A graph-regularized dynamic sampling method for cardiovascular disease diagnosis

Abstract

The application discloses a graph regularization dynamic sampling method for cardiovascular disease diagnosis, and comprises the following steps: S1, based on a classical AdaBoost framework, an adaptive weight updating mechanism based on sample confidence is designed; S2, an improved dynamic undersampling and dynamic oversampling module is integrated in the iteration process of AdaBoost; S3, after dynamic undersampling and dynamic oversampling are completed, a graph regularization dimension reduction strategy is introduced to reduce the dimension of the training samples after dynamic sampling; and S4, through the above steps, a cardiovascular disease diagnosis model is established, and the graph regularization dynamic sampling for cardiovascular disease diagnosis is realized. The application overcomes the problems that the classes in the cardiovascular disease data set are unbalanced and key features are difficult to be effectively identified.

Description

Technical Field

[0001] This invention relates to a graph regularization dynamic sampling method for the diagnosis of cardiovascular diseases. Background Technology

[0002] Cardiovascular disease remains a leading cause of death worldwide. In contemporary healthcare, machine learning has become an indispensable tool, demonstrating immense potential in the diagnosis of cardiovascular diseases and providing strong data-driven support for clinical decision-making. Currently, an increasing number of studies are focusing on building robust and efficient diagnostic models for cardiovascular diseases, based on relevant indicators, to identify and intervene in disease risk.

[0003] However, the class imbalance problem in cardiovascular disease datasets severely impacts the accuracy of diagnostic models. In this context, models tend to learn from samples of the majority class (typically healthy cases), neglecting samples of the minority class. The feature redundancy problem is particularly pronounced in cardiovascular disease datasets compared to datasets from other domains. On one hand, the included physiological health features are highly correlated; on the other hand, these features themselves often have an imbalanced class distribution. Therefore, machine learning models that fail to effectively handle imbalanced data and identify key features, while achieving high accuracy in diagnosing healthy cases, often perform poorly in diagnosing diseased cases.

[0004] Data sampling, cost sensitivity, and ensemble learning are three common strategies for addressing class imbalance. Data sampling methods aim to achieve a balance in sample size, including oversampling of the minority class and undersampling of the majority class. While sampled datasets can enhance the predictive power of machine learning models to some extent, they also face several challenges. For example, oversampling can easily lead to overfitting, undersampling may result in the loss of important information, and low-quality synthetic samples can blur the decision boundaries.

[0005] Cost-sensitive methods can be implemented in two ways: the first is to set the misclassification cost for minority class samples to be greater than the misclassification cost for majority class samples, thereby mitigating the classifier's bias; the second is to modify the class contribution to the weights, allowing the model to learn basic features directly from imbalanced data. However, the cost matrix or sample contribution is difficult to determine. Furthermore, predetermined costs or contributions remain constant throughout the training process, making it impossible to dynamically integrate them with the machine learning model.

[0006] Ensemble learning methods dynamically adjust sample weights in each base learner, thereby gradually improving the model's ability to identify minority classes during iteration. Adaptive Boosting Tree (AdaBoost) is an ensemble learning algorithm widely used in class balancing tasks. Its core idea is to weight and combine multiple base learners to form a strong learner, which has been proven to have strong generalization and flexibility.

[0007] Most existing cardiovascular disease diagnostic models use two strategies to address feature redundancy: feature selection and feature extraction. Feature selection strategies mainly include filtering methods, embedded methods, and wrapper methods, while feature extraction strategies mainly include principal component analysis and linear discriminant analysis.

[0008] Filtering methods are computationally efficient, quickly selecting the most relevant features to the target variable using statistical indicators such as variance and information gain. However, many filtering methods are based on the statistical properties of individual features, without considering the relationships between features or samples. Embedded methods, such as LASSO regression and elastic network regression, base the feature selection process on model learning. Although embedded methods exhibit superior performance in linear spaces, their ability to handle nonlinear feature combinations is limited, making it difficult to fully capture complex data patterns. Wrapping methods use the performance of machine learning algorithms as evaluation criteria, optimizing model performance directly on feature subsets. However, traversing every feature combination is an NP-hard problem, making wrapping methods computationally expensive.

