Feature selection method for multi-label prediction of diabetic complications

Abstract

The invention discloses a feature selection method for multi-label prediction of diabetic complications, and relates to the technical field of machine learning and medical data processing. The method comprises the following steps: carrying out label layered independent weight calculation on original characteristics of diabetic complications and complication label data to obtain a label weight matrix of a plurality of complications; performing graph regularization optimization based on layered independent weights on the label weight matrix of the multiple complications, training a multi-label classification model, reducing irrelevant and redundant features in the multi-label classification model, and obtaining feature subsets with importance perception, high discrimination and robustness; and screening the original data of the diabetic complications by using the optimized multi-label classification model, calculating a feature contribution matrix, and for positive prediction labels, outputting the first three features of contribution values for constructing the features of diabetic complication prediction. According to the method, key features can be screened out, and the accuracy of model prediction is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning and medical data processing, and in particular to a feature selection method for multi-label prediction of diabetic complications. Background Art

[0002] In multi-label classification tasks, each sample may belong to multiple labels simultaneously, and the importance or relevance of different labels may vary significantly. Predictive modeling of diabetic complications is one such complex task. Its input data typically contains a large number of multivariate associations between clinical indicators and complication labels. The data is often associated with multiple complication categories simultaneously, such as chronic complications such as nephropathy, retinopathy, and diabetic foot, as well as acute complications such as diabetic ketoacidosis. Chronic complications may accompany patients throughout their lives, and severe acute complications can endanger the patient's life in a short period of time. For example, diabetic ketoacidosis may also cause other complications such as cerebral edema, acute kidney injury, and renal failure.

[0003] There are four major research difficulties in the prediction of diabetic complications. First, there are a large number of diabetic complications, and there are clinical co-occurrence relationships between different diseases. For example, the clinical co-occurrence rate of nephropathy and retinopathy exceeds 60%. Traditional single-label classification models are not suitable for solving this problem, and existing deep learning models are difficult to meet clinical interpretability requirements due to their black-box nature. Second, complications have different risk priorities. Certain key complications require higher prediction weights, while complications with lower risk levels may dilute the model's attention. Third, medical data often contains a large number of high-dimensional features, many of which are redundant or irrelevant. These noise features will interfere with model training and weaken the model's interpretability. Fourth, complication labels in real-world scenarios often rely on manual annotation, which is subject to noise problems such as inconsistent annotations, omissions, or errors, further exacerbating the bias of model learning. Therefore, from the perspective of machine learning and feature engineering, there is an urgent need for an intelligent feature selection method that can effectively deal with multi-label coupling, noise interference, and feature redundancy to improve the accuracy, robustness, and interpretability of diabetic complication prediction models. Summary of the Invention

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

[0005] To solve the above technical problems, the technical solution adopted by the present invention is: a feature selection method for multi-label prediction of diabetic complications, comprising the following steps: S1: Perform label-stratified independent weight calculation on the original features and complication label data of diabetic complications to obtain the label weight matrix of multiple complications; S2: Perform graph regularization optimization based on hierarchical independent weights on the label weight matrices of multiple complications, train the multi-label classification model, reduce irrelevant and redundant features in the multi-label classification model, and obtain the required feature subset; S3: Use the optimized multi-label classification model to screen the original data of diabetic complications, calculate the feature contribution matrix, and for positive prediction labels, output the top three contribution values ​​of the features as the features for predicting diabetic complications.

[0006] The beneficial effects of adopting the above technical solution are as follows: 1) The method described in this application removes the complication coupling effect through label weight ablation evaluation and quantifies the independent contribution of labels; 2) It integrates a graph regularization mechanism guided by medical feature data to simultaneously filter redundant features and select key features during the prediction process; and 3) It introduces a label noise robustness mechanism to automatically suppress the noise interference of mislabeled complications. This allows for efficient multi-label prediction while outputting an interpretable feature subset, which can improve the accuracy of model predictions. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0008] Figure 1 is a main flow chart of the method according to an embodiment of the present invention; Figure 2 is a flow chart of a graph regularized multi-label feature selection algorithm in the method according to an embodiment of the present invention; Figure 3 This is a flow chart of the method for predicting diabetic complications of new samples in the method described in an embodiment of the present invention. DETAILED DESCRIPTION

[0009] The following is a clear and complete description of the technical solutions in the embodiments of the present invention, in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts are within the scope of protection of the present invention.

