Method for constructing cardiovascular and cerebrovascular disease classification model

Abstract

The invention provides a cardiovascular and cerebrovascular disease classification model construction method, which comprises the following steps: receiving dynamic signal data of a cardiovascular and cerebrovascular disease patient, and constructing a dynamic signal matrix; global pathological features of patients with cardiovascular and cerebrovascular diseases are collected to serve as static feature vectors, and time dimensions of the static feature vectors and the dynamic signal feature matrix are unified to construct a fusion feature matrix; constructing a common disease association network; when patient group grouping is carried out, similarity mapping from individuals to groups is carried out through similarity calculation of features of each time slice and a group feature center to generate patient group features, and the patient group features are aligned with a patient feature center matrix; and constructing a deep classifier for cardiovascular and cerebrovascular disease classification, and finally outputting a classification result by the classifier. The method has significant breakthroughs in the aspects of dynamic feature modeling, disease relevance modeling and personalized adaptation capability, and an efficient and accurate technical means is provided for cardiovascular and cerebrovascular disease classification.

Description

Technical Field

[0001] The present invention belongs to the technical field of model construction, and in particular relates to a method for constructing a cardiovascular and cerebrovascular disease classification model. Background Art

[0002] Cardiovascular and cerebrovascular diseases are the leading cause of death and disability worldwide. Their pathogenesis is complex and is often associated with multiple physiological and pathological factors, such as dyslipidemia, hypertension, diabetes, and arrhythmias. Traditional diagnostic methods for cardiovascular and cerebrovascular diseases rely primarily on static pathological indicators and limited time series data, such as blood tests (such as LDL-C and HDL-C levels) and electrocardiograms to determine disease risk. However, these methods have the following significant problems:

[0003] Dynamic characteristics are underutilized: Data such as Holter electrocardiograms (ECGs) and ambulatory blood pressure monitoring (ABPMs) contain early signs of cardiovascular and cerebrovascular disease, such as heart rate variability and blood pressure fluctuations. However, existing technologies typically use fixed time window analysis or simple average processing, failing to capture the signal's temporal dependencies and dynamic characteristics. This neglect can lead to the loss of early disease signals, which is particularly evident in chronic, progressive diseases.

[0004] Inter-disease associations are not effectively modeled: Cardiovascular and cerebrovascular diseases are often accompanied by chronic conditions such as hypertension and diabetes, creating complex comorbidities. Current classification methods typically treat these diseases as independent features and fail to effectively leverage the inter-disease associations. For example, arteriosclerosis caused by hypertension often leads to insufficient blood flow to the heart or brain, but existing models fail to fully capture these pathological associations, reducing classification accuracy.

[0005] Ignoring individual characteristics: The manifestations of cardiovascular and cerebrovascular diseases vary from person to person and are significantly influenced by age, gender, genetic background, and lifestyle. However, existing classification methods are mostly trained based on overall population data, which makes them less adaptable to specific populations (such as the elderly or high-risk patients), making it difficult for classification models to accurately diagnose diverse patient groups.

[0006] Lack of unified multimodal feature integration: Cardiovascular disease classification tasks involve dynamic data (such as ECG and ABPM), static data (such as blood lipid indicators), and comorbidity information (such as clinical records of diabetes and hypertension). Existing technologies typically model each type of feature separately, resulting in a lack of effective cross-modal fusion between features, which limits the model's diagnostic performance.

[0007] Therefore, traditional cardiovascular and cerebrovascular disease classification technologies suffer from problems such as insufficient feature utilization, poor personalized adaptability, and insufficient multimodal fusion, failing to meet the requirements of precision medicine. In this context, there is an urgent need for a classification method that can fully utilize dynamic feature changes, disease correlations, and individual characteristics to improve the accuracy and applicability of cardiovascular and cerebrovascular disease classification. Summary of the Invention

[0008] The purpose of this invention is to design a method for constructing a cardiovascular and cerebrovascular disease classification model, which has significant breakthroughs in dynamic feature modeling, disease correlation modeling and personalized adaptation capabilities, and provides an efficient and accurate technical means for the classification of cardiovascular and cerebrovascular diseases.

[0009] In order to achieve the above object, the present invention provides a method for constructing a cardiovascular and cerebrovascular disease classification model, the method comprising the following steps:

[0010] Receiving dynamic signal data of a patient with cardiovascular and cerebrovascular diseases, constructing a dynamic signal matrix, extracting time series features from the dynamic signal matrix, and constructing a dynamic signal feature matrix of the patient;

[0011] Collecting the global pathological features of patients with cardiovascular and cerebrovascular diseases as static feature vectors, and unifying the time dimension of the static feature vectors with the dynamic signal feature matrix to construct a fusion feature matrix;

[0012] Construct a comorbidity association network, take the fused feature matrix as input, and then design a feature propagation mechanism based on graph neural network to perform initial correlation feature mapping between the fused feature matrix of cardiovascular and cerebrovascular disease patients and the comorbidity association network. When node features are input, they are combined with the initial correlation feature mapping of time segments in the comorbidity association network to generate an enhanced feature matrix.

