Auxiliary pulse diagnosis tool and method based on traditional Chinese medicine pulse perception

Abstract

The invention discloses an auxiliary pulse diagnosis tool and method based on traditional Chinese medicine pulse perception, and belongs to the technical field of traditional Chinese medicine diagnosis information processing. The auxiliary pulse diagnosis tool comprises a signal acquisition module, a signal preprocessing module, a signal processing module, an identification analysis module, a pulse condition data storage module and an intelligent diagnosis module. According to the invention, real-time high-precision acquisition of three-channel pulse signal data of cun, guan and chi pulses and identification and analysis of pulse conditions of different parts are realized, and doctors are effectively assisted in pulse diagnosis; the multi-scale discriminant subsequence distance measurement method designed by the invention is used for automatic classification and recognition of interpretable pulse signals, and solves the problems that an existing pulse signal classification and recognition method mostly adopts features of a single mode, the pulse signals cannot be comprehensively described, sub-segments with discriminant ability cannot be extracted for disease prediction tasks, and the recognition efficiency is low. The abnormal pulse signal position cannot be positioned, the result is not interpretable, and clinical application is difficult.

Description

Technical Field

[0001] The present invention belongs to the technical field of traditional Chinese medicine diagnosis information processing, and in particular relates to an auxiliary pulse diagnosis tool and method based on traditional Chinese medicine pulse perception. Background Art

[0002] In TCM clinical diagnosis, TCM doctors use their fingers to feel the pulse of the radial artery at the patient's wrist and judge the patient's health status based on the pulse. Since the pulse identification criteria in traditional TCM pulse diagnosis are not clearly quantified, the doctor's diagnosis results are subjective.

[0003] Although with the application of computer technology in the field of traditional Chinese medicine diagnosis and treatment, there has been an increase in objective research on pulse diagnosis related to automatic classification and identification of pulse signals, the classification methods proposed in the research mostly use the features of a single modality, which cannot fully describe the pulse signal, and cannot extract discriminative sub-segments for disease prediction tasks, and cannot locate the position of abnormal pulse signals. The results are not interpretable and are difficult to apply in clinical practice. Summary of the invention

[0004] The purpose of the present invention is to provide an auxiliary pulse diagnosis tool and method based on traditional Chinese medicine pulse perception to address the defects or shortcomings of the existing technology in the above-mentioned background technology, realize the combination of traditional Chinese medicine and modern electronic technology, and improve the efficiency and accuracy of doctors' diagnosis.

[0005] In order to achieve the above object, the present invention adopts the following technical solution.

[0006] An auxiliary pulse diagnosis tool based on traditional Chinese medicine pulse perception, comprising:

[0007] A signal acquisition module, used to sense the patient's pulse signal, the signal acquisition module includes a pulse sensor;

[0008] A signal preprocessing module is used to amplify and filter the original pulse signal collected by the signal acquisition module to eliminate the influence of noise on the pulse signal. The signal preprocessing module includes an operational amplifier and a filter;

[0009] A signal processing module, used for deeply analyzing and extracting features of the pulse signal preprocessed by the signal preprocessing module, including a frequency domain analysis module and a time domain analysis module, wherein the frequency domain analysis module is used for extracting frequency domain features of the pulse signal, and the time domain analysis module is used for extracting time domain features of the pulse signal;

[0010] An identification and analysis module, used for classifying and identifying the characteristic parameters of the pulse signal processed by the signal processing module and performing cluster analysis using a deep learning algorithm, wherein the identification and analysis module comprises a classification and identification module and a cluster analysis module, wherein the classification and identification module is used for classifying and identifying the pulse signal, and the cluster analysis module is used for clustering the pulse signal;

[0011] A pulse data storage module, used for storing a large number of pulse signal samples under different parts and different conditions collected from a large number of clinical trials with complete annotations, wherein the pulse data storage module includes a pulse database;

[0012] The intelligent diagnosis module is used to compare and analyze the results of the analysis of the pulse signal characteristic parameters by the recognition and analysis module with the pulse signal samples stored in the pulse data storage module, so as to assist doctors in diagnosing diseases based on the pulse signal.

[0013] Furthermore, the pulse sensor is at least one of a strain sensor, an optical sensor, and a piezoelectric force sensor.

[0014] Preferably, the pulse sensor is a piezoelectric pressure sensor.

[0015] Preferably, the pulse sensor is an annular structure and is put on the patient's wrist and / or neck to collect the pulse signal.

[0016] Furthermore, the filter is a digital filter or an analog filter.

[0017] The present invention also provides a method for pulse diagnosis using the above-mentioned auxiliary pulse diagnosis tool based on traditional Chinese medicine pulse perception, comprising:

[0018] (1) Preprocessing of pulse signals

[0019] The pulse signal is collected by using the signal collection module, and the distortion in the pulse signal is eliminated by using the signal preprocessing module;

[0020] (2) Feature extraction of pulse signal

[0021] The signal processing module is used to deeply analyze the pre-processed pulse signal to extract frequency domain features and time domain features;

[0022] (3) Identification and classification of pulse signals

[0023] The pulse signal characteristic parameters processed by the signal processing module are classified, identified and clustered using a deep learning algorithm;

[0024] (4) Disease diagnosis based on pulse signals

[0025] The result of the pulse signal characteristic parameter analysis by the recognition and analysis module is compared and analyzed with the pulse signal samples stored in the pulse data storage module to assist the physician in diagnosing diseases based on the pulse signal.

[0026] Furthermore, in step (1), the distortion in the pulse signal includes high-frequency noise, baseline drift, artifacts and saturation, wherein a Fourier transform or wavelet transform method is used to eliminate high-frequency noise using a filter; a curve fitting method is used to eliminate baseline drift, an adaptive threshold method is used to eliminate saturation, and a one-dimensional convolutional neural network model is used to detect and eliminate artifacts.

[0027] Preferably, the curve fitting to eliminate baseline drift specifically includes: detecting all peak points of the pulse signal sample by an adaptive sliding window method; detecting the local minimum point in the area between two wave peaks; fitting the baseline by a cubic spline interpolation estimation method based on the set of local minimum points; and eliminating the baseline drift by subtracting the fitted drift curve from the pulse signal.

[0028] Preferably, the adaptive threshold method for eliminating saturation specifically includes: performing wavelet transform on the saturated distorted signal, establishing a threshold calculation function under the minimum mean square error of the wavelet transform coefficients, optimizing and selecting the optimal threshold in the ant colony algorithm, performing threshold processing on the wavelet coefficients of each layer, and finally performing inverse wavelet transform to obtain the signal segment after desaturation and distortion;

[0029] Preferably, the use of a one-dimensional convolutional neural network model to detect and eliminate artifacts specifically includes: automatically learning the inherent characteristics of the pulse signal waveform through a one-dimensional convolutional neural network model, classifying the randomly segmented five-second pulse signal segments into standard signal segments and artifact-distorted signal segments, and removing the artifact-distorted signal segments.

[0030] Furthermore, in step (2), extracting time domain features specifically includes: extracting extreme points or inflection points from the pulse signal waveform as pulse waveform feature points, combining the waveform feature points with their corresponding physiological meanings, and obtaining a medically interpretable analysis result.

[0031] Furthermore, in step (2), extracting frequency domain features specifically includes: extracting spectral features from the pulse signal waveform diagram by means of a fast Fourier transform algorithm, regarding the pulse signal waveform as a signal represented by a combination of a series of sine and cosine components of different frequencies and amplitudes, and then fitting the average single-cycle pulse signal waveform of each sample by the least squares method, wherein the frequency domain feature vector of the pulse signal is represented by a discrete Fourier series.

