A non-invasive blood pressure measurement method based on multi-view differential feature space

Abstract

This invention provides a non-invasive blood pressure measurement method based on a multi-view differential feature space, relating to the field of artificial intelligence technology. The method comprises: collecting PPG signals and integrating them into a raw data set; preprocessing the raw data set to obtain an initial data set; extracting view extraction features from the initial data set from three viewpoints and performing a differential operation to obtain multi-view raw differential features; constructing a multi-view differential feature adaptive mapping module; performing stacked ensemble learning on the multi-view differential feature adaptive mapping module using an AutoML algorithm to obtain a machine learning regression model; training the machine learning regression model using the view extraction features to obtain a final model; and using the final model to measure blood pressure on real PPG signals. This method achieves non-invasive blood pressure measurement through a loss function, a joint optimization strategy for classifying change trends, and a machine learning regression model, improving the accuracy of blood pressure detection and the stability of blood pressure prediction.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a non-invasive blood pressure measurement method based on multi-view differential feature space. Background Art

[0002] Currently, blood pressure measurement methods are categorized as invasive and non-invasive. The invasive method involves arterial puncture, whereby an artery is cut or punctured to insert a measuring instrument for direct measurement. While this method offers the most accurate results, it requires a high-quality medical environment. The most widely used non-invasive method is the cuff blood pressure method, which measures systolic blood pressure by inflating and deflating a cuff. However, this method offers a poor user experience, is difficult to carry around, and lacks continuous, real-time measurement capabilities.

[0003] PPG technology uses LED light sources and photodetectors to non-invasively detect changes in blood volume within human tissue as the heart beats, thereby obtaining pulse waveforms and addressing the shortcomings of the two aforementioned methods. However, due to limitations in dataset size and acquisition methods, existing research employs a randomized data partitioning approach. This partitioning approach can easily lead to overly optimistic evaluation results for the training model when faced with data from the same subject, making it difficult to truly reflect the model's generalization and robustness. On the other hand, current research still primarily focuses on predicting absolute blood pressure values, neglecting the detection and analysis of abnormal blood pressure change points. Therefore, it is imperative to design a non-invasive blood pressure measurement method based on a multi-view differential feature space. Summary of the Invention

[0004] The purpose of the present invention is to provide a non-invasive blood pressure measurement method based on a multi-view differential feature space. By extracting differential features related to blood pressure changes, and then using a joint optimization strategy of reconstruction loss, contrast loss, and classification of change trends, and at the same time realizing non-invasive blood pressure measurement based on a constructed machine learning regression model, the accuracy of blood pressure detection and the stability of blood pressure prediction can be improved.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A non-invasive blood pressure measurement method based on multi-view differential feature space includes the following steps:

[0007] Collect the human body's photoplethysmography (PPG) signals and integrate them into the original data set;

[0008] Perform data preprocessing on the original data set to obtain the initial data set; the data preprocessing operations include: signal filtering, signal segmentation and signal normalization;

[0009] Extract three different perspective extraction features of the initial data set from the time domain perspective, frequency domain perspective and physiological statistics perspective respectively, and perform differential operation on the perspective extraction features to obtain multi-perspective original differential features; the perspective extraction features include: time domain key point features, frequency domain features and physiological statistics features;

[0010] A multi-view differential feature adaptive mapping module is constructed based on the multi-view original differential features. The multi-view differential feature adaptive mapping module consists of a connected feature encoder and a view change submodule. The multi-view differential feature adaptive mapping module has built-in reconstruction regularization constraints, an adaptive contrastive learning mechanism, and a loss function.

[0011] The AutoML algorithm is used to perform stacked ensemble learning on the multi-view differential feature adaptive mapping module to obtain a machine learning regression model.

[0012] The view extraction features are integrated into the input data set, and the machine learning regression model is trained with the input data set to obtain the final model; the input data set includes a training set, a validation set, and a test set;

[0013] Blood pressure measurement is performed on the real PPG signal through the final model.

[0014] Optionally, the original dataset consists of a MIMIC III dataset and a private dataset.