[0013] When grouping patients, the similarity between the features of each time segment and the center of the group features is calculated to perform individual-to-group similarity mapping to generate patient group features. The patient group features are aligned with the patient feature center matrix. The enhanced feature matrix representation of each time segment is optimized through individual-group mapping to generate the final optimized feature matrix.

[0014] A deep classifier for cardiovascular and cerebrovascular disease classification is constructed. The optimized feature matrix is context-aggregated to generate a global feature representation. The global feature representation is used as input and input into the deep classifier for classification. The classifier ultimately outputs a classification result, which is the probability distribution of cardiovascular and cerebrovascular disease grading for each patient.

[0015] Furthermore, the size of the dynamic signal matrix is n×m, representing n time segments, each segment containing an m-dimensional signal;

[0016] The process of extracting time series features from the dynamic signal matrix and constructing the patient's dynamic signal feature matrix specifically includes:

[0017] According to the set time window length, the dynamic signal matrix is segmented by row, and each signal vector s i represents the i-th time segment;

[0018] For each signal vector s i , use a one-dimensional convolution operation with a convolution kernel size of k×m to extract the local feature vector within the time window, where the one-dimensional convolution operation is performed on n time segments one by one to generate a local feature matrix with a size of n×k, where k is the output dimension of the convolution kernel;

[0019] The local feature matrix is used as input to construct a time series context capture module. The global attention mechanism is used to extract the global features of the time series to construct a dynamic signal feature matrix after context processing. The size is still n×k, which contains the time series dependencies of each signal segment.

[0020] Furthermore, when a specific time segment, i.e., each signal vector, is calculated using the time series context capture module, the calculation is as follows:

[0021]

[0022] Among them, f i ′ is the global feature vector of the i-th time period after context capture, with a size of k; f j is the global feature vector of a specific time segment; n is the total number of time segments; α ij represents the importance weight of time segment j to time segment i, which is calculated by the following formula:

[0023]

[0024] Among them, α ij is the normalized attention weight, and ensures W t is the time-dependent weight matrix, which is used to calculate the relationship weights between time segments; is f j The transpose of

[0025] By weighted summing of features of all time segments, the context feature f i 'Integrate local and global information.

[0026] Furthermore, the time dimension of the static feature vector and the dynamic signal feature matrix are unified to construct a joint feature matrix, specifically:

[0027] Expand the static eigenvectors to the same time dimension as the dynamic signal eigenmatrix:

[0028]

[0029] Among them, the static eigenvector F static The size of ′ is n×d, and each row is consistent with the dynamic signal feature matrix F dyn ′’s time segment, the dynamic feature F dyn ′ and static characteristics F static ' is spliced in the column dimension to form a joint feature matrix F fusion , the size is n×(k+d), and each row of it is a joint feature F fusion [i,:] contains the dynamic signal features and global static features of time segment i;

[0030] When abnormal fluctuation time of cardiovascular and cerebrovascular diseases is found, the time weighted vector is introduced to the joint feature matrix F fusion Each row of is weighted:

[0031]

[0032] in, is the fusion feature matrix, γ i is the weight of time segment i, indicating its importance to the overall classification task; γ i The value of the dynamic characteristic F dyn ′ is calculated by the attention weight, and normalized to

[0033] When the received dynamic features are affected by abnormal signals, the distribution regularization term is introduced to reduce the deviation between the overall distribution of the time segment features and the mean, reduce the interference of the abnormal time segment, and make the joint feature matrix representation stable, which can be expressed as:

[0034]

[0035] in, is a regularization function used to constrain the distribution consistency of the feature matrix; is the mean of feature dimension j; λ is the regularization strength hyperparameter used to adjust the strength of distribution alignment.

[0036] Furthermore, the comorbidity association network is represented as G disease =(V,E) is a graph, V represents the nodes of the disease set, E represents the edges of the comorbidity relationship, and the edge weight e ij ∈E represents the probability of co-morbidity of diseases i and j, H disease ∈R |V|×p It is the node feature matrix, which is the disease feature. Each row contains the disease v i∈V feature representation.

[0037] Furthermore, the fusion feature matrix is projected into the feature space of the comorbidity association network, which is expressed as:

[0038]

[0039] Among them, F disease ∈R n×p is the time segment feature matrix after mapping, each row represents the initial representation of time segment i in the comorbidity feature space; W p ∈R (k+d)×p is the mapping weight matrix, which learns to project the joint features into the disease feature space;

[0040] When disease characteristics appear, the disease characteristics H disease Perform feature propagation and compare it with the time segment feature matrix F disease Combined, the spread is as follows:

[0041] H disease ′=σ(A·H disease W g )

[0042] Where A∈R |V|×|V| is the normalized adjacency matrix of disease nodes, capturing the comorbidity relationship between diseases; W g ∈R p×p is the weight matrix of graph propagation; σ(·) is the nonlinear activation function;

[0043] Then, the disease characteristics after transmission H disease ′ and time segment feature F disease Fusion to generate enhanced feature matrix:

[0044]

[0045] Where N(i) represents the set of disease nodes associated with time segment i.

