Feature selection method for multi-label prediction of diabetic complications

Abstract

The application discloses a feature selection method for multi-label prediction of diabetic complications, and relates to the technical fields of machine learning and medical data processing. The method comprises the following steps: performing label stratified independent weight calculation on original features and complication label data of the diabetic complications to obtain a plurality of label weight matrices of the complications; performing graph regularization optimization based on the stratified independent weight on the label weight matrices of the complications, training a multi-label classification model, reducing irrelevant and redundant features in the multi-label classification model, and obtaining an important feature subset which is important, has high discriminability and is robust; filtering original data of the diabetic complications by using the optimized multi-label classification model, calculating a feature contribution matrix, and outputting features with top three contribution values for positive prediction labels to construct features for prediction of the diabetic complications. The method can screen out key features and improve the prediction accuracy of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning and medical data processing, and particularly relates to a feature selection method for multi-label prediction of diabetic complications. BACKGROUND

[0002] In a multi-label classification task, each sample can belong to multiple labels at the same time, and the importance or correlation of different labels can be significantly different. The prediction modeling of diabetic complications belongs to such a complex task. The input data usually contains a large number of clinical indicators and multi-element correlations between the complication labels. The data is often associated with multiple complication categories at the same time, such as chronic complications such as nephropathy, retinopathy, diabetic foot, and acute complications such as diabetic ketoacidosis. Chronic complications can accompany the patient for life, and severe acute complications can endanger the patient's life in a short time. For example, diabetic ketoacidosis can also cause other complications such as cerebral edema, acute kidney injury, and renal failure.

[0003] In the prediction research of diabetic complications, there are four main research difficulties. Firstly, there are many diabetic complications, and there is a co-occurrence relationship between different diseases in the clinic, such as a co-occurrence rate of more than 60% between nephropathy and retinopathy. Traditional single-label classification models are not suitable for solving this problem, and existing deep learning models are difficult to meet the clinical explainability requirements due to the black box characteristics. Secondly, complications have different risk priority levels. Some key complications need to be set with higher prediction weights, and complications with low risk levels may dilute the model attention. Thirdly, medical data often contains a large number of high-dimensional features, including redundant or irrelevant features. These noise features will interfere with model training and weaken the model's explainability. Fourthly, the complication labels in real scenarios often rely on manual labeling, and there are noise problems such as inconsistent labeling, omission or error, which further exacerbate the bias of model learning. Therefore, from the perspective of machine learning and feature engineering, an intelligent feature selection method is needed to effectively deal with multi-label coupling, noise interference and feature redundancy, so as to improve the accuracy, robustness and explainability of the diabetic complication prediction model. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a feature selection method that can effectively screen key features and improve the discriminant performance and robustness of the model, in view of the problems of feature redundancy, label coupling and noise interference in existing multi-label prediction of diabetic complications.

[0005] To solve the above technical problems, the technical solution adopted by the present application is: a feature selection method for multi-label prediction of diabetic complications, comprising the following steps:

[0006] S1: label hierarchical independent weight calculation is performed on original features and complication label data of diabetic complications to obtain a label weight matrix of multiple complications;

[0007] S2: graph regularization optimization based on hierarchical independent weight is performed on the label weight matrix of multiple complications, a multi-label classification model is trained, irrelevant and redundant features in the multi-label classification model are reduced, and a required feature subset is obtained;

[0008] S3: the original data of diabetic complications is screened using the optimized multi-label classification model, a feature contribution matrix is calculated, and for a positive prediction label, the top three features with contribution values are output as the features for predicting diabetic complications.

[0009] The beneficial effects produced by the above technical solutions are as follows: 1) the method described in the application can evaluate the label weight ablation, strip the complication coupling effect, and quantify the label independent contribution; 2) the graph regularization mechanism guided by medical feature data is fused, and redundant feature filtering and key feature selection are simultaneously realized in the prediction process; 3) the label noise robustness mechanism is introduced, and the noise interference of mislabeled complications is automatically suppressed. Thus, while efficiently completing multi-label prediction, an interpretable feature subset is output, which can improve the accuracy of model prediction. BRIEF DESCRIPTION OF DRAWINGS

[0010] The application will be further described in detail below with reference to the drawings and specific embodiments.

[0011] Figure 1 is a main flowchart of the method described in the embodiments of the application;

[0012] Figure 2 is a graph regularization multi-label feature selection algorithm flowchart in the method described in the embodiments of the application;

[0013] Figure 3 is a new sample diabetic complication prediction method flowchart in the method described in the embodiments of the application. DETAILED DESCRIPTION

[0014] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0015] In the following description, a lot of specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the concept of the present application, therefore, the present application is not limited to the specific embodiments disclosed below.

