Coronary heart disease risk prediction method and device

Abstract

The embodiment of the invention provides a coronary heart disease risk prediction method and device, and belongs to the technical field of data identification and prediction. The method comprises the steps that cunkou pulse data, finger tip pulse data and electrocardiogram data of a sample object are obtained, and corresponding views are extracted; extracting cunkou pulse feature data, finger tip pulse feature data and electrocardio feature data from the corresponding views to form a heart pulse multi-view feature data set, and unifying dimensions; calculating a risk relationship weight between feature data of the feature data set input into the model by adopting a self-attention mechanism, and carrying out weight coding on the risk relationship weight; calculating the credibility of the risk evidence value of the feature data set by adopting an evidence theory to obtain a fused feature data set, and training the model; and performing category probability prediction on the heart pulse multi-view feature data set of the target object by the trained model, and outputting a coronary heart disease risk prediction result. The coronary heart disease risk prediction method improves the accuracy and prediction efficiency of the coronary heart disease risk prediction model.

Description

Technical Field

[0001] The present invention belongs to the technical field of data recognition and prediction, and in particular relates to a method and device for predicting the risk of coronary heart disease. Background Art

[0002] At present, the incidence and mortality of cardiovascular diseases in my country are increasing year by year. According to statistics, there are about 330 million cardiovascular disease patients in my country, of which about 11 million are coronary heart disease patients. The incidence and mortality of coronary heart disease continue to rise, and it has now been listed as a major public health problem in my country. How to effectively prevent, detect and treat coronary heart disease is an important issue that needs to be solved urgently.

[0003] Studies have shown that in addition to uncontrollable factors such as gender and age, smoking, obesity, high-sugar diet, and drinking are all important risk factors for coronary heart disease. In addition, some physical indicators of patients with coronary heart disease, such as pulse data and electrocardiogram data, are significantly different from those of healthy people. Therefore, how to predict the probability of coronary heart disease based on physical characteristic indicator data such as pulse data and electrocardiogram data has become a hot research issue.

[0004] With the development of big data and artificial intelligence technology, multimodal biomedical technology has provided more advanced methods for disease risk prediction and provided strong support for medical technology. In theory, more accurate prediction or diagnosis results can be obtained through deep learning of human physiological indicators by artificial intelligence. However, existing artificial intelligence technology, especially deep learning algorithms, cannot meet the processing of complex human physiological indicator information and the prediction of human diseases. In addition, different types of diseases also require targeted risk prediction models for coronary heart disease.

[0005] In view of this, a new method and device for predicting the risk of coronary heart disease is needed to improve the adaptability of artificial intelligence deep learning in processing human physiological indicator information, thereby improving the accuracy of coronary heart disease risk prediction and improving prediction efficiency. Summary of the invention

[0006] In order to solve at least one aspect of the above problems and defects in the prior art, the embodiments of the present invention provide a method and device for predicting the risk of coronary heart disease, which uses a self-attention mechanism to encode the risk relationship weight of multiple feature data related to the heart pulse and uses evidence theory to fuse multiple feature data related to the heart pulse to construct a coronary heart disease risk prediction model, and obtains a high-accuracy coronary heart disease risk prediction result by training the coronary heart disease risk prediction model. The technical solution is as follows:

[0007] According to one aspect of the present invention, a method for predicting the risk of coronary heart disease is provided.

[0008] The coronary heart disease risk prediction method includes:

[0009] Obtain the cunkou pulse data, fingertip pulse data, and electrocardiogram data of the sample object, and extract the cunkou pulse wave view, fingertip pulse wave view, and electrocardiogram view therefrom;

[0010] Extract the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data from the cunkou pulse wave view, fingertip pulse wave view, and electrocardiogram view respectively to form a multi-view heart pulse feature data set, and obtain the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data with a unified dimension through dimension transformation;

[0011] Adopt the self-attention mechanism to calculate the risk relationship weights between each feature data in the multi-view heart pulse feature data set input into the coronary heart disease risk prediction model, and encode the risk relationship weights for each feature data;

[0012] Adopt the evidence theory to calculate the trust degree of the risk evidence value of the multi-view heart pulse feature data set, obtain the fused multi-view heart pulse feature data set, and train the coronary heart disease risk prediction model;

[0013] The trained coronary heart disease risk prediction model performs class probability prediction on the multi-view heart pulse feature data set of the target object, and outputs the coronary heart disease risk prediction result of the target object.

[0014] In some embodiments, the coronary heart disease risk prediction method further includes preprocessing the multi-view heart pulse feature data set, specifically including: performing data cleaning on all the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data in the multi-view heart pulse feature data set, and the data cleaning includes deleting the data with missing heart pulse characterization features or abnormal heart pulse measurement values; randomly dividing the multi-view heart pulse feature data set after data cleaning into a training set and a test set of the coronary heart disease risk prediction model according to a preset ratio.

[0015] In some embodiments, specifically, the steps of extracting the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data from the cunkou pulse wave view, fingertip pulse wave view, and electrocardiogram view respectively to form a multi-view heart pulse feature data set, and obtaining the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data with a unified dimension through dimension transformation specifically include: performing normalization processing on each feature data in the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data, and scaling all the feature data to the same preset range; performing standardization processing on each feature data, and adjusting all the feature data to a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0016] In some embodiments, specifically, the self-attention mechanism is used to calculate the risk relationship weights between each piece of feature data in the multi-view feature dataset of heart pulses input into the coronary heart disease risk prediction model. The steps of encoding the risk relationship weights for each piece of feature data specifically include:

[0017] Map all the feature data in the multi-view feature dataset of heart pulses to the same high-dimensional space through linear transformation;

[0018] Calculate the query vectors, key vectors, and value matrices of all the feature data;

[0019] Calculate the similarity between the query vector of each piece of feature data and the key vectors of all other pieces of feature data;

[0020] Assign a risk relationship weight to each piece of feature data according to the similarity to obtain the risk relationship weight vector of each piece of feature data;

[0021] Multiply the value matrix of each piece of feature data by its risk relationship weight vector to obtain the weighted summation result of all the feature data with respect to the risk relationship weights;

[0022] Concatenate the weighted summation results of all the feature data with respect to the risk relationship weights together, and obtain the output vector of the self-attention encoding through linear transformation.

[0023] In some embodiments, specifically, the Dempster-Shafer theory is used to calculate the degree of belief of the risk evidence values of the multi-view feature dataset of heart pulses, and the steps of obtaining the fused multi-view feature dataset of heart pulses and training the coronary heart disease risk prediction model specifically include:

[0024] Take the wrist pulse feature data, finger-end pulse feature data, and electrocardiogram feature data in the multi-view feature dataset of heart pulses as risk evidences, and calculate the Dirichlet distribution parameters of the wrist pulse wave view, finger-end pulse wave view, and electrocardiogram wave view respectively;

[0025] Calculate the degree of belief and uncertainty of the corresponding view according to the Dirichlet distribution parameters of each view;

[0026] Fuse the degrees of belief and uncertainties of each view according to the combination rule to obtain the comprehensive Dirichlet distribution parameters of the multi-view feature dataset of heart pulses;

[0027] Use the cross-entropy loss function to measure the matching degree between the probability distribution of the coronary heart disease risk prediction model and the true label;

[0028] Regularize the comprehensive Dirichlet distribution parameters using the KL divergence.