[0046] Furthermore, the mapping weight α is constructed by calculating the similarity between the features of each time segment and the center of the group features. ij , expressed as:

[0047]

[0048] Among them, α ij Represents the similarity weight of time segment i mapped to group j, satisfying φ(F enhanced [i,:],G group [j,:]) is a distance function used to calculate the feature difference between time segment i and group j; Ggroup [k,:] represents the feature center of group k;

[0049] According to the mapping weight α ij , by weighted summation of group features, the optimized feature representation of time segment i is generated:

[0050]

[0051] Among them, F optimized [i,:] is the optimized feature representation of time segment i, combined with the individual features F enhanced [i,:] and its distribution in the patient population space; α ij represents the mapping weight of time segment i to group j; G group [j,:] represents the characteristic center of group j;

[0052] The feature distribution alignment constraint is introduced to align the patient group features with the patient feature center matrix, which is expressed as:

[0053]

[0054] Among them, ||F optimized [i,:]-G group [j,:]|| 2 is the characteristic distance between the optimized feature and the group center; d It is a distribution alignment regularization strength hyperparameter used to adjust the consistency of the optimized features with the group feature distribution.

[0055] Furthermore, the distance function φ(F enhanced [i,:],G group [j,:]) is defined as:

[0056] φ(F enhanced [i,:],G group [j,:])=||F enhanced [i,:]-G group [j,:]|| 2 +λ r ||G group [j,:]||1

[0057] Among them, ||F enhanced [i,:]-G group [j,:]|| 2 is the Euclidean distance between the characteristics of time segment i and group j; ||G group [j,:]||1 is the sparse regularization term of group characteristics; λ r A hyperparameter that adjusts the strength of the sparsity regularization to ensure efficient representation of group features.

[0058] Furthermore, context aggregation is performed on the time segment features to generate a global feature representation F global , which is calculated as follows:

[0059]

[0060] Among them, F global represents the global features of the patient, with a size of 1×p; ω i represents the weight of time segment i, reflecting its importance in the overall classification; F optimized [i,:] represents the optimized feature matrix; ω i The calculation of is based on the attention distribution of time segment features:

[0061]

[0062] Where w∈R p is a trainable weight vector that captures the importance of a temporal segment.

[0063] Furthermore, the classifier consists of two nonlinear feature conversion layers and a final classification layer:

[0064] Among them, the first layer reduces the dimension of the input features and introduces a regularization term that penalizes abnormal distributions:

[0065] Z1=σ(W1F global +b1)+λ r ||F global ||1

[0066] Where Z1∈R h is the output feature of the first layer, with a size of 1×h, where h is the feature dimension after dimensionality reduction; W1∈R h×p is the weight matrix of the first layer, b1∈R h is the bias vector; λ r ||F global ||1 is a regularization term that suppresses abnormally large distribution deviations in global features; σ(·) is a nonlinear activation function;

[0067] The second layer performs further feature compression and generates a probability distribution of the classification results:

[0068] Y=softmax(W2Z1+b2)

[0069] Where Y∈R c is the probability distribution vector of the classification result, c represents the number of categories; W2∈R c×h is the weight matrix of the second layer, b2∈R c is the bias vector; the softmax function converts the features into classification probabilities;

[0070] The objective function of the classifier is weighted cross entropy loss, which optimizes the classifier parameters by combining sample distribution and time importance:

[0071]

[0072] in, is the loss function of the classifier; w c [j] is the sample weight of category j, which is used to balance the proportion of samples of different categories; I(y i =j) is an indicator function that takes the value 1 when sample i belongs to category j; Y[i,j] is the predicted probability that sample i is classified as category j.

[0073] The beneficial technical effects of the present invention are at least as follows:

[0074] First, they innovatively introduce a dynamic segmented feature modeling approach, segmenting electrocardiogram (ECG) and ambulatory blood pressure (ABPM) data into time windows. They then use a convolutional neural network (CNN) to extract local features from each segment, combined with a Transformer to capture contextual dependencies over long time series. This approach accurately extracts the time series features of dynamic signals, overcoming the inadequacy of existing technologies in modeling dynamic changes.

[0075] Secondly, to address the existing problem of neglecting the correlation between diseases, this paper constructs a comorbidity association network based on actual clinical data (such as the potential correlation between diabetes and cardiovascular and cerebrovascular diseases). It then propagates the features in the comorbidity network through a graph neural network (GNN), enhancing the classification model's ability to model complex disease relationships. Compared to directly using comorbidity features as input, this method can effectively capture the potential dependencies between comorbidities, improving classification accuracy and generalization.

