Pulse analysis method and device

Abstract

A pulse analysis method and device. The method comprises the following steps: S1, acquiring three pulse datasets corresponding to Cun, Guan and Chi positions; S2, acquiring a frequency domain feature, a time domain feature and a Gaussian mixture model feature of each piece of pulse data; and S3, inputting the frequency domain feature, the time domain feature and the Gaussian mixture model feature of each piece of pulse data into a pre-trained neural network model, and outputting a predicted pulse condition classification result. Because each pulse dataset comprises multiple pieces of pulse data acquired by multiple pulse acquisition units, such multi-channel pulse data can improve the prediction accuracy compared with single-channel pulse data. The introduction of Gaussian mixture model features can further improve the accuracy of a model.

Description

Pulse analysis method and device Technical Field

[0001] The present invention relates to the technical field of pulse analysis, and in particular to a pulse analysis method and device. Background Art

[0002] During cardiovascular circulation, blood flows from the left ventricle into the aorta under the action of the heart's contraction, causing the proximal arterial wall to dilate and internal pressure to rise. When the ventricle relaxes, ejection temporarily ceases, and the artery resumes contraction due to the elasticity of the vessel. During this contraction and relaxation process, blood rapidly propagates distally along the vessel from the heart under the influence of the pressure difference, causing similar pulsations in the distal arteries, forming a pulse wave. This signal carries a wealth of information related to cardiovascular status. In traditional Chinese medicine (TCM) pulse diagnosis, physicians press on the radial artery at the wrist to sense this cyclical expansion and dilation, and classify the pulse patterns they sense into different pulse patterns. Pulse patterns can be used to determine the nature of illnesses, such as cold or heat, deficiency or excess, as well as the prognosis. In TCM, pulse patterns are important objective indicators for disease identification.

[0003] Pulse diagnosis, as a non-invasive diagnostic method, offers advantages such as ease of diagnosis, low cost, and effective results, along with a comprehensive assessment and intervention system. However, physicians vary in their skills and standards for pulse strength. Therefore, relying solely on subjective perception and experience often fails to provide an objective and accurate analysis of a patient's pulse. Therefore, there is an urgent need for a pulse analysis method and device that can objectively and accurately analyze a patient's pulse.

[0004] Summary of the Invention

[0005] This application provides a pulse analysis method and device to solve the problem in the prior art that different doctors have different medical skills and different standards for pulse strength, making it impossible to objectively and accurately analyze the patient's pulse.

[0006] To achieve the above objectives, the embodiments of the present application adopt the following technical solutions.

[0007] In one aspect, a pulse analysis method is provided, comprising the following steps:

[0008] S1. Acquire three sets of pulse data sets corresponding to the Cun, Guan, and Chi points, each set of pulse data sets including multiple pulse data acquired by multiple pulse acquisition units;

[0009] S2. Obtain frequency domain features, time domain features, and Gaussian mixture model features of each pulse data; wherein the Gaussian mixture model features are the weight, mean, and covariance matrix of each Gaussian distribution function in the Gaussian mixture model;

[0010] S3. Input the frequency domain features, time domain features and Gaussian mixture model features of each pulse data into a pre-trained neural network model, and output the predicted pulse classification results.

[0011] The pulse analysis method described above enables objective and accurate analysis of a patient's pulse. Furthermore, because each pulse data set includes multiple pulse data points acquired by multiple pulse acquisition units, this multi-channel pulse data can improve prediction accuracy compared to single-channel pulse data. Furthermore, the introduction of Gaussian mixture model features can further enhance the model's accuracy.

[0012] In order to speed up the model calculation, in some embodiments, a step of using principal component analysis to eliminate redundant frequency domain features, time domain features, and Gaussian mixture model features is further included between steps S2 and S3.

