A pulse disease recognition device based on multi-channel correlation map features

Abstract

The application relates to a pulse disease recognition device based on a multi-channel correlation graph feature, which fully utilizes the correlation between multi-channel pulse data, can solve the problem of poor classification effect caused by individual differences of pulses, converts a periodic signal to a frequency domain, contains rich frequency domain information, and through pre-experiment, it is found that the classification effect of the frequency domain periodic signal is more stable than that of the time domain periodic signal, and through global attention mechanism coding and decoding operation, the global correlation of the pulse feature is excavated; an adaptive threshold algorithm is used to automatically judge the similarity between each periodic node, a graph feature of spatial relationship change is generated for each period, the change of the pulse is reflected from time and space, the pulse disease recognition effect is improved, and the channel combination with high correlation is counted, so that the pulse change rule of different people is more comprehensively reflected.

Description

TECHNICAL FIELD

[0001] The application relates to a pulse disease recognition device based on a multi-channel correlation map feature. BACKGROUND

[0002] The human pulse is formed by the heart contracting to squeeze blood into the aorta, and is a quasi-periodic signal like the human heartbeat. Many diseases in clinical practice will affect the pulse. After collecting the human pulse signal, a series of processing and analysis can objectively reflect the disease and physiological changes. The mainstream system currently extracts the pulse signal of one wrist inch, guan and chi of the subject or the pulse signal of one acupoint. The pulse theory of traditional Chinese medicine mentions the left hand heart, liver and kidney, and the right hand lung and spleen, that is, the inch, guan and chi parts of the left and right hands correspond to different organs. In addition, diseases often affect multiple tissues and organs, and also affect the pulse. Therefore, reasonably combining multiple channel signals is better than the former in the breadth of information extraction. At the same time, in the feature extraction stage, the existing methods mainly include manually extracting time-frequency domain and entropy value features, and extracting deep features by using a neural network. After optimization, the extracted features are linearly combined and sent to a classifier. This feature vector method cannot extract the correlation between multiple acupoints, thereby missing part of the effective information. SUMMARY

[0003] In order to overcome at least one deficiency in the prior art, the application provides a pulse disease recognition device based on a multi-channel correlation map feature.

[0004] In a first aspect, a pulse disease recognition device based on a multi-channel correlation map feature is provided, comprising:

[0005] A pulse signal acquisition module is configured to acquire a pulse signal of a user to be identified.

[0006] A cycle division module is configured to divide the pulse signal of the user to be identified into cycles to obtain pulse signals of multiple cycles. Each cycle of the pulse signal includes pulse data of six channels.

[0007] A frequency domain amplitude feature matrix acquisition module is configured to perform Fourier transform on each cycle of the pulse signal, and input the transform result into a Transformer network model to output a frequency domain amplitude feature matrix of each cycle. The frequency domain amplitude feature matrix includes frequency domain amplitude feature vectors of six channels.

[0008] A correlation matrix construction module is configured to determine the correlation between any two frequency domain amplitude feature vectors of the six channels for each cycle of the frequency domain amplitude feature matrix, and construct a correlation matrix based on the correlation.

[0009] A six-channel correlation map feature construction module is configured to construct a channel correlation map feature based on the correlation matrix.

[0010] The classification module is configured to input the plurality of six-channel correlation graph features corresponding to the plurality of cycles to a graph convolutional neural network to obtain a pulse disease recognition result of the to-be-recognized user.

[0011] In an embodiment, the correlation matrix construction module is further configured to:

[0012] determine a penalty factor γ between any two channels;

[0013] determine a correlation between the frequency domain amplitude feature vectors of any two channels based on the penalty factor γ, using the following formula:

[0014]

[0015] wherein d ′ The (U, V) is the correlation between the frequency domain amplitude feature vectors U and V of any two channels, and S is the covariance matrix of the frequency domain amplitude feature vectors U and V of any two channels.

