Blood pressure measurement system and method based on oscillatory wave feature extraction and supervised fusion

Abstract

The application provides a blood pressure measurement system and method based on oscillatory wave feature extraction and supervised fusion, and relates to the technical field of wearable medical health monitoring. First, an oscillatory wave database is constructed, the oscillatory wave database comprising oscillatory wave signals and corresponding blood pressure measurement values, then the original oscillatory wave signals of a sample to be identified are obtained, and according to sensor adhesion and lead connection detection indicators, usable oscillatory wave signals and unusable oscillatory wave signals are distinguished, and the usable oscillatory wave signals are quality evaluated; the oscillatory wave signals of good quality are preprocessed, time domain, frequency domain and time-frequency features of the preprocessed oscillatory wave signals are extracted respectively, and multi-index features of the oscillatory wave signals are obtained; finally, the time domain, frequency domain and time-frequency features of the obtained oscillatory wave signals are standardized and supervised feature fusion is performed, feature weighted fusion is performed, and then non-invasive blood pressure measurement is realized.

Description

Technical Field

[0001] The present invention relates to the field of wearable medical health monitoring technology, and in particular to a blood pressure measurement system and method based on oscillation wave feature extraction and supervised fusion. Background Art

[0002] With rising health awareness, the demand for non-invasive blood pressure monitoring in the field of wearable medical health monitoring technology is increasing. Currently, traditional non-invasive blood pressure measurement methods have numerous limitations and cannot meet the needs of precision medicine. On the one hand, physiological differences between individuals, such as differences in vascular elasticity and heart rate variability, lead to large errors in measurement methods based on general models, making it difficult to accurately reflect individual blood pressure conditions. On the other hand, existing technologies lack sufficient precision in processing and analysis of oscillation wave signals, fail to fully account for these individual differences, and struggle to fully extract key features. These two factors jointly affect the accuracy and reliability of non-invasive blood pressure measurement technology, limiting its application and development in personalized medicine.

[0003] Chinese patent CN101548883A determines blood pressure by measuring the amplitude of an oscillating wave during inflation and deflation. While this method improves accuracy by correcting blood pressure using an acoustic or optical pulse monitoring device, it is essentially an expanded application of the amplitude coefficient method, insufficiently exploring the characteristics of the oscillating wave signal, and resulting in large measurement errors under complex physiological conditions. Chinese patent CN119138869A uses an oscillometric method to apply a proportional coefficient to obtain preliminary blood pressure results, then constructs a model using the extracted oscillating wave characteristic parameters to calculate the final blood pressure. However, its single feature dimension makes the model less adaptable and less reliable when faced with complex physiological signals and environmental interference.

[0004] In summary, existing technologies generally suffer from problems such as a single feature dimension, insufficient individual adaptability, and inadequate signal quality assessment, making it difficult to achieve high-precision blood pressure measurement under complex physiological conditions and dynamic environments. This invention, through multimodal feature fusion, supervised learning, and personalized matching technology, breaks through the traditional method's reliance on a single feature, effectively addressing measurement bias caused by individual waveform specificity, and improving the accuracy and robustness of blood pressure monitoring. Summary of the Invention

[0005] In view of the deficiencies of the existing technology, the present invention provides a blood pressure measurement system and method based on oscillation wave feature extraction and supervised fusion, which realizes non-invasive blood pressure measurement based on oscillation wave signals.

[0006] On the one hand, a blood pressure measurement system based on oscillation wave feature extraction and supervised fusion includes: an oscillation wave signal acquisition and available signal discrimination module, a preprocessing module, a multi-index feature extraction module, a multi-index feature standardization processing and supervised fusion module and a non-invasive blood pressure measurement module.

[0007] The oscillation wave signal acquisition and usable signal discrimination module is used to continuously acquire the original oscillation wave signal in real time, judge the oscillation wave signal based on the cuff fit and arm movement, distinguish usable oscillation wave signals from unusable oscillation wave signals, and perform quality assessment on the usable oscillation wave signals by fusing time domain and frequency domain features. The usable oscillation waves with fusion features greater than a set threshold are called oscillation wave signals of good quality, and the remaining usable oscillation waves are called oscillation wave signals of poor quality.

[0008] The preprocessing module preprocesses the oscillation wave signal of good quality to remove noise in the oscillation wave signal and improve the quality of the oscillation wave signal;

[0009] The multi-index feature extraction module extracts the time domain features of the oscillating wave signal based on the Hamilton algorithm, extracts the frequency domain features of the oscillating wave signal based on the continuous wavelet transform, and extracts the time-frequency features of the oscillating wave signal based on the Hilbert-Huang transform (HHT);

[0010] The multi-index feature standardization processing and supervised fusion module performs standardization processing on the oscillation wave signal feature data, applies independent component analysis (ICA) to reduce the dimension of the standardized oscillation wave signal feature data, and uses supervised variational autoencoders to perform feature learning and fusion on the features selected by the Fisher criterion;

[0011] The non-invasive blood pressure measurement module is based on the fusion features extracted by the supervised variational autoencoder, performs feature importance evaluation based on the reconstruction error contribution, performs weighted fusion of multi-index features based on XGBoost weight distribution, and applies the pre-trained model of non-invasive blood pressure prediction to perform non-invasive blood pressure measurement, ultimately completing blood pressure measurement.

[0012] On the other hand, a blood pressure measurement method based on oscillation wave feature extraction and supervised fusion is implemented based on the aforementioned blood pressure measurement system, comprising the following steps:

[0013] Step 1: Construct an oscillation wave database, which includes oscillation wave signals and their corresponding blood pressure measurement values;

[0014] Step 1.1: Collect the subject's oscillatory wave signal and its corresponding blood pressure measurement value as a monitoring indicator;

[0015] Step 1.2: Preprocess the oscillation wave signal in the oscillation wave database;

[0016] Step 1.3: Extract the time domain, frequency domain, and time-frequency features of the oscillatory wave signal in the oscillatory wave database, and select the supervisory features related to blood pressure measurement;

[0017] Step 1.4: Build and train a supervised variational autoencoder.

[0018] Step 1.5: Calculate feature weights based on XGBoost and the oscillatory wave signal to be analyzed.

[0019] Step 2: Obtain the original oscillation wave signal of the sample to be identified, and distinguish between usable oscillation wave signals and unusable oscillation wave signals based on sensor fit and wire connection detection indicators, and perform quality assessment on the usable oscillation wave signals;

[0020] Step 2.1: Detect the collected oscillation wave signals and distinguish between usable oscillation wave signals and unusable oscillation wave signals;

[0021] The unusable oscillation wave signal is an oscillation wave signal whose waveform is distorted or completely disappears, and which loses effective information;

[0022] Step 2.2: For the available oscillation wave signal, comprehensively consider both time domain and frequency domain indicators to evaluate the signal quality;

[0023] Step 2.2.1: Perform time domain analysis on the oscillation wave signal to determine the signal quality evaluation index in the time domain;

[0024] In the time domain analysis, the amplitude variation index, baseline drift index and pulse rate consistency index are used to evaluate the quality of the oscillation wave signal;

[0025] Specifically, the oscillation wave signal is subjected to variational mode decomposition and period segmentation to obtain a reconstructed signal and a single-period signal that characterize the signal baseline drift;

[0026] The amplitude variation index is obtained by the amplitude difference of the signal waveform in multiple cardiac cycles; the amplitude variation index is obtained by the amplitude difference of the signal waveform in multiple cardiac cycles; the amplitude variation index in the i-th cycle of the oscillation wave signal a ( i ), as shown in the following formula: ;in, and They are the first i The maximum and minimum values ​​of the waveform amplitude of each cycle;

[0027] The baseline drift index is obtained by windowing the reconstructed signal; the baseline drift index is obtained by windowing the reconstructed signal; the reconstructed signal is windowed to obtain the baseline drift index within the i-th cycle of the oscillation wave signal. b ( i ), as shown in the following formula: ;in, bWave i is the reconstructed signal of the ith period of the oscillation wave signal, bIt is the baseline standard value of the oscillation wave signal in the static state;

[0028] The pulse rate consistency index is calculated by calculating the consistency of the signal time length in multiple cycles; the pulse rate consistency index is calculated by calculating the consistency of the signal time length in multiple cycles; the cycle length of the oscillation wave signal is set to l , get the pulse rate consistency index within the i-th cycle of the oscillation wave signal , as shown in the following formula: ;in,

[0029] For the calculation of i The standard deviation of the signal cycle length, For the calculation of i The mean of the signal cycle lengths;

[0030] Step 2.2.2: Perform frequency domain analysis on the oscillation wave signal to determine the signal quality evaluation index in the frequency domain;

[0031] In the frequency domain analysis, the frequency component index is used to evaluate the quality of the oscillation wave signal;

[0032] Specifically: the frequency component index is obtained by selecting the effective frequency band of the oscillation wave signal through fast Fourier transform and calculating the power spectrum density; the frequency component index in the i-th cycle of the oscillation wave signal f ( i )for: ;in, P is the power of the oscillation wave signal, f Indicates frequency. The effective frequency range of the oscillation wave signal is selected to be 0.5 Hz ~ 10 Hz;

[0033] Step 2.2.3: Calculate the entropy value based on the signal quality evaluation indicators in the time domain and frequency domain to obtain the weight coefficient, and perform time-frequency domain multi-indicator weighted fusion on the oscillation wave signal to obtain the fusion indicator;

[0034] The fusion index SQI ( i ), as shown in the following formula: ;in 、 、 、 The evaluation indicators are 、 、 The entropy weight of

[0035] Step 2.2.4: Perform threshold discrimination on the fusion index by setting a threshold to classify the available oscillation wave signals into signals of good quality and signals of poor quality;

[0036] If the fusion index is greater than the set threshold, the oscillation wave signal can be regarded as a good quality signal, otherwise it is a poor quality signal. For the good quality signal, step 3 is performed. After filtering, denoising and baseline removal, further analysis is performed in the time domain, frequency domain and time-frequency domain. For the poor quality signal, the oscillation wave needs to be re-collected.