[0013] In some embodiments, the trained neural network model is obtained according to the following steps:

[0014] Obtain a pulse sample data set, perform pulse pattern annotation on the pulse sample data set to obtain pulse pattern classification, divide the pulse sample data set into a training set and a test set, obtain frequency domain features, time domain features, and Gaussian mixture model features of each pulse sample data in the training set and the test set, and use principal component analysis to eliminate redundant frequency domain features, time domain features, and Gaussian mixture models;

[0015] The neural network model is trained using the frequency domain features, time domain features and Gaussian mixture model features of all pulse sample data in the training set, and the accuracy of the neural network model is evaluated using the frequency domain features, time domain features and Gaussian mixture model features of all pulse sample data in the test set to obtain a trained neural network model.

[0016] In some embodiments, the pulse data includes pulse width; in step S3, the frequency domain features, time domain features, Gaussian mixture model features and pulse width of each pulse data are input into a pre-trained neural network model, and the predicted pulse classification result is output.

[0017] In some embodiments, a pulse sample data set is obtained, pulse condition annotation is performed on the pulse sample data set to obtain a pulse condition classification, the pulse sample data set is divided into a training set and a test set, frequency domain features, time domain features, and Gaussian mixture model features of each pulse sample data in the training set and the test set are obtained, and principal component analysis is used to eliminate redundant frequency domain features, time domain features, and Gaussian mixture models;

[0018] The neural network model is trained using the frequency domain features, time domain features, Gaussian mixture model features and pulse width of all pulse sample data in the training set, and the accuracy of the neural network model is evaluated using the frequency domain features, time domain features, Gaussian mixture model features and pulse width of all pulse sample data in the test set to obtain a trained neural network model.

[0019] In another aspect, a pulse analysis device is provided, comprising:

[0020] A first acquisition module is used to acquire three sets of pulse data sets corresponding to the three parts of Cun, Guan, and Chi, each set of pulse data sets including multiple pulse data acquired by multiple pulse acquisition units;

[0021] A feature extraction module for obtaining frequency domain features, time domain features, and Gaussian mixture model features of each pulse data; wherein the Gaussian mixture model features are the weight, mean, and covariance matrix of each Gaussian distribution function in the Gaussian mixture model; and

[0022] The prediction module is used to input the frequency domain features, time domain features and Gaussian mixture model features of each pulse data into a pre-trained neural network model, and output the predicted pulse classification results.

[0023] In some embodiments, the pulse analysis device further includes a feature elimination module for eliminating redundant frequency domain features, time domain features, and Gaussian mixture model features using principal component analysis.

[0024] In some embodiments, the pulse analysis device further includes a model training module for obtaining a trained neural network model according to the following steps:

[0025] Obtain a pulse sample data set, perform pulse pattern annotation on the pulse sample data set to obtain pulse pattern classification, divide the pulse sample data set into a training set and a test set, obtain frequency domain features, time domain features, and Gaussian mixture model features of each pulse sample data in the training set and the test set, and use principal component analysis to eliminate redundant frequency domain features, time domain features, and Gaussian mixture models;

[0026] The neural network model is trained using the frequency domain features, time domain features and Gaussian mixture model features of all pulse sample data in the training set, and the accuracy of the neural network model is evaluated using the frequency domain features, time domain features and Gaussian mixture model features of all pulse sample data in the test set to obtain a trained neural network model.

[0027] In some embodiments, the pulse data includes pulse width; the prediction module is used to input the frequency domain features, time domain features, Gaussian mixture model features and pulse width of each pulse data into a pre-trained neural network model, and output the predicted pulse classification result.

[0028] In some embodiments, a model training module is further included, which is used to obtain a trained neural network model according to the following steps:

[0029] Obtain a pulse sample data set, perform pulse pattern annotation on the pulse sample data set to obtain pulse pattern classification, divide the pulse sample data set into a training set and a test set, obtain frequency domain features, time domain features, and Gaussian mixture model features of each pulse sample data in the training set and the test set, and use principal component analysis to eliminate redundant frequency domain features, time domain features, and Gaussian mixture models;

[0030] The neural network model is trained using the frequency domain features, time domain features, Gaussian mixture model features and pulse width of all pulse sample data in the training set, and the accuracy of the neural network model is evaluated using the frequency domain features, time domain features, Gaussian mixture model features and pulse width of all pulse sample data in the test set to obtain a trained neural network model.