[0010] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0011] like Figure 1 As shown, the embodiment of the present invention discloses a feature selection method for multi-label prediction of diabetic complications, comprising the following steps: S1: Perform label-stratified independent weight calculation on the original features and complication label data of diabetic complications to obtain the label weight matrix of multiple complications; S2: Perform graph regularization optimization based on hierarchical independent weights on the label weight matrices of multiple complications, train the multi-label classification model, reduce irrelevant and redundant features in the multi-label classification model, and obtain an importance-aware, highly discriminative and robust feature subset; S3: Use the optimized multi-label classification model to screen the original data of diabetic complications, calculate the feature contribution matrix, and for positive prediction labels, output the top three contribution values ​​of the features as the features for predicting diabetic complications.

[0012] The following is a detailed description of the above steps based on the specific content: 1) Label layer independent weight calculation: During the training of multi-label classification models, the synergy between labels can significantly affect the overall classification performance of the model.

[0013] First, from the perspective of model optimization, different labels contribute significantly different amounts to model performance. By introducing a label weight quantization mechanism, this difference in importance can be accurately captured. This differentiated processing can improve the model's performance on key labels, meeting the need to distinguish between core and auxiliary labels that may exist in real applications. Second, from the perspective of data quality, label annotation in real-world scenarios often has quality issues. In particular, manually generated labels may contain annotation noise, subjective bias, or missing values. By assigning smaller weights to these low-quality labels, their interference with the feature selection process can be effectively reduced. Finally, from the perspective of feature relevance, different labels may rely on different feature subsets. Considering label weights can help identify complex association patterns between features and labels, allowing the feature selection process to adaptively balance the needs of different labels, avoid neglecting features that are discriminative for secondary labels, and thus discover feature subsets with greater generalization capabilities.

[0014] In the present invention, a patient record is defined as a sample, and the diabetes-related clinical indicators it contains (such as HbA1c, eGFR, fundus image features, etc.) constitute a feature vector. The complications suffered by the patient are marked as multi-label vectors: if a 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 model decision-making is heterogeneous, which can be verified through label ablation experiments: when removing a certain label causes a significant fluctuation in model performance, it indicates that the label has a high learning value; if the performance does not change significantly after removal, it may imply that the label is redundant or the information is invalid; when the performance is significantly improved after removal, it means that the complication label is very likely to be a noise label and should be removed.

[0015] However, the change in model classification performance after simply removing a label doesn't directly reflect the label's importance. This is because diabetic complications have strong clinical co-occurrence, and these labels are highly correlated. When any label is directly removed, another label can still partially compensate for the information loss, resulting in a distorted assessment of independent contribution. In this paper, the label-based recall rate is used as a performance evaluation metric. Because the medical cost of missed diagnosis of high-risk complications is far higher than that of misdiagnosis, the recall rate can quantify the risk of missed diagnosis.

[0016] To address the evaluation bias caused by label coupling, this paper proposes a noise-robust hierarchical independent weight calculation method. This method achieves accurate contribution quantification through the following steps: S1-1: Input original features and complication label data.

[0017] S1-2: For each complication label i, calculate the mutual information value between it and other labels. For example, the mutual information value between label i and label j is calculated as follows: ; Construct a label similarity matrix S, then select the top three labels most relevant to label i, and combine their values ​​to form a set ; S1-3: According to the collection The data is layered by the value combination, and within each layer s, a complete model is trained based on the multi-label classifier MLKNN , train the ablation model , calculate the recall difference within the layer as shown below: ; S1-4: Analyze the difference between the indicators. If > 0, the complication signature has predictive value. < 0, the complication label may contain potential noise information and needs to be further judged whether it is a noise label. < 0 and | |≥ , then the label is determined to be a noise label, the data column corresponding to the label is deleted, and the number of labels is updated to ,in For user-defined label thresholds, is the original number of labels, is the number of noise labels; S1-5: Weighted aggregation: weighted average of the recall differences of all layers of non-noise labels: ; S1-6: Normalize to get the final weight: .

