Personalized physiological parameter measurement method combined with biometric recognition

Abstract

The present invention provides a personalized physiological parameter measurement method combined with biometric identification. The method includes: sampling the acquired physiological signals and generating a one-dimensional sequence; reducing the noise of the original signal and realizing data segmentation; selecting and normalizing the segments that meet the requirements using a quality assessment module; obtaining a multidimensional data feature matrix through mathematical transformation; using an expandable deep network with the ability to capture long-range dependencies to realize the user's biometric identification; and extracting potential features based on a multi-scale deep fusion network to realize personalized physiological parameter measurement for different users. The method of the present invention provides a non-invasive and portable physiological parameter measurement method that can realize personalized monitoring in daily life. Its application range covers the measurement of all physiological parameters. It has certain application value in the fields of cardiovascular disease research, user biometric identification, and physiological signal processing research, and is conducive to personalized active management of health and long-term monitoring of related chronic diseases.

Description

Technical Field

[0001] The invention relates to a personalized physiological parameter measurement method combined with biometric identification. Background Art

[0002] The heart is one of the most important organs in the human body and is the power pump for blood circulation. It transports blood to various parts of the body through regular beating to meet the needs of human metabolism. A large number of systematic review studies have shown that cardiovascular disease has become the leading cause of death or disability in the global population. The risk and probability of cardiovascular disease in the population gradually increase with age. Long-term monitoring and real-time warning related to various chronic or acute heart diseases have become increasingly important. The measurement of various physiological parameters related to cardiovascular disease plays a very important role and has great practical significance for the prevention and treatment of cardiovascular disease.

[0003] In the existing physiological parameter measurement methods, the individual differences of subjects have been ignored for a long time. Due to the qualitative or quantitative changes in the body caused by individual congenital (genetic) or acquired (acquired) factors, there are significant individual differences in physiological signals between different subjects in clinical practice (for example, signal amplitude, frequency, variability, etc.), which requires the measurement system to make timely targeted adjustments according to the actual situation of the subjects to achieve personalized medicine. At present, the commonly used physiological parameter detection equipment, including electrocardiographs, continuous blood pressure monitoring systems, cardiac output monitoring systems, arteriosclerosis monitors, etc., can provide high-precision evaluation results. However, such equipment not only requires a large number of sensors to be worn during the detection process, but also the large size and expensive price of the equipment determine that it cannot meet the needs of physiological parameter measurement in daily life. In addition, such equipment generally does not have the ability to automatically adjust for different subjects, and it is difficult to adapt to the development trend of the next generation of medical systems represented by personalized medicine and based on precision medicine.

[0004] In recent years, with the vigorous development of flexible materials and edge computing platforms, wearable technology has made it possible to measure long-term physiological parameters in daily life. At the same time, breakthroughs in deep learning technology have made it possible for personal health monitoring based on physiological parameter measurement to transition from traditional hospital environments to daily life, greatly expanding the implementation scenarios of real-time health monitoring, long-term management of chronic diseases, and personalized medicine.

[0005] At present, the demand for personalized medical care is achieved through classification management through manual records by doctors or subjects. This method is inefficient and causes a huge waste of manpower, material and medical resources. Another method is to introduce other personal biometric information to achieve automated individual identity recognition, such as fingerprints, irises, facial information, etc. This method brings about problems such as increased equipment costs and increased equipment size.