[0031] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: They enable objective and accurate analysis of a patient's pulse. Furthermore, because each pulse data set includes multiple pulse data acquired by multiple pulse acquisition units, this multi-channel pulse data can improve prediction accuracy compared to single-channel pulse data. Furthermore, the introduction of Gaussian mixture model features can further improve model accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] FIG1 is a flow chart of a pulse analysis method according to an embodiment of the present application;

[0033] FIG2 is a comparison diagram of the original pulse signal and the reconstructed signal after denoising in one embodiment of the present application;

[0034] FIG3 is a schematic diagram of a pulse wave in an embodiment of the present application. DETAILED DESCRIPTION

[0035] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0036] Referring to FIG1 , a pulse analysis method includes the following steps:

[0037] S1. Acquire three sets of pulse data sets corresponding to the Cun, Guan, and Chi points, each set of pulse data sets including multiple pulse data acquired by multiple pulse acquisition units;

[0038] It should be noted that the pulse data set is a collection of preprocessed pulse data. In this embodiment, wavelet transform is used to preprocess the pulse data. The specific steps are: multi-scale wavelet decomposition of the original signal, threshold quantization based on the obtained high-frequency detail components and low-frequency approximate components, and then wavelet reconstruction to obtain the denoised signal. Figure 2 shows a comparison of the original signal and the denoised reconstructed signal.

[0039] S2. Obtain frequency domain features, time domain features, and Gaussian mixture model features for each pulse data set, and use principal component analysis to eliminate redundant frequency domain features, time domain features, and Gaussian mixture model features. The Gaussian mixture model features are the weight, mean, and covariance matrix of each Gaussian distribution function in the Gaussian mixture model.

[0040] After pre-processing, pulse signal data is subjected to time domain feature extraction to the single cycle pulse signal after resampling and averaging for each sample. As shown in Figure 3, time domain feature comprises 6 amplitude ratio features (h2 / h1 (descending tributary wave height / main wave height), h3 / h1 (tidal wave height / main wave height), h4 / h1 (descending mid-gap height / main wave height), h5 / h1 (diborrhea wave height / main wave height) and the abscissa width w at 2 / 3 of the main wave height), 6 time ratio features (t1 (main wave time) / T, t2 (descending tributary wave time) / T, t3 (tidal wave time) / T, t4 (descending mid-gap time) / T, t5 (diborrhea wave time) / T), also have mean, variance, root mean square and skewness of pulse pressure, form factor, peak factor, pulse factor, margin factor, 20 time domain features in total in time domain.

[0041] Frequency domain features need to be extracted from the filtered multi-period signal, and the power spectrum of the reconstructed pulse signal sequence needs to be estimated. The extracted frequency domain features include the maximum frequency, frequency domain variance, frequency domain range, rectified mean (the average value of the integral of the absolute value of the signal), frequency domain skewness, and frequency domain root mean square, a total of 6 frequency domain features.

[0042] A single-cycle pulse wave can be modeled as a Gaussian mixture signal. Assume that the single-cycle pulse wave data set is X = x1, x2, x3, ... x N , a single cycle pulse wave has two Gaussian distributions, and the parameter of each Gaussian distribution is θ k =(π k ,μ k ,∑ k ), where π k is the weight of the Gaussian distribution (satisfying ), μ k is the mean of the Gaussian distribution, ∑ k is the covariance matrix of the Gaussian distribution.

[0043] Then for each sample x i, its responsiveness to the kth Gaussian distribution is:

[0044] Where N(x|μ,∑) is the probability density function of the multidimensional Gaussian distribution, which represents the probability density value of x in the Gaussian distribution with mean μ and covariance matrix ∑.