[0032] Furthermore, in step (3), the classification and identification is to automatically classify and identify the interpretable pulse signal using a multi-scale discriminant subsequence distance measurement method; the multi-scale discriminant subsequence distance measurement method includes:

[0033] 1) Obtaining significant pulse signal subsegments based on the objective function of discriminating subsequence positioning;

[0034] 2) Calculate the distance measurement results of the signal sub-segment pairs under the time domain shape features and behavior correlation feature modes, use the distance measurement results of multiple discriminant signal sub-sequences as the features of the signal sub-segment pairs, and establish a multi-scale distance feature space;

[0035] 3) Pulse signal classification is realized through support vector machine.

[0036] Preferably, the objective function based on the positioning of the discriminant subsequence to obtain the significant pulse signal subsegment is to use the generalized eigenvector technology to maximize the data variance ratio of the projections between pulse signals of different categories, introduce a sparse regularization term in the objective function to limit the sparsity of the solution, introduce a continuous regularization term to guide the strong block structure solution, introduce a directional difference regularization term to ensure that the solutions under different categories are different, solve the objective function, and search and locate the significant pulse signal subsequence; specifically including:

[0037] The generalized eigenvector technique is used to find a discriminant direction u that can maximize the data variance ratio of the projection between the pulse signal in the i-th category and the pulse signal in the j-th category. Bold lowercase letters are used to represent vectors, bold uppercase letters are used to represent matrices, and non-bold letters are used to represent real numbers;

[0038]

[0039] Among them, T represents the transposition operation, C i is the sample covariance matrix of class i. To ensure the uniqueness of the solution, let u T C i u=1, then equation (1) can be rewritten as:

[0040] u=arg min u u T C j u,u T C i u=1 (2);

[0041] Among them, u T Represents the transpose of vector u. By minimizing the above objective function, the information in class i can be retained to the maximum extent, while the information in class j can be compressed, so that class i and class j can be distinguished by u. After preprocessing, c pulse signal training samples {pi,yi}i=1c are obtained, where pi=[p1,…,p t ,…,p q ] is the i-th pulse signal, p t is the value of the pulse signal at time t, q is the length of the pulse signal, y iis its corresponding category, y i ∈{1,2,…,n}, n represents the number of categories. The above problem is extended to a multi-category problem, and the discriminant direction obtained for class i is denoted as u i , then the objective function for all classes can be expressed as:

[0042]

[0043] Among them, C i is the sample covariance matrix of class i, C i is the covariance matrix of the remaining samples after removing the class i samples, and the discriminant directions u of all categories i Connect in series to get the discrimination direction v = (u1, .., u i ,...,u n )=[v1,…,v i ,…,v n ×q], q is the length of the pulse signal, and equation (3) is expressed in matrix form:

[0044]

[0045] In order to promote the sparsity of the solution, ensure the generation of a clustered strong block structure solution, and effectively capture the correlation of adjacent positions of the pulse signal sequence, the l1 norm sparse regularization term ‖v‖1 and the continuous restriction term ∑k|vk-vk+1|, k∈[1,n×q-1] are added to formula (4). The continuous restriction term ‖Dv‖1 is expressed in matrix form, where D is a constant matrix whose elements Dk,k=1, Dk,k+1=-1, and the remaining elements are all 0. At the same time, in order to ensure the difference between ui, the directional difference regularization term ∑||ui-uj||2i,j is introduced, expressed as v T Gv, where G is a constant, defined as shown in formula (5);

[0046]

[0047] Where I represents the identity matrix, and the problem of searching for and locating the discriminant subsequence is formalized as a numerical optimization problem:

[0048]

[0049] stv T Cv=1 (6)

[0050] Among them, parameter α1 controls the weight of the continuous restriction term, α2 controls the weight of the sparse regularization term, and parameter α3 controls the weight of the directional difference regularization term. Formula (6) can be rewritten as Formula (7):

[0051]

[0052] stv T Dv=z,v T Cv=1 (7)

[0053] Formula (7) represents the objective function of locating the pulse signal discriminator sequence.

[0054] Preferably, the establishment of the multi-scale distance feature space includes: intercepting S subsequences from each pulse signal sample, using these S discriminant signal sub-segments to represent a pulse signal sample, and then comparing the time domain shape distance and behavior correlation between the subsequences at the corresponding positions of the two pulse signals to obtain the modal features at the subsequence level, and then splicing the modal distances of multiple subsequences together to form a multi-scale distance feature vector. Define x i =(x i1 ,...,x il ) and x j =(x j1 ,...,x jl ) represent two pulse signal subsequences of length l, where x it Represents a subsequence x i The value at time t, x jt Represents a subsequence x j At the time t, l represents the length of the pulse signal sub-segment, 1≤t≤l. The distance vector dE(x i ,x j ) and the distance metric dB(x) based on behavioral correlation i ,x j ), calculate the distance metric dE(x) based on the time domain shape features i ,x j ), we need to calculate x i 、x j The Euclidean distance between two sub-segments is calculated as:

[0055]

[0056] The result obtained by the Euclidean distance formula has a large range of values ​​and generally needs to be normalized. Therefore, dE(x i ,x j )The final result is:

[0057]

[0058] Calculate the distance metric dB(x) based on the behavioral correlation i ,x j ), first we need to calculate x i 、x jThe first-order temporal correlation coefficient between two sub-segments is cort(x i ,x j ), and its calculation formula is:

[0059]

[0060] Where t′=t+1. cort(x i ,x j ) has a value range of [-1,1]. i ,x j )=1, indicating that the subsequence x i With x j have similar behaviors and hold similar trends. They will rise or fall at the same time, and the increase or decrease is also similar. When cort(x i ,x j )=-1, indicating that the subsequence x i With x j The two pulse signal subsequences show opposite behaviors. The rising and falling trends are exactly opposite. When cort(x i ,x j )=0, indicating subsequence x i With x j There is no correlation in terms of monotonicity. According to the first-order time correlation coefficient cort(x i ,x j ) calculates the distance based on behavioral correlation, and the calculation formula is:

[0061] d B (x i ,x j )=1-cort(x i ,x j ) (11)

[0062] dB(x i ,x j ) is smaller, indicating that the signals xi and xj are closer in waveform trend.

[0063] Through the position indication vector, S subsequences can be intercepted from each pulse signal sample. i The sth subsequence of i Indicates that the pulse signal p j The sth subsequence of j Indicates that, according to the distance vector dE(x i ,x j ) and the distance metric dB(x) based on behavioral correlation i ,x j ) to obtain the pulse signal x measured on the s-th subsequence iand x j The distance dms(x i ,x j ), where m∈{E,B}, s∈{1,…,S}, the pulse signal p can be constructed using DMS i With signal p j The multi-scale distance feature vector p ij , whose expression is p ij =(dE1(x i ,x j ),…,dms(x i ,x j ),...,dBS(x i ,x j )). After obtaining the multi-scale distance feature vector, the pulse signal is classified using the support vector machine. In order to deal with the nonlinear separable situation, the kernel function Φ(x i ) i Projected into high-dimensional space, the support vector machine objective function becomes:

[0064] y i [(W T Φ(x i ))+b]-1+ξ i ≥0,i=1,2...,n (12)

[0065] Since the optimization of the support vector machine only needs to calculate the dot product K(x i ,x j )=Φ(x i )Φ(x j ), therefore, if we can calculate K(x i ,x j ), then there is no need to calculate Φ(x i ), the kernel function used is a kernel function based on multi-scale distance feature vectors, defined as follows:

[0066] K(x i ,x j )=exp(-||p ij || 2 ) (13)