[0006] At the same time, for physiological signals, which are one-dimensional time series with significant periodicity, most traditional methods use single-scale convolutional neural networks to achieve automatic feature extraction, thus ignoring the long-range dependencies of the time series, resulting in low measurement accuracy of physiological parameters. On the other hand, some deep learning methods that use recursive neural networks or long short-term memory modules have a large number of parameters and computational costs, which limits their practical application. Summary of the invention

[0007] In order to overcome the above defects, conform to the trend of precision medicine, and meet the needs of practical applications, the purpose of the present invention is to propose a personalized physiological parameter measurement method combined with biometric identification. Under the premise of minimizing the computational complexity of the model, the present invention uses an end-to-end one-dimensional convolutional neural network to model the long-range dependency of the physiological signals collected by the wearable device based on an extensible strategy, fully extracts the feature information of different receptive fields, and realizes user biometric identification based on the physiological signal to be measured; then, according to the identified specific user individual, the exclusive multi-scale fusion network in the evaluation system is automatically allocated to perform feature extraction and regression related to physiological parameters, and the output of the estimated value of the physiological parameter to be measured is obtained, thereby realizing high-precision and personalized physiological parameter measurement.

[0008] The present invention proposes a personalized physiological parameter measurement method combined with biometric recognition, and the specific steps are as follows:

[0009] (1) The original signal is obtained through a dedicated acquisition device. The original signal is the physiological signal of different individual users. A robust signal preprocessing module is deployed on the obtained original signal to improve the signal quality of the input data: first, considering the different configurations of the acquisition device, the original signal is resampled to a fixed frequency, and a filter or a noise reduction method in digital signal processing is used to remove the noise components in the original signal (including baseline drift, power frequency interference, breathing and motion artifacts, and other physiological signals such as electrocardiogram and surface electromyography). Then, a sliding time window is used to segment the filtered signal, and a quality assessment module is used to assess the signal quality. The segmented segments that meet the requirements are selected, and finally the segmented segments are normalized to obtain clean one-dimensional physiological signal segments;

[0010] (2) For the one-dimensional physiological signal fragment obtained in step (1), use mathematical transformation to obtain its signal dynamics information or use mathematical transformation and digital signal processing technology to obtain other features to obtain a multidimensional matrix in: is the jth data segment of the i-th user after segmentation, N1 is the total number of users in the evaluation system, is the total number of segments of the data set after segmentation by the i-th user, C is the number of signal dimensions after transformation, and L represents the length of each segment of the signal after segmentation (the number of sampling points of each sample);

[0011] (3) For the multidimensional matrix obtained in step (2) According to the natural order of physiological signals, the training set is divided into specific ratios. and test set

[0012] (4) extracting feature representations for biometric identification and generating a biometric extraction model for the user, wherein the biometric extraction model includes two modules: a backbone neural network module based on a residual model and a convolution module based on an extensible strategy;

[0013] Based on the residual connection neural network module, a multi-scale network architecture is deployed according to the prior scalable convolutional layer number M to model the long-range dependency relationship of physiological signals and extract potential features. Then the multidimensional matrix obtained in step (2) is A shallow convolution module of uniform scale is input to obtain the first-level feature map S0 of the same dimension; then a convolution module based on an extensible strategy is used to squeeze and select multi-scale and long-distance physiological features, and then an optimized second-level feature map S1 that is more relevant to the physiological parameters to be measured is obtained; after the Mth layer of the scalable convolution layer, the second-level feature map S1 is distilled and represented using a fully connected layer to generate a template for user biometric recognition Among them Yi(i∈Z + , i=1...N1) represents the i-th user, and stores the template in the user pool authenticated by the biometric recognition system;

[0014] (5) Constructing a deep neural network model for estimating physiological parameter measurements, the deep neural network model is based on a residual connection network and includes three parts: a multi-scale feature fusion module, a hybrid attention mechanism, and a convolutional network layer. A multi-scale fusion convolutional network is designed according to the number of convolutional network layers I set a priori, and the multi-dimensional matrix obtained in step (2) is transformed into Input the multi-scale fusion network to obtain the first feature data three-dimensional matrix S2; use the hybrid attention mechanism to mine the second feature data three-dimensional matrix S3 related to the physiological parameters to be measured; according to the set number of convolutional network layers I, use the convolutional network layer to perform regression calculation on the second feature data three-dimensional matrix S3 to generate an estimated value of the physiological parameter, thereby obtaining the two-dimensional data matrix of the i-th user and evaluate inference models in, is the estimated value of the physiological parameter to be measured of the j-th data segment of the i-th user;