[0009] Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) are two of the most representative feature extraction methods. PCA is an unsupervised method that aims to find the subspace that maximizes the global variance of the projected data, thereby minimizing reconstruction error and preserving the global structure of the data. LDA is a supervised method that finds the discriminant subspace such that the projected data has maximum inter-class dispersion and minimum intra-class dispersion. However, neither PCA nor LDA considers the manifold structure of the data. A reasonable and effective feature extraction method should preserve the manifold structure of the original feature space as much as possible in the extracted low-dimensional subspace. Laplacian eigenmaps and local preserving projections are two classic graph-based manifold learning algorithms. The core idea of ​​these two methods is to construct a weighted undirected graph to characterize the neighborhood relationships between data, thereby preserving the local geometric properties of the original high-dimensional space during dimensionality reduction. Compared with traditional linear dimensionality reduction methods such as PCA and LDA, Laplacian eigenmaps and local preserving projections can more effectively capture the nonlinear manifold structure of the data.

[0010] Traditional cardiovascular disease diagnostic models typically separate class imbalance handling, feature extraction, and classification processes, making it impossible to dynamically adapt to changes in sample weights during AdaBoost iterations. Consequently, they cannot be considered as a whole to pursue optimal model performance.

[0011] Therefore, to address the above issues, a graph-regularized dynamic sampling method for cardiovascular disease diagnosis is provided. Summary of the Invention

[0012] The purpose of this invention is to overcome the shortcomings of existing methods and provide a graph regularized dynamic sampling method for the diagnosis of cardiovascular diseases, which overcomes the problems of class imbalance in cardiovascular disease datasets and the difficulty in effectively identifying key features.

[0013] The technical solution to achieve the above objectives is:

[0014] A graph-regularized dynamic sampling method for cardiovascular disease diagnosis includes:

[0015] Step S1: Based on the classic AdaBoost framework, design an adaptive weight update mechanism based on sample confidence.

[0016] Step S2: Integrate the improved dynamic undersampling and dynamic oversampling modules into the AdaBoost iteration process;

[0017] Step S3: After completing dynamic undersampling and dynamic oversampling, a graph regularization dimensionality reduction strategy is introduced to reduce the dimensionality of the dynamically sampled training samples.

[0018] Step S4: Through the above steps, establish a cardiovascular disease diagnostic model (GRDSAD) for graph regularized dynamic sampling in cardiovascular disease diagnosis.

[0019] Preferably, in step S1, the classic AdaBoost framework is used:

[0020] Let the input of AdaBoost be the training dataset. ,in , The output is a strong learner. ;

[0021] Before training AdaBoost begins, the weight distribution of the training data is first initialized. :

[0022] ;

[0023] In the formula, Indicates the first Initial weights for each sample;

[0024] Suppose AdaBoost consists of a total of It consists of individual base learners and needs to go through... Round of iteration, the first The weight of the wheel is denoted as Using a weighted distribution Learning from the training dataset, the first classification error rate is obtained by minimizing the classification error rate. Wheel's strong learning device :

[0025] ;

[0026] calculate Classification error rate on the training dataset :

[0027] ;

[0028] Based on the classification error rate calculated by the above formula, the first... Combinatorial coefficients corresponding to wheel-based learners Defined as:

[0029] ;

[0030] No. The weights of the round's dataset are updated according to the following formula:

[0031] ;

[0032] In the formula, This is a normalization factor used to normalize the weights, making them... It appears in the form of a probability distribution:

[0033] ;

[0034] Normalization factor Expressed as follows:

[0035] ;

[0036] Based on the new round of weights, the calculation of the combined coefficients of the repeated base learners, and the update of the dataset weight distribution, the final strong learner is obtained. :

[0037] ;

[0038] Strong learners are The weighted sum of the weak learners yields:

[0039] .

[0040] Preferably, in step S1, a weight update mechanism based on sample confidence is designed, including:

[0041] In each iteration, the prediction probability of the current base learner for each training sample is first calculated. For the th iteration... Predicted probability of each sample Defined as:

[0042] ;

[0043] Calculate the confidence level of each training sample based on the prediction probability of the current base learner. For the th... Confidence level of each sample Defined as:

[0044] ;

[0045] Sample confidence The training samples were measured Based on the degree of difficulty in being misclassified, samples from different confidence intervals were further divided into three groups according to the following rules:

[0046] ;

[0047] for and The samples in the groups are defined as easy samples and hard samples, respectively. The samples in a group are defined as boundary samples, and for each group, the average confidence level within the group is defined. for:

[0048] ;

[0049] In the formula, Grouping The number of samples in the sample;

[0050] Introducing correction factors for each group :

[0051] ;

[0052] In the formula, The hyperparameters for adjusting sigmoid smoothness;

[0053] Using the correction factor calculated by the above formula, the first The weights of each group are defined as follows:

[0054] ;

[0055] After normalization, we get:

[0056] .