[0015] Optionally, perform data preprocessing on the dataset to obtain an initial dataset, including:

[0016] The power frequency noise, high frequency noise and low frequency noise in the data set are removed by a bandpass filter to obtain a filtered signal;

[0017] The filtered signal is divided into multiple windows using the sliding window segmentation method, with a window width of 9 seconds and a window step of 6 seconds;

[0018] Normalize the window to obtain the initial data set.

[0019] Optionally, the time domain key point features include: total systolic-diastolic waveform area , waveform symmetry ratio , maximum rising slope and time location ; The calculation formulas for time domain key point features include:

[0020] ;

[0021] ;

[0022] ;

[0023] ;

[0024] in, is the PPG signal, is the current peak time, is the trough time, is the peak time of the next moment, argmax t (·) is the variable position when the objective function reaches its maximum value.

[0025] Optionally, the frequency domain features include: the energy of each layer of wavelet decomposition , spectrum centroid and the average amplitude in the neighborhood of the main frequency ; The calculation formulas for frequency domain features include:

[0026] ;

[0027] ;

[0028] ;

[0029] in, is the j-th layer wavelet detail coefficient, is the frequency, is the power spectral density, is the dominant frequency, k is the frequency index, K is the total number of frequency points in the spectrum, and n is the number of frequency points.

[0030] Optionally, the physiological statistical features include: aging index AI, skewness, kurtosis and slope deviation SDC up ; The calculation formulas for physiological statistical characteristics include:

[0031] ;

[0032] ;

[0033] ;

[0034] ;

[0035] in, is the average value of the PPG signal amplitude, is the standard deviation of the PPG signal amplitude, The amplitude of the i-th sampling point, is the average slope of the rising segment, N is the total number of sampling points, and a, b, c, d, and e are the amplitudes of the APG feature points.

[0036] Optionally, the regularization constraint can be reconstructed as: ; Among them, △f v is the original differential feature of multiple views, Z v is the output of the perspective change submodule, For the modulo operation, R v is the reconstruction error, is the function for reconstructing the original differential features.

[0037] Optionally, the loss function is expressed as: ;in, is the adaptive weight of view m and view n, L CL is the contrast loss, Z m is the output of the view m in the view change submodule, Z n is the output of the view n in the view change submodule, L cls is the trend classification loss, and are the weight coefficients of different loss functions respectively.

[0038] Optionally, the ratio of the training set, validation set, and test set is 7:1.5:1.5.

[0039] The present invention discloses the following technical effects: a non-invasive blood pressure measurement method based on a multi-view differential feature space, comprising: collecting human photoplethysmography (PPG) signals and integrating them into an original dataset; performing data preprocessing on the dataset to obtain an initial dataset; extracting features from three different perspectives of the initial dataset, namely, from the time domain, frequency domain, and physiological statistics perspectives, and performing differential operations on the perspective-extracted features to obtain multi-view original differential features; constructing a multi-view differential feature adaptive mapping module based on the multi-view original differential features; performing stacked ensemble learning on the multi-view differential feature adaptive mapping module using an AutoML algorithm to obtain a machine learning regression model; integrating the perspective-extracted features into an input dataset, and training the machine learning regression model using the input dataset to obtain a final model; and using the final model to measure blood pressure on real PPG signals. This method extracts differential features related to blood pressure changes, and then implements non-invasive blood pressure measurement based on the constructed machine learning regression model using a joint optimization strategy of reconstruction loss, contrast loss, and change trend classification, thereby improving the accuracy of blood pressure detection and the stability of blood pressure prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0041] Figure 1 This is a flow chart of the non-invasive blood pressure measurement method of the present invention;

[0042] Figure 2 Schematic diagram of key point data of PPG and its derivatives according to an embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram of the differential systolic blood pressure results according to an embodiment of the present invention;

[0044] Figure 4 This is a flowchart of the multi-view differential feature adaptive mapping module according to an embodiment of the present invention;

[0045] Figure 5 This is a diagram showing the working principle of a machine learning regression model according to an embodiment of the present invention;

[0046] Figure 6 A Bland-Altman plot of estimated systolic blood pressure (SBP) values ​​according to an embodiment of the present invention;

[0047] Figure 7 is a correlation diagram of the estimated systolic blood pressure (SBP) value according to an embodiment of the present invention;

