computing requirements of different scales. For example, the number of channels of ConvNeXt-T is configured as (96, 192, 384, 768), while that of ConvNeXt-B is (128, 256, 512, 1024). These configurations reflect the adjustment of feature depth and width by the model at different stages.

[0043] According to one embodiment of the present invention, the feature reconstruction module sequentially masks and resamples the feature representations encoded by the encoding module. By randomly eliminating features, the model obtains different feature combinations in each training process, increasing feature diversity, and then resamples the features, wherein a self-attention mechanism is used to dynamically assign different weights to each resampled feature, thereby optimizing the feature weighting process.

[0044] According to one embodiment of the present invention, the projection module is configured as a fully connected layer containing a preset number of neurons. Preferably, the projection module of the present invention is configured as a fully connected layer (also called a dense layer or a linear layer) containing 256 neurons. The main function of the projection module is to map the feature representation output by the feature reconstruction module to a low-dimensional feature vector. In the unsupervised training phase of step S3, the projection module assists the encoding module training and the feature reconstruction module to ensure that the model learns a more robust feature representation. Specifically, on the one hand, it helps to reduce overfitting and may improve the generalization ability of the model. On the other hand, it helps the model to concentrate on learning the most important information, thereby enhancing the robustness of the encoding module.

[0045] According to one embodiment of the present invention, the prediction module is composed of two fully connected layers and uses a modified linear unit as an activation function for supervised training and prediction. Preferably, the number of neurons in the two fully connected layers is 64 and 2, respectively. The prediction module receives label information during the training process and optimizes the model according to the label information. During the prediction process, the corresponding diastolic and systolic pressures are predicted according to the feature vector output by the projection module. The activation function of the prediction module is used to increase nonlinearity and help the model learn complex feature relationships. Specifically, the first fully connected layer of the prediction module (containing 64 neurons in full connection) further transforms the output of the projection head, and the second fully connected layer (containing 2 neurons in full connection) generates the final prediction result, wherein the 2 neurons in the second layer correspond to the output of the binary classification task (i.e., systolic and diastolic pressure) of the present invention.

[0046] According to one embodiment of the present invention, in step S3, key signal sampling points containing key physiological knowledge in the photoplethysmogram are extracted based on key physiological information of blood pressure, and the data enhancement operation of these signal sampling points and their neighboring signal sampling points is constrained during the data enhancement process to avoid the data enhancement operation from destroying the time dependency in the physiological signal to retain the key features in the photoplethysmogram. In order to more clearly illustrate the distribution of the key sampling points selected by the present invention in the photoplethysmogram, the following will be described in detail with reference to the diagram, as shown in FIG. Figure 2 As shown in the figure, the distribution of key signal sampling points (O, S, N, D, a, b, c, d, e, f) selected by the present invention in the photoplethysmogram is shown, wherein O is the signal sampling point corresponding to the starting point of a cardiac cycle, i.e., the beginning of the systolic pulse; S is the signal sampling point corresponding to the maximum peak value during the systolic ejection period; N is the signal sampling point corresponding to the transition from the systolic period to the diastolic period; D is the signal sampling point corresponding to the lowest pressure in the arterial system; a, b, c, d, e, f are signal sampling points related to blood pressure information selected from the accelerated photoplethysmogram signal corresponding to the photoplethysmogram, these signal sampling points include additional hemodynamic information and correspond to the signal sampling points of the significantly accelerated changes in the cardiac cycle of the photoplethysmogram, wherein the accelerated photoplethysmogram signal is the second-order derivative of the photoplethysmogram, a, b, c, d are the signal sampling points corresponding to the four morphological feature points of the systolic period; e, f are the signal sampling points corresponding to the two peak points of the diastolic period. Each photoplethysmogram is enhanced based on all its key signal sampling points by using any of the following methods: method 1, method 2, method 3, method 4, and method 5. Each method is described in detail below.

[0047] Method 1: Noise based on semantic consistency

[0048] Noise is added to the amplitude of other signal sampling points except the key signal sampling point and its adjacent signal sampling points, and the noise is configured as a value randomly generated by a normal distribution probability density function. In order to more clearly illustrate the data enhancement operation in method 1, method 1 will be described in detail with examples below.