[0037] Step 3: Preprocess the high-quality oscillation wave signal, including noise removal, filtering, and downsampling operations, to remove noise from the oscillation wave signal;

[0038] Step 3.1: Use a digital filter to remove noise from the oscillation wave signal; specifically, low-pass filtering and median filtering;

[0039] Step 3.2: Use sliding window averaging and high-pass filtering to smooth the signal and remove baseline drift and low-frequency noise;

[0040] Step 3.3: Use the decimation method to downsample, selectively retain key feature samples through interpolation and multi-order decimation;

[0041] Step 3.4: Construct a supervised feature selection unit to align the preprocessed oscillatory wave signals with the blood pressure measurements in the oscillatory wave database and mark them as supervised samples.

[0042] Step 4: Extract the time domain, frequency domain and time-frequency features of the preprocessed oscillation wave signal to obtain the multi-index features of the oscillation wave signal;

[0043] Step 4.1: Extract the time domain features of the oscillation wave signal based on the Hamilton algorithm;

[0044] Step 4.2: Extract the frequency domain features of the oscillating wave signal based on continuous wavelet transform (CWT);

[0045] By performing continuous wavelet transform (CWT) on the signal, the signal is decomposed into wavelet basis functions of different scales. The mathematical formula of CWT is as follows: ;in, t is the time variable, x(t) is the input oscillation wave signal; ψ ( t ) is the mother wavelet function; a is the scale parameter; b is the translation parameter; CWTx ( a , b ) are wavelet coefficients;

[0046] Step 4.3: Extract the time-frequency characteristics of the oscillating wave signal based on the Hilbert-Huang transform (HHT);

[0047] The Hilbert-Huang transform (HHT) includes empirical mode decomposition (EMD) and Hilbert transform, and uses empirical mode decomposition (EMD) to decompose the original blood pressure oscillation wave signal into several intrinsic mode functions (IMFs); each IMF represents a different frequency component in the signal;

[0048] Perform Hilbert transform on each IMF to calculate its instantaneous frequency and instantaneous amplitude. The formula of Hilbert transform is: ;in: represents the Cauchy principal value, is the original oscillation wave signal, is the integration variable;

[0049] Step 4.4: Supervised feature selection based on the oscillatory wave database;

[0050] For the preprocessed oscillatory wave signal, the top 20 features strongly correlated with blood pressure were screened using the mutual information and F-value of the time-frequency features calculated from the blood pressure measurements in the oscillatory wave database and SBP and DBP:

[0051] , ;in, represents mutual information, yes and The joint probability of and are the marginal probabilities of X and Y respectively, 、 is the mean value of feature j in high and low blood pressure groups, 、 is the variance.

[0052] Step 5: Standardize the time domain, frequency domain and time-frequency features of the oscillation wave signal obtained in step 4 and perform supervised feature fusion;

[0053] Step 5.1: Standardize the oscillation wave signal characteristic data;

[0054] The time domain, frequency domain and time-frequency characteristics of the oscillation wave signal are standardized as follows: ;

[0055] Where x is the original eigenvalue of the oscillating wave signal, μ is the mean value of the feature, σ is the standard deviation of the feature, and x1 is the standardized eigenvalue;

[0056] Step 5.2: Calculate the covariance matrix for the standardized oscillation wave signal feature data; perform eigenvalue decomposition on the covariance matrix to obtain a series of eigenvalues ​​and corresponding eigenvectors; use Fisher scores to sort the features and retain the top 80% of the features;

[0057] Step 5.2.1: Calculate the covariance matrix: Calculate the covariance matrix for the standardized oscillation wave signal feature data;

[0058] Step 5.2.2: Eigenvalue decomposition: Perform eigenvalue decomposition on the covariance matrix to obtain a series of eigenvalues ​​and corresponding eigenvectors;

[0059] Step 5.2.3: Feature selection: Rank the features using Fisher score and retain the top 80% of the features;

[0060] For each feature j , Fisher score F j Defined as: ;in, K is the total number of categories; n k For category k The number of samples in ; μ k,j For category k Middle j The mean of the features; μ j For the j The overall mean of the characteristics; σ k,j 2 For category k Middle j The variance of a feature.

[0061] Step 5.2.4: Data projection: Project the original oscillation wave data onto the selected principal component to obtain the data after dimensionality reduction of the oscillation wave signal characteristics.

[0062] Step 5.3: Build and train a supervised variational autoencoder and obtain fused features;

[0063] Step 5.3.1: Construct a supervised variational autoencoder and train it using the blood pressure measurements in the oscillatory wave database as a supervisory signal. The variational autoencoder consists of an encoder and a decoder. The encoder compresses the input features into a low-dimensional space, and the decoder reconstructs the original features from the low-dimensional space.

[0064] Step 5.3.2: Supervised Variational Autoencoder Training: Use the reduced feature data as the input of the supervised variational autoencoder, use the mean squared error as the loss function, and train the supervised variational autoencoder to minimize the difference between the input features and the reconstructed features;

[0065] Step 5.3.3: Feature fusion representation: After training is completed, the output of the decoder, that is, the feature representation in the low-dimensional space, is used as the fusion feature;

[0066] The joint loss function of the supervised variational autoencoder is: ;in, is the mean square error of feature reconstruction, is the mean square error of blood pressure prediction in the blood pressure data set, and β is a hyperparameter used to balance the mean square error of feature reconstruction. and the mean square error of blood pressure prediction of blood pressure values ​​in the database The relative importance of L in the joint loss function.

[0067] Step 5.4: Calculate feature weights based on XGBoost;

[0068] Step 5.4.1: Use the preprocessed and feature-extracted oscillation wave data as input and the corresponding blood pressure measurements as output to construct a training set.

[0069] Step 5.4.2: Use the training set to train the XGBoost model to obtain a pre-trained model for non-invasive blood pressure prediction;

[0070] During the training process, the mean square error (MSE) is used as the loss function, and the formula is as follows: ; where n is the number of samples, is the true blood pressure value, The blood pressure values ​​predicted by the XGBoost model.

[0071] Step 5.4.3: After training, obtain the weight of each feature and obtain the weighted feature vector;

[0072] The XGBoost model outputs the importance score of each feature, and after normalizing these scores, we get the weight of each feature. .

[0073] Step 6: Perform feature weighted fusion to achieve non-invasive blood pressure measurement;

[0074] Step 6.1: Evaluate feature importance based on Fisher score;

[0075] Step 6.1.1: Feature removal: Remove each fused feature from the input data in turn and use the remaining features to reconstruct it using the trained supervised variational autoencoder;

[0076] Step 6.1.2: Reconstruction error calculation: For each feature removed, calculate the reconstruction error and compare it with the reconstruction error of the model without removing any fusion features;

[0077] Step 6.1.3: Importance evaluation: The importance of the feature is reflected by the increase in reconstruction error after feature removal;

[0078] Step 6.2: Perform feature weighted fusion based on the feature importance evaluation results to obtain a weighted feature vector;

[0079] Step 6.2.1: Feature weight assignment;

[0080] The weight of a feature is proportional to its importance. The importance score of each feature is divided by the sum of all feature importance scores to obtain the weight of the feature.