[0067] Furthermore, the pulse signal classification is realized by the support vector machine after the support vector machine training is completed, the disease label of the pulse signal is predicted, and the pulse signal is classified and identified by combining the global information and periodic information of the pulse signal with a multi-scale periodic information fusion network, and the corresponding disease label is predicted, which specifically includes: constructing an MPIFN model, the MPIFN model is mainly composed of two encoders, an information fusion gate and an attention module, wherein the two encoders are a periodic signal encoder and a global signal convolution encoder, and the periodic signal convolution encoder takes each period of the pulse signal as input and provides a global signal convolution encoder for each period. Take the intra-cycle features, and the global convolution encoder takes the complete pulse signal as input to extract the global features of the pulse signal. In order to integrate the global features into the intra-cycle features, the information fusion gate is used to control the fusion ratio of global information and intra-cycle features, and the global features are fused with the intra-cycle features. In order to integrate the feature representations of all cycles to obtain the contextual representation of the pulse signal, the attention module is used to calculate the contribution scores of different cycles to the classification, and the cycle feature representation is weighted and summed based on the contribution scores to obtain the pulse context representation vector. The patient's demographic information is concatenated with the context vector and sent to the classification layer to identify the disease category of the pulse signal;

[0068] In the process of intra-cycle feature extraction, only the pulse signal of the current cycle is used as input, without considering the relationship between the current cycle and its previous and subsequent sequences, and ignoring the global features of the pulse signal, a fusion attention mechanism is used to fuse the global information into the cycle features:

[0069]

[0070] Among them, h i is the feature representation obtained after fusion of the i-th cycle, ⊙ is the bitwise multiplication operation, r i is the fusion ratio, the fusion ratio r i Calculated by the information fusion gate:

[0071]

[0072] Among them, W rg , W rc and b rc is the parameter of the information fusion gate, σ(·) is the sigmoid function, which limits the fusion rate to the range of [0,1] and controls the fusion ratio of periodic features and global features. i =1, then h i =h i , which means that when calculating the feature representation of period i, global information is not integrated. If r i = 0, then h i =h g, which means that when calculating the feature representation of cycle i, the information of the cycle is discarded and only the global information is passed back. This fusion method can take into account its position in the entire pulse signal and the relationship with the previous and next cycles when calculating the feature representation of the current cycle, and adaptively integrate the global information into the cycle information.

[0073] The fused feature representations of multiple cycles are integrated into the context representation through the attention energy function. First, the attention energy function of each cycle is calculated:

[0074]

[0075] Among them, W e and b e is the parameter of the attention energy function, e i is the attention energy of the ith cycle. After obtaining the attention energy of all cycles, the softmax function is used to calculate the normalized attention score of each cycle:

[0076] α=(α1,...,α i ,...,α m )

[0077] =Softmax([e1,...,e i ,...,e m ]) (17)

[0078] The purpose of normalizing the attention energy is to make the sum of the attention scores of all cycles equal to 1, and then the feature representations of all cycles are fused by weighted summation to obtain the pulse signal context representation h c :

[0079]

[0080] Through the attention mechanism, the model can effectively integrate information from different cycles. If the attention score of the i-th cycle is larger, the proportion of the cycle in the context vector will be larger, and the amount of information provided for the final classification will be more. c Then, the patient's demographic information d Incorporate h into the model c and h d After concatenation, the final feature representation of the patient's pulse signal is obtained through the fully connected layer:

[0081]

[0082] Among them, W h and b h are the parameters of the fully connected layer, is a serial concatenation operation, f is an activation function, and h is the final feature representation of the pulse signal.

[0083] A pulse signal data may correspond to one or more disease labels, which belongs to the multi-label classification problem. For n training samples {(S (i) ,y (i) )}i=1n data set, S (i) is the i-th pulse signal sample, y (i) is a vector containing only 0 and 1 values, which is used to indicate the disease category to which the pulse signal belongs. For example, yk (i) =1 means that the pulse signal has the disease label of the kth category. If yk (i) =0 means that the pulse signal does not have the disease label of the kth category, k∈{1,…,C}, C represents the number of categories in the data set. In the multi-label classification task, sigmoid is generally used as the activation function of the classification layer, and the binary cross entropy is used as the loss function. There are C neurons in the sigmoid classification layer, corresponding to C disease categories. Each neuron is regarded as a binomial distribution, which is equivalent to converting the multi-label classification problem into a binary classification problem on each label. The probability of each neuron output is the predicted probability of the corresponding disease, and the disease category is identified through the sigmoid classification layer:

[0084]

[0085] Among them, W s and b s is the parameter of the sigmoid layer, h (i) is the deep feature representation of the i-th pulse signal obtained by the multi-scale periodic information fusion network, is the disease label probability predicted based on the i-th pulse signal, is the predicted probability under the kth disease category. Since the multi-label classification problem has been transformed into a binary classification problem on each label, the binary cross entropy is used as the loss function. The dataset has a sample imbalance problem. In order to alleviate the negative impact of sample imbalance, according to the number of samples in each category, the loss weights are set for samples in different categories in the loss function. The more samples a category has, the smaller the loss weight of the samples in that category. The fewer samples a category has, the larger the loss weight of the samples in that category. The calculation formula of the loss function is shown in formula (22), where w k is the loss weight of the sample under the kth category, c k Represents the number of samples under the kth category;

[0086]

[0087]

[0088] Furthermore, in step (3), the cluster analysis uses a K-Means clustering algorithm to cluster the pulse signals of different patients with the same disease and identify the regularity therein.

[0089] Compared with the prior art, the beneficial technical effects of the present invention are:

[0090] 1. The present invention realizes the real-time and high-precision acquisition of three-channel pulse signal data of Cun, Guan and Chi pulses and the identification and analysis of pulse conditions in different parts, effectively assisting doctors in pulse diagnosis; at the same time, the large-scale pulse condition database formed also provides a valuable reference basis for traditional Chinese medicine pulse diagnosis.

[0091] 2. The multi-scale discriminant subsequence distance measurement method designed by the present invention is used for automatic classification and identification of interpretable pulse signals, which solves the problems that the existing pulse signal classification and identification methods mostly use the features of a single mode, cannot fully describe the pulse signal, cannot extract discriminative sub-segments for disease prediction tasks, cannot locate the position of abnormal pulse signals, and the results are not interpretable and difficult to apply in clinical practice. The multi-scale periodic information fusion network combines the global information and periodic information of the pulse signal to classify and identify the pulse signal, and predict the corresponding disease label, which greatly improves the efficiency of doctors in seeing patients.

[0092] 3. The present invention realizes the combination of traditional Chinese medicine and modern electronic technology, which can promote the transformation of Chinese pulse diagnosis towards digitization and standardization, and promote the development of Chinese medicine. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 This is a principle block diagram of an auxiliary pulse diagnosis tool based on traditional Chinese medicine pulse perception in Example 1 of the present invention;

[0094] Figure 2 is a flow chart of preprocessing of a pulse signal in Embodiment 2 of the present invention;

[0095] Figure 3 is a waveform diagram of a single-cycle pulse signal in Example 2 of the present invention;

[0096] Figure 4 is a pulse signal classification structure diagram based on a multi-scale discriminant subsequence distance measurement method in Example 2 of the present invention;

[0097] Figure 5 is a pulse signal subsegment extracted by using a multi-scale discriminant subsequence distance measurement method in Embodiment 2 of the present invention;

[0098] Figure 6 is a structural diagram of the MPIFN model in Example 2 of the present invention;

[0099] In the figure: 1-signal acquisition module, 2-signal preprocessing module, 3-signal processing module, 4-identification and analysis module, 5-pulse data storage module, 6-intelligent diagnosis module. DETAILED DESCRIPTION

[0100] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0101] Example 1

[0102] Please refer to the attached Figure 1 The present invention provides an auxiliary pulse diagnosis tool based on traditional Chinese medicine pulse perception, comprising:

[0103] A signal acquisition module 1, used for sensing a patient's pulse signal, wherein the signal acquisition module 1 comprises a pulse sensor;

[0104] The signal preprocessing module 2 is used to amplify and filter the original pulse signal collected by the signal acquisition module 1 to eliminate the influence of noise on the pulse signal. The signal preprocessing module 2 includes an operational amplifier and a filter;

[0105] The signal processing module 3 is used to perform in-depth analysis and extract features on the pulse signal preprocessed by the signal preprocessing module 2, and includes a frequency domain analysis module and a time domain analysis module. The frequency domain analysis module is used to extract the frequency domain features of the pulse signal, and the time domain analysis module is used to extract the time domain features of the pulse signal.