[0048] Figure 8 A Bland-Altman plot of the estimated diastolic blood pressure (DBP) according to an embodiment of the present invention;

[0049] Figure 9 This is a correlation diagram of the diastolic blood pressure (DBP) estimation value according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0051] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0052] like Figure 1 As shown, the present invention provides a non-invasive blood pressure measurement method based on multi-view differential feature space, comprising the following steps:

[0053] Step 100: Collecting human photoplethysmography (PPG) signals and integrating them into a raw data set;

[0054] Specifically, the original dataset of this embodiment consists of the MIMIC (Multi-parameter Intelligent Monitoring in Intensive Care) III dataset and a private dataset, both of which contain collected data from 2012 subjects.

[0055] The MIMIC III dataset contains 67,830 records from nearly 30,000 intensive care unit (ICU) patients, covering invasive blood pressure waveforms, electrocardiograms, respiratory signals, and PPG signals. Due to incomplete recordings of some signals, this example ultimately selected paired PPG-BP data from 1,329 patients, with data with a single recording duration exceeding 3 minutes for blood pressure estimation.

[0056] The private dataset contains PPG, ECG, and invasive blood pressure data from 683 ICU patients in seven different hospitals. The private dataset contains anonymized electronic health records of patients, including demographic information, medication history, comorbidities, etc.

[0057] It should be noted that non-invasive blood pressure estimation based on PPG can be regarded as a time series regression problem. The original input PPG signal space and the corresponding target BP space Given paired samples To be precise, for a dataset containing paired s segments, the input PPG signal The input dimension can be PPG, VPG, APG, or other physiological signals (such as ECG). Similarly, the blood pressure corresponding to the PPG signal includes two dimensions: SBP and DBP.

[0058] Step 200: performing data preprocessing on the original data set to obtain an initial data set;

[0059] Specifically, for the PPG signal in the original data set, since the original waveform is mixed with a large amount of noise and interference, and different databases have differences in signal length and amplitude, in order to ensure the consistency and reliability of the data during model training, the PPG signal is subjected to three steps: signal filtering, signal segmentation, and signal normalization.

[0060] First, a bandpass filter was used to remove the main noise sources in the PPG signal—50Hz power frequency noise, high-frequency noise generated by muscle activity, and low-frequency noise caused by baseline drift and respiration. The filter's start and cutoff frequencies were set to 1Hz and 25Hz, respectively. After filtering, the noise interference in the signal was effectively suppressed. Subsequently, to standardize the signal length and expand the data volume, a sliding window segmentation method was used to segment the signal into 9-second window widths and 6-second step lengths. For invasive blood pressure data, the average value within each window was calculated as the corresponding blood pressure label. Finally, Min-Max normalization was performed on all signal data to complete the preprocessing process, ensuring that subsequent model training was based on unified and standardized data.

[0061] Step 300: Extracting three different perspective extraction features of the initial data set from the time domain perspective, the frequency domain perspective, and the physiological statistics perspective, and performing a differential operation on the perspective extraction features to obtain multi-perspective original differential features; the perspective extraction features include: time domain key point features, frequency domain features, and physiological statistics features;

[0062] Specifically, the time domain perspective is based on the morphological structure of the PPG waveform and its derivative (VPG, APG) signals, which is manifested as a depiction of the local dynamic changes within a complete pulse cycle. This embodiment fully explores the key points of PPG and its derivatives within a cardiac cycle, including but not limited to the systolic peak, notch point, and multiple anatomically related derivative waveform feature points. By calculating the time interval, normalized position, rise / fall rate, waveform area and relative duration ratio between these points and other parameters in multiple dimensions, the hemodynamic characteristics of each cardiac cycle are accurately restored, and signal structure information with high physiological interpretability is provided for subsequent models. The time domain key point features of this embodiment include: total systolic-diastolic waveform area , waveform symmetry ratio , maximum rising slope and time location 38 specific features; the calculation formula of some time domain key point features in this embodiment is as follows:

[0063] ;

[0064] ;

[0065] ;

[0066] ;

[0067] in, is the PPG signal, is the current peak time, is the trough time, is the peak time of the next moment, argmax t (·) represents the variable position corresponding to the maximum value of the objective function. The test results of this embodiment are as follows Figure 2 shown.