[0016] In an embodiment, the correlation matrix construction module is further configured to:

[0017] determine a time domain index in the time domain and a frequency domain index in the frequency domain of the pulse data of any two channels, the time domain index including kurtosis, skewness, maximum value and minimum value, and the frequency domain index including energy, peak value, root mean square and wave crest factor;

[0018] construct an index vector corresponding to any two channels based on the time domain index and the frequency domain index corresponding to the pulse data of any two channels;

[0019] determine the difference between corresponding elements of the index vector corresponding to any two channels, and sum the absolute values of the differences as the penalty factor γ between any two channels.

[0020] In an embodiment, the six-channel correlation graph feature construction module is further configured to:

[0021] binarize the correlation matrix using an adaptive threshold algorithm to obtain a binary matrix;

[0022] construct a graph network using 6 channels as nodes and the association relationship between the nodes represented by the elements in the binary matrix as edges;

[0023] construct a six-channel correlation graph feature using the frequency domain amplitude feature vectors of the 6 channels as node features of the nodes of the graph network.

[0024] In an embodiment, the six-channel correlation graph feature construction module is further configured to:

[0025] calculate the mean of all elements in each row of the correlation matrix;

[0026] If the value of the element in the row is greater than the mean value, the value of the element is set to 1, otherwise, the value of the element is set to 0, to obtain a binary matrix.

[0027] In one embodiment, the device further comprises a graph convolutional neural network training module configured to train the graph convolutional neural network based on the collected pulse signals of normal people and users with pulse diseases, to obtain the trained graph convolutional neural network.

[0028] In one embodiment, the Transformer network model comprises an encoder and a decoder with a global attention mechanism.

[0029] Compared with the prior art, the present application has the following beneficial effects:

[0030] 1. The present application makes full use of the correlation between multi-channel pulse data, and can solve the problem of poor classification effect caused by individual differences in pulse.

[0031] 2. The periodic signal is converted to the frequency domain, which contains rich frequency domain information. Through pre-experiment, it is found that the classification effect of the frequency domain periodic signal is more stable than that of the time domain periodic signal. And through the encoding and decoding operation of the global attention mechanism, the global correlation of the pulse features is mined.

[0032] 3. An adaptive threshold algorithm is used to automatically determine the similarity between each periodic node, generating a graph feature of spatial relationship change for each cycle, which reflects the change of pulse in time and space, not only improves the effect of pulse disease recognition, but also statistics the channel combination with high correlation, and more comprehensively reflects the pulse change rule of different groups of people.

[0033] 4. The improved Mahalanobis distance calculation method is used to calculate the correlation, which not only measures the distance between two node features in space, but also considers the time-frequency domain similarity of the signal, effectively measuring the correlation degree of the pulse between different channels. BRIEF DESCRIPTION OF DRAWINGS

[0034] The present application can be better understood by referring to the description given below in conjunction with the accompanying drawings, which are incorporated in and form a part of the specification, and together with the detailed description, serve to explain the principles of the present application. In the drawings:

[0035] Figure 1 A structural block diagram of a pulse disease recognition device based on multi-channel correlation graph features according to an embodiment of the present application is shown;

[0036] Figure 2 A pulse signal period division schematic diagram is shown;

[0037] Figure 3A diagram showing six-channel correlation map features based on a correlation matrix is shown. DETAILED DESCRIPTION

[0038] In the following, exemplary embodiments of the present application will be described with reference to the accompanying drawings. In the description, not all features of a practical embodiment are described in order to keep the description clear and concise. It should be appreciated that in developing any such practical embodiment numerous implementation-specific decisions must be made and these decisions will vary from one implementation to another. Such implementation-specific decisions can include, for example, specific circuit implementations, specific hardware components, specific circuit configurations, specific programming code, specific software modules, specific program command sequences, specific algorithms, specific username and password combinations, specific user interfaces, specific procedural flows, specific data formats, specific data sizes, specific end users, specific computing environments, specific technical

[0039] It should also be noted that, in the description, only the structures closely related to the scheme according to the present application are shown in the drawings, and other details not closely related to the present application are omitted.