[0106] The recognition and analysis module 4 is used to classify and recognize the characteristic parameters of the pulse signal processed by the signal processing module 3 using a deep learning algorithm and perform cluster analysis. The recognition and analysis module 4 includes a classification recognition module and a cluster analysis module. The classification recognition module is used to classify and recognize the pulse signal, and the cluster analysis module is used to perform cluster analysis on the pulse signal.

[0107] A pulse data storage module 5, for storing a large number of pulse signal samples under different parts and different conditions collected from a large number of clinical trials with complete annotations, wherein the pulse data storage module 5 includes a pulse database;

[0108] The intelligent diagnosis module 6 is used to compare and analyze the results of the analysis of the pulse signal characteristic parameters by the recognition and analysis module 4 with the pulse signal samples stored in the pulse data storage module 5, so as to assist the doctor in diagnosing diseases based on the pulse signal.

[0109] Specifically, the pulse sensor is at least one of a strain force sensor, an optical sensor, and a piezoelectric force sensor. Preferably, the pulse sensor adopts a piezoelectric pressure sensor. The pulse sensor is an annular structure and is worn on the patient's wrist and / or neck to collect pulse signals; the filter adopts a digital filter or an analog filter.

[0110] Example 2

[0111] Please refer to the attached Figure 2 The present invention also provides a method for pulse diagnosis using the auxiliary pulse diagnosis tool based on traditional Chinese medicine pulse perception of embodiment 1, comprising:

[0112] (1) Preprocessing of pulse signals

[0113] The pulse signal is collected by the signal collection module 1, and the distortion in the pulse signal is eliminated by the signal preprocessing module 2;

[0114] (2) Feature extraction of pulse signal

[0115] The signal processing module 3 is used to deeply analyze the pre-processed pulse signal to extract frequency domain features and time domain features;

[0116] (3) Identification and classification of pulse signals

[0117] The pulse signal characteristic parameters processed by the signal processing module 3 are classified, identified and clustered using a deep learning algorithm;

[0118] (4) Disease diagnosis based on pulse signals

[0119] The result of the analysis of the pulse signal characteristic parameters by the recognition and analysis module 4 is compared and analyzed with the pulse signal samples stored in the pulse data storage module 5 to assist the physician in diagnosing diseases based on the pulse signal.

[0120] In this embodiment, the pulse sensor adopts MAX30102. The pulse signal collected by the pulse sensor often has distortion, and the distortion includes high-frequency noise, baseline drift, artifacts and saturation. Among them, high-frequency noise is generally caused by power frequency interference of the power grid. A high-pass filter with cutoff frequencies of 50Hz and 60Hz can be used to filter the high-frequency noise caused by power frequency interference, or a smoothing filter and a zero-phase filter can be used to obtain an appropriate frequency band and eliminate noise. The filter adopts Fourier transform or wavelet transform method to eliminate high-frequency noise; curve fitting is used to eliminate baseline drift, adaptive threshold method is used to eliminate saturation, and a one-dimensional convolutional neural network model is used to detect and eliminate artifacts.

[0121] Specifically, the curve fitting to eliminate baseline drift specifically includes: detecting all peak points of the pulse signal sample through an adaptive sliding window method; detecting the local minimum point in the area between two wave peaks; fitting the baseline through a cubic spline interpolation estimation method based on the set of local minimum points; and eliminating the baseline drift by subtracting the fitted drift curve from the pulse signal.

[0122] Specifically, the adaptive threshold method for eliminating saturation specifically includes: performing wavelet transform on the saturated distorted signal, establishing a threshold calculation function under the minimum mean square error of the wavelet transform coefficients, optimizing and selecting the optimal threshold in the ant colony algorithm, performing threshold processing on the wavelet coefficients of each layer, and finally performing inverse wavelet transform to obtain the signal segment after desaturation and distortion;

[0123] Specifically, the use of a one-dimensional convolutional neural network model to detect and eliminate artifacts specifically includes: automatically learning the inherent characteristics of the pulse signal waveform through a one-dimensional convolutional neural network model, classifying the randomly segmented five-second pulse signal segments into standard signal segments and artifact-distorted signal segments, and removing the artifact-distorted signal segments.

[0124] Specifically, extracting time domain features specifically includes: extracting extreme points or inflection points from the pulse signal waveform as pulse waveform feature points, combining the waveform feature points with their corresponding physiological meanings, and obtaining a medically interpretable analysis result.

[0125] Specifically, extracting frequency domain features includes: extracting spectral features from the pulse signal waveform through the fast Fourier transform algorithm, regarding the pulse signal waveform as a signal represented by a combination of a series of sine and cosine components of different frequencies and amplitudes, and then fitting the average single-cycle pulse signal waveform of each sample using the least squares method. The frequency domain feature vector of the pulse signal is represented by a discrete Fourier series.

[0126] In the time domain analysis method, the commonly used waveform parameters in the time domain include the peak height of the main wave, the peak height of the dicrotic wave, and the peak height of the tidal wave. The time domain waveform characteristics cannot fully reflect the pulse wave information, and the pulse wave has different information in different frequency bands, so the frequency domain analysis method is introduced to analyze the pulse signal. The frequency domain analysis method mainly analyzes the pulse signal through the corresponding spectrum diagram of the pulse signal, and usually uses Fourier transform to realize the conversion of the signal between the time domain and the frequency domain.

[0127] The waveform of the pulse has slight local differences due to different human health conditions, so time domain analysis of the pulse waveform is a common method for extracting pulse signal features. The time domain analysis method can analyze the relationship between the pulse amplitude and the pulse phase. This method records the extreme points and inflection points at the rise and fall of the pulse waveform and uses them as waveform feature parameters that reflect the signal specificity to distinguish different types of pulse signals.

[0128] The waveform of a single-cycle pulse signal is as follows: Figure 3 As shown in the figure, the characteristic parameters in the waveform diagram and their physiological significance are explained as follows: h1 represents the main wave amplitude, which refers to the relative distance between the pulse baseline and the main wave peak, reflecting the strength of the left ventricular ejection function; h2 represents the main wave gorge amplitude, which refers to the relative distance between the pulse baseline and the bottom of the main wave gorge; h3 represents the pre-dicrotic wave amplitude, which is the relative distance between the pulse baseline and the pre-dicrotic wave peak, reflecting the elastic state of the aortic wall; h4 represents the descending mid-valley amplitude, which is the distance between the pulse baseline and the bottom of the descending mid-valley, and is related to the peripheral resistance of the arterial blood vessels (its value increases with the increase of peripheral resistance and decreases with the decrease of peripheral resistance); h5 It represents the amplitude of the dicrotic wave, which is the relative distance between the dicrotic wave peak and the parallel baseline of the bottom of the descending middle canyon, reflecting the elastic state of the large artery wall (when the elasticity decreases, h5 will also decrease accordingly); t1 corresponds to the rapid ejection period of the left ventricle, which is the distance from the peak of the main wave to the starting point; t2 is the time interval from the starting point to the bottom of the main wave canyon; t3 is the time interval from the starting point to the peak of the pre-dicrotic wave; t4 corresponds to the left ventricular contraction period, which is the relative distance from the starting point to the bottom of the descending middle canyon; t5 corresponds to the left ventricular diastole, which is the distance from the bottom of the descending middle canyon to the end point of the pulse graph; t corresponds to a cardiac cycle of the left ventricle, which is the relative distance from the starting point to the end point.