[0068] Specifically, the frequency domain perspective uses a variety of frequency analysis techniques to perform time-frequency transformation on the signal based on the rhythmicity and energy distribution of the PPG signal. The frequency domain features of this embodiment include: , spectrum centroid and the average amplitude in the neighborhood of the main frequency etc.; the calculation formulas for some frequency domain features in this embodiment are as follows:

[0069] ;

[0070] ;

[0071] ;

[0072] in, is the j-th layer wavelet detail coefficient, is the frequency, is the power spectral density, is the dominant frequency, k is the frequency index, K is the total number of frequency points in the spectrum, and n means: at the frequency point Take n frequency points on each side, a total of 2n+1 points, to calculate the local average power.

[0073] It's important to note that the frequency-domain perspective not only incorporates the classic fast Fourier transform (FFT) to extract global frequency features such as dominant frequency, power spectral density distribution, and spectral centroid, but also incorporates discrete wavelet transform (DWT) and wavelet packet analysis techniques to extract the signal's local frequency dynamics from a multi-scale and multi-resolution perspective. By calculating energy fractions, entropy values, and energy migration trends at multiple wavelet scales, it comprehensively captures the effects of physiological rhythm regulation, dominant frequency shifts between the sympathetic and parasympathetic nervous systems, and changes in vascular compliance on spectral structure.

[0074] Specifically, the physiological statistics perspective is based on the overall distribution characteristics, morphological structural deviation, waveform asymmetry and multi-index combination, which is a joint reflection of the PPG signal from "structure + probability + logic". This embodiment quantifies the dynamic stability and structural deviation of the signal by targeting the characteristics of indicators such as the histogram density, skewness, kurtosis and coefficient of variation of the rising segment, falling segment and full cycle of the signal. At the same time, the combination of the slope deviation curve (SDC) further assists the subsequent model in identifying abnormal blood pressure. The physiological statistics features of this embodiment include: aging index AI, skewness, kurtosis and slope deviation SDC up 137 specific features; the calculation formulas for some physiological statistical features of this embodiment are as follows:

[0075] ;

[0076] ;

[0077] ;

[0078] ;

[0079] in, is the average value of the PPG signal amplitude, is the standard deviation of the PPG signal amplitude, The amplitude of the i-th sampling point, is the average slope of the rising segment, N is the total number of sampling points, and a, b, c, d, and e are the amplitudes of the APG feature points.

[0080] Furthermore, after completing the multi-view feature extraction, a differential mechanism is designed to construct the multi-view original differential features. The feature vector of each subject's initial or baseline time point is used as the reference feature, and the feature vector of any subsequent time point is expressed as the difference between the reference feature and the reference feature. Correspondingly, the blood pressure value at the initial time is used as the baseline blood pressure, the predicted content is the blood pressure change value, and the absolute value of the blood pressure at each time point is recorded as the true blood pressure value. The test results of this embodiment are as follows: Figure 3 shown.

[0081] It should be noted that this differential method has three advantages: first, it effectively eliminates static individual differences, such as baseline blood pressure, resting heart rate and other non-dynamic factors; second, it significantly enhances the subsequent model's ability to model the instantaneous fluctuation trend (increase / decrease) of blood pressure; third, the characteristic changes of the same subject at different times can be mapped into dynamic change trajectories under a unified scale.

[0082] Step 400: Construct a multi-view differential feature adaptive mapping module based on the multi-view original differential features; the multi-view differential feature adaptive mapping module consists of a connected feature encoder and a view change submodule; the multi-view differential feature adaptive mapping module has built-in reconstruction regularization constraints, adaptive contrast learning mechanism and blood pressure change trend classification tasks; the specific workflow of this embodiment is as follows Figure 4 shown.