[0076] Furthermore, to address the problem of insufficient classification adaptability for different populations (e.g., elderly and young patients), this paper introduces population feature generation and individual-population alignment methods. By generating population feature labels through cluster analysis and embedding individual patient features into the population feature space, the classification model's adaptability to specific populations is optimized, addressing the problem of existing models' insufficient utilization of personalized features.

[0077] Finally, the present invention integrates dynamic signal features, static pathological features, and comorbidity enhancement features through a unified feature fusion framework. The weights of these features are dynamically adjusted through an attention mechanism. This cross-modal fusion approach fully exploits the information redundancy of multidimensional features, addressing the inadequate multimodal integration capabilities of existing classification techniques. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] The present invention is further described with reference to the accompanying drawings. However, the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative effort.

[0079] Figure 1 This is a flow chart of a method for constructing a cardiovascular and cerebrovascular disease classification model according to the present invention. DETAILED DESCRIPTION

[0080] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0081] In one or more embodiments, Figure 1 As shown, a method for constructing a cardiovascular and cerebrovascular disease classification model is disclosed, and the method includes the following steps:

[0082] 101. Receive dynamic signal data of a patient with cardiovascular and cerebrovascular diseases, construct a dynamic signal matrix, extract time series features from the dynamic signal matrix, and construct a dynamic signal feature matrix for the patient.

[0083] Specifically, the dynamic signal data matrix S is received dyn , where S dyn The size of is n×m, representing n time segments, each of which contains an m-dimensional signal (such as the sampling points of an electrocardiogram or ambulatory blood pressure). The data has been normalized and denoised.

[0084] Furthermore, according to the set time window length T w , the signal matrix S dyn Time segmentation is performed by row, and each signal vector s i Represents the i-th time segment:

[0085] s i =[s i1 ,s i2 ,…,s iTw ]

[0086] Among them, s i is the signal vector of the i-th time segment. w Indicates the length of the time window (e.g. the number of sampling points corresponding to 5 seconds). The segmented signal segments form a set {s1, s2, ..., s n}.

[0087] Among them, for each time segment s i, use a one-dimensional convolution operation with a convolution kernel size of k×m to extract the local feature vector f within the time window i , and its calculation formula is:

[0088] f i =σ(W c s i +b c )

[0089] Among them, f i W is the local feature vector extracted in the i-th time period, with a size of k. c is the convolution weight matrix of size k×m, which is used to capture the signal pattern within the time window. c is a bias term with a size of k × 1. σ(·) is a nonlinear activation function (such as ReLU), which introduces nonlinear capabilities to enhance feature expression.

[0090] Among them, the convolution operation is performed on n time segments one by one to generate the local feature matrix F dyn , of size n×k, where k is the output dimension of the convolution kernel.

[0091] Furthermore, F dyn As input, a time series context capture module (TSCM) is constructed to extract the global features of the time series using the global attention mechanism. i The context feature f i ' is calculated by the following formula:

[0092]

[0093] Among them, f i ′ is the global feature vector of the i-th time period after context capture, with a size of k. ij represents the importance weight of time segment j to time segment i, which is calculated by the following formula:

[0094]

[0095] Among them, α ij is the normalized attention weight, ensuring W t is a time-dependent weight matrix used to calculate the relationship weights between time segments. By weighted summing the features of all time segments, the context feature f i ′ combines local and global information.

[0096] Finally, the dynamic feature matrix F after context processing dyn ′, the size is still n×k, including the time series dependency of each signal segment. dyn′ is the output of this step and is provided to the next step of static feature fusion. dyn ′ not only retains the temporal local features but also integrates the global contextual relationships, providing rich dynamic feature expressions for subsequent steps.

[0097] 102. Collect global pathological features of patients with cardiovascular and cerebrovascular diseases as static feature vectors, and unify the time dimension of the static feature vectors with the dynamic signal feature matrix to construct a fusion feature matrix.

[0098] Specifically, the output dynamic signal feature matrix F of step 1 is received dyn ′, of size n×k, where n represents the number of time segments and k represents the feature dimension of each segment; and the static feature vector F static , of size 1×d, contains the patient's global pathological characteristics (such as blood lipid levels, genetic factors, etc.). Static feature F static is a time-invariant global representation.

[0099] The goal of this step is to transform the dynamic feature matrix F dyn ′ and static characteristics F static Together, they form a temporal and global feature representation, providing a unified input for subsequent comorbidity relationship enhancement.

[0100] Furthermore, the dynamic feature F dyn Each row of ′ corresponds to a specific time segment, and the static feature F static There is no time dimension. To solve this mismatch problem, F static Expanded to F dyn 'Same time dimension:

[0101]

[0102] Among them F static The size of ′ is n×d, and each row is consistent with the dynamic feature F dyn ’ have one-to-one correspondence with the time segments.

[0103] Next, the dynamic feature F dyn ′ and static characteristics F static ' is spliced in the column dimension to form a joint feature matrix F fusion , the size is n×(k+d), each row of which is F fusion [i,:] contains the dynamic features of time segment i and the global static features.