[0049] According to an embodiment of the present invention, the method 1 implements data enhancement by the following steps: Step 1, initialize the array m[j]=1, where j∈[0,N], N is the total number of sampling points of the photoplethysmogram; Step 2, if j∈[i FP -a,i FP +a], then m[j] is set to 0, where FP∈{O,N,S,D,a,b,c,d,e,f}, a is a preset value, and the value of a determines the range of signal sampling points adjacent to the key signal sampling point. When a is 3, it means j∈[i FP -3,i FP +3] interval, m[j] is set to 0; Step 3, traverse each signal sampling point of the photoplethysmogram, and calculate s′[i]=n(0,σ)*m[i]+s[i], where s′[i] is the updated amplitude of the i signal sampling point, s[i] is the original amplitude of the i signal sampling point, and n(0,σ) is a pre-configured normal distribution probability density function with a distribution mean of 0 and a standard deviation of σ.

[0050] According to one embodiment of the present invention, in the operations in step 1 to step 3 of method 1, when calculating the updated amplitudes of the key signal sampling points and their adjacent signal sampling points, the value of m[i] is 0, the amplitudes of these signal sampling points remain unchanged, and the m[i] corresponding to other signal sampling points is 1, and noise is added to the original amplitude. When method 1 is used to perform data enhancement on the photoplethysmogram, the amplitude and change trend of the key sampling points are retained and noise is added to other sampling points, so as to increase data diversity while avoiding feature loss caused by data enhancement.

[0051] Method 2: Semantically consistent masking

[0052] A portion of the photoplethysmogram waveform is randomly intercepted, and the amplitudes of the signal sampling nodes other than the key signal sampling point and its adjacent signal sampling points in the intercepted portion of the waveform are adjusted to 0. In order to more clearly illustrate the data enhancement operation in method 2, method 2 will be described in detail with examples below.

[0053] According to one embodiment of the present invention, the method 2 implements data enhancement through the following steps: step 1, initialize the array m[j]=1, where j∈[0,N]; step 2, generate a random number t in the interval [0,Nb]; step 3, set the value of m[j] of j∈[t,t+150] to 0, if j∈[iFP-a,iFP+a], then set m[j] to 1, where FP∈{O,N,S,D,a,b,c,d,e,f}, b is a preset value, and the value of b determines the range of signal sampling points adjacent to the key signal sampling point; step 4, traverse each signal sampling point of the photoplethysmogram, and the updated amplitude of each signal sampling point is m[i]*s[i], where s[i] is the original amplitude of the i-th signal sampling point of the photoplethysmogram.

[0054] According to one embodiment of the present invention, in the operations in step 1 to step 4 of the method 2, part of the sampling interval in the photoplethysmogram is randomly intercepted, the key feature information in the interval is retained, and other features of the interval are masked to achieve data enhancement. Method 2 uses feature masking to perform data enhancement, and retains the features of the signal sampling points containing key physiological knowledge during the feature masking process to solve the problem of key feature loss of the photoplethysmogram caused by feature masking.

[0055] Method 3: Time shift based on semantic consistency

[0056] The label-free photoplethysmogram is rolled and shifted to multiple signal sampling points. In order to more clearly illustrate the data enhancement operation in the method 3, the method 3 will be described in detail with reference to an example.

[0057] According to an embodiment of the present invention, the method 3 implements data enhancement through the following steps: Step 1, determining the number of signal sampling points corresponding to the start of the cardiac systolic pulse, and the index of the signal sampling point corresponding to the start of each cardiac systolic pulse; Step 2, configuring the number of signal sampling points of the rolling displacement to be: Wherein, T represents the number of signal sampling points of rolling displacement, M represents the total number of signal sampling points corresponding to the beginning of the systolic pulse, and O t represents the index of the signal sampling point corresponding to the start of the t-th systolic pulse, O t-1 Represents the index of the signal sampling point corresponding to the beginning of the t-1th cardiac systolic pulse; Step 3, rolling the original photoplethysmogram by T signal sampling points.

[0058] According to one embodiment of the present invention, in the operations in step 1 to step 3 of method 3, the number of signal sampling points contained in each cardiac cycle is obtained based on the signal sampling point corresponding to the start of each cardiac systolic pulse and the average value is taken, and the average value is used as the rolling displacement distance. Method 3 performs rolling displacement on the data based on the average cardiac cycle of the photoplethysmogram can better maintain the consistency and periodicity of the physiological signal, thereby improving the generalization ability and prediction accuracy of the model.