[0045] According to the above responsiveness, the new Gaussian weight, Gaussian mean and Gaussian covariance matrix of each Gaussian distribution can be calculated:

[0046] 1. Gaussian weight π k The estimated value of is:

[0047] That is, for each sample x i The mean of the responsiveness to the K-th Gaussian distribution.

[0048] 2. Gaussian mean μ k The estimated value of is:

[0049] That is, for each sample x i The weighted average is multiplied by the weight of its response to the K-th Gaussian distribution.

[0050] 3. Gaussian covariance matrix ∑ k The estimated value of is:

[0051] That is, for each sample x i Subtract the corresponding mean μ k Then multiply it by the weighted covariance matrix of its response to the K-th Gaussian distribution.

[0052] Two Gaussian functions can be used to model a single-cycle pulse waveform, including Gaussian weight, Gaussian mean, and Gaussian covariance, a total of 6 Gaussian mixture model parameter features.

[0053] S3. Input the frequency domain features, time domain features and Gaussian mixture model features of each pulse data into a pre-trained neural network model, and output the predicted pulse classification results.

[0054] Among them, the trained neural network model can be obtained according to the following steps:

[0055] Obtain a pulse sample dataset. This dataset is also a preprocessed dataset. The preprocessing method is consistent with the preprocessing method of the pulse dataset. The pulse sample dataset should include data on all pulse conditions, and the number of samples corresponding to each pulse condition should be no less than one thousand people.

[0056] The pulse sample dataset is annotated to obtain pulse classification (pulse categories include 28 pulse types: floating, sinking, slow, rapid, slippery, astringent, weak, strong, long, short, strong, weak, tight, slow, stringy, wiry, leathery, firm, moist, weak, scattered, thin, hidden, moving, hurried, knotted, intermittent, and large). The pulse sample dataset is divided into a training set and a test set. The frequency domain features, time domain features, and Gaussian mixture model features of each pulse sample data in the training set and the test set are obtained, and the principal component analysis method is used to eliminate redundant frequency domain features, time domain features, and Gaussian mixture models.

[0057] The neural network model is trained using the frequency domain, time domain, and Gaussian mixture model features of all pulse sample data in the training set. The accuracy of the neural network model is evaluated using the frequency domain, time domain, and Gaussian mixture model features of all pulse sample data in the test set, resulting in a trained neural network model. When evaluating the accuracy of the neural network model, the Adam optimizer is used to update the parameters of the neural network model based on the first-order and second-order moment estimates of the gradient of the neural network model's loss function until the accuracy reaches a preset threshold (e.g., 95%).

[0058] Example 2

[0059] This embodiment is basically the same as the first embodiment, except that the pulse data includes pulse width in addition to pulse wave data, and the pulse width is also input into the neural network model (including training, testing and prediction) as one of the features.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A pulse analysis method comprising the following steps: S1, acquiring three groups of pulse data sets corresponding to the three parts of Cun, Guan and Chi, each group of pulse data sets including a plurality of pulse data acquired by a plurality of pulse acquisition units; S2, obtaining frequency domain features, time domain features and Gaussian mixture model features of each pulse data; wherein the Gaussian mixture model features are the weight, mean and covariance matrix of each Gaussian distribution function in the Gaussian mixture model; S3. Input the frequency domain features, time domain features and Gaussian mixture model features of each pulse data into a pre-trained neural network model, and output the predicted pulse classification result.

2. The pulse analysis method according to claim 1, characterized in that: The step between step S2 and step S3 also includes a step of using principal component analysis to eliminate redundant frequency domain features, time domain features and Gaussian mixture model features.

3. The pulse analysis method according to claim 2, characterized in that: Follow the steps below to get the trained neural network model: Obtain a pulse sample data set, annotate the pulse sample data set to obtain pulse classification, divide the pulse sample data set into a training set and a test set, obtain frequency domain features, time domain features and Gaussian mixture model features of each pulse sample data in the training set and the test set, and use principal component analysis to eliminate redundant frequency domain features, time domain features and Gaussian mixture models; The neural network model is trained using the frequency domain features, time domain features and Gaussian mixture model features of all pulse sample data in the training set, and the accuracy of the neural network model is evaluated using the frequency domain features, time domain features and Gaussian mixture model features of all pulse sample data in the test set to obtain a trained neural network model.