[0040] It should be understood that the present application is not limited to the described embodiments only because of the following description with reference to the drawings. In this context, the embodiments can be combined with each other, features can be replaced or borrowed between different embodiments, and one or more features can be omitted in one embodiment.

[0041] The present application provides a pulse disease recognition device based on multi-channel correlation map features, Figure 1 A structural block diagram of a pulse disease recognition device based on multi-channel correlation map features according to an embodiment of the present application is shown, referring to Figure 1 The device comprises:

[0042] A pulse signal acquisition module 11 is configured to acquire a pulse signal of a user to be recognized. Here, a multi-channel piezoelectric pulse acquisition device can be used to acquire the pulse signal of the user to be recognized, and the acquired pulse signal includes pulse data of six channels of the Cun, Guan and Chi parts of the left and right hands.

[0043] A cycle division module 12 is configured to divide the pulse signal of the user to be recognized into cycles to obtain pulse signals of multiple cycles, and each cycle of the pulse signal includes pulse data of six channels. Here, a div cycle division function can be used to divide the pulse signal into cycles, and the div cycle division function divides the pulse signal into cycles based on the minimum amplitude point (valley) of the pulse signal. After the cycle division, each cycle includes pulse data of six channels in the same time period. Figure 2 A pulse signal cycle division diagram is shown.

[0044] The frequency domain amplitude feature matrix acquisition module 13 is configured to perform Fourier transform on each cycle of pulse signals, and input the transform result into a Transformer network model to output a frequency domain amplitude feature matrix of each cycle, the frequency domain amplitude feature matrix including six frequency domain amplitude feature vectors.

[0045] The correlation matrix construction module 14 is configured to determine the correlation between any two frequency domain amplitude feature vectors of the six frequency domain amplitude feature vectors for each cycle of frequency domain amplitude feature matrix, and construct a correlation matrix based on the correlation. The Mahalanobis distance calculation method can be used to calculate the correlation between any two frequency domain amplitude feature vectors, and finally obtain a 6*6 correlation matrix.

[0046] The six-channel correlation graph feature construction module 15 is configured to construct a channel correlation graph feature based on the correlation matrix.

[0047] The classification module 16 is configured to input a plurality of six-channel correlation graph features corresponding to a plurality of cycles into a graph convolutional neural network to obtain a pulse disease recognition result of the to-be-identified user. The pulse disease recognition result of the to-be-identified user refers to that the to-be-identified user is classified as a patient with a pulse disease or a normal person, and the pulse disease may be, for example, hypertension. The graph convolutional neural network is a trained graph convolutional neural network. After the six-channel correlation graph feature corresponding to each cycle is input into the graph convolutional neural network, the to-be-identified user is output as a patient with a pulse disease or a normal person for the cycle. When the probability of the to-be-identified user being a patient with a pulse disease in the pulse disease recognition result output according to the six-channel correlation graph features corresponding to a plurality of cycles exceeds a set proportion, the to-be-identified user is considered as a patient with a pulse disease.

[0048] In the above embodiments of the present application, the correlation between multi-channel pulse data is fully utilized, and the problem of poor classification effect caused by pulse individual difference is solved. The cycle signal is converted to the frequency domain, which contains rich frequency domain information. Through pre-experiment, it is found that the classification effect of the frequency domain cycle signal is more stable than that of the time domain cycle signal. Through the encoding and decoding operations of the global attention mechanism, the global correlation of the pulse feature is mined. The adaptive threshold algorithm is used to automatically determine the similarity between each cycle node, and a graph feature with spatial relationship change is generated for each cycle. The change of the pulse is reflected from time and space, which not only improves the pulse disease recognition effect, but also counts the channel combination with high correlation, and more comprehensively reflects the pulse change rule of different populations.