[0057] Preferably, in step S2, the improved dynamic undersampling module includes:

[0058] After obtaining the normalized weights for each group, the model undersamples the majority class samples, calculating the target number of majority class samples in each group according to the following formula. :

[0059] ;

[0060] In the formula, For minority class samples Quantity, Indicates rounding up. The imbalance rate of the training set is defined by the following formula:

[0061] ;

[0062] In the formula, For majority class samples The quantity.

[0063] Preferably, in step S2, the improved dynamic oversampling module includes:

[0064] For minority class samples A dynamic oversampling strategy based on Gaussian mixture model clustering is adopted, assuming... Depend on The mixed data distribution is composed of a mixture of Gaussian distribution components. Represented as:

[0065] ;

[0066] In the formula, For the first The mixing weights of the components satisfy , The mean is covariance is The multivariate Gaussian distribution;

[0067] Determine the optimal number of clusters for Gaussian mixture model clustering using the Bayesian information criterion. The Bayesian Information Criterion (BIC) is defined as follows:

[0068] ;

[0069] In the formula, For the maximum likelihood estimation of the model's log-likelihood, For the number of parameters, The total number of samples;

[0070] After clustering using Gaussian mixture model, Assign them to the appropriate clusters based on the maximum a posteriori probability of Gaussian mixture model clustering. Then, new samples are synthesized in each cluster using linear interpolation. Let the clusters be... The number of samples in Its proportion in the entire minority class sample is:

[0071] ;

[0072] Then cluster Number of target samples within for:

[0073] ;

[0074] Randomly select two points in the cluster. and New sample points are generated through linear interpolation. :

[0075] .

[0076] Preferably, in step S3, after completing dynamic undersampling and dynamic oversampling, the current base learner will obtain a new dataset. A graph regularization dimensionality reduction strategy was introduced.

[0077] Define an undirected weighted graph It is used to measure the local similarity between training samples. The construction depends on supervision information:

[0078] ;

[0079] In the formula, Indicates the thermal core bandwidth parameter;

[0080] The similarity graph constructed based on the above formula The degree matrix of the constructed graph Each element is:

[0081] ;

[0082] Therefore, the standard Laplace matrix is ​​defined as:

[0083] ;

[0084] To preserve the local geometric structure of the data during dimensionality reduction while suppressing overfitting and noise, a regularization term is introduced to solve the following generalized eigenvalue problem:

[0085] ;

[0086] In the formula, This is the sampled data matrix. It is the identity matrix. For regularization parameters, Let be the projection matrix to be learned;

[0087] The optimization problem is transformed into working with the matrix:

[0088] ;

[0089] Eigenvalue decomposition, and selection of the corresponding minimum The eigenvectors form the projection matrix. ;

[0090] Finally, by mapping the original data to a low-dimensional subspace, a new feature set is obtained:

[0091] .

[0092] The beneficial effects of this invention are: GRDSAD of this invention integrates dynamic oversampling based on clustering and dynamic undersampling based on sample confidence in the iteration process of AdaBoost, as well as graph-based regularized local preserving projection to reduce the dimensionality of the sampled dataset. Unlike other mainstream Boosting variants, GRDSAD not only focuses on class balance and feature extraction, but also on how to provide clearer decision boundaries for the base learner and ensure that the low-dimensional subspace after dimensionality reduction can better preserve the manifold structure of the data. Attached Figure Description

[0093] Figure 1 This is a flowchart of a graph regularization dynamic sampling method for cardiovascular disease diagnosis according to the present invention;

[0094] Figure 2 This is a schematic diagram of the cardiovascular disease diagnostic model (GRDSAD) in this invention;

[0095] Figure 3 This is a schematic diagram of traditional undersampling;

[0096] Figure 4 This is a schematic diagram of dynamic undersampling in this invention;

[0097] Figure 5 This is a schematic diagram of traditional oversampling;

[0098] Figure 6 This is a schematic diagram of dynamic oversampling comparison in this invention. Detailed Implementation

[0099] The technical solution of the present invention will now be clearly and completely described in conjunction with the accompanying drawings. In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0100] The invention will now be further described with reference to the accompanying drawings.