[0129] Specifically, the cluster analysis uses the K-Means clustering algorithm to cluster the pulse signals of different patients with the same disease and identify the regularity therein.

[0130] Please refer to the attached Figure 4 The classification and recognition is to automatically classify and recognize the interpretable pulse signal using a multi-scale discriminant subsequence distance measurement method; the multi-scale discriminant subsequence distance measurement method includes:

[0131] 1) Obtaining significant pulse signal subsegments based on the objective function of discriminating subsequence positioning;

[0132] 2) Calculate the distance measurement results of the signal sub-segment pairs under the time domain shape features and behavior correlation feature modes, use the distance measurement results of multiple discriminant signal sub-sequences as the features of the signal sub-segment pairs, and establish a multi-scale distance feature space;

[0133] 3) Pulse signal classification is realized through support vector machine.

[0134] Please refer to the attached Figure 5 , Figure 5 The original pulse signal, the extracted discriminant direction vector and the discriminant subsequence are shown, where each row represents the extraction result for a category, from left to right are the original pulse signal, the extracted discriminant direction vector and the discriminant subsequence.

[0135] Specifically, the objective function based on the positioning of the discriminant subsequence is used to obtain the significant pulse signal subsegment by using the generalized eigenvector technology to maximize the data variance ratio of the projections between pulse signals of different categories, introduce a sparse regularization term in the objective function to limit the sparsity of the solution, introduce a continuous regularization term to guide the strong block structure solution, introduce a directional difference regularization term to ensure that the solutions under different categories are different, solve the objective function, and search and locate the significant pulse signal subsequence; specifically including:

[0136] The generalized eigenvector technique is used to find a discriminant direction u that can maximize the data variance ratio of the projection between the pulse signal in the i-th category and the pulse signal in the j-th category. Bold lowercase letters are used to represent vectors, bold uppercase letters are used to represent matrices, and non-bold letters are used to represent real numbers;

[0137]

[0138] Among them, T represents the transposition operation, C i is the sample covariance matrix of class i. To ensure the uniqueness of the solution, let u T C i u=1, then equation (1) can be rewritten as:

[0139] u=arg min u u T C j u,u T C i u=1 (2);

[0140] Among them, u T Represents the transpose of vector u. By minimizing the above objective function, the information in class i can be retained to the maximum extent, while the information in class j can be compressed, so that class i and class j can be distinguished by u. After preprocessing, c pulse signal training samples {pi,yi}i=1c are obtained, where pi=[p1,…,p t ,…,p q ] is the i-th pulse signal, p t is the value of the pulse signal at time t, q is the length of the pulse signal, y i is its corresponding category, y i ∈{1,2,…,n}, n represents the number of categories. The above problem is extended to a multi-category problem, and the discriminant direction obtained for class i is denoted as u i , then the objective function for all classes can be expressed as:

[0141]

[0142] Among them, C i is the sample covariance matrix of class i, C iis the covariance matrix of the remaining samples after removing the class i samples, and the discriminant directions u of all categories i Connect in series to get the discrimination direction v = (u1, .., u i ,...,u n )=[v1,…,v i ,…,v n ×q], q is the length of the pulse signal, and equation (3) is expressed in matrix form:

[0143]

[0144] In order to promote the sparsity of the solution, ensure the generation of a clustered strong block structure solution, and effectively capture the correlation of adjacent positions of the pulse signal sequence, the l1 norm sparse regularization term ‖v‖1 and the continuous restriction term ∑k|vk-vk+1|, k∈[1,n×q-1] are added to formula (4). The continuous restriction term ‖Dv‖1 is expressed in matrix form, where D is a constant matrix whose elements Dk,k=1, Dk,k+1=-1, and the remaining elements are all 0. At the same time, in order to ensure the difference between ui, the directional difference regularization term ∑||ui-uj||2i,j is introduced, expressed as v T Gv, where G is a constant, defined as shown in formula (5);

[0145]

[0146] Where I represents the identity matrix, and the problem of searching for and locating the discriminant subsequence is formalized as a numerical optimization problem:

[0147]

[0148] stv T Cv=1 (6)

[0149] Among them, parameter α1 controls the weight of the continuous restriction term, α2 controls the weight of the sparse regularization term, and parameter α3 controls the weight of the directional difference regularization term. Formula (6) can be rewritten as Formula (7):

[0150]

[0151] stv T Dv=z,v T Cv=1 (7)

[0152] Formula (7) represents the objective function of locating the pulse signal discriminator sequence.

[0153] Preferably, the establishment of the multi-scale distance feature space includes: intercepting S subsequences from each pulse signal sample, using these S discriminant signal sub-segments to represent a pulse signal sample, and then comparing the time domain shape distance and behavior correlation between the subsequences at the corresponding positions of the two pulse signals to obtain the modal features at the subsequence level, and then splicing the modal distances of multiple subsequences together to form a multi-scale distance feature vector. Define x i =(x i1 ,...,x il ) and x j =(x j1 ,...,x jl ) represent two pulse signal subsequences of length l, where x it Represents a subsequence x i The value at time t, x jt Represents a subsequence x j At the time t, l represents the length of the pulse signal sub-segment, 1≤t≤l. The distance vector dE(x i ,x j ) and the distance metric dB(x) based on behavioral correlation i ,x j ), calculate the distance metric dE(x) based on the time domain shape features i ,x j ), we need to calculate x i 、x j The Euclidean distance between two sub-segments is calculated as:

[0154]

[0155] The result obtained by the Euclidean distance formula has a large range of values ​​and generally needs to be normalized. Therefore, dE(x i ,x j )The final result is:

[0156]

[0157] Calculate the distance metric dB(x) based on the behavioral correlation i ,x j ), first we need to calculate x i 、x j The first-order temporal correlation coefficient between two sub-segments is cort(x i ,x j ), and its calculation formula is:

[0158]

[0159] Where t′=t+1. cort(x i,x j ) has a value range of [-1,1]. i ,x j )=1, indicating that the subsequence x i With x j have similar behaviors and hold similar trends. They will rise or fall at the same time, and the increase or decrease is also similar. When cort(x i ,x j )=-1, indicating that the subsequence x i With x j The two pulse signal subsequences show opposite behaviors. The rising and falling trends are exactly opposite. When cort(x i ,x j )=0, indicating subsequence x i With x j There is no correlation in terms of monotonicity. According to the first-order time correlation coefficient cort(x i ,x j ) calculates the distance based on behavioral correlation, and the calculation formula is:

[0160] d B (x i ,x j )=1-cort(x i ,x j ) (11)

[0161] dB(x i ,x j ) is smaller, indicating that the signals xi and xj are closer in waveform trend.