[0049] In one embodiment, considering that the Mahalanobis distance calculation method adopted in the process of constructing the correlation matrix cannot effectively measure the correlation degree of pulses between different channel pieces, the Mahalanobis distance calculation method is improved in this embodiment, specifically, the correlation matrix construction module is also used for:

[0050] determining a penalty factor γ between any two channels;

[0051] calculating the correlation between the frequency domain amplitude feature vectors of any two channels based on the penalty factor γ, using the following formula:

[0052]

[0053] wherein d ′ The (U, V) is the correlation between the frequency domain amplitude feature vectors U and V of any two channels, and S is the covariance matrix of the frequency domain amplitude feature vectors U and V of any two channels.

[0054] Specifically, in the above embodiment, the method for determining the penalty factor γ between any two channels is as follows:

[0055] determining the time domain indicators of the pulse data of any two channels in the time domain and the frequency domain indicators in the frequency domain, the time domain indicators including kurtosis, skewness, maximum value and minimum value, and the frequency domain indicators including energy, peak value, root mean square and wave crest factor;

[0056] constructing the indicator vector corresponding to any two channels based on the time domain indicators and the frequency domain indicators corresponding to the pulse data of any two channels;

[0057] calculating the difference of corresponding elements of the indicator vector corresponding to any two channels, and calculating the sum of absolute values of the difference as the penalty factor γ between any two channels.

[0058] In this embodiment, in the process of calculating the correlation, not only the distance between two node features is measured in space, but also the time-frequency domain similarity of the signal is considered, effectively measuring the correlation degree of pulses between different channel pieces.

[0059] In one embodiment, the six-channel correlation graph feature construction module is also used for:

[0060] using an adaptive threshold algorithm to binarize the correlation matrix to obtain a binary matrix; here, the elements in the binary matrix are 0 or 1.

[0061] Taking the six channels as nodes and the association relationship between the nodes represented by the elements in the binary matrix as edges, a graph network is constructed; here, if the element in the binary matrix in the graph network is 1, there is an edge connection between the two nodes represented by the element, and if the element is 0, there is no edge connection between the two nodes represented by the element.

[0062] The six-channel frequency domain amplitude feature vector is taken as the node feature of the node of the graph network, and a six-channel correlation graph feature is constructed. Figure 3 An illustration of obtaining the six-channel correlation graph feature based on the correlation matrix is shown.

[0063] Specifically, in the above embodiment, the correlation matrix is binarized by using an adaptive threshold algorithm to obtain a binary matrix, and the method used is as follows:

[0064] For each row in the correlation matrix, the mean of all elements in the row is calculated.

[0065] If the value of an element in the row is greater than the mean, the value of the element is set to 1, otherwise the value of the element is set to 0, to obtain a binary matrix.

[0066] In one embodiment, the device further comprises a graph convolutional neural network training module for training the graph convolutional neural network based on the collected pulse signals of normal people and users with pulse diseases, to obtain a trained graph convolutional neural network.

[0067] Here, in the model training module, the collected pulse signals of normal people and users with pulse diseases are subjected to the cycle division module 12, the frequency domain amplitude feature vector acquisition module 13, the correlation matrix construction module 14, and the six-channel correlation graph feature construction module 15 of the aforementioned embodiments, to obtain a six-channel correlation graph feature, which is input to the graph convolutional neural network for iterative training, and finally a trained graph convolutional neural network is obtained.

[0068] In summary, the present application has the following technical effects:

[0069] 1. The present application makes full use of the correlation between multi-channel pulse data, and can solve the problem of poor classification effect caused by individual differences in pulse;

[0070] 2. The periodic signal is converted to the frequency domain, which contains rich frequency domain information. Through pre-experiments, it is found that the classification effect of the frequency domain periodic signal is more stable than that of the time domain periodic signal, and through the encoding and decoding operations of the global attention mechanism, the global correlation of the pulse features is mined.