[0029] In some embodiments, specifically, the steps of the trained coronary heart disease risk prediction model for predicting the class probabilities of the multi-view feature dataset of the target object's heart pulse and outputting the coronary heart disease risk prediction result of the target object specifically include:

[0030] Obtain the cunkou pulse data, fingertip pulse data, and electrocardiogram data of the target object, and extract the cunkou pulse wave view, fingertip pulse wave view, and electrocardiogram view therefrom;

[0031] Extract the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data from the cunkou pulse wave view, fingertip pulse wave view, and electrocardiogram view respectively to form a multi-view feature dataset of the heart pulse, and obtain the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data with a unified dimension through dimension conversion;

[0032] Input the multi-view feature dataset of the target object's heart pulse into the coronary heart disease risk prediction model, encode the risk relationship weights for each feature data in the multi-view feature dataset of the heart pulse in the coronary heart disease risk prediction model, and calculate the corresponding risk evidence values;

[0033] Adopt the DS evidence theory to fuse the risk evidence values of the feature data of each view to obtain the comprehensive Dirichlet distribution parameter;

[0034] Calculate the binary classification probability distribution of the target object through the comprehensive Dirichlet distribution parameter and output the coronary heart disease risk prediction result.

[0035] According to another aspect of the present invention, there is also provided a coronary heart disease risk prediction device for predicting whether a target object has coronary heart disease using the coronary heart disease risk prediction method described in the above aspects and obtaining a prediction result.

[0036] The coronary heart disease risk prediction device includes:

[0037] An acquisition module configured to acquire the cunkou pulse data, fingertip pulse data, and electrocardiogram data of the sample object, and extract the cunkou pulse wave view, fingertip pulse wave view, and electrocardiogram view therefrom;

[0038] A feature extraction module configured to extract the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data from the cunkou pulse wave view, fingertip pulse wave view, and electrocardiogram view respectively to form a multi-view feature dataset of the heart pulse, and obtain the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data with a unified dimension through dimension conversion;

[0039] A self-attention encoding module configured to calculate the risk relationship weights between each feature data in the multi-view feature dataset of the heart pulse input into the coronary heart disease risk prediction model using the self-attention mechanism, and encode the risk relationship weights for each feature data;

[0040] A feature fusion module, configured to calculate the trust degree of the risk evidence value of the multi-view feature dataset of the heart pulse by using the evidence theory, obtain the fused multi-view feature dataset of the heart pulse, and train a coronary heart disease risk prediction model;

[0041] A prediction module, configured to use the trained coronary heart disease risk prediction model to perform class probability prediction on the multi-view feature dataset of the heart pulse of the target object, and output the coronary heart disease risk prediction result of the target object.

[0042] In some embodiments, specifically, the self-attention encoding module is further configured to:

[0043] Map all the feature data in the multi-view feature dataset of the heart pulse to the same high-dimensional space through a linear transformation;

[0044] Calculate the query vector, key vector, and value matrix of all the feature data;

[0045] Calculate the similarity between the query vector of each feature data and the key vectors of all other feature data;

[0046] Assign a risk relationship weight to each feature data according to the similarity to obtain the risk relationship weight vector of each feature data;

[0047] Multiply the value matrix of each feature data by its risk relationship weight vector to obtain the weighted summation result of all the feature data with respect to the risk relationship weight;

[0048] Concatenate the weighted summation results of all the feature data with respect to the risk relationship weight together, and obtain the output vector of the self-attention encoding through a linear transformation.

[0049] In some embodiments, specifically, the feature fusion module is further configured to:

[0050] Take the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data in the multi-view feature dataset of the heart pulse as risk evidences, and calculate the Dirichlet distribution parameters of the cunkou pulse wave view, fingertip pulse wave view, and electrocardiogram view respectively;

[0051] Calculate the trust degree and uncertainty of the corresponding view according to the Dirichlet distribution parameter of each view;

[0052] Fuse the trust degrees and uncertainties of each view according to the synthesis rule to obtain the comprehensive Dirichlet distribution parameter of the multi-view feature dataset of the heart pulse;

[0053] Use the cross-entropy loss function to measure the matching degree between the probability distribution of the coronary heart disease risk prediction model and the true label;

[0054] Regularize the parameters of the comprehensive Dirichlet distribution using the KL divergence.

[0055] In some embodiments, specifically, the prediction module is further configured to:

[0056] Obtain the wrist pulse data, fingertip pulse data, and electrocardiogram data of the target object, and extract the wrist pulse wave view, fingertip pulse wave view, and electrocardiogram view therefrom;

[0057] Extract wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data from the wrist pulse wave view, fingertip pulse wave view, and electrocardiogram view respectively to form a multi-view heart pulse feature dataset, and obtain wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data with a unified dimension through dimension conversion;

[0058] Input the multi-view heart pulse feature dataset of the target object into the coronary heart disease risk prediction model, and perform risk relationship weight encoding on each feature data in the multi-view heart pulse feature dataset in the coronary heart disease risk prediction model and calculate the corresponding risk evidence value;

[0059] Fuse the risk evidence values of the feature data of each view using the DS evidence theory to obtain the parameters of the comprehensive Dirichlet distribution;

[0060] Calculate the binary classification probability distribution of the target object through the parameters of the comprehensive Dirichlet distribution and output the coronary heart disease risk prediction result.

[0061] The coronary heart disease risk prediction method and device provided by the embodiments of the present invention have at least one or a part of at least one of the following advantages:

[0062] (1) The coronary heart disease risk prediction method and device provided by the embodiments of the present invention perform adaptive comprehensive feature fusion on multi-source multi-view data of wrist pulse, fingertip pulse, and electrocardiogram to construct a coronary heart disease risk prediction model, realizing high-efficiency and high-precision prediction of the coronary heart disease risk probability of the target object;

[0063] (2) The coronary heart disease risk prediction method and device provided by the embodiments of the present invention use the self-attention mechanism to assign and encode the risk relationship weights of wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data, and use the DS evidence theory to perform comprehensive feature fusion on all the above feature data and then train the coronary heart disease risk prediction model, improving the efficiency of deep learning of the coronary heart disease risk prediction model, enhancing the stability of the coronary heart disease risk prediction model, and improving the accuracy of coronary heart disease risk prediction;

[0064] (3) The coronary heart disease risk prediction method and device provided by the embodiments of the present invention perform risk relationship weight encoding on each feature data in the multi-view feature dataset of the heart pulse for the coronary heart disease risk prediction model through the self-attention mechanism, which can fully mine the deep association information about the coronary heart disease risk among each feature data within the global scope;

[0065] (4) The coronary heart disease risk prediction method and device provided by the embodiments of the present invention fuse the evidence values of the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data according to the Dirichlet distribution parameters of each view in the multi-view feature dataset of the heart pulse through the DS evidence theory, so as to realize the comprehensive feature fusion of multi-views and their multi-category feature data, which helps to improve the training efficiency of the coronary heart disease risk prediction model and the accuracy of the coronary heart disease risk prediction;

[0066] (5) The coronary heart disease risk prediction method and device provided by the embodiments of the present invention add the cross-entropy loss function and KL divergence regularization during the training process of the coronary heart disease risk prediction model to optimize the coronary heart disease risk prediction model, ensure the rationality of the distribution of the evidence values of the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data, and avoid the overconfidence of the coronary heart disease risk prediction model;

[0067] (6) The coronary heart disease risk prediction method and device provided by the embodiments of the present invention use the multi-view feature dataset of the heart pulse obtained by feature splicing from the cunkou pulse data, fingertip pulse data, and electrocardiogram data to provide rich, complete, and comprehensive basic data of human physiological indicators related to coronary heart disease for training the coronary heart disease risk prediction model, which helps to improve the prediction accuracy of the coronary heart disease risk prediction model;

[0068] (7) The coronary heart disease risk prediction method and device provided by the embodiments of the present invention obtain deep association information closely related to the coronary heart disease risk through comprehensive feature fusion of multi-source data and multi-source views, effectively enhancing the comprehensive learning ability of the coronary heart disease risk prediction model for data from different sources, and providing a more comprehensive judgment basis for its coronary heart disease risk prediction results; at the same time, the coronary heart disease risk prediction results of this coronary heart disease risk prediction model can, on the one hand, provide a basis for coronary heart disease diagnosis for clinicians, and on the other hand, also provide technical support for decision-making in the medical diagnosis process. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] These and / or other aspects and advantages of the present invention will become apparent and be readily understood from the following description of the preferred embodiments in conjunction with the accompanying drawings, in which:

[0070] Figure 1 It is a schematic flowchart of the steps of the coronary heart disease risk prediction method according to Embodiment 1 of the present invention;

[0071] Figure 2 A wrist pulse wave view generated from the wrist pulse data according to an embodiment of the present invention;

[0072] Figure 3 A fingertip pulse wave view generated from the fingertip pulse data according to an embodiment of the present invention;

[0073] Figure 4 A structural schematic diagram of a coronary heart disease risk prediction device according to the second embodiment of the present invention. Detailed implementation manners

[0074] The technical solutions of the present invention will be further specifically described below through embodiments in conjunction with the accompanying drawings. In the specification, the same or similar reference numerals indicate the same or similar components. The description of the embodiments of the present invention with reference to the accompanying drawings is intended to explain the overall inventive concept of the present invention and should not be construed as a limitation on the present invention.