[0018] 2. Graph regularization optimization method based on layered independent weights: The aforementioned calculations can yield a label weight matrix for multiple complications. Next, we need to model the complex associations between diabetes clinical feature data and labels. The core of the HIW-GRMLFS (Hierarchical Independent Weightbased Graph Regularized Multi-Label Feature Selection) algorithm is to calibrate the label relationship graph using hierarchical independent weights, and use this to guide the graph regularization process, enabling feature selection to adaptively prioritize important labels while suppressing interference from noisy labels and redundant associations. Ultimately, we obtain a feature subset that is importance-aware, highly discriminative, and robust. The specific calculation process is as follows: Input: Feature matrix ; Label matrix after removing noise ; Weight vector ; Specifically, the following steps are included, and the S2 specifically includes the following steps: S2-1: Constructing label weight matrix ; in is the complication label weight vector obtained by layered independent weight calculation, ∈ [0,1] is the independent importance of the i-th complication label; S2-2: Constructing a label-weighted similarity matrix ,in ; S2-3: Calculate the degree matrix , The diagonal elements of The sum of the elements in each row; S2-4: Generate Laplacian Matrix ; S2-5: Use the gradient descent algorithm to optimize the following optimization problem. The specific optimization steps are shown in steps S2-5-1 to S2-5-4: ; in, The dimension is The feature importance matrix of is the characteristic dimension, is the number of tags; The dimension is The label matrix represents The true label value of each sample; , is a matrix Each row Sum of norms; 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 forces the weights of some rows to approach zero as a whole, thereby automatically filtering out features that are irrelevant or redundant to label prediction; the third term ensures that adjacent features in the feature map have similar weight distributions.

[0019] The S2-5 specifically includes 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 using the gradient descent algorithm , for the number of iterations t from 1 to : (1) Calculate the gradient: ; in is a diagonal matrix whose diagonal elements are ; (2) Update parameters: ; (3) Convergence condition judgment: If it meets , then stop the iteration, otherwise repeat steps (1)-(3); S2-5-3: Calculate feature importance scores , obtain feature ranking; S2-5-4: Arrange all features in descending order to obtain the top Important feature index .

[0020] Output: Iterate the optimal feature-label weight matrix; complication label weight vector; forward An important feature index.

[0021] The flowchart of the graph regularization optimization method based on layered independent weights is as follows Figure 2 As shown, it is divided into 3 stages.

[0022] Phase 1 involves calculating independent weights for complication labels. First, the patient feature matrix and complication label matrix data are input. Mutual information is used to quantify the degree of association between labels and generate a label similarity matrix. A stratified ablation evaluation is then performed, looping through each label: the three most relevant labels are selected to form the relevant label set; the original data is divided into multiple stratified subsets based on all possible combinations of the relevant label set; a full model containing all labels is trained on each stratified subset, followed by an ablation model with the current label removed, and the performance difference between the two models is calculated. Noisy labels are removed based on the sign of the model performance difference and the label noise threshold. Finally, the weighted results are fused, and the performance difference values ​​are weighted and summarized by the sample proportions of the stratified subsets. The difference values ​​for all labels are normalized to generate the final label weight matrix.

[0023] Phase 2 is the graph-regularized multi-label feature selection process. First, the weight matrix is ​​element-wise multiplied with the original label similarity matrix to generate an importance-weighted label similarity matrix. A graph Laplacian matrix is ​​then constructed based on the weighted similarity matrix. The objective optimization function is then defined to minimize label prediction error, encourage the weights of irrelevant features to return to zero, and enforce that the feature weight distribution is consistent with the weighted label graph structure.

[0024] Phase 3 is an iterative optimization process. First, the feature importance matrix is ​​randomly initialized, and then the matrix is ​​iteratively updated through the gradient descent method until convergence. Finally, the importance score of each feature is calculated, the features are sorted from high to low according to the score, and the top-ranked feature subset is output.

[0025] 3. Prediction Method of Diabetes Complications in New Patient Samples The weight vector of the diabetic complication label is obtained by a layered label-independent weight calculation method, and a layered independent weight-based graph regularization optimization framework is used to train a multi-label classification model and reduce irrelevant and redundant features. Next, we will introduce how to use the data of new patients to predict the diabetic complications they may suffer from. The flowchart of the method for predicting diabetic complications of new patient samples is as follows: Figure 3 The specific steps are as follows: Input: New patient feature vector ; Parameters output during the model training phase , , .