[0059] Method 4: Replacement based on semantic consistency

[0060] The unlabeled photoelectric volumetric pulse wave is divided into a plurality of sub-waves based on the acquired signal sampling points corresponding to the start of all cardiac systolic pulses, and all the sub-waves are randomly rearranged, wherein each sub-wave is a photoelectric volumetric pulse wave between two adjacent signal sampling points at the start of cardiac systolic pulses. In order to more clearly illustrate the data enhancement operation in method 4, method 4 will be described in detail with examples below.

[0061] According to one embodiment of the present invention, the method 4 realizes data enhancement through the following steps: step 1, dividing the photoelectric volumetric pulse wave into multiple sub-waves based on the signal sampling points corresponding to the start of all cardiac systolic pulses, wherein each sub-wave is the waveform between two adjacent signal sampling points of the start of the cardiac systolic pulse of the photoelectric volumetric pulse wave; step 2, randomly arranging all the sub-waves to obtain the enhanced photoelectric volumetric pulse wave.

[0062] According to one embodiment of the present invention, in step 1 to step 3 of method 3, the photoplethysmogram is divided into a plurality of sub-waves corresponding to the cardiac cycle and then rearranged, so as to maintain the temporal consistency of cardiac events, simulate physiological changes more realistically, and improve the model's ability to recognize the periodic characteristics of the cardiac cycle.

[0063] Method 5: Semantically consistent substitution and noise addition

[0064] The method 1 and the method 4 are used in sequence to perform data enhancement operations on the label-free photoplethysmogram.

[0065] In order to more clearly illustrate the difference between the five data enhancement methods proposed in the present invention and the classical data enhancement method, the inventors respectively processed the same section of photoplethysmogram using the five data enhancement methods proposed in the present invention and the traditional data enhancement method. The results are as follows: Figure 3As shown, from the results in the analysis diagram, it can be concluded that the signal after data enhancement using the semantically consistent noise and the semantically consistent mask proposed in the present invention can ensure that the amplitude of the feature point does not change, and the signal after data enhancement using the semantically consistent time shift and the semantically consistent replacement proposed in the present invention can maintain the timing consistency of heartbeat events and can more realistically simulate physiological changes. Therefore, the data enhancement method proposed in the present invention ensures that the semantic information of the signal after data enhancement is consistent with that of the original signal as much as possible.

[0066] According to an embodiment of the present invention, the preset first loss function is used to calculate the contrast loss between two feature vectors corresponding to two photoplethysmograms in each positive sample pair, and the preset first loss function is configured as:

[0067]

[0068] in, represents the contrastive learning loss, z i 、z j represents the feature vectors of the original unlabeled photoplethysmogram in a positive sample pair and the corresponding unlabeled photoplethysmogram after data enhancement, sim(·) is a function for calculating the similarity of two feature vectors, N represents the number of positive sample pairs processed simultaneously in a training process, exp(·) represents an exponential function, τ represents a hyperparameter, preferably, in the present invention, τ is configured to be 0.1, z k It represents the kth eigenvector among all the eigenvectors corresponding to the positive sample pairs processed simultaneously in a training process. f(k,i) is the indicator function. When k≠i, the value of f(k,i) is 1, and when k=i, the value of f(k,i) is 0.

[0069] According to an embodiment of the present invention, the preset second loss function is configured as:

[0070]

[0071] Among them, l MAE represents the mean absolute error loss, n represents the number of labeled photoplethysmograms used to train the blood pressure prediction model, represents the predicted systolic and diastolic blood pressures corresponding to the i-th labeled photoplethysmogram, Denotes the label of the i-th labeled photoplethysmogram.

[0072] According to one embodiment of the present invention, the blood pressure prediction model construction method can be divided into three stages: Figure 4As shown in the figure, the three stages shown in the figure include: the first stage, unsupervised pre-training using a large amount of unlabeled data; the second stage, supervised pre-training using a small amount of labeled data; the third stage, personalized training using the labeled data of the target object. Among them, the first stage is the unsupervised pre-training stage, which is the basis of the blood pressure prediction model construction method, and includes four basic components: a semantically consistent data enhancement library, a deep model-based encoder, a semantically consistent feature reconstruction, and a contrastive loss. The semantically consistent data enhancement library uses the physiological knowledge related to blood pressure in the PPG signal to generate positive sample pairs to ensure that semantic consistency is maintained during the data enhancement process. The encoder consists of a series of convolutional neural network layers that capture spatial and temporal features from the PPG signal and output high-level semantics. In order to enhance the diversity of the encoder output, the semantically consistent feature reconstruction performs feature resampling and reweighting through random masks and self-attention mechanisms respectively. The output of the feature reconstructor is input into the projection head, which maps the reconstructed features into the feature space and calculates the loss based on the contrastive learning loss function. The second stage is the supervised pre-training stage, in which a small amount of labeled data is used to fine-tune the model to obtain a supervised pre-training model consistent with the framework of other contrastive learning methods. Since there is a significant distribution difference between the data used for supervised pre-training and the personalized data collected on wearable devices, in the third stage, the personalized supervised training stage, we follow the settings of previous blood pressure prediction work and further fine-tune the model for each individual using a small amount of personalized PPG signals to improve the accuracy of the model.