[0162] Through the position indication vector, S subsequences can be intercepted from each pulse signal sample. i The sth subsequence of i Indicates that the pulse signal p j The sth subsequence of j Indicates that, according to the distance vector dE(x i ,x j ) and the distance metric dB(x) based on behavioral correlation i ,x j ) to obtain the pulse signal x measured on the s-th subsequence i and x j The distance dms(x i ,x j ), where m∈{E,B}, s∈{1,…,S}, the pulse signal p can be constructed using DMS i With signal p j The multi-scale distance feature vector p ij , whose expression is pij =(dE1(x i ,x j ),…,dms(x i ,x j ),...,dBS(x i ,x j )). After obtaining the multi-scale distance feature vector, the pulse signal is classified using the support vector machine. In order to deal with the nonlinear separable situation, the kernel function Φ(x i ) i Projected into high-dimensional space, the support vector machine objective function becomes:

[0163] y i [(W T Φ(x i ))+b]-1+ξ i ≥0,i=1,2...,n (12)

[0164] Since the optimization of the support vector machine only needs to calculate the dot product K(x i ,x j )=Φ(x i )Φ(x j ), therefore, if we can calculate K(x i ,x j ), then there is no need to calculate Φ(x i ), the kernel function used is a kernel function based on multi-scale distance feature vectors, defined as follows:

[0165] K(x i ,x j )=exp(-||p ij || 2 ) (13)

[0166] Further, the pulse signal classification is realized by the support vector machine after the support vector machine training is completed, the pulse signal is predicted to have a disease label, a multi-scale periodic information fusion network is used to combine the global information and periodic information of the pulse signal, the pulse signal is classified and identified, and the corresponding disease label is predicted, which specifically includes: constructing an MPIFN model, the MPIFN model is as follows: Figure 6As shown, it is mainly composed of two encoders, an information fusion gate and an attention module. The two encoders are a periodic signal encoder and a global signal convolution encoder. The periodic signal convolution encoder takes each period of the pulse signal as input and extracts intra-period features for each period respectively, while the global convolution encoder takes the complete pulse signal as input and extracts the global features of the pulse signal. In order to integrate the global features into the intra-period features, the information fusion gate is used to control the fusion ratio of the global information and the intra-period features, and the global features are fused with the intra-period features. In order to integrate the feature representations of all periods to obtain the contextual representation of the pulse signal, the attention module is used to calculate the contribution scores of different periods to the classification, and the periodic feature representations are weighted summed based on the contribution scores to obtain the pulse context representation vector. The patient's demographic information is concatenated with the context vector and sent to the classification layer to identify the disease category of the pulse signal.

[0167] In the process of intra-cycle feature extraction, only the pulse signal of the current cycle is used as input, without considering the relationship between the current cycle and its previous and subsequent sequences, and ignoring the global features of the pulse signal, a fusion attention mechanism is used to fuse the global information into the cycle features:

[0168]

[0169] Among them, h i is the feature representation obtained after fusion of the i-th cycle, ⊙ is the bitwise multiplication operation, r i is the fusion ratio, the fusion ratio r i Calculated by the information fusion gate:

[0170]

[0171] Among them, W rg , W rc and b rc is the parameter of the information fusion gate, σ(·) is the sigmoid function, which limits the fusion rate to the range of [0,1] and controls the fusion ratio of periodic features and global features. i =1, then h i =h i , which means that when calculating the feature representation of period i, global information is not integrated. If r i = 0, then h i =h g , which means that when calculating the feature representation of cycle i, the information of the cycle is discarded and only the global information is passed back. This fusion method can take into account its position in the entire pulse signal and the relationship with the previous and next cycles when calculating the feature representation of the current cycle, and adaptively integrate the global information into the cycle information.

[0172] The fused feature representations of multiple cycles are integrated into the context representation through the attention energy function. First, the attention energy function of each cycle is calculated:

[0173]

[0174] Among them, W e and b e is the parameter of the attention energy function, e i is the attention energy of the ith cycle. After obtaining the attention energy of all cycles, the softmax function is used to calculate the normalized attention score of each cycle:

[0175] α=(α1,...,α i ,...,α m )

[0176] =Soft max([e1,...,e i ,...,e m ]) (17)

[0177] The purpose of normalizing the attention energy is to make the sum of the attention scores of all cycles equal to 1, and then the feature representations of all cycles are fused by weighted summation to obtain the pulse signal context representation h c :

[0178]

[0179] Through the attention mechanism, the model can effectively integrate information from different cycles. If the attention score of the i-th cycle is larger, the proportion of the cycle in the context vector will be larger, and the amount of information provided for the final classification will be more. c Then, the patient's demographic information d Incorporate h into the model c and h d After concatenation, the final feature representation of the patient's pulse signal is obtained through the fully connected layer:

[0180]

[0181] Among them, W h and b h are the parameters of the fully connected layer, is a serial concatenation operation, f is an activation function, and h is the final feature representation of the pulse signal.

[0182] A pulse signal data may correspond to one or more disease labels, which belongs to the multi-label classification problem. For n training samples {(S (i) ,y(i) )}i=1n data set, S (i) is the i-th pulse signal sample, y (i) is a vector containing only 0 and 1 values, which is used to indicate the disease category to which the pulse signal belongs. For example, yk (i) =1 means that the pulse signal has the disease label of the kth category. If yk (i) =0 means that the pulse signal does not have the disease label of the kth category, k∈{1,…,C}, C represents the number of categories in the data set. In the multi-label classification task, sigmoid is generally used as the activation function of the classification layer, and the binary cross entropy is used as the loss function. There are C neurons in the sigmoid classification layer, corresponding to C disease categories. Each neuron is regarded as a binomial distribution, which is equivalent to converting the multi-label classification problem into a binary classification problem on each label. The probability of each neuron output is the predicted probability of the corresponding disease, and the disease category is identified through the sigmoid classification layer:

[0183]

[0184] Among them, W s and b s is the parameter of the sigmoid layer, h (i) is the deep feature representation of the i-th pulse signal obtained by the multi-scale periodic information fusion network, is the disease label probability predicted based on the i-th pulse signal, is the predicted probability under the kth disease category. Since the multi-label classification problem has been transformed into a binary classification problem on each label, the binary cross entropy is used as the loss function. The dataset has a sample imbalance problem. In order to alleviate the negative impact of sample imbalance, according to the number of samples in each category, the loss weights are set for samples in different categories in the loss function. The more samples a category has, the smaller the loss weight of the samples in that category. The fewer samples a category has, the larger the loss weight of the samples in that category. The calculation formula of the loss function is shown in formula (22), where w k is the loss weight of the sample under the kth category, c k Represents the number of samples under the kth category;

[0185]

[0186]

[0187] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

Claims

1. An auxiliary pulse diagnosis tool based on traditional Chinese medicine pulse perception, It is characterized in that include: A signal acquisition module, used for sensing the patient's pulse signal, wherein the signal acquisition module comprises a pulse sensor; A signal preprocessing module is used to amplify and filter the original pulse signal collected by the signal acquisition module to eliminate the influence of noise on the pulse signal. The signal preprocessing module includes an operational amplifier and a filter; A signal processing module, used for deeply analyzing and extracting features of the pulse signal preprocessed by the signal preprocessing module, including a frequency domain analysis module and a time domain analysis module, wherein the frequency domain analysis module is used for extracting frequency domain features of the pulse signal, and the time domain analysis module is used for extracting time domain features of the pulse signal; An identification and analysis module, used for classifying and identifying the characteristic parameters of the pulse signal processed by the signal processing module and performing cluster analysis using a deep learning algorithm, wherein the identification and analysis module comprises a classification and identification module and a cluster analysis module, wherein the classification and identification module is used for classifying and identifying the pulse signal, and the cluster analysis module is used for clustering the pulse signal; A pulse data storage module, used for storing a large number of pulse signal samples under different parts and different conditions collected from a large number of clinical trials with complete annotations, wherein the pulse data storage module includes a pulse database; The intelligent diagnosis module is used to compare and analyze the results of the analysis of the pulse signal characteristic parameters by the recognition and analysis module with the pulse signal samples stored in the pulse data storage module, so as to assist doctors in diagnosing diseases based on the pulse signal.

2. The auxiliary pulse diagnosis tool based on TCM pulse perception according to claim 1, Features: The pulse sensor is at least one of a strain sensor, an optical sensor, and a piezoelectric force sensor.

3. The auxiliary pulse diagnosis tool based on TCM pulse perception according to claim 2, Features: The pulse sensor adopts a piezoelectric pressure sensor, which is an annular structure and is put on the patient's wrist and / or neck to collect the pulse signal; the filter adopts a digital filter or an analog filter.