[0101] like Figure 1 , 2 As shown, a graph regularized dynamic sampling method for cardiovascular disease diagnosis includes:

[0102] Step S1: Based on the classic AdaBoost framework, design an adaptive weight update mechanism based on sample confidence.

[0103] In this embodiment, the classic AdaBoost framework is used:

[0104] Let the input of AdaBoost be the training dataset. ,in , The output is a strong learner. ;

[0105] Before training AdaBoost begins, the weight distribution of the training data is first initialized. :

[0106] ;

[0107] In the formula, Indicates the first Initial weights for each sample;

[0108] Suppose AdaBoost consists of a total of It consists of individual base learners and needs to go through... Round iteration, in this invention Set to 30, the first The weight of the wheel is denoted as Using a weighted distribution Learning from the training dataset, the first classification error rate is obtained by minimizing the classification error rate. Wheel's strong learning device :

[0109] ;

[0110] calculate Classification error rate on the training dataset :

[0111] ;

[0112] Based on the classification error rate calculated by the above formula, the first... Combinatorial coefficients corresponding to wheel-based learners Defined as:

[0113] ;

[0114] No. The weights of the round's dataset are updated according to the following formula:

[0115] ;

[0116] In the formula, This is a normalization factor used to normalize the weights, making them... It appears in the form of a probability distribution:

[0117] ;

[0118] Normalization factor Expressed as follows:

[0119] ;

[0120] Based on the new round of weights, the calculation of the combined coefficients of the repeated base learners, and the update of the dataset weight distribution, the final strong learner is obtained. :

[0121] ;

[0122] Strong learners are The weighted sum of the weak learners yields:

[0123] .

[0124] In this embodiment, a weight update mechanism based on sample confidence is designed, including:

[0125] In each iteration, the prediction probability of the current base learner for each training sample is first calculated. For the th iteration... Predicted probability of each sample Defined as:

[0126] ;

[0127] Calculate the confidence level of each training sample based on the prediction probability of the current base learner. For the th... Confidence level of each sample Defined as:

[0128] ;

[0129] Sample confidence The training samples were measured Based on the degree of difficulty in being misclassified, samples from different confidence intervals were further divided into three groups according to the following rules:

[0130] ;

[0131] for and The samples in the groups are defined as easy samples and hard samples, respectively. The samples in a group are defined as boundary samples, and for each group, the average confidence level within the group is defined. for:

[0132] ;

[0133] In the formula, Grouping The number of samples in the sample;

[0134] In each iteration, the same sample may be assigned to different groups due to the different predicted probabilities of the base learner. This process is dynamic and adaptive. What remains constant is that during the base learner update process, GRDSAD always focuses on the boundary sample groups and introduces a correction coefficient for each group. :

[0135] ;

[0136] In the formula, The hyperparameters for adjusting sigmoid smoothness;

[0137] Specifically, the correction magnitude for non-boundary sample grouping increases with the number of iterations, gradually focusing on difficult samples; while the correction magnitude for boundary sample grouping is smoothed using the sigmoid function to avoid misjudgment. The correction coefficient calculated using the above formula is used for the... The weights of each group are defined as follows:

[0138] ;

[0139] After normalization, we get:

[0140] .

[0141] Step S2 integrates the improved dynamic undersampling and dynamic oversampling modules into the AdaBoost iteration process.

[0142] In this embodiment, the improved dynamic undersampling module includes:

[0143] After obtaining the normalized weights for each group, the model undersamples the majority class samples, calculating the target number of majority class samples in each group according to the following formula. :

[0144] ;

[0145] In the formula, For minority class samples Quantity, Indicates rounding up. The imbalance rate of the training set is defined by the following formula:

[0146] ;

[0147] In the formula, For majority class samples The quantity.