[0075] In recent years, in many scientific and medical fields, attempts have been made to use artificial intelligence machine learning methods to identify, analyze, and learn a certain amount of event data, and then predict the probability of a specific event through a specific prediction model. However, when existing learning prediction models are used to process events such as medical diagnoses and disease probabilities, due to the deep and complex correlations between human physiological index data, the effect and efficiency of their deep learning are limited. Therefore, the embodiments of the present invention provide a prediction method and device based on feature fusion and deep learning for predicting the risk of coronary heart disease, so as to construct a coronary heart disease risk prediction model with high stability, high accuracy, and high prediction efficiency, thereby efficiently and accurately predicting whether a target object has coronary heart disease and the probability of its risk of having coronary heart disease, and providing technical support for data analysis in the field of coronary heart disease and related medical research fields for medical detection and disease treatment.

[0076] Embodiment 1

[0077] Refer to Figure 1 , which shows a coronary heart disease risk prediction method provided in this embodiment. The coronary heart disease risk prediction method constructs a coronary heart disease risk prediction model (hereinafter referred to as the model) through the following specific steps and uses the model to predict whether a target object has coronary heart disease or the probability of its coronary heart disease.

[0078] Step S100: Obtain the wrist pulse data, fingertip pulse data, and electrocardiogram data of a sample object, and extract the wrist pulse wave view, fingertip pulse wave view, and electrocardiogram view therefrom.

[0079] In one example, specifically, a certain number of sample objects are pre - collected. The number of sample objects should be at least greater than 100, and preferably at least greater than 300, to ensure the comprehensiveness and representativeness of the pulse and electrocardiogram data collected for the model. Further, for each sample object, three types of indicators, namely cun - kou pulse data, fingertip pulse data, and electrocardiogram data, are collected and the corresponding data is obtained.

[0080] In one example, alternatively, important characterization parameters of the cun - kou pulse data, fingertip pulse data, and electrocardiogram data of the sample objects are extracted by a collection instrument (such as a pulse - diagnosis instrument, an electrocardiograph, etc.) or by a data - processing device (such as a computer, a data - processing server, etc.) and corresponding cun - kou pulse wave views, fingertip pulse wave views, and electrocardiogram wave views are generated. Alternatively, a pulse - diagnosis instrument can be used to collect important characterization parameters of the cun - kou pulse or fingertip pulse and directly generate a cun - kou pulse wave view or a fingertip pulse wave view in the pulse - diagnosis instrument; an electrocardiograph is used to collect important characterization parameters in the electrocardiogram data and directly generate an electrocardiogram wave view in the electrocardiograph. Alternatively, the above - mentioned important characterization parameters can be collected and recorded and stored first by a pulse - diagnosis instrument, an electrocardiograph, etc., and then corresponding views are generated by other data - processing devices (analyzers, microprocessors, computers, servers, etc.). For example, first, a pulse - diagnosis instrument is used to collect and record and store cun - kou pulse data related to coronary heart disease, and then a heart - pulse analysis instrument is used to analyze and process the cun - kou pulse data to obtain a cun - kou pulse wave view. Those skilled in the art can understand that this example only provides some illustrative examples and cannot be used as a limitation of the present invention.

[0081] In one example, referring to Figure 2 , a cun - kou pulse wave view and its expressed important characterization parameters are shown. Alternatively, a pulse - diagnosis instrument can be used to collect cun - kou pulse data at the cun - kou position of the sample object (the pulse - diagnosis site of the radial artery on the inner side of the two hands' radius bones, divided into three parts: cun, guan, and chi, that is, the traditional Chinese medicine pulse - taking position). Then, a cun - kou pulse wave view as shown in Figure 2 can be obtained through a pulse - diagnosis instrument or other data - analysis and calculation devices. Corresponding to the cun - kou pulse wave view, as shown in Table 1, it includes the following important characterization parameters. Of course, those skilled in the art can understand that the device through which the cun - kou pulse wave view is obtained and which characterization parameters are selected as the characteristic data of the subsequent coronary heart disease risk prediction model can be selected according to the actual situation. This example only provides some illustrative examples and cannot be used as a limitation of the present invention. The data collection and image acquisition methods of the fingertip pulse wave view and the electrocardiogram wave view are similar to the method used for the cun - kou pulse wave view and will not be elaborated further later.

[0082] Table 1 Characterization parameters and their meanings in the cun - kou pulse wave view

[0083]

[0084] In one example, referring to Figure 3 , a fingertip pulse wave view and important characteristic parameters expressed thereby are shown. Alternatively, a fingertip pulse oximeter can be used to collect fingertip pulse data by clamping the fingertip of a sample subject. Thereafter, a fingertip pulse wave view as shown in Figure 3 can be obtained by the fingertip pulse oximeter or other data analysis and calculation devices. Corresponding to the fingertip pulse wave view, as shown in Table 2, the following important characteristic parameters are included.

[0085] Table 2 Characteristic Parameters and Their Meanings in the Fingertip Pulse Wave View

[0086] Characterization parameter Meaning T = tF(s) Duration of a complete cycle of a single pulse wave (from point A to point A') <![CDATA[Tab=t1F(s)]]> Duration from the starting point of the pulse wave to the main wave (from point A to point B) <![CDATA[Hb=h1F]]> Amplitude of the main wave (amplitude of the wave peak at point B)

[0087] In one example, alternatively, an electrocardiograph can be used to collect electrocardiogram data of a sample subject. Thereafter, an electrocardiogram wave view can be obtained by the electrocardiograph or other data analysis and calculation devices. Referring to Table 3, a part of important characteristic parameters included in the electrocardiogram wave view are listed.

[0088] Table 3 Characteristic Parameters and Their Meanings in the Electrocardiogram Wave View

[0089]

[0090] After performing physical sign data collection, data extraction, and generating views in the form of charts, etc. on each sample subject through the above process, the wrist pulse wave view, fingertip pulse wave view, and electrocardiogram wave view of all sample subjects are obtained. In various embodiments of the present invention, the data basis input into the model is these three types of views and the important characteristic parameters they contain. Of course, those skilled in the art can understand that the basic data input into the model can be adjusted according to the prediction target, prediction requirements, etc. of the model, including but not limited to adjusting the types and quantities of views, the types and quantities of characteristic parameters in the views, etc. This example is only an illustrative example, and those skilled in the art should not understand it as a limitation to the present invention.

[0091] Step S200: Extract wrist pulse characteristic data, fingertip pulse characteristic data, and electrocardiogram characteristic data from the wrist pulse wave view, fingertip pulse wave view, and electrocardiogram wave view respectively to form a multi-view characteristic data set of the heart pulse, and obtain wrist pulse characteristic data, fingertip pulse characteristic data, and electrocardiogram characteristic data with a unified dimension through dimension conversion.