[0071] 3. The adaptive threshold algorithm is used to automatically determine the similarity between each cycle node, generating a graph feature that changes in space for each cycle, which reflects the changes in pulse from space and time, not only improving the effect of pulse disease recognition, but also statistically analyzing the channel combinations with high correlation, and more comprehensively reflecting the pulse change rules of different populations.

[0072] 4、The correlation is calculated by improved Mahalanobis distance calculation method, which not only measures the distance between the features of two nodes in space, but also considers the time-frequency domain similarity of signals, effectively measuring the correlation degree of pulses between different channels.

[0073] The above merely describes various embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A pulse disease identification device based on multi-channel correlation map features, characterized in that, include: The pulse signal acquisition module is used to acquire the pulse signal of the user to be identified; The period division module is used to divide the pulse signal of the user to be identified into multiple periods to obtain pulse signals of multiple periods; each period of pulse signal includes pulse data from 6 channels. The frequency domain amplitude feature matrix acquisition module is used to perform Fourier transform on the pulse signal of each cycle, input the transform result into the Transformer network model, and output the frequency domain amplitude feature matrix of each cycle. The frequency domain amplitude feature matrix includes frequency domain amplitude feature vectors of 6 channels. The correlation matrix construction module is used to determine the correlation between the frequency domain amplitude feature vectors of any two channels among the frequency domain amplitude feature vectors of the six channels for each period's frequency domain amplitude feature matrix, and to construct a correlation matrix based on the correlation. A six-channel correlation graph feature construction module is used to construct channel correlation graph features based on the correlation matrix; The classification module is used to input multiple six-channel correlation graph features corresponding to multiple cycles into a graph convolutional neural network to obtain the pulse disease identification result of the user to be identified.

2. The apparatus as claimed in claim 1, characterized in that, The correlation matrix construction module is also used for: Determine the penalty factor γ between any two channels; The correlation between the frequency domain amplitude eigenvectors of any two channels is calculated based on the penalty factor γ using the following formula: Where, d ′ (U,V) represents the correlation between the frequency domain amplitude eigenvectors U and V of any two channels, and S represents the covariance matrix of the frequency domain amplitude eigenvectors U and V of any two channels.

3. The apparatus as described in claim 2, characterized in that, The correlation matrix construction module is also used for: Determine the time-domain indices and frequency-domain indices of pulse data from any two channels in the time domain. The time-domain indices include kurtosis, skewness, maximum value, and minimum value. The frequency-domain indices include energy, peak value, root mean square (RMS), and crest factor. Construct an index vector corresponding to any two channels based on the time-domain index and the frequency-domain index corresponding to the pulse data of any two channels; Find the difference between the corresponding elements of the index vectors corresponding to any two channels, and sum the absolute values ​​of the differences as the penalty factor γ between any two channels.

4. The apparatus as claimed in claim 1, characterized in that, The six-channel correlation map feature construction module is also used for: The correlation matrix is ​​binarized using an adaptive thresholding algorithm to obtain a binary matrix; A graph network is constructed using 6 channels as nodes and the relationships between nodes represented by the elements in the binary matrix as edges. The frequency domain amplitude feature vectors of the six channels are used as the node features of the nodes in the graph network to construct a six-channel correlation graph feature.

5. The apparatus as described in claim 4, characterized in that, The six-channel correlation map feature construction module is also used for: For each row in the correlation matrix, calculate the mean of all elements in that row; If the value of an element in a row is greater than the mean, the value of the element is set to 1; otherwise, the value of the element is set to 0, resulting in a binary matrix.

6. The apparatus as claimed in claim 1, characterized in that, The device also includes a graph convolutional neural network training module, which is used to train the graph convolutional neural network based on the pulse signals of normal people and users with pulse diseases, so as to obtain the trained graph convolutional neural network.

7. The apparatus as claimed in claim 1, characterized in that, The Transformer network model includes an encoder and a decoder with a global attention mechanism.

Images