[0104] Specifically, in cardiovascular data, different time segments have different importance for disease classification. For example, some abnormal peaks of dynamic blood pressure may be more critical, while other time periods may be relatively stable and irrelevant. Therefore, a time weighting vector γ is introduced to the joint feature matrix Ffusion Each row of is weighted:

[0105]

[0106] Among them, γ i is the weight of time segment i, indicating its importance to the overall classification task; γ i The value of the dynamic characteristic F dyn ′ is calculated by the attention weight, and normalized to Temporal weighted dynamic adjustment can enhance the contribution of important time segments and reduce the interference of minor time segments, thereby optimizing the expressive power of joint features.

[0107] Furthermore, in cardiovascular data, dynamic features may be affected by abnormal signals (such as noise or equipment errors), causing the features of a specific time segment to deviate from the overall distribution. Based on this, the following distribution regularization term is introduced:

[0108]

[0109] in, is the mean of feature dimension j; λ is a regularization hyperparameter that adjusts the strength of distribution alignment. This regularization term reduces the deviation of the overall distribution of time segment features from the mean, reducing the interference of abnormal time segments and making the joint feature representation more stable.

[0110] Output optimized feature matrix The size is n×(k+d), which contains the joint representation of dynamic features and static features, and enhances its expressive power through time weighting and distribution regularization. It is the input for the subsequent enhancement of comorbid relationships.

[0111] 103. Construct a comorbidity association network, take the fused feature matrix as input, and then design a feature propagation mechanism based on graph neural network to perform initial correlation feature mapping between the fused feature matrix of patients with cardiovascular and cerebrovascular diseases and the comorbidity association network. When there is node feature input, it will be combined with the initial correlation feature mapping of the time segment in the comorbidity association network to generate an enhanced feature matrix.

[0112] Specifically, receive the fusion feature matrix output from step 102 The size is n×(k+d), each row represents the dynamic-static joint features of a time segment, and is input into the comorbidity association network G disease ,in:

[0113] G disease =(V,E) is a graph, V represents the nodes of the disease set, E represents the edges of the comorbidity relationship, and the edge weight eij ∈E represents the probability of co-morbidity of diseases i and j;

[0114] is the node feature matrix, each row contains the disease v i ∈V feature representation (such as epidemiological statistics and molecular biology indicators).

[0115] Furthermore, in order to associate the patient time series features with the comorbidity network, a feature propagation mechanism based on graph neural network (GNN) is designed. First, the fusion feature matrix Projected into the feature space of the comorbidity network, the initial correlation feature mapping is performed using the following formula:

[0116]

[0117] Among them, F disease ∈R n×p is the time segment feature matrix after mapping, each row represents the initial representation of time segment i in the comorbidity feature space; W p ∈R (k+d)×p is the mapping weight matrix that learns to project the joint features into the disease feature space.

[0118] Furthermore, in the comorbidity network G disease On the disease characteristics H disease Perform feature propagation and map the feature F of the time segment disease Combined. The propagation formula is as follows:

[0119] H disease ′=σ(A·H disease W g )

[0120] Where A∈R |V|×|V| is the normalized adjacency matrix of disease nodes, capturing the comorbidity relationship between diseases; W g ∈R p×p is the weight matrix of graph propagation; σ(·) is the nonlinear activation function (such as ReLU).

[0121] Then, the disease characteristics after transmission H disease ′ and time segment feature F disease Fusion to generate enhanced feature matrix:

[0122]

[0123] Where N(i) represents the set of disease nodes associated with time segment i.

[0124] Furthermore, the comorbidity enhancement feature may be unevenly distributed due to the heterogeneity of the comorbidity network (such as sparse nodes or strong weights of certain comorbidity relationships). To address this issue, the F enhanced Each row of is normalized, and a regularization term is introduced to suppress overfitting:

[0125]

[0126] in, is the mean of the time segment features; λ is the regularization strength hyperparameter used to adjust the global feature consistency.

[0127] Output the final feature matrix F enhanced , with a size of n×p, and each row contains a joint representation of dynamic, static and comorbidity enhancement, providing high-dimensional input for the next step of group feature optimization.

[0128] 104. When grouping patients, the similarity between the features of each time segment and the center of the group features is calculated to perform individual-to-group similarity mapping to generate patient group features. The patient group features are aligned with the patient feature center matrix, and the enhanced feature matrix representation of each time segment is optimized through individual-group mapping to generate the final optimized feature matrix.

[0129] Specifically, the patient group characteristic center matrix G group : The size is m×p, where m represents the number of predefined patient groups (such as high-risk patient group, medium-risk group, low-risk group), and each row G group [j,:] represents the feature center of group j, which contains the average feature representation of the group.

[0130] Goal: To transform the time segment feature F enhanced and group characteristics G group Alignment is performed and the feature representation of each time segment is optimized through individual-group mapping to generate a feature matrix that is more group-adaptive.