[0092] In one example, specifically, several characteristic parameters included in each type of view of all sample subjects form a characteristic data set corresponding to that view. That is, a wrist pulse characteristic data set is formed corresponding to the wrist pulse wave view a fingertip pulse characteristic data set is formed corresponding to the fingertip pulse wave view Form an electrocardiogram feature data set corresponding to the electrocardiogram waveform view Concatenate the above three groups of feature data sets to form a multi-view feature data set of the heart pulse. The expression of this multi-view feature data set of the heart pulse is as follows:

[0093]

[0094] In one example, further, classify all sample objects and label them according to the classification. One category is the sample objects without coronary heart disease, which can be labeled as O exemplarily; the other category is the sample objects with coronary heart disease, which can be labeled as 1 exemplarily. At this time, the annotation set of all sample objects can be expressed as Y i ={0, 1}. Then, the expression of the multi-view feature data set of the heart pulse including the annotation of all objects is {X i , Y i} N .

[0095] In one example, further, perform data cleaning on the cun-kou pulse feature data, finger-tip pulse feature data, and electrocardiogram feature data in the multi-view feature data set of the heart pulse. Alternatively, data cleaning includes deleting the data with missing heart pulse characterization features or abnormal heart pulse measurement values. For example but not limited to, the data that exceeds its effective measurement range in the characterization parameters of each view (cun-kou pulse wave view, finger-tip pulse wave view, and electrocardiogram view), the data with incomplete characterization parameters during the acquisition period, the data with missing in important positions of the view, etc. In this way, the influence of noise data on subsequent data processing, feature fusion, and model training can be reduced, the quality of the multi-view feature data set of the heart pulse can be guaranteed, and it is helpful to improve the accuracy of the model.

[0096] In one example, when cleaning the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data in the multi-view cardiac feature dataset, exemplarily, the isnull() function can be first used to check each specific feature in the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data one by one to determine whether there are null values in the specific data values of each specific feature. For example, when checking one by one through the isnull() function, situations where the specific value of the time value t1 in the cunkou pulse feature data of a certain sample object is not collected by the sensor or the data collected by the sensor is not stored will be screened and recorded by the isnull() function. Next, for any sample object with missing features detected by the isnull() function, regardless of which features are missing, the number of missing features, and the distribution positions of the missing features, a strict elimination strategy is executed. Exemplarily, the drop() function can be used to completely remove the sample object with missing features, so as to eliminate the interference of potential feature null values in the dataset. Exemplarily, the sample objects with obvious outliers in the feature data can also be screened and eliminated by adjusting the condition values of the isnull() function, the drop() function, or other functions with data screening and cleaning functions. Through the data cleaning process, it can be effectively ensured that each feature of each sample object in the multi-view cardiac feature dataset at this time has integrity and authenticity, avoiding the introduction of random noise or systematic bias into the multi-view cardiac feature dataset due to filling operations for missing values, outliers, and other error values deviating from the detected data, thereby providing a reliable and high-quality data basis for subsequent model learning.

[0097] In one example, the multi-view cardiac feature dataset after cleaning is divided into the training set and the test set of the model according to a preset ratio. Alternatively, the preset ratio can be adjusted according to the training or prediction requirements of the model. For example, the ratio of the training set to the test set is 9:1, 8:2, 7:3, 6:4, etc. Preferably, in this embodiment, the ratio of the training set to the test set is set to 8:2. At the same time, when dividing which sample objects into the training set and which into the test set, it is randomly divided to ensure an equal data sample basis for the training of the model and its prediction of coronary heart disease risk.

[0098] In one example, after the division of the training set and the test set of the model is completed, the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data therein need to be preprocessed. Alternatively, the preprocessing includes normalization processing and standardization processing. After normalization and standardization processing, it can be ensured that all the above-mentioned feature data are converted into cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data with a unified dimension.

[0099] In one example, specifically, the normalization process scales all the wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data to the same range, such as from 0 to 1. This can eliminate the scale differences among these feature data, thereby avoiding the impact of dimensional differences of some feature data on the stability and accuracy of the model.

[0100] In one example, specifically, the standardization process adjusts all the wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data to a standard normal distribution with a mean of 0 and a standard deviation of 1 to improve the convergence speed and stability of the model during training.

[0101] Step S300: Calculate the risk relationship weights between each feature data in the multi-view heart pulse feature dataset input to the coronary heart disease risk prediction model using the self-attention mechanism, and encode the risk relationship weights for each feature data.

[0102] In one example, after the multi-view heart pulse data including wrist pulse data, fingertip pulse data, and electrocardiogram data collected from all sample objects and the corresponding multi-view heart pulse views of the wrist pulse wave view, fingertip pulse wave view, and electrocardiogram view are preprocessed (including data cleaning, normalization process, and standardization process), a multi-view heart pulse feature dataset including wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data that can be input to the model is obtained. Before inputting the multi-view heart pulse feature dataset into the model, it is also necessary to identify and classify different views and their corresponding different categories of feature data to ensure that each feature data can be correctly input to the corresponding marked position in the model. During the loading process of the wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data, each feature data of each view will be assigned a corresponding view identifier and view alignment will be performed.

[0103] In one example, the self-attention mechanism is adopted to construct a self-attention encoder in the model for extracting and encoding each feature data of each view. The self-attention mechanism allows the model to automatically learn the correlation between feature data at different positions when processing sequence data, thereby capturing deeper correlation information between these feature data globally. That is to say, for each feature data in the sequence data, calculate the correlation between this feature data and all other feature data, and generate a weighted vector accordingly. This weighted vector can be regarded as the representation of this feature data in the context of the current sequence data. In this way, the model can automatically learn the dependence relationship between feature data at different positions in the sequence data, so as to obtain information between feature data with deeper correlation meaning. Compared with traditional recurrent neural networks or convolutional neural networks, the self-attention mechanism can effectively capture the dependence relationship between long-distance feature data in sequence data, thereby obtaining global dependence relationship information of sequence data.

[0104] In one example, specifically, after the construction of the multi-view feature dataset of heart pulse, the self-attention mechanism is used for feature extraction and encoding. Alternatively, the encoder consists of a multi-layer perceptron (MLP) and a multi-head attention mechanism (MultiHead Attention), which are used to extract and encode each feature data of the corresponding cun-kou pulse wave features, finger-tip pulse wave features, and electrocardiogram features in the cun-kou pulse wave view, finger-tip pulse wave view, and electrocardiogram view, so as to obtain deep feature information about the risk of coronary heart disease for each view and its respective feature data. The self-attention mechanism captures a more effective feature representation with deep correlation meaning regarding the risk of coronary heart disease by calculating the similarity between input feature data and automatically adjusting the risk relationship weights of each feature data.

[0105] In one example, the specific calculation process of using the self-attention mechanism for feature extraction and encoding includes the following steps:

[0106] Step S31: Map all feature data in the multi-view feature dataset of heart pulse to the same high-dimensional space through a linear transformation.

[0107] Specifically, the expression of the input feature is X ∈ R n×d , where n is the number of features and d is the feature dimension. The linear transformation can be expressed as: X' = WX + b; where W ∈ R d'×d is the weight matrix, which represents the risk relationship weight regarding coronary heart disease here, b is the bias term, and X' is each mapped feature data.

[0108] Step S32: Calculate the query vector, key vector, and value matrix of all feature data.

[0109] Specifically, for each of the wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data after being mapped in step S31, calculate its query vector, key vector, and value matrix respectively. Perform a linear transformation on the mapped feature X' to obtain Q, K, and V, corresponding to the query vector, key vector, and value matrix respectively, and the risk relationship weight matrix W ∈ R d'×d is shared during the calculation of the three.