[0081] Step 6.2.2: Perform feature weighted fusion based on weight distribution;

[0082] Apply the calculated weight to each eigenvector to obtain a weighted eigenvector;

[0083] Step 6.3: Perform non-invasive blood pressure measurement based on weighted fusion of multiple index features;

[0084] The weighted feature vector obtained in step 6.2 and the feature weight of XGBoost obtained in step 5.4 are used as the feature template of the current object to be detected; the weighted cosine similarity algorithm and Mahalanobis distance algorithm are used to perform SBP, DBP and MAP recognition and final output in the pre-stored oscillation wave feature database;

[0085] Step 6.3.1: Calculate the weighted cosine similarity and weighted Mahalanobis distance between the sample to be identified and the SBP waveform, DBP waveform, and MAP waveform samples;

[0086] Specifically, the weighted feature vector is used in combination with XGBoost weights to calculate the weighted cosine similarity and weighted Mahalanobis distance between the sample to be identified and each known SBP waveform, DBP waveform, and MAP waveform sample;

[0087] Step 6.3.2: Based on the weighted cosine similarity and weighted Mahalanobis distance between the sample to be identified and each known SBP waveform, DBP waveform, and MAP waveform sample, assign corresponding weight values ​​and calculate the weighted similarity;

[0088] For the sample A to be identified and the known waveform sample B, the weighted cosine similarity is as follows: ;in is the weighted cosine similarity between sample A and sample B; is the feature weight output by the XGBoost model trained on existing data; m is the number of feature dimensions; represents weighted dot product; is the norm of the weighted vector A, is the norm of the weight vector B.

[0089] The weighted Mahalanobis distance is described by the following formula: ;in The weighted Mahalanobis distance between sample A and sample B; W is the weight matrix, which is a diagonal matrix in the form of: ;

[0090] Step 6.3.3: Identify SBP, DBP, and MAP based on weighted similarity;

[0091] By comparing the weighted similarity between the sample to be identified and all known SBP waveforms, DBP waveforms and MAP waveform samples, the SBP, DBP and MAP corresponding to the sample to be identified with the lowest weighted similarity are selected as the final result.

[0092] The beneficial effects of adopting the above technical solution are: the present invention provides a blood pressure measurement system and method based on oscillation wave feature extraction and supervised fusion, by constructing a large oscillation wave database, storing a large number of subjects' oscillation wave signals and blood pressure measurement values ​​as reference standards, and extracting multi-index features to comprehensively analyze the signal from the time domain, frequency domain, and time-frequency domain. By using the supervised variational autoencoder and XGBoost algorithm, not only the features and patterns related to blood pressure are learned, but also the accurate feature weights are calculated. Non-invasive blood pressure measurement is performed by combining cosine similarity and Mahalanobis distance, and the signal quality is strictly evaluated and screened. This series of operations effectively overcomes the measurement deviation caused by individual waveform specificity, greatly improves the accuracy and robustness of blood pressure monitoring, and injects new vitality into the development of non-invasive blood pressure measurement technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 A structural block diagram of a blood pressure measurement system based on oscillation wave feature extraction and supervised fusion provided by an embodiment of the present invention;

[0094] Figure 2 A flow chart of a blood pressure measurement method based on oscillation wave feature extraction and supervised fusion provided by an embodiment of the present invention;

[0095] Figure 3 A flow chart of the quality assessment of an "usable" oscillation wave signal provided by an embodiment of the present invention;

[0096] Figure 4 Flowchart for extracting time domain, frequency domain and time-frequency features of oscillation wave signals provided by an embodiment of the present invention;

[0097] Figure 5 A complete waveform diagram of the oscillation wave signal provided by an embodiment of the present invention;

[0098] Figure 6 A schematic diagram of the time domain characteristics of an oscillation wave signal provided by an embodiment of the present invention;

[0099] Figure 7 Flowchart of the standardized processing of multi-category feature data of oscillation wave signals and the fusion of supervisory features provided by an embodiment of the present invention;

[0100] Figure 8 This is a flow chart of weighted fusion of oscillation wave signal features provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0101] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0102] On the one hand, a blood pressure measurement system based on oscillation wave feature extraction and supervised fusion, e.g. Figure 1 As shown, it includes: oscillation wave signal acquisition and available signal discrimination module, preprocessing module, multi-index feature extraction module, multi-index feature standardization processing and supervision fusion module and non-invasive blood pressure measurement module;

[0103] The oscillation wave signal acquisition and usable signal discrimination module is used to continuously acquire original oscillation wave signals in real time, judge the oscillation wave signals based on the cuff fit and arm movement, distinguish usable oscillation wave signals from unusable oscillation wave signals, and evaluate the quality of usable oscillation wave signals by fusing time domain and frequency domain features. Usable oscillation waves with fusion features greater than a set threshold are called oscillation wave signals with good quality, and the remaining usable oscillation waves are called oscillation wave signals with poor quality.

[0104] In this embodiment, the oscillation wave signal acquisition and usable signal discrimination module discriminates the acquired original oscillation wave signal into usable or unusable signals based on the cuff fit and arm movement. Unusable oscillation wave signals are caused by interference caused by other reasons such as the cuff being too tight or loose, the user's arm moving or exerting force during the measurement, etc., which leads to distortion or complete disappearance of the oscillation wave waveform. Unusable oscillation wave signals cannot be simply restored because they have lost effective information and therefore need to be re-acquired. Usable oscillation wave signals indicate that their signal quality has not been completely obliterated by noise or interference and can be further distinguished as good or bad quality. For usable oscillation wave signals, the signal quality is evaluated by comprehensively considering indicators in both the time domain and frequency domain. In the time domain analysis, the amplitude change degree indicator, the baseline drift degree indicator and the pulse rate consistency degree indicator are used to evaluate the signal quality. In the frequency domain analysis, the frequency component indicator is used to evaluate the signal quality. The weight coefficient obtained by calculating the entropy value of each indicator is used to weightedly fuse the multi-index features of the oscillation wave signal in the time-frequency domain to obtain a fusion index, and threshold discrimination is performed based on the fusion index. The available oscillation wave signals are classified into signals of good quality and signals of poor quality; the signals of good quality are further analyzed, while the signals of poor quality need to be re-collected.

[0105] The preprocessing module preprocesses the oscillation wave signal with good quality, removes the noise in the oscillation wave signal, and improves the quality of the oscillation wave signal; the multi-index feature extraction module extracts the time domain features of the oscillation wave signal based on the Hamilton algorithm, extracts the frequency domain features of the oscillation wave signal based on the continuous wavelet transform, and extracts the time-frequency features of the oscillation wave signal based on the Hilbert-Huang transform (HHT); the multi-index feature standardization processing and supervision fusion module performs standardization processing on the feature data of the oscillation wave signal; after the feature of the oscillation wave signal is extracted, the extracted feature data is first standardized to eliminate the dimension and scale differences; then, the Fisher criterion is applied. Score) performs feature selection on the standardized feature data. By calculating the discriminative power between each feature and the target variable (such as systolic blood pressure and diastolic blood pressure), the most discriminative feature subset is selected, thereby reducing feature dimensionality and improving the accuracy and efficiency of the blood pressure estimation model. A variational autoencoder is used to further learn the selected features. The non-invasive blood pressure measurement module, based on the features extracted by the variational autoencoder, assesses feature importance based on reconstruction error contribution and performs multi-metric weighted feature fusion based on feature weight assignment. To more comprehensively measure the similarity between oscillation wave signal features, the cosine similarity and Mahalanobis distance metrics are combined by setting weight parameters. Cosine similarity measures the directional consistency of feature vectors, while Mahalanobis distance measures the dissimilarity of feature vectors while considering data distribution. By optimizing the weight parameters, a multi-angle assessment of feature similarity can be achieved, thereby improving the accuracy and robustness of the blood pressure estimation model.

[0106] In this embodiment, an oscillatory wave database is established. First, the oscillatory wave signal device is correctly placed on the subject's upper arm and secured with a cuff to ensure accuracy and comfort. The cuff gradually increases pressure, recording the subject's oscillatory wave data as it is pressurized. This data is then stored and analyzed to derive blood pressure measurements. The database must contain 3,000 oscillatory wave data sets with recorded blood pressure values ​​as a reference for subsequent machine learning.

[0107] In this embodiment, the preprocessing module performs filtering and denoising on high-quality oscillatory wave signals to eliminate noise and improve signal quality. First, a digital filter, sliding window averaging, and high-pass filtering are used to smooth the signal and remove baseline drift, low-frequency noise, high-frequency noise, and other interfering components such as sudden noise. This reduces noise interference in the signal and makes subsequent processing more accurate and reliable. The preprocessed oscillatory wave signal is aligned with blood pressure data in a database and marked as a supervised sample for subsequent supervised learning.