4. The pulse analysis method according to claim 2, characterized in that: The pulse data includes pulse width; in step S3, the frequency domain features, time domain features, Gaussian mixture model features and pulse width of each pulse data are input into a pre-trained neural network model, and the predicted pulse classification result is output.

5. The pulse analysis method according to claim 4, characterized in that: Follow the steps below to get the trained neural network model: Obtain a pulse sample data set, annotate the pulse sample data set to obtain pulse classification, divide the pulse sample data set into a training set and a test set, obtain frequency domain features, time domain features and Gaussian mixture model features of each pulse sample data in the training set and the test set, and use principal component analysis to eliminate redundant frequency domain features, time domain features and Gaussian mixture models; The neural network model is trained using the frequency domain features, time domain features, Gaussian mixture model features and pulse width of all pulse sample data in the training set, and the accuracy of the neural network model is evaluated using the frequency domain features, time domain features, Gaussian mixture model features and pulse width of all pulse sample data in the test set to obtain a trained neural network model.

6. A pulse analysis device, characterized in that: include: The first acquisition module is used to acquire three sets of pulse data sets corresponding to the three parts of Cun, Guan and Chi. The data set includes a plurality of pulse data acquired by a plurality of pulse acquisition units; A feature extraction module, used to obtain frequency domain features, time domain features and Gaussian mixture model features of each pulse data; wherein the Gaussian mixture model features are the weight, mean and covariance matrix of each Gaussian distribution function in the Gaussian mixture model; and The prediction module is used to input the frequency domain features, time domain features and Gaussian mixture model features of each pulse data into a pre-trained neural network model, and output the predicted pulse classification results.

7. The pulse analysis device according to claim 6, characterized in that: It also includes a feature elimination module for eliminating redundant frequency domain features, time domain features and Gaussian mixture model features by using principal component analysis.

8. The pulse analysis device according to claim 7, characterized in that: It also includes a model training module for obtaining a trained neural network model according to the following steps: Obtain a pulse sample data set, annotate the pulse sample data set to obtain pulse classification, divide the pulse sample data set into a training set and a test set, obtain frequency domain features, time domain features and Gaussian mixture model features of each pulse sample data in the training set and the test set, and use principal component analysis to eliminate redundant frequency domain features, time domain features and Gaussian mixture models; The neural network model is trained using the frequency domain features, time domain features and Gaussian mixture model features of all pulse sample data in the training set, and the accuracy of the neural network model is evaluated using the frequency domain features, time domain features and Gaussian mixture model features of all pulse sample data in the test set to obtain a trained neural network model.

9. The pulse analysis device according to claim 6, characterized in that: The pulse data includes pulse width; the prediction module is used to input the frequency domain features, time domain features, Gaussian mixture model features and pulse width of each pulse data into a pre-trained neural network model, and output the predicted pulse classification result.

10. The pulse analysis device according to claim 9, characterized in that: It also includes a model training module for obtaining a trained neural network model according to the following steps: Obtain a pulse sample data set, annotate the pulse sample data set to obtain pulse classification, divide the pulse sample data set into a training set and a test set, obtain frequency domain features, time domain features and Gaussian mixture model features of each pulse sample data in the training set and the test set, and use principal component analysis to eliminate redundant frequency domain features, time domain features and Gaussian mixture models; The neural network model is trained using the frequency domain features, time domain features, Gaussian mixture model features and pulse width of all pulse sample data in the training set, and the accuracy of the neural network model is evaluated using the frequency domain features, time domain features, Gaussian mixture model features and pulse width of all pulse sample data in the test set to obtain a trained neural network model.

Images