[0148] GRDSAD will be in the Random undersampling in each group If a majority class sample is insufficient, backsampling with replacement is allowed, and the remaining number is added to all groups. Figure 3 As shown, traditional undersampling methods arbitrarily remove majority class samples, easily leading to information loss. Dynamic undersampling methods construct differentiated objectives based on correction factors, effectively reducing the model's emphasis on non-boundary samples while enhancing its attention to boundary samples, such as... Figure 4 As shown, the dynamic undersampling method removes more majority class samples that are close to the decision boundary.

[0149] In this embodiment, the improved dynamic oversampling module includes:

[0150] First, for minority samples A dynamic oversampling strategy based on Gaussian mixture model clustering is adopted, assuming... Depend on The mixed data distribution is composed of a mixture of Gaussian distribution components. Represented as:

[0151] ;

[0152] In the formula, For the first The mixing weights of the components satisfy , The mean is covariance is The multivariate Gaussian distribution;

[0153] Determine the optimal number of clusters for Gaussian mixture model clustering using the Bayesian information criterion. The Bayesian Information Criterion (BIC) is defined as follows:

[0154] ;

[0155] In the formula, For the maximum likelihood estimation of the model's log-likelihood, For the number of parameters, The total number of samples;

[0156] After clustering using Gaussian mixture model, Assign them to the appropriate clusters based on the maximum a posteriori probability of Gaussian mixture model clustering. Then, new samples are synthesized in each cluster using linear interpolation. Let the clusters be... The number of samples in Its proportion in the entire minority class sample is:

[0157] ;

[0158] Then cluster Number of target samples within for:

[0159] ;

[0160] Randomly select two points in the cluster. and New sample points are generated through linear interpolation. :

[0161] .

[0162] like Figure 5 As shown, traditional oversampling methods linearly generate synthetic minority class samples based on distance, completely ignoring the sample distribution of the minority class; for example... Figure 6 As shown, the dynamic oversampling method clusters samples according to their distribution and generates new samples in each cluster through linear interpolation. Therefore, GRDSAD enhances minority class identification by generating high-quality synthetic samples, while reducing noise and hard instances, thus clarifying the decision boundary.

[0163] Step S3: After completing dynamic undersampling and dynamic oversampling, a graph regularization dimensionality reduction strategy is introduced to reduce the dimensionality of the dynamically sampled training samples.

[0164] In this embodiment, after completing dynamic undersampling and dynamic oversampling, the current base learner will obtain a new dataset. A graph regularization dimensionality reduction strategy was introduced.

[0165] Graph regularization is essentially based on the local preserving projection algorithm, which is a typical graph embedding method. It aims to preserve the local manifold structure of the original data by minimizing the distance between similar samples in the embedding space. In GRDSAD, this invention adds a discriminative graph construction strategy and regularization coefficients to the local preserving projection framework to guide the sample structure in the low-dimensional subspace to be more conducive to classification tasks.

[0166] Define an undirected weighted graph It is used to measure the local similarity between training samples. The construction depends on supervision information:

[0167] ;

[0168] In the formula, The graph represents the heat core bandwidth parameter. The way this graph is constructed ensures that only samples of the same type have positive similarity, thereby enhancing the discriminative power after dimensionality reduction.

[0169] The similarity graph constructed based on the above formula The degree matrix of the constructed graph Each element is:

[0170] ;

[0171] Therefore, the standard Laplace matrix is ​​defined as:

[0172] ;

[0173] To preserve the local geometric structure of the data during dimensionality reduction while suppressing overfitting and noise, a regularization term is introduced to solve the following generalized eigenvalue problem:

[0174] ;

[0175] In the formula, This is the sampled data matrix. It is the identity matrix. For regularization parameters, Let be the projection matrix to be learned;

[0176] The optimization problem is transformed into working with the matrix:

[0177] ;

[0178] Eigenvalue decomposition, and selection of the corresponding minimum The eigenvectors form the projection matrix. ;

[0179] Finally, by mapping the original data to a low-dimensional subspace, a new feature set is obtained:

[0180] .

[0181] Using graph-based regularized local preserving projection as a feature dimensionality reduction method, a similarity graph is first dynamically constructed in each base learner according to the sample weights. Then, a high-dimensional to low-dimensional feature space mapping is achieved through Laplace transform, thereby preserving key information of minority class samples in a subspace with unbalanced class distribution, and finally generating a clearer classification decision boundary.

[0182] Step S4: Through the above steps, establish a cardiovascular disease diagnostic model (GRDSAD) for graph regularized dynamic sampling in cardiovascular disease diagnosis.