[0108] The multi-index feature extraction module extracts the time domain features of the oscillation wave signal based on the Hamilton algorithm, extracts the frequency domain features of the oscillation wave signal based on the continuous wavelet transform, and extracts the time-frequency features of the oscillation wave signal based on the Hilbert-Huang transform (HHT);

[0109] The multi-index feature standardization processing and supervised fusion module performs standardization processing on the oscillation wave signal feature data, applies independent component analysis (ICA) to reduce the dimension of the standardized oscillation wave signal feature data, and uses supervised variational autoencoders to perform feature learning and fusion on the features selected by the Fisher criterion;

[0110] In this embodiment, the multi-index feature extraction module extracts time-domain, frequency-domain and time-frequency features from the pre-processed oscillatory wave signals, obtaining multiple index features of the oscillatory wave signals. The time-domain features of the oscillatory wave signals are extracted based on the Hamilton feature point extraction algorithm. The Hamilton algorithm is used to detect the starting point of the waveform cycle, and then the oscillatory wave waveform is segmented beat by beat to extract the time-domain waveform contour features. The frequency-domain features of the oscillatory wave signals are extracted based on the continuous wavelet transform. The frequency distribution characteristics of the signal are used to divide the frequency bands, and the spectral energy ratio of different sub-band frequency bands is used as the frequency-domain feature of the oscillatory wave signal. The time-frequency graph features of the oscillatory wave signals are extracted based on the Hilbert-Huang transform. The Hilbert-Huang transform (HHT) is performed on the oscillatory wave signal using a sliding window method. First, the intrinsic mode function (IMF) of the signal is obtained through empirical mode decomposition (EMD), and then the Hilbert transform is applied to obtain the time-frequency distribution. Next, the main energy distribution area in the time-frequency graph is detected, and the edge contour closed curve of the area is extracted. Based on the closed curve, the maximum value detection algorithm is used to identify the 8 local maximum values of the local peak points of the energy on the time-frequency distribution graph. The selection basis is the sorting of the energy values. In this embodiment, the multi-index feature standardization processing and supervised fusion module performs standardization processing and supervised feature fusion on the multi-index feature data of the oscillatory wave signal. The multi-index feature standardization of the oscillatory wave signal converts all features to the same scale, ensuring that all features contribute roughly the same to the objective function, and avoiding some features from dominating in parameter updates. Independent components analysis (ICA) is used for the standardized oscillatory wave signal feature data. It aims to reduce the dimension of features while preserving the independent information in the data. ICA first preprocesses the standardized feature data to ensure that the data meets the analysis requirements. Then, the ICA algorithm is used to extract independent components from the mixed signal. This process does not involve the calculation of the covariance matrix, but directly applies the algorithm to separate the independent components from the original signal. By maximizing the non-Gaussianity, ICA can effectively identify multiple source signals. After obtaining the independent components, the N independent components that are most important for analysis can be selected according to actual needs. These components represent the main signal sources in the original data and can reflect the key features of the data. Finally, the original data is projected onto the selected independent components to obtain the reduced dimension feature data. Fisher's rule is applied to select features from the standardized oscillatory wave signal feature data, and a supervised variational autoencoder is used to further learn and fuse the ICA reduced dimension features. The variational autoencoder structure consists of an encoder and a decoder. The encoder is responsible for compressing the input features into a low-dimensional latent space and generating the distribution parameters (usually mean and variance) of the points in this space. The decoder attempts to reconstruct the original input features from this latent space.In the decoder, the generated output is a feature representation sampled from the latent space, which can be used as the basis for further analysis or fusion. Unlike traditional autoencoders, variational autoencoders introduce probabilistic properties in the latent space, giving them greater flexibility in generating models and feature representations.

[0111] The non-invasive blood pressure measurement module is based on the fusion features extracted by the supervised variational autoencoder, performs feature importance evaluation based on the reconstruction error contribution, performs weighted fusion of multi-index features based on XGBoost weight distribution, and applies the pre-trained model of non-invasive blood pressure prediction to perform non-invasive blood pressure measurement, ultimately completing blood pressure measurement.

[0112] In this embodiment, the non-invasive blood pressure measurement module performs feature weighted fusion; based on the fusion features extracted from the supervised variational autoencoder, the feature importance is evaluated based on the contribution of the reconstruction error; the change in the autoencoder reconstruction error after each fusion feature is removed is analyzed, and the greater the contribution of the feature to the reconstruction error, the more important the feature is for the representation of the data, which facilitates the evaluation of the importance of the fusion feature. For the data after each feature is removed, the reconstruction error is calculated and compared with the reconstruction error of the original model (when no features are removed). The increase in the reconstruction error after the feature is removed reflects the importance of the feature, and the feature with a significant increase in the reconstruction error is considered to be more important to the model and data representation. In the process of feature fusion and similarity matching, the weight of each feature is determined by using the XGBoost model trained based on the blood pressure and waveform data in the database. , to optimize weighted similarity calculations for non-invasive blood pressure measurement. Cosine similarity and Mahalanobis distance metrics are used for blood pressure measurement and final output. By comparing the weighted similarity between the sample to be identified and all known SBP and DBP waveforms, a lower weighted similarity indicates a closer match between the waveform and SBP / DBP, thus increasing the confidence level of the blood pressure measurement. This effectively utilizes the information in the weighted feature vectors and selects the final SBP and DBP results for the sample to be identified with the lowest weighted similarity.

[0113] On the other hand, a blood pressure measurement method based on oscillation wave feature extraction and supervised fusion is implemented based on the aforementioned blood pressure measurement system, such as Figure 2 As shown, the following steps are included:

[0114] Step 1: Construct an oscillation wave database, which includes oscillation wave signals and their corresponding blood pressure measurement values;

[0115] In this embodiment, the oscillation wave signals and their corresponding blood pressure measurements of at least 3,000 subjects are stored and processed as reference standards;

[0116] Step 1.1: Collect oscillation wave signals and their corresponding blood pressure measurements from at least 3,000 subjects as monitoring indicators, which serve as the reference standard for subsequent blood pressure prediction;

[0117] Step 1.2: Preprocess the oscillation wave signal in the oscillation wave database;

[0118] Step 1.3: Extract the time domain, frequency domain, and time-frequency features of the oscillatory wave signal in the oscillatory wave database, and select the supervisory features related to blood pressure measurement;

[0119] Step 1.4: Build and train a supervised variational autoencoder.

[0120] Step 1.5: Calculate feature weights based on XGBoost and the oscillatory wave signal to be analyzed.

[0121] Step 2: Obtain the original oscillation wave signal of the sample to be identified, and distinguish between usable oscillation wave signals and unusable oscillation wave signals based on sensor fit and wire connection detection indicators, and perform quality assessment on the usable oscillation wave signals;

[0122] Step 2.1: Detect the collected oscillation wave signals and distinguish between usable oscillation wave signals and unusable oscillation wave signals;

[0123] The unusable oscillation wave signal is an oscillation wave signal whose waveform is distorted or completely disappears, and which loses effective information;

[0124] In this embodiment, the "unusable" oscillation wave signal refers to interference caused by the cuff being too tight or loose, the user's arm movement or force during measurement, and other reasons, which leads to the distortion or complete disappearance of the oscillation wave signal waveform. The "unusable" signal cannot be simply restored because it has lost effective information and needs to be re-collected. The "usable" signal indicates that its signal quality has not been completely obliterated by noise or interference and can be further distinguished between good and poor quality. Good signals can be directly used for interpretation, while poor signals need to be re-collected.

[0125] Step 2.2: For the available oscillation wave signal, comprehensively consider the indicators in both time domain and frequency domain to evaluate the signal quality, such as Figure 3 As shown;

[0126] Step 2.2.1: Perform time domain analysis on the oscillation wave signal to determine the signal quality evaluation index in the time domain;

[0127] In the time domain analysis, the amplitude variation index, baseline drift index and pulse rate consistency index are used to evaluate the quality of the oscillation wave signal;

[0128] Specifically, the oscillation wave signal is subjected to variational mode decomposition and period segmentation to obtain a reconstructed signal and a single-period signal that characterize the signal baseline drift;

[0129] The amplitude variation index is obtained by the amplitude difference of the signal waveform in multiple cardiac cycles; the amplitude variation index is obtained by the amplitude difference of the signal waveform in multiple cardiac cycles; the amplitude variation index in the i-th cycle of the oscillation wave signal a ( i ), as shown in the following formula: ;in, and They are the first i The maximum and minimum values ​​of the waveform amplitude of each cycle;

[0130] The baseline drift index is obtained by windowing the reconstructed signal; the baseline drift index is obtained by windowing the reconstructed signal; the reconstructed signal is windowed to obtain the baseline drift index within the i-th cycle of the oscillation wave signal. b ( i ), as shown in the following formula: ;in, bWave i is the reconstructed signal of the ith period of the oscillation wave signal, b It is the baseline standard value of the oscillation wave signal in the static state;

[0131] The pulse rate consistency index is calculated by calculating the consistency of the signal time length in multiple cycles; the pulse rate consistency index is calculated by calculating the consistency of the signal time length in multiple cycles; the cycle length of the oscillation wave signal is set to , get the pulse rate consistency index within the i-th cycle of the oscillation wave signal , as shown in the following formula: ;in, For the calculation of i The standard deviation of the signal cycle length, ) is the calculated mean value of the length of the i-th signal cycle;

[0132] Step 2.2.2: Perform frequency domain analysis on the oscillation wave signal to determine the signal quality evaluation index in the frequency domain;

[0133] In the frequency domain analysis, the frequency component index is used to evaluate the quality of the oscillation wave signal;

[0134] Specifically: the frequency component index is obtained by selecting the effective frequency band of the oscillation wave signal through fast Fourier transform and calculating the power spectrum density; the frequency component index in the i-th cycle of the oscillation wave signal f ( i )for: ;in, P is the power of the oscillation wave signal, f Indicates frequency. The effective frequency range of the oscillation wave signal is selected to be 0.5 Hz ~ 10 Hz;

[0135] Step 2.2.3: Calculate the entropy value based on the signal quality evaluation indicators in the time domain and frequency domain to obtain the weight coefficient, and perform time-frequency domain multi-indicator weighted fusion on the oscillation wave signal to obtain the fusion indicator;

[0136] The fusion index SQI ( i ), as shown in the following formula: ;in 、 、 、 The evaluation indicators are 、 、 The entropy weight of

[0137] Step 2.2.4: Perform threshold discrimination on the fusion index by setting a threshold to classify the available oscillation wave signals into signals of good quality and signals of poor quality;

[0138] If the fusion index is greater than the set threshold, the oscillation wave signal can be regarded as a good quality signal, otherwise it is a poor quality signal. For the good quality signal, step 3 is performed. After filtering, denoising and baseline removal, further analysis is performed in the time domain, frequency domain and time-frequency domain. For the poor quality signal, the oscillation wave needs to be re-collected.