[0110] Step S33: Calculate the similarity between the query vector of each feature data and the key vectors of all other feature data. Alternatively, the similarity can be obtained through dot product calculation or scaled dot product calculation.

[0111] Step S34: Assign a risk relationship weight to each feature data according to the similarity to obtain the risk relationship weight vector of each feature data.

[0112] Specifically, assign a risk relationship weight to each of the wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data according to the similarity calculation result obtained in step S33. Alternatively, this risk relationship weight can be obtained through the softmax function, where the sum of all elements of the risk relationship weight vector is 1. The softmax function can use, but is not limited to, the following expression:

[0113]

[0114] where d k is the dimension of the key.

[0115] Step S35: Multiply the value matrix of each feature data by its risk relationship weight vector to obtain the weighted summation result of all feature data with respect to the risk relationship weight.

[0116] Step S36: Concatenate the weighted summation results of all feature data with respect to the risk relationship weight together, and obtain the output vector of the self-attention encoding through linear transformation.

[0117] Alternatively, the fusion of the wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data of each view can be achieved through a linear layer transformation. The expression of the linear layer transformation is: H = Linear(A); where Linear is a linear transformation layer, A is the self-attention output, and H is the final fused feature representation.

[0118] In one example, for a specific coronary heart disease risk prediction task, alternatively, three groups of Transformer network structures can be deployed in parallel to process feature data from different perspectives respectively. For example, the first Transformer network structure processes the wrist pulse feature data, the second Transformer network structure processes the fingertip pulse feature data, and the third Transformer network structure processes the electrocardiogram feature data. These feature data from different perspectives are used to separately extract the wrist pulse features, fingertip pulse features, and electrocardiogram features through the Transformer network structures that work independently. Further alternatively, it is also possible to set up dedicated modeling for each type of feature data within their respective Transformer network structures. For example, a wrist pulse feature model is established in the first Transformer network structure, a fingertip pulse feature model is established in the second Transformer network structure, and an electrocardiogram feature model is established in the third Transformer network structure. When the feature data from different perspectives are fused, these feature models from different perspectives can also be fused, which helps to further improve the efficiency of model learning and thus improve the accuracy of the model in predicting the risk of coronary heart disease.

[0119] In one example, the multi-head self-attention mechanism is set in the Transformer network structure to capture the long-range dependencies between the feature data from different perspectives (actually, the different physiological signals of an individual through data representation) in the coronary heart disease risk prediction task. Such long-range dependencies can occur in the dependencies directly or indirectly related to the coronary heart disease risk between some feature data in a set of feature data from a certain perspective (such as the fingertip pulse feature data) (such as the dependency between Tab and Hb in Table 2). It can also occur in the dependencies directly or indirectly related to the coronary heart disease risk between some feature data (such as Tab in Table 2) in a set of feature data from a certain perspective (such as the fingertip pulse feature data) and some other feature data (such as HRV in Table 3) in a set of feature data from another perspective (such as the electrocardiogram feature data).

[0120] Therefore, different from traditional feature learning methods, the Transformer network structure is better at processing long sequence data by focusing on the long-range dependencies between feature data, thus effectively capturing the deep and complex dependencies between multi-dimensional features. For example, there may be different time delays or phase correspondence relationships between some feature values in the wrist pulse feature data and some other feature values in the fingertip pulse feature data, thus having feature dependencies, and these feature dependencies need to be captured and analyzed by the self-attention mechanism and used for subsequent model building and model learning.

[0121] In one example, further, in the Transformer network structure, exemplarily, layer normalization and residual connections are also provided, so that the output of each layer can be refined, which helps to stabilize the model training process and avoid the problem of gradient disappearance, thereby enhancing the model's learning ability for coronary heart disease-related features. Especially for medical or disease data, layer normalization also helps to accelerate the convergence of model learning. At the same time, the residual connection helps to maintain the stability of model training when the model processes features with deep correlations and complexity, ensuring that feature information can be effectively transmitted and updated.

[0122] In one example, further, at the end of the Transformer network structure, alternatively, a global average pooling layer can also be used to compress the output fused feature map to the feature embedding dimension. After the fused feature extraction is completed, the global average pooling layer is used to pool the fused features output by the multi-view cardiac pulse feature dataset. For example, this process can compress the feature maps of each view into a fixed preset-size feature vector, providing a unified feature input for subsequent classification and prediction tasks. Since in the coronary heart disease risk prediction task, the feature data often comes from multiple perspectives, and there are many specific feature data directly or indirectly related to coronary heart disease in the feature data of each perspective. Therefore, the model needs to integrate data from different sources (such as pulse signals at different positions) into a compact representation for the model to more efficiently predict the coronary heart disease risk through deep learning.

[0123] In summary, the self-attention mechanism can dynamically adjust the weights according to the similarity between each feature data, so that the model can automatically focus on the features most relevant to the coronary heart disease risk. For example, there may be important similarities between the cun-kou pulse and the fingertip pulse, and the electrocardiogram data may provide important risk signals. By calculating the similarity between the query vector and the key vector, the model can automatically adjust the weights according to the relationship between the feature data, thereby achieving efficient feature fusion.

[0124] Step S400: Calculate the trust degree of the risk evidence value of the multi-view cardiac pulse feature dataset using the evidence theory, obtain the fused multi-view cardiac pulse feature dataset, and train the coronary heart disease risk prediction model.

[0125] In one example, the Dempster / Shafer evidence theory (hereinafter referred to as the DS evidence theory) is adopted to fuse the information of the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data in the multi-view feature dataset of heart pulses for the training and prediction of the model. In the DS evidence theory, the Dempster combination rule is to fuse the features or information of multiple entities (such as different prediction result requirements, different types of collected data, output results of different weight calculations, etc.). By calculating the confidence and / or mass of the Dirichlet distribution parameters, weight assignment and synthesis are performed, and finally the fused feature representation is obtained.

[0126] Generally, the DS evidence theory assigns probabilities to each hypothesis in its identification framework (such as the set of each view in the coronary heart disease risk prediction model or the set of feature data corresponding to each view), which is defined as the basic probability assignment (BPA, Basic Probability Assignment) or the basic belief assignment (BBA, Basic Belief Assignment). Then calculations can be performed through, for example, the mass assignment function. Among them, the mass function value (such as probability or trust degree) of each hypothesis is between 0 and 1.

[0127] In one example, specifically, each view is defined as v in the model i , and each view is derived from the cunkou pulse wave view, fingertip pulse wave view, and electrocardiogram wave view of all sample objects input into the model. Then the corresponding feature data of each view v i is used as the risk evidence e i . Specifically, the risk evidence e i can be regarded as a mapping from the feature data of each view input into the model to the Dirichlet distribution parameter α i , where the risk evidence e i is used to calculate the Dirichlet distribution parameter α i , and its expression is α i = e i + 1.

[0128] For the Dirichlet distribution parameter α i of each view, its belief mass b i and uncertainty u i are calculated respectively, and the expressions of b i and u i are as follows:

[0129]

[0130] Among them, E i = α i -1 represents the risk evidence vector, K represents the number of classification categories, represents the Dirichlet distribution intensity.

[0131] According to the Dempster combination rule, the belief degrees b i and uncertainties u i of each view are fused.

[0132] First, calculate the combined belief degree between two of the views (the corresponding belief degrees are b1 and b2, and the uncertainties are u1 and u2 respectively): Among them, C is the conflict measure, defined as:

[0133] After that, similarly, gradually fuse the Dirichlet distribution parameters of each view, and finally obtain the comprehensive Dirichlet distribution parameters.

[0134] The uncertainty u a and the Dirichlet distribution intensity S a after fusion can be calculated through the following expressions:

[0135]

[0136] Finally, the expression of the comprehensive Dirichlet distribution parameter α a after fusion is:

[0137] α a = b a · S a + 1.