[0131] Specifically, given the complexity of cardiovascular and cerebrovascular diseases, different patients may belong to different group feature spaces (e.g., patients with hypertension comorbidity may be closer to a specific high-risk group). To this end, the mapping weight α is constructed by calculating the similarity between the features of each time segment and the center of the group feature. ij , the formula is as follows:

[0132]

[0133] Among them, α ij Represents the similarity weight of time segment i mapped to group j, satisfying φ(F enhanced [i,:],G group[j,:]) is a distance function used to calculate the feature difference between time segment i and group j, defined as:

[0134] φ(F enhanced [i,:],G group [j,:])=||F enhanced [i,:]-G group [j,:]|| 2 +λ r ||G group [j,:]||1

[0135] Among them, ||F enhanced [i,:]-G group [j,:]|| 2 is the Euclidean distance between the characteristics of time segment i and group j; ||G group [j,:]||1 is the sparse regularization term of group characteristics; λ r A hyperparameter that adjusts the strength of the sparsity regularization to ensure efficient representation of group features.

[0136] Furthermore, according to the mapping weight α ij , by weighted summation of group features, the optimized feature representation of time segment i is generated:

[0137]

[0138] Among them, F optimized [i,:] is the optimized feature representation of time segment i, combined with the individual features F enhanced [i,:] and its distribution in the patient population space; α ij G represents the mapping weight of time segment i to group j; group [j,:] represents the characteristic center of group j.

[0139] Among them, in cardiovascular time series data, time segment features usually have a certain degree of stability. optimized When there are large jumps between adjacent time segments, a temporal consistency regularization term is designed:

[0140]

[0141] Among them, ||F optimized [i+1,:]-F optimized [i,:]|| 2 is the characteristic change of adjacent time segments; c is a hyperparameter of the regularization strength, which controls the weight of the temporal consistency optimization.

[0142] To improve the consistency between time segment features and population distribution, feature distribution alignment constraints are further introduced:

[0143]

[0144] Among them, ||F optimized [i,:]-G group [j,:]|| 2 is the characteristic distance between the optimized feature and the group center; d It is a distribution alignment regularization strength hyperparameter used to adjust the consistency of the optimized features with the group feature distribution.

[0145] Output the final optimized feature matrix F optimized , of size n×p, with each row representing the individual-group optimized feature representation of a time segment. This matrix combines individual features, group feature distribution, and time segment consistency optimization to form high-quality features that are more suitable for classification tasks, providing reliable input for the next step of classifier design.

[0146] 105. Construct a deep classifier for cardiovascular and cerebrovascular disease classification, perform context aggregation on the optimized feature matrix to generate a global feature representation, use the global feature representation as input, and input it into the deep classifier for classification. The classifier finally outputs the classification result, which is the probability distribution of cardiovascular and cerebrovascular disease grading for each patient.

[0147] Specifically, the time segment F optimized [i,:] has time series information, but the classification task requires judging the overall status of the patient. To this end, we first perform context aggregation on the time segment features to generate a global feature representation F global , which is calculated as follows:

[0148]

[0149] Among them, F global Represents the global features of the patient, with a size of 1×p; ω i represents the weight of time segment i, reflecting its importance in the overall classification; ω i The calculation of is based on the attention distribution of time segment features:

[0150]

[0151] Where w∈R p is a trainable weight vector that captures the importance of a temporal segment.

[0152] Furthermore, the global feature F global Input classifier, which consists of two layers of nonlinear feature conversion layers and a final classification layer:

[0153] The first layer reduces the dimensionality of the input features and introduces a regularization term that penalizes abnormal distributions:

[0154] Z1=σ(W1F global +b1)+λ r ||F global ||1

[0155] Where Z1∈R h is the output feature of the first layer, with a size of 1×h, where h is the feature dimension after dimensionality reduction; W1∈R h×p is the weight matrix of the first layer, is the bias vector; λ r ||F global ||1 is a regularization term that suppresses abnormally large distribution deviations in global features; σ(·) is a nonlinear activation function (such as ReLU).

[0156] The second layer performs further feature compression and generates a probability distribution of the classification results:

[0157] Y=softmax(W2Z1+b2)

[0158] Where Y∈R c is the probability distribution vector of the classification result, c represents the number of categories (such as cardiovascular and cerebrovascular disease classification as mild, moderate, and severe); W2∈R c×h is the weight matrix of the second layer, b2∈R c is the bias vector; the softmax function converts features into classification probabilities.

[0159] Furthermore, the objective function of classification is defined as weighted cross entropy loss, and the classifier parameters are optimized in combination with sample distribution and time importance:

[0160]

[0161] in, is the loss function of the classifier; w c [j] is the sample weight of category j, which is used to balance the proportion of samples of different categories; is an indicator function that takes the value 1 when sample i belongs to category j; Y[i,j] is the predicted probability that sample i is classified as category j.

[0162] Finally, the classifier outputs the classification result Y, which is the probability distribution of each patient's cardiovascular and cerebrovascular disease classification (such as mild, moderate, or severe). By selecting the category label corresponding to the maximum probability, the final diagnosis of the patient's condition is completed.