[0016] As shown in Figure 1 The embodiment of the present application discloses a feature selection method for multi-label prediction of diabetic complications, comprising the following steps:

[0017] S1: label stratified independent weight calculation is performed on the original features and complication label data of diabetic complications, and a plurality of label weight matrices of complications are obtained;

[0018] S2: graph regularization optimization based on stratified independent weight is performed on the plurality of label weight matrices of complications, a multi-label classification model is trained, irrelevant and redundant features in the multi-label classification model are reduced, and an important perception, high discriminability and robust feature subset is obtained;

[0019] S3: the original data of diabetic complications is screened using the optimized multi-label classification model, a feature contribution matrix is calculated, and for the positive prediction label, the top three features of the contribution value are output as the features of the diabetic complication prediction.

[0020] The above steps will be described in detail in combination with specific contents as follows:

[0021] 1) Label stratified independent weight calculation:

[0022] In the multi-label classification model training process, the synergistic effect between labels will significantly affect the overall classification performance of the model.

[0023] First, from the perspective of model optimization, the contribution of different labels to the performance of the model is obviously different. By introducing a label weight quantization mechanism, this importance difference can be accurately described. This differentiated processing can improve the performance of the model on key labels, which meets the needs of the possible distinction between core labels and auxiliary labels in actual applications. Secondly, from the perspective of data quality, label labeling in real-world scenarios often has quality problems, especially manually generated labels may contain labeling noise, subjective bias or missing values. By giving smaller weights to such low-quality labels, the interference of the feature selection process can be effectively reduced. Finally, from the perspective of feature correlation, different labels may depend on different feature subsets. Considering label weights can help identify the complex association patterns between features and labels, so that the feature selection process can adaptively balance the needs of different labels, avoid ignoring features that are discriminative for secondary labels, and thus discover feature subsets with better generalization ability.

[0024] In the present application, a patient record is defined as a sample, which contains diabetes-related clinical indicators (such as HbA1c, eGFR, fundus image features, etc.) to form a feature vector. The complications suffered by the patient are labeled as a multi-label vector: if a certain complication such as diabetic nephropathy is diagnosed, the corresponding label is marked as 1; otherwise, it is 0. The contribution of different complication labels to the model decision is heterogeneous, which can be verified by label ablation experiments: when removing a certain label causes a significant fluctuation in model performance, it indicates that the label has a higher learning value; if there is no significant change after removal, it may imply that the label is redundant or ineffective; if the performance is improved significantly after removal, it means that the complication label is most likely a noise label and should be removed.

[0025] However, the change in model classification performance after directly removing a certain label cannot directly reflect the importance of the label, because there is a strong clinical co-occurrence between diabetes complications, and such labels have strong correlation. When any label is directly removed, the other label can still partially compensate for the loss of information, resulting in distorted independent contribution evaluation. In the present application, the recall rate based on the label is used as the performance evaluation indicator, because the medical cost of missed diagnosis of high-risk complications is much higher than that of misdiagnosis, and the recall rate can quantify the risk of missed diagnosis.

[0026] To solve the evaluation bias caused by label coupling, the present application proposes a noise-robust hierarchical independent weight calculation method. This method realizes accurate contribution quantification through the following steps:

[0027] S1-1: input the original features and complication label data.

[0028] S1-2: for each complication label i, calculate its mutual information value with other labels, for example, the mutual information value of label i and label j is calculated as follows:

[0029] ;

[0030] Construct a label similarity matrix S, then select the top 3 labels most relevant to label i, and combine their values to form a set ;

[0031] S1-3: according to the set , the data is divided into layers, and in each layer s, a complete model is trained based on the multi-label classifier MLKNN , an ablation model is trained , and the intra-layer recall rate difference is calculated as follows:

[0032] ;

[0033] S1-4: analyze the sign of the indicator difference, if 0, the complication label has predictive value, if < 0, the complication label may have potential noise information, and it is needed to further determine whether it is a noise label. If < 0 and | , it is determined that the label is a noise label, the data column corresponding to the label is deleted, and the number of labels is updated to , wherein is a user-defined label threshold, is the number of original labels, is the number of noise labels;

[0034] S1-5: Weighted aggregation: weighted average of the recall rate difference of all layers of non-noise labels: ;

[0035] S1-6: Normalization to obtain the final weight: .