[0015] (6) The training set obtained in step (3) The scalable model input into step (4) is iteratively trained and optimized to obtain a reliable and highly secure authentication user template. Input the user authentication template recognition inference model for testing to verify its accuracy in identifying the user;

[0016] (7) The training set obtained in step (3) Input into the multi-scale feature fusion model in step (5) for training and iterative optimization, and obtain physiological parameter evaluation models suitable for different user individuals. Input the generated evaluation model for performance testing to verify its accuracy in evaluating physiological parameters;

[0017] (8) For user Y identified in step (4) i Automatically match the personalized physiological parameter evaluation inference model generated by the user in step (5) Based on the physiological signals to be measured, biometric identification of user identity and personalized evaluation of physiological parameters are realized.

[0018] In the present invention, the physiological parameters include but are not limited to one or more of heart rate, blood pressure, respiratory rate, cardiac function index or arteriosclerosis index.

[0019] In the present invention, the physiological signals of different individual users in step (1) are physiological signals containing physiological and pathological information of the cardiovascular system, mainly including one or more of the following types: electrocardiogram signals, pulse wave signals, cardiac shock signals, seismocardiogram signals or heart sound signals.

[0020] In the present invention, the signal quality assessment method adopted by the quality assessment module used in step (1) includes one or more statistical characteristics such as skewness, kurtosis, sample entropy and morphological characteristics such as area under the waveform line and waveform slope variability.

[0021] In the present invention, the mathematical transformation method for dimensional expansion of the one-dimensional physiological signal segment in step (2) includes any of the following types: differential operation, integral operation, Fourier series decomposition, wavelet transform, empirical mode decomposition, singular value decomposition and variational mode decomposition.

[0022] In the present invention, the scalable convolution module used to capture the long-range dependencies of physiological signals in step (4) includes methods for implementing the scalable strategy including superimposing multi-scale convolution kernels, non-localized weighting, global attention mechanism and feature channel attention mechanism.

[0023] In the present invention, the multi-scale feature fusion module in step (5) can further mine feature information related to the physiological parameters to be estimated, and the fusion method includes multi-dimensional fusion at the data layer, multi-scale fusion at the feature layer, and multi-weight fusion at the decision layer.

[0024] In the present invention, the training set described in steps (6) and (7) is used to generate and save the inference model, and the test set is used to verify the performance indicators of the biometric recognition and physiological parameter measurement model on an unknown data set.

[0025] In the present invention, the step (8) is based on the automatic matching of the existing personalized physiological parameter evaluation model in the system based on the identified specific user; if the identified individual is a new user that does not exist in the system, a universal model that does not distinguish between users is used to give the result of the physiological parameter estimation measurement, and the collected signals are recorded and accumulated for the subsequent generation of the personalized physiological parameter evaluation model.

[0026] The present invention has the following beneficial effects:

[0027] 1. The present invention fully considers the individual differences of physiological signals and automatically allocates exclusive personalized physiological parameter evaluation models based on different subjects to be tested, which is conducive to improving the accuracy of its measurement and promoting the development of personalized medicine and precision medicine;

[0028] 2. The present invention applies mathematical transformation to physiological signals to achieve data dimension expansion under the premise of low computational complexity, and can mine the potential information of physiological signals in different spatiotemporal perspectives and different time-frequency conversion domains, thereby obtaining a more complete feature representation and achieving more accurate physiological parameter measurement;

[0029] 3. The present invention uses convolution operations based on scalable strategies to capture long-range dependencies of physiological signals and characterize the rich information of convolution operations under different receptive fields, which is conducive to improving the accuracy and security of user identification. At the same time, it can serve as the basis for automatically generating and matching personalized models for different users;

[0030] 4. The present invention utilizes multi-scale fusion technology to fully extract potential information of different scales in the original physiological signals, providing reliable personal health data evaluation and measurement results, thereby ensuring the accuracy of physiological parameter measurement;