[0163] The above disclosures are merely some preferred embodiments of the present invention, and certainly cannot be used to limit the scope of the present invention. A person skilled in the art will understand that all or part of the processes of the above embodiments and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A method for constructing a cardiovascular and cerebrovascular disease classification model, characterized in that: The method comprises the following steps: Receiving dynamic signal data of a patient with cardiovascular and cerebrovascular diseases, constructing a dynamic signal matrix, extracting time series features from the dynamic signal matrix, and constructing a dynamic signal feature matrix of the patient; Collecting the global pathological features of patients with cardiovascular and cerebrovascular diseases as static feature vectors, and unifying the time dimension of the static feature vectors with the dynamic signal feature matrix to construct a fusion feature matrix; Construct a comorbidity association network, take the fused feature matrix as input, and then design a feature propagation mechanism based on graph neural network to perform initial correlation feature mapping between the fused feature matrix of cardiovascular and cerebrovascular disease patients and the comorbidity association network. When node features are input, they are combined with the initial correlation feature mapping of time segments in the comorbidity association network to generate an enhanced feature matrix. When grouping patients, the similarity between the features of each time segment and the center of the group features is calculated to perform individual-to-group similarity mapping to generate patient group features. The patient group features are aligned with the patient feature center matrix. The enhanced feature matrix representation of each time segment is optimized through individual-group mapping to generate the final optimized feature matrix. A deep classifier for cardiovascular and cerebrovascular disease classification is constructed. The optimized feature matrix is context-aggregated to generate a global feature representation. The global feature representation is used as input and input into the deep classifier for classification. The classifier ultimately outputs a classification result, which is the probability distribution of cardiovascular and cerebrovascular disease grading for each patient.

2. The method for constructing a cardiovascular and cerebrovascular disease classification model according to claim 1, characterized in that: The size of the dynamic signal matrix is n×m, representing n time segments, each segment containing an m-dimensional signal; The process of extracting time series features from the dynamic signal matrix and constructing the patient's dynamic signal feature matrix specifically includes: According to the set time window length, the dynamic signal matrix is segmented by row, and each signal vector s i represents the i-th time segment; For each signal vector s i , use a one-dimensional convolution operation with a convolution kernel size of k×m to extract the local feature vector within the time window, where the one-dimensional convolution operation is performed on n time segments one by one to generate a local feature matrix with a size of n×k, where k is the output dimension of the convolution kernel; The local feature matrix is used as input to construct a time series context capture module. The global attention mechanism is used to extract the global features of the time series to construct a dynamic signal feature matrix after context processing. The size is still n×k, which contains the time series dependencies of each signal segment.

3. The method for constructing a cardiovascular and cerebrovascular disease classification model according to claim 2, characterized in that: When a specific time segment, that is, each signal vector, is calculated using the time series context capture module, the calculation is as follows: Among them, f i ′ is the global feature vector of the i-th time period after context capture, with a size of k; f j is the global feature vector of a specific time segment; n is the total number of time segments; α ij represents the importance weight of time segment j to time segment i, which is calculated by the following formula: Among them, α ij is the normalized attention weight, and ensures W t is the time-dependent weight matrix, which is used to calculate the relationship weights between time segments; f i T is f j The transpose of By weighted summing of features of all time segments, the context feature f i 'Integrate local and global information.

4. The method for constructing a cardiovascular and cerebrovascular disease classification model according to claim 1, characterized in that: The time dimension of the static feature vector and the dynamic signal feature matrix are unified to construct a joint feature matrix, specifically: Expand the static eigenvectors to the same time dimension as the dynamic signal eigenmatrix: Among them, the static eigenvector F static The size of ′ is n×d, and each row is consistent with the dynamic signal feature matrix F dyn ′’s time segment, the dynamic feature F dyn ′ and static characteristics F static ' is spliced in the column dimension to form a joint feature matrix F fusiom , the size is n×(k+d), and each row of it is a joint feature F fusion [i,:] contains the dynamic signal features and global static features of time segment i; When abnormal fluctuation time of cardiovascular and cerebrovascular diseases is found, the time weighted vector is introduced to the joint feature matrix F fusion Each row of is weighted: in, is the fusion feature matrix, γ i is the weight of time segment i, indicating its importance to the overall classification task; γ i The value of the dynamic characteristic F dyn ′ is calculated by the attention weight, and normalized to When the received dynamic features are affected by abnormal signals, the distribution regularization term is introduced to reduce the deviation between the overall distribution of the time segment features and the mean, reduce the interference of the abnormal time segment, and make the joint feature matrix representation stable, which can be expressed as: in, is a regularization function used to constrain the distribution consistency of the feature matrix; is the mean of feature dimension j; λ is the regularization strength hyperparameter used to adjust the strength of distribution alignment.