[0139] Step 3: Preprocess the high-quality oscillation wave signal, including noise removal, filtering, and downsampling operations to remove noise from the oscillation wave signal. Oscillating wave signals contain multiple sources of noise, including environmental noise, light scattering, motion artifacts, baseline drift, and sensor noise. These noise components can reduce the accuracy and reliability of the signal. Therefore, noise removal and filtering are necessary during the preprocessing process to reduce the impact of noise and extract clean and accurate oscillation wave signal features.

[0140] Step 3.1: Use a digital filter to remove noise from the oscillation wave signal; specifically, low-pass filtering and median filtering are used to filter out interference components such as high-frequency noise and other sudden noise, thereby reducing noise interference in the signal and making subsequent processing more accurate and reliable;

[0141] Step 3.2: Use sliding window averaging and high-pass filtering to filter and smooth the signal and remove baseline drift and low-frequency noise; improve the stability and reliability of the signal and further improve the stability of the oscillation wave signal;

[0142] Step 3.3: Use the decimation method to downsample, selectively retain key feature samples through interpolation and multi-order decimation; reduce the amount of data and computational complexity, while maximizing the retention of important information and reducing the computational load;

[0143] Step 3.4: Construct a supervised feature selection unit to align the preprocessed oscillatory wave signals with the blood pressure measurements in the oscillatory wave database and mark them as supervised samples for subsequent supervised learning.

[0144] Step 4: Extract the time domain, frequency domain and time-frequency features of the pre-processed oscillation wave signal to obtain the multi-index features of the oscillation wave signal, such as Figure 4 As shown;

[0145] Step 4.1: Extract the time domain features of the oscillation wave signal based on the Hamilton algorithm;

[0146] The waveform of the oscillation wave signal mainly includes two waveforms: the main wave and the dicrotic wave. The amplitude, slope and time characteristics of the waveform contain a lot of physiological information of the human body, which can effectively reflect the rhythm pattern and functional state of the cardiovascular system, and contain important physiological information and clinical significance. The complete oscillation wave waveform is shown in the figure below. Figure 5 As shown. The Hamilton algorithm is used to extract the waveform features in the time domain from the single-cycle waveform of the oscillating wave signal: first, the first-order difference of the sparse signal (i.e., calculating the derivative of the signal) is performed to reflect the rate of change of the signal at each sampling point; the derivative signal is squared, which can further amplify large waveform changes in the signal (such as QRS waves) while suppressing smaller changes and noise; a fixed-length sliding window is applied to the squared signal for integration to obtain the energy distribution of the signal over a period of time. The size of the moving window is usually set according to the specific signal sampling frequency and the duration of the target waveform. This step can smooth the signal and highlight the key features in the oscillating wave signal; the position of the feature point is determined by analyzing the local maximum, minimum and change slope of the signal; the detected feature point position and the corresponding time domain feature information (such as the time point, amplitude, period, partial time, slope of the peak and trough) are output for subsequent use;

[0147] Step 4.2: Extract the frequency domain features of the oscillating wave signal based on continuous wavelet transform (CWT);

[0148] The frequency distribution characteristics of the signal are used to divide the frequency bands, and the spectral energy ratios of different sub-bands are used as the frequency domain characteristics of the oscillatory wave signal. The frequency domain characteristics of the oscillatory wave signal, by analyzing the distribution and components of the signal in the frequency domain, reveal the complexity and nuances of the dynamic regulation of the cardiovascular system. These characteristics reflect the response of the heart and vascular system to changes in physiological and pathological conditions. The specific method for extracting the frequency domain characteristics of the oscillatory wave signal based on the continuous wavelet transform (CWT) is as follows:

[0149] The signal is decomposed into wavelet basis functions of different scales by continuous wavelet transform (CWT) on the signal. The mathematical formula of CWT is as follows: ; Where, t is a time variable, x(t) is the input oscillatory wave signal; ψ ( t ) is the mother wavelet function; a is a scale parameter that controls the compression or expansion of the wavelet function, corresponding to the frequency; b is a translation parameter that represents the position of the wavelet on the time axis; CWTx ( a , b ) is the wavelet coefficient, reflecting the frequency information of the signal at different time points and scales;

[0150] The wavelet coefficient matrix calculated by CWT can reflect the changes of the frequency components of the signal in time.

[0151] By analyzing the wavelet coefficients at different scales, the multi-scale frequency domain features of the signal can be extracted, including:

[0152] Dominant frequency: corresponding to the main frequency component of the signal oscillation, which can be determined by the wavelet coefficients at different scales;

[0153] Instantaneous frequency: the frequency change of the signal at different time points, which can be reflected by CWT;

[0154] Energy distribution: the sum of the squares of the wavelet coefficients at different scales can reflect the energy distribution of the signal, which is used to capture the intensity of different frequency components;

[0155] Local spectral characteristics: by analyzing the wavelet coefficients at a specific time point, the frequency domain features of specific events (such as QRS wave, P wave, T wave in ECG signal) can be identified;

[0156] Step 4.3: Extract the time-frequency features of the oscillatory wave signal based on Hilbert-Huang transform (HHT);

[0157] The time-frequency graph feature of the oscillatory wave signal provides a method to simultaneously examine the changes of the signal in time and frequency, revealing the complexity and multidimensionality of the oscillatory wave dynamic changes. This analysis can capture the frequency component changes of the oscillatory wave at different time points, providing more abundant and detailed information about the state of the cardiovascular system. The specific method of extracting the time-frequency graph feature of the oscillatory wave signal based on Hilbert-Huang transform (HHT) is as follows:

[0158] The Hilbert-Huang transform (HHT) includes empirical mode decomposition (EMD) and the Hilbert transform. Using EMD, the original blood pressure oscillation signal is decomposed into several intrinsic mode functions (IMFs). Each IMF represents a different frequency component in the signal. The decomposition steps are as follows:

[0159] (1) Envelope fitting: Using local extreme points, the upper and lower envelopes of the signal are constructed, usually obtained through spline interpolation;

[0160] (2) Extracting the initial IMF: Remove the mean of the envelope from the original signal to obtain the candidate IMF, and iterate repeatedly until it meets the IMF condition, that is:

[0161] The number of zero-crossing points and extreme points of IMF is equal or the difference does not exceed 1;

[0162] In the entire signal, the local average value of the envelope curve of the IMF is zero;

[0163] (3) Remove IMF and repeat the steps: remove the IMF from the original signal and then continue to decompose the residual signal until all modal functions are extracted.

[0164] The result of EMD decomposition is a set of IMFs, which are regarded as components of the signal at different frequency levels.

[0165] Perform Hilbert transform on each IMF to calculate its instantaneous frequency and instantaneous amplitude. The formula of Hilbert transform is: ;in: represents the Cauchy principal value, is the original oscillation wave signal, is the integration variable.

[0166] Based on the time-frequency diagram, such as Figure 6 As shown, the following key features can be extracted:

[0167] (1) Instantaneous frequency change: observe the dynamic changes of frequency components in the blood pressure oscillation wave signal over time. For example, high-frequency components are usually associated with rapidly changing cardiovascular activity, while low-frequency components may reflect the long-term trend of blood pressure;

[0168] (2) Energy distribution: The energy concentration areas in the time-frequency graph correspond to important events in the signal. By analyzing these energy peak areas, the frequency domain characteristics of key events (such as heartbeat or blood pressure fluctuations) can be identified;

[0169] (3) Frequency bandwidth: By observing the changes in frequency bandwidth, we can infer the changes in blood pressure fluctuations in different time periods;

[0170] (4) Time-frequency aggregation: Analyze the aggregation of signals in the time-frequency graph to determine whether the blood pressure signal exhibits stable periodic characteristics;

[0171] Step 4.4: Supervised feature selection based on the oscillatory wave database;

[0172] For the preprocessed oscillatory wave signal, the top 20 features strongly correlated with blood pressure were screened using the mutual information and F-value of the time-frequency features calculated from the blood pressure measurements in the oscillatory wave database and SBP and DBP: , ;in, represents mutual information, yes and The joint probability of and are the marginal probabilities of X and Y respectively, 、 is the mean value of feature j in high and low blood pressure groups, 、 is the variance.