[0183] In the ischemic heart disease follow-up diagnostic dataset, "7-year follow-up" refers to the dataset reviewed seven years later, and "10-year follow-up" refers to the dataset reviewed ten years later. This invention uses different random seeds to divide the training and test sets in a ratio of 70%-30%, and after five consecutive runs, records the mean and standard deviation of the geometric mean and area under the curve (ROC) metrics as evaluation results. The geometric mean refers to the geometric mean of the performance metrics that identify all positive and negative examples, and the area under the curve is defined as the area enclosed by the ROC curve and the coordinate axes.

[0184] Both the "7-year follow-up" and "10-year follow-up" datasets contain 3987 samples and 18 features. The difference lies in the imbalance rate: the "7-year follow-up" dataset has an imbalance rate of 5.34, while the "10-year follow-up" dataset has an imbalance rate of 4.51. Tables 1 and 2 present the results of the comparative and ablation experiments for these two datasets, respectively.

[0185]

[0186] Table 1. Comparative Experiment Results of Ischemic Heart Disease Datasets

[0187] As shown in Table 1, GRDSAD significantly outperforms other models on both datasets, demonstrating the most outstanding performance. Specifically, on the "7-year follow-up" dataset, GRDSAD's geometric mean index is 9.7% higher than the second-best decision tree and 13.1% higher than the average of other models. Its area under the curve (AUC) is 4.4% higher than the second-best gradient boosting tree and 6.3% higher than the average of other models. On the "10-year follow-up" dataset, GRDSAD's geometric mean index is 9.7% higher than the second-best decision tree and 13.0% higher than the average of other models. Its AUC is 5.1% higher than the second-best gradient boosting tree and 6.8% higher than the average of other models. This is because other traditional machine learning models do not consider the effects of class imbalance and feature redundancy, while GRDSAD, through dynamic mixture sampling and graph regularization, provides each base learner with a clearer decision boundary and a more discriminative low-dimensional subspace input.

[0188]

[0189] Table 2 Ablation Experiment Results of Ischemic Heart Disease Dataset

[0190] As shown in Table 2, because clustering is performed beforehand based on the density of minority class samples, the newly synthesized samples from dynamic oversampling better conform to the distribution within the clusters. Therefore, it can more effectively improve the model's ability to identify minority classes compared to random oversampling. Although dynamic undersampling does not show an advantage over random undersampling, its combination with dynamic oversampling can construct a clearer decision boundary, thereby improving the geometric mean and area under the curve metrics. After adding the graph regularization module, the model performance is significantly improved because the dimensionality reduction method that preserves the data manifold structure suppresses the overfitting caused by feature redundancy to the base learner. Although the geometric mean of GRDSAD on the "7-year follow-up" dataset is slightly lower than that of random oversampling + random undersampling, GRDSAD still performs outstandingly on both ischemic heart disease datasets overall. Therefore, the model proposed in this invention can more accurately identify disease cases in cardiovascular disease diagnosis tasks.