5. The method for constructing a cardiovascular and cerebrovascular disease classification model according to claim 1, characterized in that: The comorbidity association network is represented by G disease =(V,E) is a graph, V represents the nodes of the disease set, E represents the edges of the comorbidity relationship, and the edge weight e ij ∈E represents the probability of co-morbidity of diseases i and j, H disease ∈R |V|×p It is the node feature matrix, which is the disease feature. Each row contains the disease v i ∈V feature representation.

6. The method for constructing a cardiovascular and cerebrovascular disease classification model according to claim 5, characterized in that: Project the fusion feature matrix into the feature space of the comorbidity association network, expressed as: Among them, F disease ∈R n×p is the time segment feature matrix after mapping, each row represents the initial representation of time segment i in the comorbidity feature space; W p ∈R (k+d)×p is the mapping weight matrix, which learns to project the joint features into the disease feature space; When disease characteristics appear, the disease characteristics H disease Perform feature propagation and compare it with the time segment feature matrix F disease Combined, the spread is as follows: H disease ′=σ(A·H disease ·W g ) Where A∈R |V|×|V| is the normalized adjacency matrix of disease nodes, capturing the comorbidity relationship between diseases; W g ∈R p×p is the weight matrix of graph propagation; σ(·) is the nonlinear activation function; Then, the disease characteristics after transmission H disease ′ and time segment feature F disease Fusion to generate enhanced feature matrix: Where N(i) represents the set of disease nodes associated with time segment i.

7. The method for constructing a cardiovascular and cerebrovascular disease classification model according to claim 1, characterized in that: The mapping weight α is constructed by calculating the similarity between the features of each time segment and the center of the group features. ij , expressed as: Among them, α ij Represents the similarity weight of time segment i mapped to group j, satisfying φ(F enhanced [i,:],G group [j,:]) is a distance function used to calculate the feature difference between time segment i and group j; G group [k,:] represents the feature center of group k; According to the mapping weight α ij , by weighted summation of group features, the optimized feature representation of time segment i is generated: Among them, F optimized [i,:] is the optimized feature representation of time segment i, combined with the individual features F enhanced [i,:] and its distribution in the patient population space; α ij G represents the mapping weight of time segment i to group j; group [j,:] represents the characteristic center of group j; The feature distribution alignment constraint is introduced to align the patient group features with the patient feature center matrix, which is expressed as: Among them, ||F optimized [i,:]-G group [j,:]|| 2 is the characteristic distance between the optimized feature and the group center; d It is a distribution alignment regularization strength hyperparameter used to adjust the consistency of the optimized features with the group feature distribution.

8. The method for constructing a cardiovascular and cerebrovascular disease classification model according to claim 7, characterized in that: The distance function φ(F enhanced [i,:],G group [j,:]) is defined as: φ(F enhanced [i,:],G group [j,:])=||F enhanced [i,:]-G group [j,:]|| 2 +λ r ||G group [j,:]||1 Among them, ||F enhanced [i,:]-G group [j,:]|| 2 is the Euclidean distance between the characteristics of time segment i and group j; ||G group [j,:]||1 is the sparse regularization term of group characteristics; λ r A hyperparameter that adjusts the strength of the sparsity regularization to ensure efficient representation of group features.

9. The method for constructing a cardiovascular and cerebrovascular disease classification model according to claim 1, characterized in that: Perform context aggregation on time segment features to generate global feature representation F global , which is calculated as follows: Among them, F global Represents the global features of the patient, with a size of 1×p; ω i represents the weight of time segment i, reflecting its importance in the overall classification; F optimized [i,:] represents the optimized feature matrix; ω i The calculation of is based on the attention distribution of time segment features: Where w∈R p is a trainable weight vector that captures the importance of a temporal segment.

10. The method for constructing a cardiovascular and cerebrovascular disease classification model according to claim 9, characterized in that: The classifier consists of two nonlinear feature conversion layers and a final classification layer composition: Among them, the first layer reduces the dimension of the input features and introduces a regularization term that penalizes abnormal distributions: Z1=σ(W1F global +b1)+λ r ||F global ||1 Where Z1∈R h is the output feature of the first layer, with a size of 1×h, where h is the feature dimension after dimensionality reduction; W1∈R h×p is the weight matrix of the first layer, b1∈R h is the bias vector; λ r ||F g1obal ||1 is a regularization term that suppresses abnormally large distribution deviations in global features; σ(·) is a nonlinear activation function; The second layer performs further feature compression and generates a probability distribution of the classification results: Y=softmax(W2Z1+b2) Where Y∈R c is the probability distribution vector of the classification result, c represents the number of categories; W2∈R c×h is the weight matrix of the second layer, b2∈R c is the bias vector; the softmax function converts the features into classification probabilities; The objective function of the classifier is weighted cross entropy loss, which optimizes the classifier parameters by combining sample distribution and time importance: in, is the loss function of the classifier; w c [j] is the sample weight of category j, which is used to balance the proportion of samples of different categories; I(y i =j) is an indicator function that takes the value 1 when sample i belongs to category j; Y[i,j] is the predicted probability that sample i is classified as category j.

Images