[0031] 5. The convolution operations of the present invention are all based on one-dimensional convolution operations, sharing the same residual connection network structure. While ensuring low computational complexity and real-time performance, the use of scalable measurement and multi-scale fusion can ensure the accuracy of recognition, the safety of the system and the accuracy of physiological parameter measurement;

[0032] 6. The method of the present invention provides a non-invasive and portable method for measuring physiological parameters, which can realize personalized monitoring in daily life. Its application scope covers the measurement of all physiological parameters. It has certain application value in the fields of cardiovascular disease research, user biometric identification and physiological signal processing research. At the same time, it is conducive to personalized active management of health and long-term accurate monitoring of related chronic diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of the original pulse wave signal (PPG) and the preprocessed pulse wave signal, where the horizontal axis is time and the vertical axis is signal amplitude. (a) Schematic diagram of the comparison of the pulse wave signal before and after noise reduction by the preprocessing module; (b) The VPG and APG signals obtained by applying the first-order difference and second-order difference operations to the clean pulse wave signal after preprocessing, that is, the multidimensional matrix obtained after mathematical transformation, is used as the user biometric recognition network and physiological parameter estimation model.

[0034] Figure 2 A schematic diagram of a biometric identification system based on pulse wave signals in Example 1 of the present invention is shown.

[0035] Figure 3 A schematic diagram of the biometric recognition model in Example 1 of the present invention capturing long-range dependencies is given, wherein A is the feature extraction process of superimposing multi-scale convolution kernels, B is the feature output process of non-localized operations, and C is the feature output process of the global attention mechanism.

[0036] Figure 4 The initialization settings of the multi-scale fusion physiological parameter measurement network in Example 1 of the present invention, the architecture and execution process of the multi-scale feature fusion convolutional layer with the number of scales set to 3 are demonstrated.

[0037] Figure 5 Correlation analysis between the output value and the true value of the blood pressure measurement model implemented in Example 1 of the present invention. (a) Correlation analysis of diastolic pressure, (b) Correlation analysis of systolic pressure; (c) Bland-Altman analysis of diastolic pressure, (d) Bland-Altman analysis of systolic pressure. Figure 4 In the coordinates, the horizontal axis is the true value and the vertical axis is the output value of the model. DETAILED DESCRIPTION

[0038] The method and application of the present invention are further described below in conjunction with the accompanying drawings and embodiments. These embodiments do not limit the present invention; any structural, method, or functional changes made by a person skilled in the art based on these embodiments are included in the protection scope of the present invention.

[0039] Example 1: The personalized physiological parameter measurement method combined with biometric identification of the present invention is applied to the diastolic and systolic blood pressure measurement tasks based on pulse waves, and multiple public databases are used for performance testing. The PRRB, BIDMC and VitalDB databases are selected to perform cross-performance verification across multiple databases on the biometric identification model based on the scalable strategy. The above three databases all contain pulse wave signals (PPG) with a sampling rate of 125Hz. The BIDMC and VitalDB databases also contain electrocardiogram signals and arterial blood pressure signals (ABP), both of which have a sampling rate of 125Hz. They are also used to test the physiological parameter measurement model of the multi-scale fusion network. The PPG signal is used to measure the blood pressure value, and the ABP signal is used as the true value to be compared with the blood pressure value measured by the model. The personalized physiological parameter measurement method combined with biometric identification of the present invention is adopted, and the specific steps are as follows:

[0040] (1) Figure 1 The original pulse wave signal shown in the upper part of (a) is firstly used to remove the noise such as power frequency interference and baseline drift in the original signal by wavelet transform; then the pulse wave signal is segmented using a sliding time window with a width of 8 seconds, and then the skewness is used as an evaluation index to evaluate the signal quality of all data segments (remove significant motion artifacts), and finally the data segments that meet the requirements are subjected to zero mean normalization processing to obtain a clean and evenly segmented pulse wave signal, as shown in FIG. Figure 1 As shown in the lower part of (a).