4. A method for pulse diagnosis using the auxiliary pulse diagnosis tool based on traditional Chinese medicine pulse perception as described in any one of claims 1 to 3, include: (1) Preprocessing of pulse signals The pulse signal is collected by using the signal collection module, and the distortion in the pulse signal is eliminated by using the signal preprocessing module; The distortion in the pulse signal includes high-frequency noise, baseline drift, artifacts and saturation, wherein the high-frequency noise is eliminated by using a filter using a Fourier transform or a wavelet transform method; the baseline drift is eliminated by using a curve fitting method, the saturation is eliminated by using an adaptive threshold method, and the artifacts are detected and eliminated by using a one-dimensional convolutional neural network model; The curve fitting method for eliminating baseline drift specifically includes: detecting all peak points of the pulse signal sample by an adaptive sliding window method; detecting a local minimum point in the area between two wave peaks; fitting a baseline by a cubic spline interpolation estimation method according to a set of local minimum points; eliminating baseline drift by subtracting the fitted drift curve from the pulse signal; The adaptive threshold method for eliminating saturation specifically includes: performing wavelet transform on the saturated distorted signal, establishing a threshold calculation function under the minimum mean square error of the wavelet transform coefficients, optimizing and selecting the optimal threshold in the ant colony algorithm, performing threshold processing on the wavelet coefficients of each layer, and finally performing inverse wavelet transform to obtain the signal segment after desaturation and distortion; The method of using a one-dimensional convolutional neural network model to detect and eliminate artifacts specifically includes: automatically learning the inherent characteristics of the pulse signal waveform through the one-dimensional convolutional neural network model, classifying the randomly divided five-second pulse signal segments into standard signal segments and artifact-distorted signal segments, and removing the artifact-distorted signal segments; (2) Feature extraction of pulse signal The signal processing module is used to deeply analyze the pre-processed pulse signal to extract frequency domain features and time domain features; Extracting time domain features specifically includes: extracting extreme points or inflection points from the pulse signal waveform as pulse waveform feature points, combining the waveform feature points with their corresponding physiological meanings, and obtaining a medically interpretable analysis result; Extracting frequency domain features specifically includes: extracting spectrum features from the pulse signal waveform through the fast Fourier transform algorithm, treating the pulse signal waveform as a signal represented by a combination of a series of sine and cosine components of different frequencies and amplitudes, and then fitting the average single-cycle pulse signal waveform of each sample using the least squares method. The frequency domain feature vector of the pulse signal is represented by a discrete Fourier series; (3) Identification and classification of pulse signals The pulse signal characteristic parameters processed by the signal processing module are classified, identified and clustered using a deep learning algorithm; (4) Disease diagnosis based on pulse signals The result of the pulse signal characteristic parameter analysis by the recognition and analysis module is compared and analyzed with the pulse signal samples stored in the pulse data storage module to assist the physician in diagnosing diseases based on the pulse signal.

5. The method according to claim 4, Features: The classification and recognition is to automatically classify and recognize the interpretable pulse signal using a multi-scale discriminant subsequence distance measurement method; The multi-scale discriminant subsequence distance measurement method comprises: 1) Obtaining significant pulse signal subsegments based on the objective function of discriminating subsequence positioning; 2) Calculate the distance measurement results of the signal sub-segment pairs under the time domain shape features and behavior correlation feature modes, use the distance measurement results of multiple discriminant signal sub-sequences as the features of the signal sub-segment pairs, and establish a multi-scale distance feature space; 3) Pulse signal classification is realized through support vector machine.

6. The method according to claim 5, Features: The objective function based on the positioning of the discriminant subsequence is used to obtain the significant pulse signal subsegment by using the generalized eigenvector technology to maximize the data variance ratio of the projections between pulse signals of different categories, introduce a sparse regularization term in the objective function to limit the sparsity of the solution, introduce a continuous regularization term to guide the strong block structure solution, introduce a directional difference regularization term to ensure that the solutions under different categories are different, solve the objective function, and search and locate the significant pulse signal subsequence; specifically including: The generalized eigenvector technique is used to find a discriminant direction u that can maximize the data variance ratio of the projection between the pulse signal in the i-th category and the pulse signal in the j-th category. Bold lowercase letters are used to represent vectors, bold uppercase letters are used to represent matrices, and non-bold letters are used to represent real numbers; Among them, T represents the transposition operation, C i is the sample covariance matrix of class i. To ensure the uniqueness of the solution, let u T C i u=1, then equation (1) can be rewritten as: u=argmin u he T C j u,u T C i u=1 (2)? Among them, u T Represents the transpose of vector u. By minimizing the above objective function, the information in class i can be retained to the maximum extent, while the information in class j can be compressed, so that class i and class j can be distinguished by u. After preprocessing, c pulse signal training samples {pi, yi}i = 1c are obtained, where pi = [p 1 ,…,p t ,…,p q ] is the i-th pulse signal, p t is the value of the pulse signal at time t, q is the length of the pulse signal, y i is its corresponding category, y i ∈{1,2,…,n}, n represents the number of categories. The above problem is extended to a multi-category problem, and the discriminant direction obtained for class i is denoted as u i , then the objective function for all classes can be expressed as: Among them, C i is the sample covariance matrix of class i, C i is the covariance matrix of the remaining samples after removing the class i samples, and the discriminant directions u of all categories i Connect in series to get the discrimination direction v = (u 1 ,..,u i ,...,u n )=[v 1 ,…,v i ,…,v n ×q], q is the length of the pulse signal, and equation (3) is expressed in matrix form: In order to promote the sparsity of the solution, ensure the generation of a clustered strong block structure solution, and effectively capture the correlation between adjacent positions of the pulse signal sequence, the l1 norm sparse regularization term ‖v‖1 and the continuous restriction term ∑k|vk-vk+1|, k∈[1,n×q-1] are added to formula (4). The continuous restriction term ‖Dv‖1 is expressed in matrix form, where D is a constant matrix whose elements Dk,k=1, Dk,k+1=-1, and the remaining elements are all 0. At the same time, in order to ensure the difference between ui, the directional difference regularization term ∑||ui-uj||2i,j is introduced, expressed as v T Gv, where G is a constant, defined as shown in formula (5); Where I represents the identity matrix, and the problem of searching for and locating the discriminant subsequence is formalized as a numerical optimization problem: s.t.v T Cv=1 (6) Among them, the parameter α 1 Controls the weight of the continuous restriction term, α 2 Control the weight of the sparse regularization term, parameter α 3 The weight of the control direction difference regularization term. Formula (6) can be rewritten as formula (7): s.t.v T Dv=z,v T Cv=1 (7) Formula (7) represents the objective function of locating the pulse signal discriminator sequence.