[0191] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A graph-regularized dynamic sampling method for cardiovascular disease diagnosis, characterized in that, include: Step S1: Based on the classic AdaBoost framework, design an adaptive weight update mechanism based on sample confidence. Step S2: Integrate the improved dynamic undersampling and dynamic oversampling modules into the AdaBoost iteration process; Step S3: After completing dynamic undersampling and dynamic oversampling, a graph regularization dimensionality reduction strategy is introduced to reduce the dimensionality of the dynamically sampled training samples. Step S4: Through the above steps, establish a cardiovascular disease diagnostic model for graph regularized dynamic sampling in cardiovascular disease diagnosis. In step S1, a weight update mechanism based on sample confidence is designed, including: In each iteration, the prediction probability of the current base learner for each training sample is first calculated. For the th iteration... Predicted probability of each sample Defined as: ； in, To base the strong learner on the first The predicted output for each sample, the strong learner : ;No. Wheel's strong learning device : ; The first base learner; Combinatorial coefficients corresponding to wheel-based learners : ; Classification error rate on the training dataset : ; AdaBoost takes the training dataset as input. ,in , ; No. The weights of the round's dataset are updated according to the following formula: ； In the formula, For normalization factor, ; Calculate the confidence level of each training sample based on the prediction probability of the current base learner. For the th... Confidence level of each sample Defined as: ； Sample confidence The training samples were measured Based on the degree of difficulty in being misclassified, samples from different confidence intervals were further divided into three groups according to the following rules: ； for and The samples in the groups are defined as easy samples and hard samples, respectively. The samples in a group are defined as boundary samples, and for each group, the average confidence level within the group is defined. for: ； In the formula, Grouping The number of samples in the sample; Introducing correction factors for each group : ； In the formula, The hyperparameters for adjusting sigmoid smoothness; Using the correction factor calculated by the above formula, the first The weights of each group are defined as follows: ； After normalization, we get: ； In step S2, the improved dynamic undersampling module includes: After obtaining the normalized weights for each group, the model undersamples the majority class samples, calculating the target number of majority class samples in each group according to the following formula. : ； In the formula, For minority class samples Quantity, Indicates rounding up. The imbalance rate of the training set is defined by the following formula: ； In the formula, For majority class samples Quantity; In step S2, the improved dynamic oversampling module includes: For minority class samples A dynamic oversampling strategy based on Gaussian mixture model clustering is adopted, assuming... Depend on The mixed data distribution is composed of a mixture of Gaussian distribution components. Represented as: ； In the formula, For the first The mixing weights of the components satisfy , The mean is The covariance is The multivariate Gaussian distribution; Determine the optimal number of clusters for Gaussian mixture model clustering using the Bayesian information criterion. The Bayesian Information Criterion (BIC) is defined as follows: ； In the formula, For the maximum likelihood estimation of the model's log-likelihood, For the number of parameters, The total number of samples; After clustering using Gaussian mixture model, Assign them to the appropriate clusters based on the maximum a posteriori probability of Gaussian mixture model clustering. Then, new samples are synthesized in each cluster using linear interpolation. Let the clusters be... The number of samples in Its proportion in the entire minority class sample is: ； Then cluster Number of target samples within for: ； Randomly select two points in the cluster. and New sample points are generated through linear interpolation. : 。 2. The graph regularization dynamic sampling method for cardiovascular disease diagnosis according to claim 1, characterized in that, In step S1, the classic AdaBoost framework is used. Let the input of AdaBoost be the training dataset. ,in , The output is a strong learner. ; Before training AdaBoost begins, the weight distribution of the training data is first initialized. : ； In the formula, Indicates the first Initial weights for each sample; Suppose AdaBoost consists of a total of It consists of individual base learners and needs to go through... Round of iteration, the first The weight of the wheel is denoted as Using a weighted distribution Learning from the training dataset, the first classification error rate is obtained by minimizing the classification error rate. Wheel's strong learning device : ； calculate Classification error rate on the training dataset : ； Based on the classification error rate calculated by the above formula, the first... Combinatorial coefficients corresponding to wheel-based learners Defined as: ； No. The weights of the round's dataset are updated according to the following formula: ； In the formula, This is a normalization factor used to normalize the weights, making them... It appears in the form of a probability distribution: ； Normalization factor Expressed as follows: ； Based on the new round of weights, the calculation of the combined coefficients of the repeated base learners, and the update of the dataset weight distribution, the final strong learner is obtained. : ； Strong learners are The weighted sum of the weak learners yields: 。 3. The graph regularization dynamic sampling method for cardiovascular disease diagnosis according to claim 1, characterized in that, In step S3, after completing dynamic undersampling and dynamic oversampling, the current base learner will obtain a new dataset. A graph regularization dimensionality reduction strategy was introduced. Define an undirected weighted graph It is used to measure the local similarity between training samples. The construction depends on supervision information: ； In the formula, Indicates the thermal core bandwidth parameter; The similarity graph constructed based on the above formula The degree matrix of the constructed graph Each element is: ； Therefore, the standard Laplace matrix is ​​defined as: ； To preserve the local geometric structure of the data during dimensionality reduction while suppressing overfitting and noise, a regularization term is introduced to solve the following generalized eigenvalue problem: ； In the formula, This is the sampled data matrix. It is the identity matrix. For regularization parameters, Let be the projection matrix to be learned; The generalized eigenvalue problem is transformed into applying matrices as follows: ； Eigenvalue decomposition, and selection of the corresponding minimum The eigenvectors form the projection matrix. ; Finally, by mapping the original data to a low-dimensional subspace, a new feature set is obtained: 。

Images