[0083] The input of the multi-view differential feature adaptive mapping module is the multi-view original differential feature , is the view number, and the input is encoded into a latent space representation (the output of the view change submodule) through the feature encoder and the view change submodule. The process expression is ,in is the feature encoder, is the perspective change module, Represents the parameters of the module training. Then, in order to prevent the loss of information in the mapping process of the view feature, a reconstruction regularization constraint is introduced in the feature encoder to reversely reconstruct the module input to form a regularization constraint. The expression of the reconstruction regularization constraint is:

[0084] ;

[0085] Among them, △f v is the original differential feature of multiple views, Z v is the output of the perspective change submodule, For the modulo operation, is the differential feature after reconstruction, R v is the reconstruction error, Represents the function that reconstructs the original differential features from the latent space. The contrast objective is optimized by reconstructing the regularization constraint, and the robustness and physiological interpretability of the input are also improved. In order to obtain a better differential feature space Z v ,This embodiment also designs a multi-task optimization part, which is divided into ,two parts: contrast learning and blood pressure change trend classification.

[0086] Specifically, the multi-view differential feature adaptive mapping module also introduces an adaptive contrastive learning mechanism. This constructs positive sample pairs (same data, different viewpoints) and negative sample pairs (different data, different viewpoints) between every two viewpoints, and dynamically adjusts the alignment strength of different viewpoint pairs through an adaptive weighting mechanism. The contrast loss function of the adaptive contrastive learning mechanism is expressed as:

[0087] ;

[0088] in, is the adaptive weight of view m and view n, LCL is the contrast loss. For two perspectives, The larger the value is, the more information the two perspectives contain, which means they are high-quality perspective pairs. Conversely, the smaller the value is, the less information they contain, which means they are low-quality perspective pairs. The value of will be increased, in the process, enhancing the contrast between high-quality view pairs and weakening the contrast between low-quality view pairs.

[0089] This embodiment designs three adaptive weighting strategies based on feature representation difference metrics: Jensen-Shannon divergence (JSD), maximum mean difference (MMD), and class mutual information (CMI). The JSD weight measures view consistency by calculating the symmetric divergence of feature distributions between views. The MMD weight uses a kernel function to calculate the maximum mean difference between sample distributions. The CMI weight uses K-Means clustering to generate pseudo-labels and estimates the mutual information between two view pseudo-labels, thereby measuring semantic consistency. The expression is as follows:

[0090] ;

[0091] ;

[0092] ;

[0093] For JSD divergence Calculation, exp(·) represents the exponential function, Z m， Z n represents the differential feature space representation of the mth and nth viewpoints, and denotes the i-th element in the m-th and n-th distributions, respectively, and Represents the probability value of the i-th position in the m-th and n-th distributions respectively, N represents the dimension of the differential feature space, for MMD way, is the scale factor that controls the change of similarity, and k(x,y) is the kernel function. way, and They represent the labels obtained after K-Means clustering, K represents the number of clusters of the category, represents the probability of the i-th category pseudo label, Expressed as the entropy of the label distribution, is the joint probability, is the marginal probability, represents the mutual information between two pseudo labels.

[0094] The contrast loss function of this embodiment supports three forms of expression, which are expressed as follows:

[0095] ;

[0096] ;

[0097] ;

[0098] Among them, sim(·) represents the cosine similarity, τ represents the temperature coefficient, which is used to control the similarity scaling degree. The smaller the value, the stronger the discrimination. q represents the query sample, k + is the positive sample pair for this query sample, k - is a negative sample pair.

[0099] To further enhance the ability to distinguish the direction of blood pressure change trend (increase / decrease), the second part of the multi-task optimization is to predict The positive or negative (i.e. trend direction) of . At the same time, the binary cross entropy loss (trend classification loss) is introduced, and the expression is:

[0100] ;

[0101] Among them, y is the true label, It represents the probability that the input sample label is 1, 0 represents an increase in blood pressure, and 1 represents a decrease in blood pressure.

[0102] This loss uses explicit directional supervision to enable the representation space to not only focus on amplitude differences but also learn the semantic boundaries of the direction of change, significantly improving the ability to identify abnormal blood pressure trends. The final total loss function consists of three parts: contrast loss, reconstruction loss, and trend classification loss, and is expressed as:

[0103] ;

[0104] Among them, Z m is the output of the view m in the view change submodule, Z n is the output of the view n in the view change submodule, L cls is the trend classification loss, and are the weight coefficients of different loss functions respectively.