[0036] 2. Hierarchical independent weight-based graph regularization optimization method:

[0037] Through the foregoing calculation, a plurality of label weight matrices of complications can be obtained. Next, modeling of the complex association between diabetes clinical feature data and labels is needed. The core of the HIW-GRMLFS (Hierarchical Independent Weight-based Graph Regularized Multi-Label Feature Selection) algorithm is to calibrate the label relationship graph through hierarchical independent weights, and to guide the graph regularization process, so that the feature selection adaptively prioritizes the needs of important labels, while suppressing the interference of noise labels and redundant associations, and finally obtains an importance-aware, high-discriminative and robust feature subset. The specific calculation process is as follows:

[0038] Input: feature matrix ;

[0039] Label matrix after removing noise ;

[0040] Weight vector ;

[0041] Specifically, it includes the following steps, and the S2 specifically includes the following steps:

[0042] S2-1: Constructing a label weight matrix ;

[0043] Wherein is a complication label weight vector calculated by hierarchical independent weights, ∈ [0,1] is the independent importance of the i-th complication label;

[0044] S2-2: Constructing a label-weighted similarity matrix , wherein ;

[0045] S2-3: Computing a degree matrix , the diagonal elements of the sum of the elements in each row;

[0046] S2-4: Generating a Laplacian matrix ;

[0047] S2-5: Optimizing the following optimization problem using a gradient descent algorithm, and the specific optimization steps are shown in steps S2-5-1 to S2-5-4:

[0048] ;

[0049] wherein is a feature importance matrix of dimension , is the feature dimension, is the number of labels; is a label matrix of dimension , representing the true label values of samples; , is the sum of the norms of each row of the matrix ; is the coefficient of the regularization term, is the coefficient of the graph regularization term; The first term of the formula minimizes the difference between the predicted label and the true label, ensuring that the model has high prediction accuracy. The second term encourages the overall weight of certain rows to approach zero, thereby automatically filtering out features that are irrelevant or redundant for label prediction. The third term causes adjacent features in the feature graph to have similar weight distributions.

[0050] The S2-5 specifically includes the following steps:

[0051] S2-5-1: Randomly initializing the matrix , setting the learning rate , and the maximum number of iterations ;

[0052] S2-5-2: Iteratively updating the matrix using a gradient descent algorithm, for the iteration number t from 1 to :

[0053] (1) Calculate the gradient: ;

[0054] wherein is a diagonal matrix with diagonal elements ;

[0055] (2) Update parameters: ;

[0056] (3) Convergence condition judgment: if , stop iteration, otherwise repeat steps (1)-(3);

[0057] S2-5-3: Calculate feature importance score , obtain feature ranking;

[0058] S2-5-4: Arrange all features in descending order, obtain the top important feature index .

[0059] Output: the optimal feature-label weight matrix of iteration; the complication label weight vector; the top important feature index.

[0060] The flow chart of the hierarchical independent weight-based graph regularization optimization method is shown in Figure 2 , which is divided into 3 stages.

[0061] Stage 1 is the calculation of complication label hierarchical independent weight. First, input the patient feature matrix and complication label matrix data, and generate the label similarity matrix by quantifying the correlation between labels through mutual information. Then, perform hierarchical ablation evaluation, and loop the operation for each label: select the 3 labels most relevant to the label to form a relevant label set; according to all value combinations of the relevant label set, divide the original data into multiple hierarchical subsets; train a complete model containing all labels on each hierarchical subset, train an ablation model removing the current label, and calculate the performance difference between the two models. According to the sign of the model performance difference and the label noise threshold, remove the noise label. Finally, fuse the weight results, weight the performance difference values according to the sample proportion of the hierarchical subsets, and normalize the difference values of all labels to generate the final label weight matrix.

[0062] Stage 2 is the process of graph regularization multi-label feature selection. First, multiply the weight matrix and the original label similarity matrix by elements to generate an importance-weighted label similarity matrix, and construct a graph Laplacian matrix based on the weighted similarity matrix. Then define the objective optimization function, i.e. minimize the label prediction error, promote the weight of irrelevant features to zero, and force the feature weight distribution to be consistent with the weighted label graph structure.

[0063] Stage 3 is an iterative optimization process, which first randomly initializes the feature importance matrix, then iteratively updates the matrix by gradient descent method until convergence, and finally calculates the importance score of each feature, sorts the features from high to low according to the score, and outputs the top-ranked feature subset.

[0064] 3. New patient sample diabetes complication prediction method

[0065] The weight vector of the diabetes complication label is obtained by a label-independent weight calculation method based on stratification, and a graph regularization optimization framework based on stratification-independent weight is used, which can train a multi-label classification model and reduce irrelevant and redundant features. Next, how to use the data of new patients to predict the possible complications of diabetes is introduced. The flow chart of the new patient sample diabetes complication prediction method is shown in Figure 3 , and the specific steps are as follows:

[0066] Input: new patient feature vector ; parameters output by the model training stage , , .

[0067] The new patient sample diabetes complication prediction method specifically includes the following steps:

[0068] S3-1, prediction step:

[0069] Predict the label score: ;

[0070] S3-2, generate prediction label:

[0071] For complication labels from to :

[0072] ;

[0073] Where is the label selection threshold, and users can set a lower value for the high-risk complication threshold to reduce the missed diagnosis rate;

[0074] S3-3, complication label associated feature explainability analysis:

[0075] Calculate the feature contribution matrix:

[0076] ;

[0077] Where is a diagonal matrix with patient feature values as diagonal elements, and each element represents the independent contribution of feature i to complication j;

[0078] For each positive prediction label j, output the top three contributing values of its features:

[0079] .