[0073] According to one embodiment of the present invention, a blood pressure prediction method is proposed, which is used to predict the corresponding systolic and diastolic pressures of a target object based on its photoplethysmogram. The method comprises: obtaining the photoplethysmogram of the target object; predicting the systolic and diastolic pressures of the target object based on the photoplethysmogram of the target object using a target blood pressure prediction model constructed using the above-mentioned blood pressure prediction model construction method.

[0074] According to one embodiment of the present invention, a blood pressure prediction system is proposed, characterized in that the system includes: a feature encoding module, used to encode the photoplethysmogram into a feature representation; a feature reconstruction module, used to resample the feature representation to obtain a new feature representation of the photoplethysmogram; a projection module, used to convert the new feature representation of the photoplethysmogram into its corresponding low-dimensional feature vector; and a prediction module, used to predict the systolic and diastolic pressures corresponding to the photoplethysmogram based on the low-dimensional feature vector of the photoplethysmogram.

[0075] In order to verify the beneficial effects of the present invention, the inventors designed a comparative experiment for evaluation, which will be described in detail below in conjunction with the experimental results of the comparative experiment designed by the inventors.

[0076] According to one embodiment of the present invention, a comparative experiment compares the blood pressure prediction model construction method (PhysCL) proposed by the present invention with the SOTA (state-of-the-art) comparative learning baseline method (such as SimCLR, VICREG, TS-TCC and TF-C) to demonstrate the beneficial effects of the present invention. The comparative experiment uses the data of 50 subjects in the MIMICⅢ data set for pre-training, that is, the initial model is subjected to unsupervised training and supervised training to obtain the initial blood pressure prediction model, and then other data sets are used to perform personalized supervised training on the initial blood pressure prediction model, wherein the methods tested in the comparative experiment are distributed and the data set is distributed in the following manner, and the experimental setting is that 88% of the data is unlabeled data for unsupervised pre-training, 2% of the data is labeled data for supervised pre-training, and the remaining 10% of the data is used for verification in the pre-training stage. For each individual, 5, 10, and 25 samples are used to personalize the initial blood pressure prediction model to improve the performance of the method. The size of the validation set sval idat ion for each individual is set to 1024 (MIMICⅢ), 1024 (MIMICⅣ), and 360 (UQVS), respectively. The personalized model is evaluated on the remaining data of the individual and the mean absolute error between individuals is recorded.

[0077] PhysCL is compared with SOTA (state-of-the-art) contrastive learning baseline methods (such as SimCLR, VICREG, TS-TCC and TF-C). The results of the comparison between PhysCL and the classic contrastive learning baseline methods are shown in Tables 1 and 2. Table 1 shows the systolic blood pressure prediction accuracy of each method, and Table 2 shows the diastolic blood pressure prediction accuracy of each method. The symbols used in Tables 1 and 2 are explained as follows: labelis the percentage of labeled data in supervised pre-training, "5", "10", and "25" are the number of personalized labeled data, "MIMIC III→MIMIC III" represents pre-training the initial model on the MIMIC III dataset and performing personalized supervised training on the MIMIC III dataset; "MIMIC III→MIMIC IV" represents pre-training the model on the MIMIC III dataset and performing personalized supervised training on the MIMIC IV dataset; "MIMIC III→UQVS" represents pre-training the model on the MIMIC III dataset and performing personalized supervised training on the UQVS dataset. By analyzing the data in Tables 1 and 2, PhysCL can achieve the lowest blood pressure MAE prediction accuracy under 9 different experimental settings of 3 datasets compared with other comparative learning methods. Among them, using 2% labeled data from MIMIC III for supervised pre-training and using 25 samples from each individual from MIMIC III for personalization, PhysCL achieved prediction accuracies of 7.600 mmHg and 3.940 mmHg in terms of systolic MAE and diastolic MAE, respectively.