[0173] Step 5: Standardize the time domain, frequency domain and time-frequency features of the oscillation wave signal obtained in step 4 and perform supervision feature fusion, such as Figure 7 As shown;

[0174] Step 5.1: Standardize the oscillation wave signal characteristic data;

[0175] The various features of an oscillatory wave signal have different dimensions and numerical ranges. Standardization eliminates the impact of dimension by converting all features to the same scale, preventing model training from favoring features with large numerical ranges. Standardizing features also ensures that all features contribute roughly the same to the objective function, preventing certain features from dominating parameter updates. Furthermore, standardization helps prevent model overfitting and improves the model's ability to generalize to unseen data. The time domain, frequency domain, and time-frequency features of the oscillatory wave signal are standardized as follows: ; Where x is the original eigenvalue of the oscillating wave signal, μ is the mean value of the feature, σ is the standard deviation of the feature, and x1 is the standardized eigenvalue;

[0176] Step 5.2: Calculate the covariance matrix for the standardized oscillation wave signal feature data; perform eigenvalue decomposition on the covariance matrix to obtain a series of eigenvalues ​​and corresponding eigenvectors; use Fisher scores to sort the features and retain the top 80% of the features;

[0177] Step 5.2.1: Calculate the covariance matrix: Calculate the covariance matrix for the standardized oscillation wave signal feature data;

[0178] Step 5.2.2: Eigenvalue decomposition: Perform eigenvalue decomposition on the covariance matrix to obtain a series of eigenvalues ​​and corresponding eigenvectors;

[0179] Step 5.2.3: Feature selection: Rank the features using Fisher score and retain the top 80% of the features;

[0180] For each feature j , Fisher score F j Defined as: ;in, K is the total number of categories; n k For category k The number of samples in ; μ k,j For category k Middle j The mean of the features; μ j For the j The overall mean of the feature (the mean across all categories); σ k,j 2 For category k Middle j The variance of the feature;

[0181] Step 5.2.4: Data projection: Project the original oscillation wave data onto the selected principal component to obtain the data after dimensionality reduction of the oscillation wave signal features;

[0182] Step 5.3: Build a supervised variational autoencoder, train the model, and obtain fusion features;

[0183] Step 5.3.1: Construct a supervised variational autoencoder and train it using the blood pressure measurements in the oscillatory wave database as a supervisory signal. The variational autoencoder consists of an encoder and a decoder. The encoder compresses the input features into a low-dimensional space, and the decoder reconstructs the original features from the low-dimensional space.

[0184] A supervised variational autoencoder (SVAE) is a model that combines a variational autoencoder with supervised learning. While learning the underlying representation of the data, it also uses a supervisory signal (here, blood pressure data from a database) to optimize the model's parameters to better perform specific tasks, such as blood pressure prediction. By using blood pressure data from the database as a supervisory signal, the model can learn features and patterns related to blood pressure, thereby improving the accuracy and reliability of blood pressure predictions.

[0185] Step 5.3.2: Supervised Variational Autoencoder Training: Use the reduced feature data as the input of the supervised variational autoencoder, use the mean squared error as the loss function, and train the supervised variational autoencoder to minimize the difference between the input features and the reconstructed features;

[0186] Step 5.3.3: Feature fusion representation: After training is completed, the output of the decoder, that is, the feature representation in the low-dimensional space, is used as the fusion feature;

[0187] The joint loss function of the supervised variational autoencoder is: ;in, is the mean square error of feature reconstruction, is the mean square error of blood pressure prediction in the blood pressure data set, and β is a hyperparameter used to balance the mean square error of feature reconstruction. and the mean square error of blood pressure prediction of blood pressure values ​​in the database The relative importance in the joint loss function L;

[0188] Step 5.4: Calculate feature weights based on XGBoost. To more accurately measure the importance of each feature in non-invasive blood pressure measurement, this paper uses the XGBoost algorithm to calculate feature weights based on the blood pressure and waveform data in the database. The specific steps are as follows:

[0189] Step 5.4.1: Use the preprocessed and feature-extracted oscillation wave data as input and the corresponding blood pressure measurements (systolic blood pressure (SBP) and diastolic blood pressure (DBP)) as output to construct a training set.

[0190] Step 5.4.2: Train the XGBoost model using the training set.

[0191] XGBoost is a gradient boosting tree algorithm that iteratively trains multiple decision trees to gradually optimize the model's predictive performance. During the training process, the mean squared error (MSE) is used as the loss function, as shown in the following formula: ; where n is the number of samples, is the true blood pressure value, The blood pressure value predicted by the XGBoost model;

[0192] Step 5.4.3: After training, obtain the weight of each feature and obtain the weighted feature vector;

[0193] The XGBoost model outputs the importance score of each feature, and after normalizing these scores, the weight of each feature is obtained. ; These weights reflect the relative importance of each feature in predicting blood pressure values.

[0194] Step 6: Perform feature weighted fusion to achieve non-invasive blood pressure measurement; Figure 8 As shown in Figure 5; based on the fusion features extracted from the supervised variational autoencoder in step 5.3, these features already contain the key information of the original data, but with lower dimensions and more concentrated information.

[0195] Feature importance is evaluated based on the Fisher score, analyzing the change in the variational autoencoder reconstruction error after each fusion feature is removed. The greater the contribution of a feature to the reconstruction error, the more important the feature is for representing the data, which facilitates the evaluation of the importance of fusion features.

[0196] Step 6.1: Evaluate feature importance based on the Fisher score. First, perform feature removal. Remove each fused feature from the input data in turn and reconstruct the remaining features using the trained variational autoencoder. For each feature removed, calculate the reconstruction error and compare it with the reconstruction error of the original model. The increase in reconstruction error after feature removal reflects the importance of the feature.

[0197] Step 6.1.1: Feature removal: Remove each fused feature from the input data in turn and use the remaining features to reconstruct it using the trained supervised variational autoencoder;

[0198] Step 6.1.2: Reconstruction error calculation: For each feature removed, calculate the reconstruction error and compare it with the reconstruction error of the model without removing any fusion features;

[0199] Step 6.1.3: Importance evaluation: The importance of the feature is reflected by the increase in reconstruction error after feature removal;

[0200] Step 6.2: Perform weighted feature fusion based on the feature importance assessment results to obtain a weighted feature vector. The weight of a feature is proportional to its importance. Divide the importance score of each feature by the sum of all feature importance scores to obtain the weight of the feature, thereby assigning feature weights. Perform weighted feature fusion based on the weight assignment. Apply the calculated weight to each feature vector to obtain a weighted feature vector.

[0201] Step 6.2.1: Feature weight assignment;

[0202] The weight of a feature is proportional to its importance. The importance score of each feature is divided by the sum of all feature importance scores to obtain the weight of the feature.

[0203] Step 6.2.2: Perform feature weighted fusion based on weight distribution;

[0204] Apply the calculated weight to each eigenvector to obtain a weighted eigenvector;

[0205] Step 6.3: Perform non-invasive blood pressure measurement based on weighted fusion of multiple index features;

[0206] The weighted feature vector obtained in step 6.2 and the feature weight of XGBoost obtained in step 5.4 are used as the feature template of the current object to be detected; the weighted cosine similarity algorithm and Mahalanobis distance algorithm are used to perform SBP, DBP and MAP recognition and final output in the pre-stored oscillation wave feature database;

[0207] Step 6.3.1: Calculate the weighted cosine similarity and weighted Mahalanobis distance between the sample to be identified and the SBP waveform, DBP waveform, and MAP waveform samples;

[0208] Specifically, the weighted feature vector is used in combination with XGBoost weights to calculate the weighted cosine similarity and weighted Mahalanobis distance between the sample to be identified and each known SBP waveform, DBP waveform, and MAP waveform sample;

[0209] Step 6.3.2: Based on the weighted cosine similarity and weighted Mahalanobis distance between the sample to be identified and each known SBP waveform, DBP waveform, and MAP waveform sample, assign corresponding weight values ​​and calculate the weighted similarity;

[0210] For the sample A to be identified and the known waveform sample B, the weighted cosine similarity is as follows: ;in is the weighted cosine similarity between sample A and sample B; is the feature weight output by the XGBoost model trained on existing data. By introducing XGBoost weights, the importance of each feature in blood pressure measurement can be more accurately reflected, thereby improving the accuracy of similarity matching. m is the number of feature dimensions. represents the weighted dot product, taking into account the weight of each feature; is the norm of the weighted vector A, is the norm of the weight vector B;

[0211] The weighted Mahalanobis distance is described by the following formula: ;in The weighted Mahalanobis distance between sample A and sample B; W is the weight matrix, which is a diagonal matrix in the form of: ;

[0212] Step 6.3.3: Identify SBP, DBP, and MAP based on weighted similarity;

[0213] By comparing the weighted similarity between the sample to be identified and all known SBP waveform, DBP waveform and MAP waveform samples, the lower the weighted similarity, the closer the waveform is to SBP / DBP / MAP, and thus the higher the reliability of blood pressure measurement. The information in the weighted feature vector is effectively utilized, and the SBP, DBP and MAP corresponding to the sample to be identified with the lowest weighted similarity are selected as the final result.