[0105] Furthermore, in the training process of the multi-view differential feature adaptive mapping module, the encoder-contrast weight alternating update strategy is adopted. Update once , the remaining steps update Parameters such as , make the structure of the feature space gradually have the ability of discrimination, consistency and direction perception in the dynamic optimization process. After training, all Splice together to obtain the final differential feature space representation .

[0106] It should be noted that the multi-view differential feature adaptive mapping module, on the one hand, uses adaptive weighted contrast learning to autonomously adjust the alignment strength between features from different viewpoints, thereby highlighting the information between important viewpoints and suppressing the interference of irrelevant or weakly correlated viewpoints on the feature representation. On the other hand, a reconstruction regularization term is added to ensure that the feature representation does not suffer from information degradation during the contrast optimization process, maintaining the integrity and interpretability of the representation. At the same time, to address the core requirements of the blood pressure prediction task (i.e., positive and negative trends in blood pressure changes), a direction classification branch is introduced to further enhance sensitivity to the "rising / falling" direction, significantly improving the ability to identify and predict abnormal blood pressure fluctuation trends. The integration of this triple mechanism not only optimizes the feature representation in terms of structural consistency across viewpoints, but also achieves a qualitative improvement in the directionality of trend prediction and physiological discrimination capabilities, making it more in line with the needs of actual clinical monitoring.

[0107] Step 500: Perform stacked ensemble learning on the multi-view differential feature adaptive mapping module using the AutoML algorithm to obtain a machine learning regression model;

[0108] like Figure 5 As shown, this embodiment uses AutoML to define the model framework. The AutoML algorithm is a highly scalable and fully automated algorithm that can automatically train a large number of candidate models, and finally aggregate the best-performing configurations to form a collection. The specific implementation process of this embodiment is as follows: First, a meta-level learner is trained to find the best combination of base-level learners. Then, AutoML enters the base-level regression phase and performs random searches in different model sets to produce a diverse base learner and an influential collection when paired with a stacking method. In the meta-level learning phase, the meta-level learner is set to be trained using the k-fold cross-validation prediction value of the base learner, and then tested on a data set outside the fold.

[0109] Step 600: Integrate the view extraction features into an input dataset, and train the machine learning regression model using the input dataset to obtain a final model; the input dataset includes a training set, a validation set, and a test set;

[0110] Specifically, the input dataset was divided into training, validation, and test sets in a ratio of 7:1.5:1.5. The training and validation sets were used to train the machine learning regression model and select parameters, while the test set was used to ultimately verify the generalization ability of the retained optimal machine learning regression model. During training, the Adam optimizer was used to update parameters with a learning rate of 1e-3, a weight decay of 0.999, and a momentum of 0.8.

[0111] Step 700: Measure blood pressure based on the real PPG signal using the final model.

[0112] The results of the blood pressure data estimation error analysis of the initial data set in this embodiment are as follows: Figures 6 to 9 The horizontal axis represents the mean blood pressure between the estimated and reference invasive blood pressure measurements, while the vertical axis shows the difference between the two measurements. The middle dashed line represents the mean difference in blood pressure, while the upper and lower dashed lines represent the limits of the 95% confidence interval for the mean difference.

[0113] The beneficial effects of the present invention are as follows:

[0114] 1) The input to the machine learning regression model consists solely of photoplethysmography (PPG), so the acquisition circuit only needs to collect PPG signals. Compared to traditional methods based on pulse wave velocity, this eliminates the need for collecting ECG signals. This makes it easier to integrate into devices like wristbands, eliminating the need for blood pressure measurement equipment like cuffs and making the device more portable.

[0115] 2) This method's multi-perspective differential feature space framework significantly improves the model's sensitivity and accuracy in predicting abnormal blood pressure fluctuations. It can more promptly and effectively identify and warn of abnormal blood pressure trends, helping users take appropriate intervention measures early and effectively reduce the risk of cardiovascular events.

[0116] 3) The adaptive contrastive learning method proposed in this method can effectively overcome the representation degradation problem existing in traditional contrastive learning algorithms. By dynamically adjusting the learning weights between different viewpoints, the feature extraction process is made more targeted, further improving the generalization performance and stability of the model.

[0117] 4) This method can autonomously and efficiently complete the optimization selection of model structure and parameters, significantly reducing the reliance on expert experience during the development process, lowering the algorithm development threshold and implementation costs, and improving the model's measurement performance and deployment efficiency.