[0138] In an example, the coronary heart disease prediction data set after annotation, data cleaning and preprocessing (that is, the set of cunkou pulse wave views, fingertip pulse wave views and electrocardiogram views including all sample objects and their corresponding cunkou pulse characteristic data, fingertip pulse characteristic data and electrocardiogram characteristic data) is input into the model as the basic data for model training. And, in the model, the risk relationship weights of the cunkou pulse characteristic data, fingertip pulse characteristic data and electrocardiogram characteristic data are extracted through the self-attention encoder for risk relationship weight coding, and the risk evidence values of each view are predicted through the classifier. Also, for each view, its risk evidence value is represented as the Dirichlet distribution parameter α through the Dirichlet distribution model, and then the DS evidence theory is used to fuse the Dirichlet distribution parameters α of each view into the global comprehensive Dirichlet distribution parameter α a, thus finally achieving the comprehensive integration of all feature data (including all cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data in the multi-view feature dataset of heart vessels). The model performs deep learning and model training based on the comprehensively integrated feature data and the cunkou pulse wave views, fingertip pulse wave views, and electrocardiogram views of all sample objects after comprehensive integration.

[0139] In the initial stage of model training, the selection of different training parameters and training formulas may not fully adapt to the complexity of the feature data and the complex requirements of the task for coronary heart disease risk prediction, resulting in the prediction accuracy not reaching the ideal precision at this time. Alternatively, the model establishment process and / or learning process can be optimized through the following listed examples.

[0140] In one example, for instance, the method of adjusting model hyperparameters can be used. In the initial training stage, hyperparameters such as the learning rate, batch size, and number of network layers of the model may not be fully optimized. By iteratively experimenting to adjust these hyperparameters related to model training, the convergence speed of the model can be accelerated and the prediction accuracy can be improved. For example, appropriately reducing the learning rate can, to a certain extent, help the model converge more smoothly and avoid excessive fluctuations.

[0141] In one example, for instance, the feature extraction process can also be optimized. A more refined coding method or a more suitable self-attention mechanism may be required during the feature extraction process. By purposefully adjusting the model structure hyperparameters of the feature extraction module (such as the self-attention encoder), optimizing the accuracy of feature extraction helps improve the fusion effect of risk evidence in the DS evidence theory, thereby gradually enhancing the model prediction accuracy.

[0142] In one example, during the training process of the model, the cross-entropy loss function combined with KL divergence regularization can also be used to optimize the model. Among them, the cross-entropy loss function is used to measure the matching degree between the predicted probability distribution of the model and the true label, and the KL divergence is used to regularize the Dirichlet distribution parameter α, which can improve the robustness and accuracy of the model, ensure the rationality of the distribution of risk evidence values of all feature data, and prevent the model from being overconfident during prediction.

[0143] In one example, specifically, the expression of the cross-entropy loss function is as follows:

[0144] L = ∑ i L CE (α i , y) + λKL(α i ) + L CE (α a , y) + λKL(α a );

[0145] where L CE represents the cross - entropy loss, KL represents the KL divergence, λ is the regularization term coefficient, and y is the true label. The model achieves the training objective by minimizing this cross - entropy loss function. In each training iteration, the model updates the risk relationship weights of each feature data through backpropagation and gradient descent, thereby gradually optimizing the process of extracting each feature data of each view and the subsequent risk evidence fusion.

[0146] In one example, for instance, the fusion process of the Dempster synthesis rule can also be gradually optimized. During the fusion process of the feature data in the multi - view feature dataset of the heart pulse, the implementation of the Dempster synthesis rule may have a problem of excessive conflict degree in the initial stage, affecting the feature fusion effect. By adjusting the trust degree and uncertainty weights of some important feature data in the cun - kou pulse view, fingertip pulse view, or electrocardiogram, the fusion strategy is gradually optimized, thereby reducing the conflict degree and improving the accuracy of the final combined Dirichlet distribution parameters.

[0147] The above - mentioned specific examples of how to optimize the model establishment process and the model learning process are only for providing some commonly used, relatively easy - to - implement, and relatively effective examples verified through iterative experiments. Those skilled in the art should not understand it as a limitation of the present invention.

[0148] Step S500: The trained coronary heart disease risk prediction model performs class probability prediction on the multi - view feature dataset of the heart pulse of the target object, and outputs the coronary heart disease risk prediction result of the target object.

[0149] In one example, the trained model can be used to predict the coronary heart disease risk of the target object. The steps of this coronary heart disease risk prediction specifically include:

[0150] Step S51: Obtain the cun - kou pulse data, fingertip pulse data, and electrocardiogram data of the target object, and extract the cun - kou pulse wave view, fingertip pulse wave view, and electrocardiogram wave view from them.

[0151] Step S52: Respectively extract the cun - kou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data from the cun - kou pulse wave view, fingertip pulse wave view, and electrocardiogram wave view to form a multi - view feature dataset of the heart pulse, and obtain the cun - kou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data with a unified dimension through dimension conversion.

[0152] In one example, the data processing procedures of the above-mentioned steps S51 and S52 can adopt the same data preprocessing method as the processing procedures of the wrist pulse wave view, fingertip pulse wave view, and electrocardiogram view of all sample objects when constructing the model, which will not be elaborated here. After data preprocessing, a multi-view feature dataset of the target object's heart pulse can be obtained.

[0153] Step S53: Input the multi-view feature dataset of the target object's heart pulse into the coronary heart disease risk prediction model, and perform risk relationship weight encoding on each feature data in the multi-view feature dataset of the heart pulse in the coronary heart disease risk prediction model and calculate the corresponding risk evidence value.

[0154] Step S54: Use the DS evidence theory to fuse the risk evidence values of the feature data of each view to obtain the comprehensive Dirichlet distribution parameter α a .

[0155] In one example, the comprehensive fusion process of the feature data based on the DS evidence theory in the above-mentioned steps S53 and S54 is the same as the comprehensive fusion process of the feature data when constructing the model. The specific calculation steps and calculation function expressions can refer to the content of the aforementioned model construction and model training, which will not be elaborated here.

[0156] Step S55: Calculate the binary classification probability distribution of the target object through the comprehensive Dirichlet distribution parameter α a and output the coronary heart disease risk prediction result.

[0157] In one example, specifically, the binary classification of the target object refers to two results of the probability prediction output. One result is that the target object has coronary heart disease, and the other result is that the target object does not have coronary heart disease.

[0158] In one example, further, classifying and making a decision on the binary classification probability distribution of the target object can obtain the final coronary heart disease risk prediction result, that is, whether the target object has coronary heart disease. Alternatively, the classification decision can be achieved by calculating the cross-entropy loss of the classification head, and then the fused wrist pulse wave view, fingertip pulse wave view, and electrocardiogram view, as well as the risk evidence values of the fused wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data corresponding thereto, are represented as the final output (coronary heart disease risk prediction result). Specifically, the softmax activation function can be used to normalize the output coronary heart disease risk prediction result to obtain the binary classification coronary heart disease risk prediction probability, and its expression is:

[0159] p(y=k|α a )=softmax(W h ·α a +bh );

[0160] where α a represents the expression of the risk evidence value of the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data fused by the DS evidence theory, that is, the comprehensive Dirichlet distribution parameter; W h and b h are the weights and biases of the classifier. The model evaluates the risk of the target object suffering from coronary heart disease through this binary classification prediction probability and outputs the coronary heart disease risk prediction result.

[0161] In one example, assume that p0 represents the probability that the target object does not have coronary heart disease, and p1 represents the probability that the target object has coronary heart disease. Based on the binary classification prediction probability p output by the softmax activation function, the category with the highest probability is selected as the final coronary heart disease risk prediction result. If p0 > p1, the target object does not have coronary heart disease; otherwise, the target object has a risk of suffering from coronary heart disease.