[0214] By comparing the weighted similarity between the sample to be identified and the waveform samples in all known databases, the blood pressure value corresponding to the sample with the highest similarity is selected as the final result, and the stability is ensured by consistency verification of three consecutive cardiac cycles.

[0215] The above description is merely an illustration of the preferred embodiments of the present disclosure and the technical principles employed. Those skilled in the art should understand that the scope of the invention encompassed by the embodiments of the present disclosure is not limited to technical solutions formed by specific combinations of the aforementioned technical features. It also encompasses other technical solutions formed by any combination of the aforementioned technical features or their equivalents, without departing from the aforementioned inventive concept. For example, a technical solution formed by replacing the aforementioned features with (but not limited to) technical features with similar functions disclosed in the embodiments of the present disclosure.

Claims

1. A blood pressure measurement system based on oscillation wave feature extraction and supervised fusion, characterized in that: include: Oscillation wave signal acquisition and available signal discrimination module, preprocessing module, multi-index feature extraction module, multi-index feature standardization processing and supervision fusion module and non-invasive blood pressure measurement module; The oscillation wave signal acquisition and usable signal discrimination module is used to continuously acquire the original oscillation wave signal in real time, judge the oscillation wave signal based on the cuff fit and arm movement, and distinguish usable oscillation wave signals from unusable oscillation wave signals, wherein the unusable oscillation wave signal is a signal with a distorted waveform or completely disappeared; classify the usable oscillation wave signals into signals of good quality and signals of poor quality; wherein the quality of the usable oscillation wave signals is evaluated by fusing time domain and frequency domain features, and the usable oscillation waves with fusion features greater than a set threshold are called oscillation wave signals of good quality, and the remaining usable oscillation waves are called oscillation wave signals of poor quality; The preprocessing module preprocesses the oscillation wave signal of good quality to remove noise in the oscillation wave signal and improve the quality of the oscillation wave signal; The multi-index feature extraction module extracts the time domain features of the oscillating wave signal based on the Hamilton algorithm, extracts the frequency domain features of the oscillating wave signal based on the continuous wavelet transform, and extracts the time-frequency features of the oscillating wave signal based on the Hilbert-Huang transform (HHT); The multi-index feature standardization processing and supervised fusion module performs standardization processing on the oscillation wave signal feature data, applies independent component analysis (ICA) to reduce the dimension of the standardized oscillation wave signal feature data, and uses a supervised variational autoencoder to perform feature learning and fusion on the features selected by the Fisher criterion. Specifically, after the oscillation wave signal features are extracted, the extracted feature data are first standardized, and then the Fisher criterion is applied to the standardized feature data for feature selection. By calculating the discriminative ability between each feature and the target variable, the most discriminative feature subset is screened out, and the selected features are further feature learned using the variational autoencoder. The non-invasive blood pressure measurement module is based on the fusion features extracted by the supervised variational autoencoder, performs feature importance assessment based on the contribution of reconstruction error, performs weighted fusion of multi-index features based on XGBoost weight allocation, and applies a pre-trained model for non-invasive blood pressure prediction to perform non-invasive blood pressure measurement, ultimately completing blood pressure measurement. Specifically, the module combines two measurement methods, cosine similarity and Mahalanobis distance, by setting weight parameters. Cosine similarity is used to measure the directional consistency of feature vectors, while Mahalanobis distance is used to measure the differences of feature vectors under the condition of considering data distribution. By optimizing the weight parameters, a multi-angle evaluation of feature similarity is achieved. The blood pressure measurement system based on oscillation wave feature extraction and supervised fusion is used to implement a blood pressure measurement method based on oscillation wave feature extraction and supervised fusion, comprising the following steps: Step 1: Construct an oscillation wave database, which includes oscillation wave signals and their corresponding blood pressure measurement values; Step 1.1: Collect the subject's oscillatory wave signal and its corresponding blood pressure measurement value as a monitoring indicator; Step 1.2: Preprocess the oscillation wave signal in the oscillation wave database; Step 1.3: Extract the time domain, frequency domain, and time-frequency features of the oscillatory wave signal in the oscillatory wave database, and select the supervisory features related to blood pressure measurement; Step 1.4: Build and train a supervised variational autoencoder. Step 1.5: Calculate feature weights based on XGBoost and the oscillatory wave signal to be analyzed; Step 2: Obtain the original oscillation wave signal of the sample to be identified, and distinguish between usable oscillation wave signals and unusable oscillation wave signals based on sensor fit and wire connection detection indicators, and perform quality assessment on the usable oscillation wave signals; Step 2.1: Detect the collected oscillation wave signals and distinguish between usable oscillation wave signals and unusable oscillation wave signals; The unusable oscillation wave signal is an oscillation wave signal whose waveform is distorted or completely disappears, and which loses effective information; Step 2.2: For the available oscillation wave signal, comprehensively consider both time domain and frequency domain indicators to evaluate the signal quality; Step 2.2.1: Perform time domain analysis on the oscillation wave signal to determine the signal quality evaluation index in the time domain; In the time domain analysis, the amplitude variation index, baseline drift index and pulse rate consistency index are used to evaluate the quality of the oscillation wave signal; Specifically, the oscillation wave signal is subjected to variational mode decomposition and period segmentation to obtain a reconstructed signal and a single-period signal that characterize the signal baseline drift; The amplitude variation index is obtained by the amplitude difference of the signal waveform in multiple cardiac cycles; the amplitude variation index is obtained by the amplitude difference of the signal waveform in multiple cardiac cycles; the amplitude variation index a(i) in the i-th cycle of the oscillation wave signal is shown in the following formula: a(i)=|max(Wave i )-min(Wave i ) (1) Among them, max(Wave i ) and min(Wave i ) are respectively the maximum and minimum values ​​of the waveform amplitude of the ith cycle of the oscillation wave signal after filtering; The baseline drift index is obtained by windowing the reconstructed signal; the baseline drift index is obtained by windowing the reconstructed signal; the reconstructed signal is windowed to obtain the baseline drift index b(i) within the i-th cycle of the oscillation wave signal, as shown in the following formula: b(i)=max(|bWave i |)-b (2) Among them, bWave i is the reconstructed signal of the ith period of the oscillation wave signal, and b is the baseline standard value of the oscillation wave signal in the static state; The pulse rate consistency index is calculated by calculating the consistency of the signal time length in multiple cycles; the pulse rate consistency index is calculated by calculating the consistency of the signal time length in multiple cycles; the cycle length of the oscillation wave signal is set to length (Wave i ), the pulse rate consistency index r(i) within the i-th cycle of the oscillation wave signal is obtained as shown in the following formula: r(i)=(SD(length(Wave i ) / Mean(length(Wave i )))·100% (3) Among them, SD(length(Wave i )) is the standard deviation of the calculated length of the i-th signal cycle, Mean(length(Wave i )) is the calculated mean value of the length of the i-th signal cycle; Step 2.2.2: Perform frequency domain analysis on the oscillation wave signal to determine the signal quality evaluation index in the frequency domain; In the frequency domain analysis, the frequency component index is used to evaluate the quality of the oscillation wave signal; Specifically, the effective frequency band of the oscillation wave signal is selected by fast Fourier transform to calculate the power spectrum density to obtain the frequency component index; the frequency component index f(i) in the i-th cycle of the oscillation wave signal is: Wherein, P is the power of the oscillation wave signal, f represents the frequency, and the effective frequency range of the oscillation wave signal is selected to be 0.5Hz~10Hz; Step 2.2.3: Calculate the entropy value based on the signal quality evaluation indicators in the time domain and frequency domain to obtain the weight coefficient, and perform time-frequency domain multi-indicator weighted fusion on the oscillation wave signal to obtain the fusion indicator; The fusion index SQI(i) is shown in the following formula: SQI(i)=w a (i)·a(i)+w b (i)·b(i)+w r (i)·r(i)+w f (i)·f(i) (5) where w a (i) w b (i) w r (i) w f (i) are the entropy weights of the evaluation indicators a(i), b(i), r(i), and f(i); Step 2.2.4: Perform threshold discrimination on the fusion index by setting a threshold to classify the available oscillation wave signals into signals of good quality and signals of poor quality; If the fusion index is greater than the set threshold, the oscillation wave signal is considered a good quality signal; otherwise, it is a poor quality signal. For good quality signals, step 3 is performed. After filtering, denoising, and baseline