[0078] Table 1

[0079]

[0080] Table 2

[0081]

[0082]

[0083] The inventors experimentally compared the effect of the number of training samples used in the personalized supervised training stage on the model prediction accuracy and obtained the following results: Figure 5 The results show that when PhysCL is trained with 5 and 10 samples for personalized supervised training on the MIMICⅢ and MIMICⅣ datasets, the prediction accuracy can reach SOTA. When PhysCL is trained with 5 samples for personalized supervised training on the UQVS dataset, the prediction accuracy can reach SOTA. Although BPCRNN can also achieve SOTA performance by training the model with 25 training samples, in the real world, it is challenging and complex to collect labeled samples from individuals. Therefore, PhysCL can achieve better performance with fewer personalized samples, which is more meaningful for practical applications.

[0084] In order to prove the influence of data enhancement and feature reconstruction based on key blood pressure information proposed in the present invention on the constructed blood pressure prediction model, the inventors conducted ablation experiments to verify the prediction accuracy of the blood pressure model constructed by using SCAB (data enhancement based on key blood pressure information), SCFR (feature reconstruction), SCAB and SCFR at the same time, and not using SCAB and SCFR. The experimental results are as follows: Figure 6 As shown, the results show that when the model is pre-trained using 2% labeled data, data augmentation based on key information of blood pressure helps to improve the experimental results in all experimental settings, and feature reconstruction helps to improve the experimental results in almost all experiments.

[0085] It should be noted that although the above describes the various steps in a specific order, it does not mean that the various steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently or even in a different order as long as the required functions can be achieved.

[0086] The present invention may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present invention.

[0087] Computer readable storage medium can be a tangible device that holds and stores instructions used by an instruction execution device. Computer readable storage medium can include, for example, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples (non-exhaustive list) of computer readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, for example, a punch card or a protruding structure in a groove on which instructions are stored, and any suitable combination thereof.

[0088] The embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A method for constructing a blood pressure prediction model, wherein the blood pressure prediction model is used to predict the corresponding systolic and diastolic blood pressure of a target object based on its photoplethysmogram, characterized in that: The method comprises: Step S1, obtaining a first training set including a plurality of unlabeled photoplethysmograms, a second training set including a plurality of labeled photoplethysmograms, and a third training set including a plurality of labeled photoplethysmograms of a target object, wherein the label of each photoplethysmogram is its corresponding systolic pressure and diastolic pressure; Step S2, constructing an initial model, the initial model comprising a feature encoding module, a feature reconstruction module, a projection module, and a prediction module, the feature encoding module is used to encode the photoplethysmogram into a feature representation, the feature reconstruction module is used to resample the feature representation to obtain a new feature representation of the photoplethysmogram, the projection module is used to convert the new feature representation of the photoplethysmogram into its corresponding low-dimensional feature vector, and the prediction module is used to predict the systolic pressure and diastolic pressure corresponding to the photoplethysmogram based on the low-dimensional feature vector of the photoplethysmogram; Step S3, based on the key physiological information of blood pressure, data enhancement is performed on the photoplethysmogram in the first training set to obtain multiple positive sample pairs, each positive sample pair includes the original photoplethysmogram and the photoplethysmogram after data enhancement, and the initial model is trained by contrastive learning based on all the positive sample pairs, and the parameters of the feature encoding module, the feature reconstruction module, and the projection module are adjusted based on a preset first loss function; Step S4, performing supervised training on the model trained in step S3 based on the second training set and adjusting the model parameters based on a preset second loss function to obtain an initial blood pressure prediction model; Step S5: performing personalized supervised training on the initial blood pressure prediction model based on the third training set and adjusting the model parameters based on a preset second loss function to obtain a target blood pressure prediction model.

2. The method according to claim 1, characterized in that The feature encoding module is based on the ConvNeXt network architecture and consists of four consecutive convolutional neural networks, wherein the number of convolutional blocks of the four consecutive convolutional neural networks is configured as (2, 2, 6, 2) in sequence, and the number of channels of the four consecutive convolutional neural networks is configured as (64, 128, 256, 512) in sequence.

3. The method according to claim 1, characterized in that The feature reconstruction module is configured as follows: The feature representation encoded by the encoding module is reconstructed by random feature masking and feature resampling in turn to obtain a new feature representation, and the weight of feature resampling is dynamically adjusted based on the self-attention mechanism.

4. The method according to claim 1, characterized in that: The projection module is a fully connected layer including a preset number of neurons.