[0118] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0119] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A non-invasive blood pressure measurement method based on multi-view differential feature space, characterized in that: The steps include: Collect the human body's photoplethysmography (PPG) signals and integrate them into the original data set; Performing data preprocessing on the original data set to obtain an initial data set; The data preprocessing operations include: signal filtering, signal segmentation and signal normalization; Extract three different perspective extraction features of the initial data set from the time domain perspective, frequency domain perspective and physiological statistics perspective respectively, and perform differential operation on the perspective extraction features to obtain multi-perspective original differential features; the perspective extraction features include: time domain key point features, frequency domain features and physiological statistics features; the time domain key point features include: total systolic-diastolic waveform area , waveform symmetry ratio , maximum rising slope and time location ; The frequency domain features include: each layer of energy of wavelet decomposition , spectrum centroid and the average amplitude in the neighborhood of the main frequency The physiological statistical characteristics include: aging index AI, skewness, kurtosis and slope deviation SDC up ; A multi-view differential feature adaptive mapping module is constructed based on the multi-view original differential features; the multi-view differential feature adaptive mapping module is composed of a connected feature encoder and a view change submodule; the multi-view differential feature adaptive mapping module has a built-in reconstruction regularization constraint, an adaptive contrast learning mechanism and a loss function; the expression of the reconstruction regularization constraint is: ; Among them, △f v is the original differential feature of multiple views, Z v is the output of the view change submodule, For the modulo operation, R v is the reconstruction error, is a function for reconstructing the original differential features; the expression of the loss function is: ;in, Perspective m and perspective n The adaptive weight, L CL is the contrast loss, Z m Perspective m At the output of the view change submodule, Z n Perspective n At the output of the view change submodule, L cls is the trend classification loss, and are the weight coefficients of different loss functions respectively; Performing stacked ensemble learning on the multi-view differential feature adaptive mapping module using the AutoML algorithm to obtain a machine learning regression model; Integrating the perspective extraction features into an input data set, and training the machine learning regression model using the input data set to obtain a final model; the input data set includes a training set, a validation set, and a test set; Blood pressure measurement is performed on the real PPG signal using the final model.

2. The non-invasive blood pressure measurement method based on multi-view differential feature space according to claim 1, characterized in that: The original dataset consists of the MIMIC III dataset and a private dataset.

3. The non-invasive blood pressure measurement method based on multi-view differential feature space according to claim 1, characterized in that: Performing data preprocessing on the data set to obtain an initial data set includes: removing power frequency noise, high frequency noise, and low frequency noise from the data set by using a bandpass filter to obtain a filtered signal; The filtered signal is divided into a plurality of windows by a sliding window segmentation method, wherein the window width of the window is 9 seconds and the step length of the window is 6 seconds; The window is normalized to obtain the initial data set.

4. The non-invasive blood pressure measurement method based on multi-view differential feature space according to claim 1, characterized in that: The calculation formula of the time domain key point feature includes: ； ； ； ； in, is the PPG signal, is the current peak time, is the trough time, is the peak time of the next moment, argmax t (·) is the variable position when the objective function reaches its maximum value.

5. The non-invasive blood pressure measurement method based on multi-view differential feature space according to claim 1, characterized in that: The calculation formula of the frequency domain feature includes: ； ； ； in, is the j-th layer wavelet detail coefficient, is the frequency, is the power spectral density, is the dominant frequency, k is the frequency index, K is the total number of frequency points in the spectrum, and n is the number of frequency points.

6. The non-invasive blood pressure measurement method based on multi-view differential feature space according to claim 1, characterized in that: The calculation formula of the physiological statistical characteristics includes: ； ； ； ； in, is the average value of the PPG signal amplitude, is the standard deviation of the PPG signal amplitude, No. i The amplitude of the sampling points, is the average slope of the rising segment, N is the total number of sampling points, a, b, c, d, and e All are APG feature point amplitudes.

7. The non-invasive blood pressure measurement method based on multi-view differential feature space according to claim 1, characterized in that: The division ratio of the training set, the validation set and the test set is: 7:1.5:1.5.

Images