removal, further analysis is performed in the time domain, frequency domain, and time-frequency domain. For poor quality signals, the oscillation wave needs to be re-collected. Step 3: Preprocess the high-quality oscillation wave signal, including noise removal, filtering, and downsampling operations, to remove noise from the oscillation wave signal; Step 3.1: Use a digital filter to remove noise from the oscillation wave signal; specifically, low-pass filtering and median filtering; Step 3.2: Use sliding window averaging and high-pass filtering to smooth the signal and remove baseline drift and low-frequency noise; Step 3.3: Use the decimation method to downsample, selectively retain key feature samples through interpolation and multi-order decimation; Step 3.4: Construct a supervised feature selection unit to align the preprocessed oscillatory wave signal with the blood pressure measurement values ​​in the oscillatory wave database and mark them as supervised samples; Step 4: Extract the time domain, frequency domain and time-frequency features of the preprocessed oscillation wave signal to obtain the multi-index features of the oscillation wave signal; Step 4.1: Extract the time domain features of the oscillation wave signal based on the Hamilton algorithm; Step 4.2: Extract the frequency domain features of the oscillating wave signal based on continuous wavelet transform (CWT); By performing continuous wavelet transform (CWT) on the signal, the signal is decomposed into wavelet basis functions of different scales; the mathematical formula of CWT is as follows: Where t is the time variable, x(t) is the input oscillating wave signal; ψ(t) is the mother wavelet function; a is the scale parameter; b is the translation parameter; CWTx(a,b) is the wavelet coefficient; Step 4.3: Extract the time-frequency characteristics of the oscillating wave signal based on the Hilbert-Huang transform (HHT); The Hilbert-Huang transform (HHT) includes empirical mode decomposition (EMD) and Hilbert transform, and uses empirical mode decomposition (EMD) to decompose the original blood pressure oscillation wave signal into several intrinsic mode functions (IMFs); each IMF represents a different frequency component in the signal; Perform Hilbert transform on each IMF to calculate its instantaneous frequency and instantaneous amplitude. The formula of Hilbert transform is: Where: PV represents the Cauchy principal value, x(τ) is the original oscillation wave signal, and τ is the integral variable; Step 4.4: Supervised feature selection based on the oscillatory wave database; For the preprocessed oscillatory wave signal, the top 20 features strongly correlated with blood pressure were screened using the mutual information and F-value of the time-frequency features calculated from the blood pressure measurements in the oscillatory wave database and SBP and DBP: Where I(X,Y) represents mutual information, P(x,y) is the joint probability when X=x and Y=y, P(x) and P(y) are the marginal probabilities of X and Y respectively. is the mean value of feature j in high and low blood pressure groups, is the variance; Step 5: Standardize the time domain, frequency domain and time-frequency features of the oscillation wave signal obtained in step 4 and perform supervised feature fusion; Step 5.1: Standardize the oscillation wave signal characteristic data; The time domain, frequency domain and time-frequency characteristics of the oscillation wave signal are standardized as follows: Where x is the original eigenvalue of the oscillating wave signal, μ is the mean value of the feature, σ is the standard deviation of the feature, and x1 is the standardized eigenvalue; Step 5.2: Calculate the covariance matrix for the standardized oscillation wave signal feature data; perform eigenvalue decomposition on the covariance matrix to obtain a series of eigenvalues ​​and corresponding eigenvectors; use Fisher scores to sort the features and retain the top 80% of the features; Step 5.2.1: Calculate the covariance matrix: Calculate the covariance matrix for the standardized oscillation wave signal feature data; Step 5.2.2: Eigenvalue decomposition: Perform eigenvalue decomposition on the covariance matrix to obtain a series of eigenvalues ​​and corresponding eigenvectors; Step 5.2.3: Feature selection: Rank the features using Fisher score and retain the top 80% of the features. For each feature j, the Fisher score F j Defined as: Where K is the total number of categories; n k is the number of samples in category k; μ k,j is the mean of the jth feature in category k; μ j is the overall mean of the jth feature; σ k,j 2 is the variance of the jth feature in category k; Step 5.2.4: Data projection: Project the original oscillation wave data onto the selected principal component to obtain the data after dimensionality reduction of the oscillation wave signal features; Step 5.3: Build and train a supervised variational autoencoder and obtain fused features; Step 5.3.1: Construct a supervised variational autoencoder and train it using the blood pressure measurements in the oscillatory wave database as a supervisory signal. The variational autoencoder consists of an encoder and a decoder. The encoder compresses the input features into a low-dimensional space, and the decoder reconstructs the original features from the low-dimensional space. Step 5.3.2: Supervised Variational Autoencoder Training: Use the reduced feature data as the input of the supervised variational autoencoder, use the mean squared error as the loss function, and train the supervised variational autoencoder to minimize the difference between the input features and the reconstructed features; Step 5.3.3: Feature fusion representation: After training is completed, the output of the decoder, that is, the feature representation in the low-dimensional space, is used as the fusion feature; The joint loss function of the supervised variational autoencoder is: L=L recon +β·L cuff (12) Among them, L recon is the mean square error of feature reconstruction, L cuff is the mean square error of blood pressure prediction for blood pressure values ​​in the database, and β is a hyperparameter used to balance the mean square error of feature reconstruction L recon and the blood pressure value in the database, the mean square error of blood pressure prediction L cuff The relative importance in the joint loss function L; Step 5.4: Calculate feature weights based on XGBoost; Step 5.4.1: Use the preprocessed and feature-extracted oscillation wave data as input and the corresponding blood pressure measurements as output to construct a training set. Step 5.4.2: Use the training set to train the XGBoost model to obtain a pre-trained model for non-invasive blood pressure prediction; During the training process, the mean square error (MSE) is used as the loss function, and the formula is as follows: Where n is the number of samples, y i is the true blood pressure value, The blood pressure value predicted by the XGBoost model; Step 5.4.3: After training, obtain the weight of each feature and obtain the weighted feature vector; The XGBoost model outputs the importance score of each feature, and after normalizing these scores, the weight ω of each feature is obtained. i ; Step 6: Perform feature weighted fusion to achieve non-invasive blood pressure measurement; Step 6.1: Evaluate feature importance based on Fisher score; Step 6.1.1: Feature removal: Remove each fused feature from the input data in turn and use the remaining features to reconstruct it using the trained supervised variational autoencoder; Step 6.1.2: Reconstruction error calculation: For each feature removed, calculate the reconstruction error and compare it with the reconstruction error of the model without removing any fusion features; Step 6.1.3: Importance evaluation: The importance of the feature is reflected by the increase in reconstruction error after feature removal; Step 6.2: Perform feature weighted fusion based on the feature importance evaluation results to obtain a weighted feature vector; Step 6.2.1: Feature weight assignment; The weight of a feature is proportional to its importance. The importance score of each feature is divided by the sum of all feature importance scores to obtain the weight of the feature. Step 6.2.2: Perform feature weighted fusion based on weight distribution; Apply the calculated weight to each eigenvector to obtain a weighted eigenvector; Step 6.3: Perform non-invasive blood pressure measurement based on weighted fusion of multiple index features; The weighted feature vector obtained in step 6.2 and the feature weight of XGBoost obtained in step 5.4 are used as the feature template of the current object to be detected; the weighted cosine similarity algorithm and Mahalanobis distance algorithm are used to perform SBP, DBP and MAP recognition and final output in the pre-stored oscillation wave feature database; Step 6.3.1: Calculate the weighted cosine similarity and weighted Mahalanobis distance between the sample to be identified and the SBP waveform, DBP waveform, and MAP waveform samples; Specifically, the weighted feature vector is used in combination with XGBoost weights to calculate the weighted cosine similarity and weighted Mahalanobis distance between the sample to be identified and each known SBP waveform, DBP waveform, and MAP waveform sample; Step 6.3.2: Based on the weighted cosine similarity and weighted Mahalanobis distance between the sample to be identified and each known SBP waveform, DBP waveform, and MAP waveform sample, assign corresponding weight values ​​and calculate the weighted similarity; For the sample A to be identified and the known waveform sample B, the weighted cosine similarity is as follows: Where WCS(A,B) is the weighted cosine similarity between sample A and sample B; ω i is the feature weight output by the XGBoost model trained on existing data; m is the number of feature dimensions; represents weighted dot product; is the norm of the weighted vector A, is the norm of the weight vector B; The weighted Mahalanobis distance is described by the following formula: Among them D W (A, B) is the weighted Mahalanobis distance between sample A and sample B; W is the weight matrix, which is a diagonal matrix in the form of: Step 6.3.3: Identify SBP, DBP, and MAP based on weighted similarity; By comparing the weighted similarity between the sample to be identified and all known SBP waveforms, DBP waveforms and MAP waveform samples, the SBP, DBP and MAP corresponding to the sample to be identified with the lowest weighted similarity are selected as the final result.

Images