[0080] In summary, the method can screen out key features for predicting diabetic complications, thereby improving the accuracy of model prediction.

Claims

1. A feature selection method for multi-label prediction of diabetic complications, characterized in that Comprise the following steps: S1: label hierarchical independent weight calculation is carried out to original feature and complication label data of diabetic complications, and label weight matrix of multiple complications is obtained; S2: graph regularization optimization based on hierarchical independent weight is carried out to label weight matrix of multiple complications, a multi-label classification model is trained, irrelevant and redundant features in the multi-label classification model are cut down, and a required feature subset is obtained; S3: the original data of diabetic complications is screened using the optimized multi-label classification model, a feature contribution matrix is calculated, for positive prediction label, the top three features of contribution value are output as the prediction features of diabetic complications; The S1 specifically comprises the following steps: S1-1: input original feature and complication label data; S1-2: for each complication label i, the mutual information value between it and other label j is calculated: ； A label similarity matrix S is constructed, then the top 3 labels most relevant to label i are selected, and their values are combined to form a set Ci; S1-3: stratify the data according to value combination of set Ci, in each layer s, train complete model based on multi-label classifier MLKNN and train ablation model , calculate the difference in recall rate within the layer, as shown in the following formula: ； S1-4: Analyze the sign of differences in indicators, if... > 0, the complication label has predictive value; if < 0, the complication label has potential noise information and needs further judgment to determine whether it is a noise label; If < 0 and | ≥ , the label is determined as a noise label, the data column corresponding to the label is deleted, and the number of labels is updated to , wherein is a user-defined label threshold, k is the number of original labels, is the number of noise labels; S1-5: weighted summary: Recall difference-weighted average over all layers for non-noisy labels: ; S1-6: Normalization to get the maximum weight: .

2. The feature selection method for multi-label prediction of diabetic complications as claimed in claim 1 wherein, The S2 specifically comprises the following steps: The label relationship graph is calibrated through hierarchical independent weight, and the graph regularization process is guided thereby, so that feature selection adaptively gives priority to meeting the requirements of important labels, while suppressing the interference of noise labels and redundant correlations, and finally a required feature subset is obtained.

3. The feature selection method for multi-label prediction of diabetic complications as claimed in claim 1 wherein, The S2 specifically comprises the following steps: S2-1 : Constructing the tag weight matrix ; wherein is a vector of complication label weights calculated by layer-wise independent weight computation, is the independent importance of the i-th complication label; S2-2: Constructing a label-weighted similarity matrix wherein ; S2-3: Compute degree matrix , the diagonal elements of the sum of the elements of each row S2-4: generating a Laplacian matrix ; S2-5: a gradient descent algorithm is used to optimize the following optimization problem: ； wherein, is a feature importance matrix of dimension , is the feature dimension, is the number of labels; is a label matrix of dimension , represents the true label values of , is the matrix whose norm of each row is is the coefficient of the regularization term, is the coefficient of the graph regularization term; the first term of the formula minimizes the difference between the predicted labels and the true labels, the second term encourages the weights of certain rows to be overall close to zero, and the third term makes the features that are adjacent in the feature graph have similar weight distribution.

4. The feature selection method for multi-label prediction of diabetic complications as claimed in claim 3, wherein, The S2-5 specifically comprises the following steps: S2-5-1: Randomly initialize the matrix , set the learning rate , maximum number of iterations ; S2-5-2: iteratively update the matrix with gradient descent algorithm for iteration number t from 1 to : (1) Compute the gradient: ; wherein is a diagonal matrix with diagonal elements ; (2) Update parameters: ; (3) Convergence condition judgment: if then stop iteration, otherwise repeat steps (1)-(3); S2-5-3: Calculate feature importance scores obtain a feature ranking; S2-5-4: Rank all features in descending order, obtain top important feature indices .

5. The feature selection method for multi-label prediction of diabetic complications as claimed in claim 1 wherein, The S3 specifically comprises the following steps: S3-1, prediction step: Predicted label score: ; S3-2, prediction label generation: From to : complication label ； wherein, The threshold value is selected for the label, and the high-risk complication threshold value is set to a lower value to avoid the missed diagnosis rate. S3-3, complication label associated feature explainability analysis: Calculate the feature contribution matrix: ； wherein, is a diagonal matrix with patient characteristic values as diagonal elements, each element represents the independent contribution of characteristic i to complication j; For each positive prediction label j, the top three contribution value features are output: 。

Images