7. The method according to claim 6, Features: The establishment of the multi-scale distance feature space includes: extracting S subsequences from each pulse signal sample, using the S discriminant signal subsegments to represent a pulse signal sample, and then comparing the time domain shape distance and behavior correlation between the subsequences at the corresponding positions of the two pulse signals to obtain the modal features at the subsequence level, and then splicing the modal distances of multiple subsequences together to form a multi-scale distance feature vector, defining x i =(x i1 ,...,x il ) and x j =(x j1 ,...,x jl ) represent two pulse signal subsequences of length l, where x it Represents a subsequence x i The value at time t, x jt Represents a subsequence x j At the time t, l represents the length of the pulse signal sub-segment, 1≤t≤l. The distance vector dE(x i ,x j ) and the distance metric dB(x) based on behavioral correlation i ,x j ), calculate the distance metric dE(x) based on the time domain shape features i ,x j ), we need to calculate x i 、x j The Euclidean distance between two sub-segments is calculated as: The result obtained by the Euclidean distance formula has a large range of values ​​and generally needs to be normalized. Therefore, dE(x i ,x j )The final result is: Calculate the distance metric dB(x) based on the behavioral correlation i ,x j ), first we need to calculate x i 、x j The first-order temporal correlation coefficient between two sub-segments is cort(x i ,x j ), and its calculation formula is: Where t′=t+1. cort(x i ,x j ) has a value range of [-1,1]. i ,x j )=1, indicating that the subsequence x i With x j have similar behaviors and hold similar trends. They will rise or fall at the same time, and the increase or decrease is also similar. When cort(x i ,x j )=-1, indicating that the subsequence x i With x j The two pulse signal subsequences show opposite behaviors. The rising and falling trends are exactly opposite. When cort(x i ,x j )=0, indicating the subsequence x i With x j There is no correlation in terms of monotonicity. According to the first-order time correlation coefficient cort(x i ,x j ) calculates the distance based on behavioral correlation, and the calculation formula is: d B (x i ,x j )=1-cort(x i ,x j ) (11) dB(x i ,x j ) is smaller, indicating that the signal x i With x j The closer the waveform trend is. Through the position indication vector, S subsequences can be intercepted from each pulse signal sample. i The sth subsequence of i Indicates that the pulse signal p j The sth subsequence of j Indicates that, according to the distance vector dE(x i ,x j ) and the distance metric dB(x) based on behavioral correlation i ,x j ) to obtain the pulse signal x measured on the s-th subsequence i and x j The distance dms(x i ,x j ), where m∈{E,B}, s∈{1,…,S}, the pulse signal p can be constructed using DMS i With signal p j The multi-scale distance feature vector p ij , whose expression is p ij =(dE1(x i ,x j ),…,dms(x i ,x j ),...,dBS(x i ,x j )). After obtaining the multi-scale distance feature vector, the pulse signal is classified using the support vector machine. In order to handle the nonlinear separable situation, the kernel function Φ(x i ) i Projected into high-dimensional space, the support vector machine objective function becomes: y i [(W T Φ(x i ))+b]-1+ξ i ≥0,i=1,2...,n (12) Since the optimization of the support vector machine only needs to calculate the dot product K(x i ,x j )=Φ(x i )Φ(x j ), therefore, if we can calculate K(x i ,x j ), then there is no need to calculate Φ(x i ), the kernel function used is a kernel function based on multi-scale distance feature vectors, defined as follows: K(x i ,x j )=exp(-||p ij || 2 ) (13) 8. The method according to claim 7, Features: The pulse signal classification is realized by support vector machine after the support vector machine training is completed, the pulse signal is predicted for the disease label, the pulse signal is classified and identified by using multi-scale periodic information fusion network combined with the global information and periodic information of the pulse signal, and the corresponding disease label is predicted, which specifically includes: constructing MPIFN model, the MPIFN model is mainly composed of two encoders, information fusion gate and attention module, wherein the two encoders are periodic signal encoder and global signal convolution encoder, the periodic signal convolution encoder takes each period of the pulse signal as input, extracts the periodic information of each period respectively, The global convolutional encoder takes the complete pulse signal as input to extract the global features of the pulse signal. In order to integrate the global features into the intra-cycle features, the information fusion gate is used to control the fusion ratio of global information and intra-cycle features, and the global features are fused with the intra-cycle features. In order to integrate the feature representations of all cycles to obtain the contextual representation of the pulse signal, the attention module is used to calculate the contribution scores of different cycles to the classification. The cycle feature representation is weighted and summed based on the contribution scores to obtain the pulse context representation vector. The patient's demographic information is concatenated with the context vector and sent to the classification layer to identify the disease category of the pulse signal. In the process of intra-cycle feature extraction, only the pulse signal of the current cycle is used as input, without considering the relationship between the current cycle and its previous and subsequent sequences, and ignoring the global features of the pulse signal, a fusion attention mechanism is used to fuse the global information into the cycle features: Among them, h i is the feature representation obtained after fusion of the i-th cycle, ⊙ is the bitwise multiplication operation, r i is the fusion ratio, the fusion ratio r i Calculated by the information fusion gate: Among them, W rg , W rc and b rc is the parameter of the information fusion gate, σ(·) is the sigmoid function, which limits the fusion rate to the range of [0,1] and controls the fusion ratio of periodic features and global features. i =1, then h i =h i , which means that when calculating the feature representation of period i, global information is not integrated. If r i = 0, then h i =h g , which means that when calculating the feature representation of cycle i, the information of the cycle is discarded and only the global information is passed back. This fusion method can take into account its position in the entire pulse signal and the relationship with the previous and next cycles when calculating the feature representation of the current cycle, and adaptively integrate the global information into the cycle information. The fused feature representations of multiple cycles are integrated into the context representation through the attention energy function. First, the attention energy function of each cycle is calculated: Among them, W e and b e is the parameter of the attention energy function, e i is the attention energy of the ith cycle. After obtaining the attention energy of all cycles, the softmax function is used to calculate the normalized attention score of each cycle: α＝(α 1 ,...,α i ,...,α m ) =Softmax([e 1 ,...,e i ,...,e m ]) (17) The purpose of normalizing the attention energy is to make the sum of the attention scores of all cycles equal to 1. Then, the feature representations of all cycles are fused by weighted summation to obtain the pulse signal context representation h c : Through the attention mechanism, the model can effectively integrate information from different cycles. If the attention score of the i-th cycle is larger, the proportion of the cycle in the context vector will be larger, and the amount of information provided for the final classification will be more. c Then, the patient's demographic information d Incorporate h into the model c and h d After concatenation, the final feature representation of the patient's pulse signal is obtained through the fully connected layer: Among them, W h and b h is the parameter of the fully connected layer, ⊕ is the serial concatenation operation, f is the activation function, and h is the final feature representation of the pulse signal. A pulse signal data may correspond to one or more disease labels, which belongs to the multi-label classification problem. For n training samples {(S (i) ,y (i) )}i=1n data set, S (i) is the i-th pulse signal sample, y (i) is a vector containing only 0 and 1 values, which is used to indicate the disease category to which the pulse signal belongs. For example, yk (i) =1 means that the pulse signal has the disease label of the kth category. If yk (i) =0 means that the pulse signal does not have the disease label of the kth category, k∈{1,…,C}, C represents the number of categories in the data set. In the multi-label classification task, sigmoid is generally used as the activation function of the classification layer, and the binary cross entropy is used as the loss function. There are C neurons in the sigmoid classification layer, corresponding to C disease categories. Each neuron is regarded as a binomial distribution, which is equivalent to converting the multi-label classification problem into a binary classification problem on each label. The probability of each neuron output is the predicted probability of the corresponding disease, and the disease category is identified through the sigmoid classification layer: Among them, W s and b s is the parameter of the sigmoid layer, h (i) is the deep feature representation of the i-th pulse signal obtained by the multi-scale periodic information fusion network, is the disease label probability predicted based on the i-th pulse signal, is the predicted probability under the kth disease category. Since the multi-label classification problem has been transformed into a binary classification problem on each label, the binary cross entropy is used as the loss function. The dataset has a sample imbalance problem. In order to alleviate the negative impact of sample imbalance, according to the number of samples in each category, the loss weights are set for samples in different categories in the loss function. The more samples a category has, the smaller the loss weight of the samples in that category. The fewer samples a category has, the larger the loss weight of the samples in that category. The calculation formula of the loss function is shown in formula (22), where w k is the loss weight of the sample under the kth category, c k Represents the number of samples under the kth category; 9. The method according to claim 4, Features: In step (3), the cluster analysis uses the K-Means clustering algorithm to cluster the pulse signals of different patients with the same disease and identify the regularity therein.

Images