5. The method according to claim 1, characterized in that The prediction module consists of two fully connected layers and uses rectified linear units as activation functions.

6. The method according to claim 1, characterized in that The following method is used to enhance the data of each photoplethysmogram in the first training set based on the key physiological information of blood pressure: Acquire all key signal sampling points in the photoplethysmogram containing key physiological information of blood pressure, wherein the key signal sampling points include signal sampling points corresponding to the start of all systolic pulses, signal sampling points corresponding to the maximum peak value during the ejection period of all systolic blood, signal sampling points corresponding to the transition from systolic to diastolic blood, signal sampling points corresponding to the lowest pressure in all arterial systems, and all morphological feature points of the accelerated photoplethysmogram corresponding to the photoplethysmogram signal, and signal sampling points corresponding to the diastolic peak, wherein the accelerated photoplethysmogram is the second-order derivative of the photoplethysmogram; Based on the acquired signal sampling points, the photoplethysmogram data is enhanced by any of the following methods: Mode 1, which is a data enhancement operation of adding noise to the amplitude of other signal sampling points in the photoplethysmogram except the key signal sampling point and its adjacent signal sampling points, wherein the noise is a value randomly generated by a pre-configured normal distribution probability density function; Method 2, which is a data enhancement operation that adjusts the amplitudes corresponding to the signal sampling points other than the key signal sampling points and their adjacent signal sampling points within a part of the sampling interval of the photoplethysmogram to 0 and keeps the amplitudes of other signal sampling points unchanged; Method 3 is a data enhancement operation of rolling the photoplethysmogram to shift multiple signal sampling points, wherein the number of signal sampling points of the rolling shift is determined by the following method: Wherein, T represents the number of signal sampling points of rolling displacement, M represents the total number of signal sampling points corresponding to the beginning of the systolic pulse, and O t represents the tth heart The index of the corresponding signal sampling point, O t-1 represents the index of the signal sampling point corresponding to the start of the t-1th systolic pulse; Method 4, which is to divide the photoelectric volume pulse wave into multiple sub-waves, and perform a data enhancement operation on all the sub-waves in a random arrangement, wherein each sub-wave is a waveform between signal sampling points corresponding to the start of two adjacent cardiac systolic pulses; Method 5: using the operations in Method 1 and Method 4 in sequence to perform data enhancement on the photoplethysmogram.

7. The method according to claim 1, characterized in that The preset first loss function is configured as: in, represents the contrastive learning loss, z i 、z j represents the original photoplethysmogram in a positive sample pair and the corresponding feature vector of the photoplethysmogram after data enhancement, sim(·) is the function for calculating the similarity of two feature vectors, N is the number of positive sample pairs processed simultaneously in one training process, exp(·) represents the exponential function, τ represents the hyperparameter, z k It represents the kth eigenvector among the 2N eigenvectors corresponding to the N positive sample pairs processed simultaneously in a training process. f(k,i) is the indicator function. When k≠i, the value of f(k,i) is 1, and when k=i, the value of f(k,i) is 0.

8. The method according to claim 1, characterized in that The preset second loss function is configured as: Among them, l MAE represents the mean absolute error loss, n represents the number of photoplethysmograms in the third training set, represents the predicted systolic and diastolic pressures corresponding to the i-th photoplethysmogram in the third training set, Represents the true systolic pressure and true diastolic pressure corresponding to the i-th photoplethysmogram in the third training set.

9. A blood pressure prediction method for predicting the corresponding systolic and diastolic blood pressure of a target object based on its photoplethysmogram, characterized in that: The method comprises: Acquire a photoplethysmogram of a target object; The target blood pressure prediction model constructed by the method as described in any one of claims 1 to 8 predicts the systolic and diastolic blood pressure of the target object based on the photoplethysmogram of the target object.

10. A blood pressure prediction system, characterized in that: The system comprises: A feature encoding module, for encoding the photoplethysmogram into a feature representation; A feature reconstruction module, for resampling the feature representation to obtain a new feature representation of the photoplethysmogram; A projection module, used to convert the new feature representation of the photoplethysmogram into its corresponding low-dimensional feature vector; The prediction module is used to predict the systolic pressure and diastolic pressure corresponding to the photoplethysmogram based on the low-dimensional feature vector of the photoplethysmogram.

11. An electronic device, characterized in that: include: one or more processors; as well as A memory, wherein the memory is used to store executable instructions; The one or more processors are configured to implement the steps of the method of any one of claims 1 to 9 by executing the executable instructions.

Images