[0026] The method for predicting diabetic complications of new patient samples specifically comprises the following steps: S3-1, prediction step: Predicted label scores: ; S3-2, prediction label generation: The complication labeling arrive : ; in When selecting a threshold for a label, users can set the high-risk complication threshold to a lower value to reduce the missed diagnosis rate; S3-3, Interpretability analysis of features associated with complication labels: Calculate the feature contribution matrix: ; in is a diagonal matrix with patient eigenvalues ​​as diagonal elements, each element represents the independent contribution of feature i to complication j; For each positive prediction label j, output the features of the top three contributing values: .

[0027] 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 by The steps include: S1: Perform label-stratified independent weight calculation on the original features and complication label data of diabetic complications to obtain the label weight matrix of multiple complications; S2: Perform graph regularization optimization based on hierarchical independent weights on the label weight matrices of multiple complications, train the multi-label classification model, reduce irrelevant and redundant features in the multi-label classification model, and obtain the required feature subset; S3: Use the optimized multi-label classification model to screen the original data of diabetic complications, calculate the feature contribution matrix, and for positive prediction labels, output the top three contribution values ​​of the features as the features for predicting diabetic complications.

2. The feature selection method for multi-label prediction of diabetic complications according to claim 1, wherein: The S1 specifically includes the following steps: S1-1: Input original features and complication label data; S1-2: For each complication label i, calculate the mutual information value between it and other labels j: ； Construct a label similarity matrix S, then select the top three labels most relevant to label i, and combine their values ​​to form a set Ci; S1-3: Stratify the data according to the combination of values ​​in the set Ci, and train a complete model based on the multi-label classifier MLKNN in each layer s And train the ablation model , calculate the recall difference within the layer as shown below: ； S1-4: Analyze the difference between the indicators. If > 0, the complication signature has predictive value; if < 0, the complication label may contain potential noise information and needs to be further judged whether it is a noise label; like < 0 and | | ≥ , then the label is determined to be a noise label, the data column corresponding to the label is deleted, and the number of labels is updated to ,in is the user-defined label threshold, k is the number of original labels, is the number of noise labels; S1-5: Weighted aggregation: Weighted average of the recall differences of all layers of non-noise labels: ; S1-6: Normalize to get the final weight: .

3. The feature selection method for multi-label prediction of diabetic complications according to claim 1, wherein: The S2 specifically includes the following steps: By calibrating the label relationship graph through hierarchical independent weights and guiding the graph regularization process, feature selection can adaptively prioritize the needs of important labels while suppressing the interference of noisy labels and redundant associations, and finally obtain the required feature subset.

4. The feature selection method for multi-label prediction of diabetic complications according to claim 1, wherein: The S2 specifically includes the following steps: S2-1: Constructing label weight matrix ; in is the complication label weight vector obtained by layered independent weight calculation, ∈ [0,1] is the independent importance of the i-th complication label; S2-2: Constructing a label-weighted similarity matrix ,in ; S2-3: Calculate the degree matrix , The diagonal elements of The sum of the elements in each row; S2-4: Generate Laplacian Matrix ; S2-5: Use the gradient descent algorithm to optimize the following optimization problem: ； in, The dimension is The feature importance matrix of is the characteristic dimension, is the number of tags; The dimension is The label matrix represents The true label value of each sample; , is a matrix Each row Sum of norms; 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, the second term encourages the weights of some rows to be close to zero as a whole, and the third term makes adjacent features in the feature map have similar weight distribution.

5. The feature selection method for multi-label prediction of diabetic complications according to claim 4, characterized in that: The S2-5 specifically includes 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 using the gradient descent algorithm , for the number of iterations t from 1 to : (1) Calculate the gradient: ; in is a diagonal matrix whose diagonal elements are ; (2) Update parameters: ; (3) Convergence condition judgment: If it meets , then stop the iteration, otherwise repeat steps (1)-(3); S2-5-3: Calculate feature importance scores , obtain feature ranking; S2-5-4: Arrange all features in descending order to obtain the top Important feature index .

6. The feature selection method for multi-label prediction of diabetic complications according to claim 1, wherein: The S3 specifically includes the following steps: S3-1, prediction step: Predicted label scores: ; S3-2, prediction label generation: The complication labeling arrive : ； in, When selecting a threshold for a label, users can set the high-risk complication threshold to a lower value to avoid missed diagnosis; S3-3, Interpretability analysis of features associated with complication labels: Calculate the feature contribution matrix: ； in, is a diagonal matrix with patient eigenvalues ​​as diagonal elements, each element represents the independent contribution of feature i to complication j; For each positive prediction label j, output the features of the top three contributing values: 。

Images