[0041] (2) Figure 1 As shown in (b), the first-order difference and second-order difference are applied to the uniformly segmented data to expand the dimension, and the pulse wave velocity signal (VPG) and pulse wave acceleration signal (APG) are obtained, resulting in a three-dimensional input matrix [3,1000].

[0042] (3) The three databases are divided into training sets and test sets in a ratio of 8:2 to form three independent data sets; feature representations for biometric recognition are extracted and a user's biometric matching template is generated: Figure 2 As shown, the process of the biometric recognition system based on pulse wave signals mainly includes three main parts: a physiological signal acquisition module, a physiological feature extraction module and a recognition module based on user templates. That is, the signals obtained in steps (1) and (2) of this embodiment are input into the training model of biometric extraction to generate user templates to be matched in the authentication user pool; biometric recognition mainly includes two stages, stage I is the registration and authentication operation of new users in the system, and stage II is the process of matching existing users in the system for identification and decision-making; then the authenticated users will be automatically assigned to their personalized models for estimation and measurement of systolic and diastolic blood pressure.

[0043] (4) Extract feature representations for biometric identification and generate a user's biometric matching template. The feature extraction model contains two main modules, a backbone neural network based on a residual model and a convolutional module based on an extensible strategy. The backbone neural network based on a residual model is used as a shared model and consists of six steps, namely, the first convolutional representation layer with an output of [X, 32, 1000], the maximum pooling layer with an output of [X, 32, 1000], the residual connection layer with a final output of [X, 256, 63], the mean pooling layer with an output of [X, 256, 1], the fully connected layer with an output of [X, 256], and the softmax layer with an output of [X, 256] (where X is the number of samples in the generated model), thereby obtaining a biometric output template for user identification; the residual connection layer is composed of multiple residual modules based on an extensible strategy. The extensible strategy used in the training model for biometric extraction is as follows: Figure 3 As shown, it includes three different angles: stacking multi-scale convolution kernels, non-localization operations, and attention mechanisms.

[0044] (5) The accuracy (ACC) and equal error rate (EER) of user biometric recognition in this embodiment on three public data sets (PRRB, BIDMC and VitalDB) are shown in Table 1, where the higher the ACC, the better the performance, and the lower the EER, the better the performance. SKNet, Res2Net and ROCKET are designed based on superimposed multi-scale convolution kernels, where SKNet is a branched multi-scale convolution, Res2Net is a multi-scale convolution between network layers, and ROCKET is based on random multi-scale convolution kernels for feature extraction; NL and SNL use non-local operations, where NL is a non-local operation for calculating global range weights, and SNL is a non-local operation based on spectral features; SENet and A 2 Refer to the attention mechanism implementation, where SENet is the convolutional channel attention mechanism, A 2 It is a dual attention mechanism. The results show that the scalable mechanism can achieve higher recognition accuracy and lower error rate.

[0045] Table 1 is a performance comparison table of biometric recognition based on three scalable strategies on three public datasets. The evaluation indicators are accuracy (ACC) and equal error rate (EER).

[0046] Table 1:

[0047]

[0048] (6) Figure 4As shown in FIG. 1 , the construction of the multi-scale fusion physiological signal measurement network model consists of five steps, and the five steps include 1, 2, 3, 3, and 3 multi-scale fusion convolution layers respectively. The first multi-scale convolution operation in each step realizes downsampling. The specific structure of the multi-scale fusion convolution layer in this embodiment is as follows: Figure 5 As shown, ConvK represents a convolution layer with a convolution kernel size of K (in this embodiment, K is 1, 7, 9, and 11, respectively). In order to reduce the amount of calculation, the number of scales of the multi-scale fusion convolution layer is set to 3, and the convolution layers with sizes of 7, 9, and 11 are implemented based on the dilated convolution operation. The original input matrix of the multi-scale fusion network model is [X, 3, 1000] (where X is the number of samples of the generated model). The shallow feature matrix [X, 256, 32] is obtained through convolution with a convolution kernel size of 1. Then, using the attention module, the systolic and diastolic subtask convolution networks obtain the feature matrix [X, 256, 32] of the last convolution layer through two convolution layers with a channel number of 512, an average pooling layer, and a fully connected layer, and then the final physiological parameter matrix [X, 2] is obtained through the hybrid attention module and the softmax function. Since this embodiment is a dynamic measurement mode, the extracted blood pressure estimates constitute a one-dimensional sequence, a systolic pressure sequence [Xs, 1] and a diastolic pressure sequence [Xd, 1], where s represents the number of systolic pressure measurements and d represents the number of diastolic pressure measurements.