[0162] In one example, referring to Table 4, it shows a comparison of the accuracy of coronary heart disease risk prediction obtained by the coronary heart disease risk prediction model trained by the coronary heart disease risk prediction method of this embodiment and the accuracy of coronary heart disease risk prediction obtained by other traditional machine learning algorithms.

[0163] Table 4 Comparison of prediction accuracies between the coronary heart disease risk prediction method of the present invention and traditional machine learning algorithms

[0164]

[0165]

[0166] The accuracy of the prediction result obtained by the coronary heart disease risk prediction method provided by the embodiment of the present invention and the trained coronary heart disease risk prediction model is significantly higher than that of other traditional machine learning algorithms, which helps to obtain a more accurate result of predicting the coronary heart disease risk of the target object.

[0167] Embodiment 2

[0168] Referring to Figure 4 , it shows a coronary heart disease risk prediction device 100 provided by this embodiment, which is used to predict whether the target object has coronary heart disease by using the coronary heart disease risk prediction method described in the above embodiments and obtain a prediction result. The coronary heart disease risk prediction device 100 is composed of 5 mutually related data processing and analysis modules, namely an acquisition module 10, a feature extraction module 20, a self-attention encoding module 30, a feature fusion module 40, and a prediction module 50. The specific function configurations of the above modules and their calculation and processing processes for data are described as follows:

[0169] In one example, the acquisition module 10 is configured to acquire the cunkou pulse data, fingertip pulse data, and electrocardiogram data of a sample object, and extract the cunkou pulse wave view, fingertip pulse wave view, and electrocardiogram view therefrom.

[0170] In one example, the feature extraction module 20 is configured to extract cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data from the cunkou pulse wave view, fingertip pulse wave view, and electrocardiogram view respectively to form a multi-view feature dataset of the heart pulse, and obtain the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data with a unified dimension through dimension transformation.

[0171] In one example, the self-attention encoding module 30 is configured to calculate the relationship weights between each feature data in the multi-view feature dataset of the heart pulse input to the coronary heart disease risk prediction model by using the self-attention mechanism, and perform weight encoding on each feature data.

[0172] In one example, the feature fusion module 40 is configured to calculate the trust degree of the risk evidence value of the multi-view feature dataset of the heart pulse by using the evidence theory, obtain the fused multi-view feature dataset of the heart pulse, and train the coronary heart disease risk prediction model.

[0173] In one example, the prediction module 50 is configured to use the trained coronary heart disease risk prediction model to perform class probability prediction on the multi-view feature dataset of the heart pulse of the target object, and output the coronary heart disease risk prediction result of the target object.

[0174] It should be noted that the coronary heart disease risk prediction device 100 provided in this embodiment can be used to execute the technical solutions of the above-mentioned embodiments of the coronary heart disease risk prediction method. Its implementation principle and technical effects are the same as or similar to those described above, and will not be elaborated here.

[0175] The coronary heart disease risk prediction method and device provided by the embodiments of the present invention have at least one or a part of at least one of the following advantages:

[0176] (1) The coronary heart disease risk prediction method and device provided by the embodiments of the present invention perform adaptive comprehensive feature fusion on multi-source multi-view data of the cunkou pulse, fingertip pulse, and electrocardiogram to construct a coronary heart disease risk prediction model, realizing high-efficiency and high-precision prediction of the coronary heart disease risk probability of the target object;

[0177] (2) The coronary heart disease risk prediction method and device provided by the embodiments of the present invention use a self-attention mechanism to allocate and encode the risk relationship weights of the wrist pulse characteristic data, fingertip pulse characteristic data, and electrocardiogram characteristic data, and use the DS evidence theory to perform comprehensive feature fusion on all the above-mentioned characteristic data, and then train the coronary heart disease risk prediction model, improving the efficiency of deep learning of the coronary heart disease risk prediction model, enhancing the stability of the coronary heart disease risk prediction model and the accuracy of coronary heart disease risk prediction;

[0178] (3) The coronary heart disease risk prediction method and device provided by the embodiments of the present invention use a self-attention mechanism to perform risk relationship weight encoding on each characteristic data in the multi-view characteristic data set of the heart pulse in the coronary heart disease risk prediction model, which can fully explore the deep association information about coronary heart disease risk among each characteristic data within the global scope;

[0179] (4) The coronary heart disease risk prediction method and device provided by the embodiments of the present invention use the DS evidence theory to fuse the evidence values of the wrist pulse characteristic data, fingertip pulse characteristic data, and electrocardiogram characteristic data according to the Dirichlet distribution parameters of each view in the multi-view characteristic data set of the heart pulse, so as to realize the comprehensive feature fusion of multi-views and their multi-category characteristic data, which helps to improve the training efficiency of the coronary heart disease risk prediction model and the accuracy of coronary heart disease risk prediction;

[0180] (5) The coronary heart disease risk prediction method and device provided by the embodiments of the present invention add a cross-entropy loss function and KL divergence regularization during the training process of the coronary heart disease risk prediction model to optimize the coronary heart disease risk prediction model, ensuring the rationality of the distribution of the evidence values of the wrist pulse characteristic data, fingertip pulse characteristic data, and electrocardiogram characteristic data, and avoiding overconfidence of the coronary heart disease risk prediction model;

[0181] (6) The coronary heart disease risk prediction method and device provided by the embodiments of the present invention use the multi-view characteristic data set of the heart pulse obtained by feature splicing of the wrist pulse data, fingertip pulse data, and electrocardiogram data to provide rich, complete, and comprehensive basic data of human physiological indicators related to coronary heart disease focus for training the coronary heart disease risk prediction model, which helps to improve the prediction accuracy of the coronary heart disease risk prediction model;

[0182] (7) The coronary heart disease risk prediction method and device provided by the embodiments of the present invention obtain deep association information closely related to coronary heart disease risk through comprehensive feature fusion of multi-source data and multi-source views, effectively enhancing the comprehensive learning ability of the coronary heart disease risk prediction model for different source data, and providing a more comprehensive judgment basis for its coronary heart disease risk prediction results; at the same time, the coronary heart disease risk prediction results of the coronary heart disease risk prediction model can, on the one hand, provide a diagnosis basis for clinical doctors for coronary heart disease, and on the other hand, also provide technical support for decision-making in the medical diagnosis process.

[0183] While some embodiments of the present general inventive concept have been shown and described, those of ordinary skill in the art will understand that changes may be made to these embodiments without departing from the principles and spirit of the present general inventive concept, and the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A method for predicting the risk of coronary heart disease, characterized in that, The coronary heart disease risk prediction method includes: Obtain the wrist pulse data, fingertip pulse data, and electrocardiogram data of the sample object, and extract the wrist pulse wave view, fingertip pulse wave view, and electrocardiogram view therefrom; Extract wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data from the wrist pulse wave view, fingertip pulse wave view, and electrocardiogram view respectively to form a multi-view heart pulse feature dataset, and obtain wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data with a unified dimension through dimension transformation; Use the self-attention mechanism to calculate the risk relationship weights between each feature data in the multi-view heart pulse feature dataset input to the coronary heart disease risk prediction model, and encode the risk relationship weights for each feature data; Use the evidence theory to calculate the trust degree of the risk evidence value of the multi-view heart pulse feature dataset, obtain the fused multi-view heart pulse feature dataset, and train the coronary heart disease risk prediction model; The trained coronary heart disease risk prediction model performs class probability prediction on the multi-view heart pulse feature dataset of the target object, and outputs the coronary heart disease risk prediction result of the target object.

2. The coronary heart disease risk prediction method according to claim 1, wherein Preprocess the multi-view heart pulse feature dataset, and the preprocessing includes: Perform data cleaning on all wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data in the multi-view heart pulse feature dataset, and the data cleaning includes deleting data with missing heart pulse characterization features or abnormal heart pulse measurement values; Randomly divide the multi-view heart pulse feature dataset after data cleaning into a training set and a test set of the coronary heart disease risk prediction model according to a preset ratio.