[0049] (7) The performance test results of this embodiment are as follows Figure 5 As shown: Figure 5 (a) and (b) reflect the correlation between the measured values ​​and true values ​​of systolic and diastolic blood pressure of the physiological parameter measurement model on the test set; Figure 5 (c) and (d) reflect the Bland-Altman analysis between the measured values ​​and true values ​​of systolic and diastolic blood pressure of the physiological parameter measurement model on the test set. The results show that the physiological parameter measurement model constructed in this embodiment can accurately measure systolic and diastolic blood pressure, and can effectively generate a personalized physiological parameter evaluation model for a single user.

[0050] (8) In this embodiment, the system obtains the pulse wave signal of the user, and after a robust and efficient data preprocessing module, it is first input into a biometric recognition network based on a scalable strategy to identify the user to whom the pulse wave signal belongs, and then automatically allocates a personalized model in the system to measure and evaluate physiological parameters. This embodiment shows that the method of the present invention fully considers the individual differences of physiological signals under the premise of low computational complexity, thereby improving the accuracy of physiological parameter measurement.

Claims

1. A personalized physiological parameter measurement method combined with biometric recognition, characterized in that: The specific steps are as follows: (1) The original signal is obtained through the acquisition device, and the original signal is the physiological signal of different individual users. A signal preprocessing module is deployed on the obtained original signal to improve the signal quality of the input data: first, considering the different configurations of the acquisition device, the original signal is resampled to a fixed frequency, and a filter or a noise reduction method in digital signal processing is used to remove the noise components in the original signal. The noise components include baseline drift, power frequency interference, breathing and motion artifacts, and other physiological signals such as electrocardiogram and surface electromyography. Then, the filtered signal is segmented using a sliding time window, and the signal quality is evaluated using a quality assessment module. The segmented segments that meet the requirements are selected, and finally the segmented segments are normalized to obtain clean one-dimensional physiological signal segments; (2) For the one-dimensional physiological signal fragment obtained in step (1), use mathematical transformation to obtain its signal dynamics information, or use mathematical transformation and digital signal processing technology to obtain other features to obtain a multidimensional matrix in: is the jth data segment of the i-th user after segmentation, N1 is the total number of users in the evaluation system, is the total number of segments of the data set after segmentation by the i-th user, C is the number of signal dimensions after transformation, and L represents the length of each segment of the signal or the number of sampling points of each sample after signal segmentation; (3) For the multidimensional matrix obtained in step (2) According to the natural order of physiological signals, the training set is divided according to the ratio and test set (4) extracting feature representations for biometric identification and generating a biometric extraction model for the user, wherein the biometric extraction model includes a backbone neural network module based on a residual model and a convolution module based on an extensible strategy; Based on the residual connection neural network module, a multi-scale network architecture is deployed according to the prior scalable convolutional layer number M to model the long-range dependency relationship of physiological signals and extract potential features. Then the multidimensional matrix obtained in step (2) is A shallow convolution module of uniform scale is input to obtain the first-level feature map S0 of the same dimension; then a convolution module based on an extensible strategy is used to squeeze and select multi-scale and long-distance physiological features, and then an optimized second-level feature map S1 that is more relevant to the physiological parameters to be measured is obtained; after the Mth layer of the scalable convolution layer, the second-level feature map S1 is distilled and represented using a fully connected layer to generate a template for user biometric recognition where Y i (i∈Z + , i=1...N1) represents the i-th user, and stores the template in the user pool authenticated by the biometric recognition system; (5) Constructing a deep neural network model for estimating physiological parameter measurements, the deep neural network model is based on a residual connection network and includes three parts: a multi-scale feature fusion module, a hybrid attention mechanism, and a convolutional network layer. A multi-scale fusion convolutional network is designed according to the number of convolutional network layers I set a priori, and the multi-dimensional matrix obtained in step (2) is converted into Input the multi-scale fusion network to obtain the first feature data three-dimensional matrix S2; use the hybrid attention mechanism to mine the second feature data three-dimensional matrix S3 related to the physiological parameters to be measured; according to the set number of convolutional network layers I, use the convolutional network layer to perform regression calculation on the second feature data three-dimensional matrix S3 to generate an estimated value of the physiological parameter, thereby obtaining the two-dimensional data matrix of the i-th user and evaluate inference models in, is the estimated value of the physiological parameter to be measured of the j-th data segment of the i-th user; (6) The training set obtained in step (3) The scalable model input into step (4) is iteratively trained and optimized to obtain a reliable and highly secure authentication user template. Input the user authentication template recognition inference model for testing to verify its accuracy in identifying the user; (7) The training set obtained in step (3) Input into the multi-scale feature fusion model in step (5) for training and iterative optimization, and obtain physiological parameter evaluation models suitable for different user individuals. Input the generated evaluation model for performance testing to verify its accuracy in evaluating physiological parameters; (8) For user Y identified in step (4) i Automatically match the personalized physiological parameter evaluation inference model generated by the user in step (5) Based on the physiological signals to be measured, biometric identification of user identity and personalized evaluation of physiological parameters are realized.

2. The method according to claim 1, characterized in that The physiological parameters are one or more of heart rate, blood pressure, respiratory rate, cardiac function index or arteriosclerosis index.

3. The method according to claim 1, characterized in that The physiological signals of different individual users in step (1) are physiological signals containing physiological and pathological information of the cardiovascular system, specifically: one or more of an electrocardiogram signal, a pulse wave signal, a cardiac shock signal, an epicardiogram signal or a heart sound signal.

4. The method according to claim 1, characterized in that: The signal quality assessment method adopted by the quality assessment module used in step (1) includes one or more of the characteristics of skewness, kurtosis, sample entropy, area under the waveform line or waveform slope variability.

5. The method according to claim 1, characterized in that In step (2), a mathematical transformation method is used to obtain a multidimensional matrix for the one-dimensional physiological signal segment, and the mathematical transformation method includes any one of differential operation, integral operation, Fourier series decomposition, wavelet transform, empirical mode decomposition, singular value decomposition or variational mode decomposition.

6. The method according to claim 1, characterized in that The method described in step (4) for capturing the long-range dependency of physiological signals based on the scalable convolution module and realizing the scalable strategy includes any one of superimposing multi-scale convolution kernels, non-localized weighting, global attention mechanism and feature channel attention mechanism.

7. The method according to claim 1, characterized in that The multi-scale feature fusion module in step (5) can further mine feature information related to the physiological parameters to be estimated, specifically one or more of multi-dimensional fusion at the data layer, multi-scale fusion at the feature layer, or multi-weight fusion at the decision layer.

8. The method according to claim 1, characterized in that: The training set described in steps (6) and (7) is used to generate and save the inference model, and its training method includes one or more of supervised learning, semi-supervised learning, weakly supervised learning or unsupervised learning.

9. The method according to claim 1, characterized in that: In step (8), the personalized physiological parameter evaluation model existing in the system is automatically matched based on the identified specific user, and the personalized model generation method includes one or more of training, pre-training model, transfer learning or style transfer based on each user.

Images