3. The coronary heart disease risk prediction method according to claim 1, wherein The steps of extracting wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data from the wrist pulse wave view, fingertip pulse wave view, and electrocardiogram view respectively to form a multi-view heart pulse feature dataset, and obtaining wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data with a unified dimension through dimension transformation specifically include: Perform normalization processing on each feature data in the wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data, and scale all feature data to the same preset range; Perform standardization processing on each feature data, and adjust all feature data to a standard normal distribution with a mean of 0 and a standard deviation of 1.

4. The coronary heart disease risk prediction method according to claim 1, wherein The steps of using the self-attention mechanism to calculate the risk relationship weights between each feature data in the multi-view heart pulse feature dataset input to the coronary heart disease risk prediction model, and encoding the risk relationship weights for each feature data specifically include: Map all feature data in the multi-view heart pulse feature dataset to the same high-dimensional space through linear transformation; Calculate the query vector, key vector, and value matrix of all feature data; Calculate the similarity between the query vector of each feature data and the key vectors of all other feature data; Assign a risk relationship weight to each feature data according to the similarity to obtain the risk relationship weight vector of each feature data; Multiply the value matrix of each feature data by its risk relationship weight vector to obtain the weighted summation result of all feature data with respect to the risk relationship weight; Concatenate the weighted summation results of all feature data with respect to the risk relationship weight, and obtain the output vector of the self-attention encoding through linear transformation.

5. The coronary heart disease risk prediction method according to claim 1, wherein The steps of calculating the confidence of the risk evidence value of the multi-view feature dataset of the heart pulse using the evidence theory, obtaining the fused multi-view feature dataset of the heart pulse, and training the coronary heart disease risk prediction model specifically include: Take the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data in the multi-view feature dataset of the heart pulse as risk evidences, and calculate the Dirichlet distribution parameters of the cunkou pulse wave view, fingertip pulse wave view, and electrocardiogram wave view respectively; Calculate the confidence and uncertainty of the corresponding view according to the Dirichlet distribution parameter of each view; Fuse the confidence and uncertainty of each view according to the combination rule to obtain the comprehensive Dirichlet distribution parameter of the multi-view feature dataset of the heart pulse; Use the cross-entropy loss function to measure the matching degree between the probability distribution of the coronary heart disease risk prediction model and the true label; Regularize the comprehensive Dirichlet distribution parameter using KL divergence.

6. The coronary heart disease risk prediction method according to claim 1, wherein The steps of the trained coronary heart disease risk prediction model performing class probability prediction on the multi-view feature dataset of the heart pulse of the target object and outputting the coronary heart disease risk prediction result of the target object specifically include: Obtain the cunkou pulse data, fingertip pulse data, and electrocardiogram data of the target object, and extract the cunkou pulse wave view, fingertip pulse wave view, and electrocardiogram wave view therefrom; Extract the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data from the cunkou pulse wave view, fingertip pulse wave view, and electrocardiogram wave view respectively to form a multi-view feature dataset of the heart pulse, and obtain the cunkou pulse feature data, fingertip pulse feature data, and electrocardiogram feature data with a unified dimension through dimension conversion; Input the multi-view feature dataset of the heart pulse of the target object into the coronary heart disease risk prediction model, perform risk relationship weight encoding on each feature data in the multi-view feature dataset of the heart pulse in the coronary heart disease risk prediction model, and calculate the corresponding risk evidence value; Fuse the risk evidence values of the feature data of each view using the DS evidence theory to obtain the comprehensive Dirichlet distribution parameter; Calculate the binary classification probability distribution of the target object through the comprehensive Dirichlet distribution parameter and output the coronary heart disease risk prediction result.

7. A coronary heart disease risk prediction device, which is used to predict whether a target object has coronary heart disease by using the coronary heart disease risk prediction method according to any one of claims 1-6 and obtain a coronary heart disease risk prediction result, and is characterized in that, The coronary heart disease risk prediction device includes: An acquisition module configured to acquire the cunkou pulse data, fingertip pulse data, and electrocardiogram data of the sample object, and extract the cunkou pulse wave view, fingertip pulse wave view, and electrocardiogram wave view therefrom; A feature extraction module, configured to extract wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data from a wrist pulse wave view, a fingertip pulse wave view, and an electrocardiogram view respectively to form a multi-view heart pulse feature dataset, and obtain wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data with a unified dimension through dimension conversion; A self-attention encoding module, configured to calculate the risk relationship weights between each feature data in the multi-view heart pulse feature dataset input to the coronary heart disease risk prediction model by using a self-attention mechanism, and encode the risk relationship weights for each feature data; A feature fusion module, configured to calculate the trust degree of the risk evidence value of the multi-view heart pulse feature dataset by using the evidence theory, obtain the fused multi-view heart pulse feature dataset, and train the coronary heart disease risk prediction model; A prediction module, configured to use the trained coronary heart disease risk prediction model to perform class probability prediction on the multi-view heart pulse feature dataset of the target object, and output the coronary heart disease risk prediction result of the target object.

8. The coronary heart disease risk prediction device according to claim 7, wherein the self-attention encoding module is further configured to: map all the feature data in the multi-view heart pulse feature dataset to the same high-dimensional space through a linear transformation; calculate the query vector, key vector, and value matrix of all the feature data; calculate the similarity between the query vector of each feature data and the key vectors of all other feature data; assign a risk relationship weight to each feature data according to the similarity to obtain a risk relationship weight vector for each feature data; multiply the value matrix of each feature data by its risk relationship weight vector to obtain the weighted summation result of all the feature data with respect to the risk relationship weight; concatenate the weighted summation results of all the feature data with respect to the risk relationship weight together, and obtain the output vector of the self-attention encoding through a linear transformation.

9. The coronary heart disease risk prediction device according to claim 7, wherein the feature fusion module is further configured to: take the wrist pulse feature data, fingertip pulse feature data, and electrocardiogram feature data in the multi-view heart pulse feature dataset as risk evidences, and calculate the Dirichlet distribution parameters of the wrist pulse wave view, the fingertip pulse wave view, and the electrocardiogram view respectively; calculate the trust degree and uncertainty of the corresponding view according to the Dirichlet distribution parameter of each view; fuse the trust degrees and uncertainties of each view according to the synthesis rule to obtain the comprehensive Dirichlet distribution parameter of the multi-view heart pulse feature dataset; use the cross-entropy loss function to measure the matching degree between the probability distribution of the coronary heart disease risk prediction model and the true label; perform regularization on the comprehensive Dirichlet distribution parameter by using the KL divergence.

10. The coronary heart disease risk prediction device according to claim 7, wherein the prediction module is further configured to: obtain the wrist pulse data, fingertip pulse data, and electrocardiogram data of the target object, and extract the wrist pulse wave view, the fingertip pulse wave view, and the electrocardiogram view therefrom; Extract the wrist pulse characteristic data, fingertip pulse characteristic data, and electrocardiogram characteristic data from the wrist pulse wave view, fingertip pulse wave view, and electrocardiogram view respectively to form a multi-view heart pulse characteristic data set, and obtain the wrist pulse characteristic data, fingertip pulse characteristic data, and electrocardiogram characteristic data with a unified dimension through dimensionality conversion; Input the multi-view heart pulse characteristic data set of the target object into the coronary heart disease risk prediction model, and perform risk relationship weight coding on each characteristic data in the multi-view heart pulse characteristic data set in the coronary heart disease risk prediction model and calculate the corresponding risk evidence value; Use the DS evidence theory to fuse the risk evidence values of the characteristic data of each view to obtain the comprehensive Dirichlet distribution parameter; Calculate the binary classification probability distribution of the target object through the comprehensive Dirichlet distribution parameter and output the coronary heart disease risk prediction result.

Images