A method and system for multi-stage feature screening and construction of a breast cancer diagnosis model

Abstract

The application discloses a kind of multi-stage feature screening and construction method and system of breast cancer diagnostic model.The breast cancer transcriptome sample data is obtained, logarithmic conversion and standardization processing are carried out on expression amount, and are divided into training set and independent test set according to 7:3 proportion stratification;Difference expression preliminary screening is carried out using Mann-Whitney U nonparametric test in training set;Parallelly execute LASSO, SVM-RFE and XGBoost three kinds of feature selection, obtain three candidate feature sets;Intersection is obtained to consensus gene to three candidate feature sets, and consistency index evaluation is carried out to consensus gene, and gene with consistency index reaching preset threshold is selected as core feature set;XGBoost diagnostic model is trained with core feature set, and performance is evaluated using independent test set, and SHAP method is used to explain feature contribution.The application improves the robustness of feature screening and the interpretability of model by multi-stage complementary screening and consistency evaluation.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and medical data processing technology, specifically to a method and system for multi-stage feature screening and construction of a breast cancer diagnostic model. Background Technology

[0002] Breast cancer is one of the most common malignant tumors in women, and early and accurate diagnosis is of great significance for improving prognosis and reducing mortality. Traditional imaging examinations and pathological biopsies have shortcomings such as high subjectivity, invasiveness, and the tendency to miss early and small lesions.

[0003] With the development of high-throughput sequencing technologies such as RNA-Seq, breast cancer transcriptome data can provide a wealth of potential diagnostic biomarkers. However, this type of data typically has a feature dimensionality far exceeding the sample size, and also suffers from problems such as skewed data distribution, high levels of noise and redundant features, and imbalanced class ratios. Directly inputting high-dimensional features into classification models can easily lead to the curse of dimensionality and overfitting, resulting in insufficient model generalization ability.

[0004] Existing technologies include feature selection schemes that combine differential expression analysis with machine learning methods. Examples include directly training a classifier using all original gene features, employing a single filter-based differential analysis to select genes, or comparing the predictive performance of only a few classification models. However, these schemes still have the following shortcomings: First, directly inputting high-dimensional features into the model can easily lead to the curse of dimensionality and overfitting; second, a single feature selector cannot simultaneously cover statistical significance, linear sparsity, feature synergy, and nonlinear interaction information; third, some processes do not strictly limit feature selection to be completed only within the training set, posing a risk of data leakage; and fourth, the model results lack mechanistic explanations, making it difficult to explain the contribution of key genes to diagnostic judgment. Summary of the Invention

[0005] This invention provides a multi-stage feature screening and construction method and system for a breast cancer diagnostic model. The method includes the following steps: acquiring breast cancer transcriptome sample data; performing logarithmic transformation and standardization on the expression levels of the transcriptome sample data; dividing the processed sample data into a training set and an independent test set according to a preset ratio; and using the Mann-Whitney U nonparametric test in the training set to perform initial screening for differential expression and obtain a differentially expressed gene set.

[0006] Furthermore, using the differentially expressed gene set as input, LASSO feature selection, SVM-RFE feature selection, and XGBoost feature selection are performed in parallel to obtain the first, second, and third candidate feature sets, respectively. LASSO feature selection employs a logistic regression model with L1 regularization, optimizing the regularization strength through hierarchical 5-fold cross-validation and ROC-AUC objective to obtain the first candidate feature set. SVM-RFE feature selection is based on a linear kernel support vector machine and a recursive feature elimination mechanism, combining 5-fold cross-validation and grid search to optimize the penalty coefficient C, and setting class weights to adapt to sample imbalance, to obtain the second candidate feature set. XGBoost feature selection constructs an XGBoost model and uses stochastic hyperparameter search to optimize hyperparameters including n_estimators, max_depth, learning_rate, subsample, colsample_bytree, gamma, reg_alpha, and reg_lambda, using the model's built-in feature importance to obtain the third candidate feature set.

[0007] Furthermore, the intersection of the first candidate feature set, the second candidate feature set, and the third candidate feature set is taken to obtain the consensus gene. The consensus gene is then evaluated using a consistency index, which is the ratio of the number of times the consensus gene is selected among the three feature selection results to the total number of selections using the three methods. Consensus genes whose consistency index reaches a preset threshold are selected as the core feature set. The preset threshold for the consistency index is preferably two-thirds.

[0008] Furthermore, the diagnostic classification model is trained using the core feature set as input, and its performance is evaluated using the independent test set to output the breast cancer discrimination result. The preferred diagnostic classification model is the XGBoost model, whose hyperparameters include: subsample=0.9, reg_lambda=2, reg_alpha=0, n_estimators=50.

[0009] Furthermore, the method includes using SHAP to analyze the contribution of each feature in the core feature set to the model output, and outputting the SHAP contribution value of each feature to the discrimination result of a single sample. The sample data is randomly divided into stratified groups in a 7:3 ratio based on the health / breast cancer label, so that the sample class ratio of the training set and the independent test set is consistent with the original data; all standardized parameter calculations, differential expression analysis and feature selection are limited to the training set.

[0010] This invention also provides a breast cancer diagnostic system, comprising: a data input interface for receiving breast cancer transcriptome sample data; a preprocessing module for performing logarithmic transformation and standardization on the expression levels of the transcriptome sample data; a feature selection module for selecting a core feature set from the transcriptome sample data according to the above method; a model inference module for using the core feature set as input to call a pre-trained diagnostic classification model to output breast cancer discrimination results; and an interpretability output module for calculating and outputting the SHAP contribution value of each feature to the current sample discrimination result.

[0011] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0012] The present invention has the following beneficial effects:

[0013] First, by using the Mann-Whitney U nonparametric test for initial screening and then employing a multi-stage complementary screening approach with three feature selection methods based on different principles—LASSO, SVM-RFE, and XGBoost—the impact of single algorithm bias and redundant noise features is reduced, effectively mitigating the curse of dimensionality and overfitting problems caused by directly inputting high-dimensional features into the model.

[0014] Second, by strictly dividing the training set and the test set into hierarchical parts, all standardized parameter calculations, differential expression analysis, and feature selection are limited to the training set, avoiding the risk of data leakage and making the evaluation results of the independent test set closer to the true generalization performance.

[0015] Third, by evaluating the consistency index of consensus genes and selecting genes that reach the preset threshold as the core feature set, we ensure that the selected gene features are robust in multiple screening methods, reducing the risk of accidental misselection due to parameter settings of a single selector or data noise. From 60,659 gene expression features, COL10A1 was finally selected as the core consensus gene, which significantly reduced the number of features and made them biologically interpretable.

[0016] Fourth, the SHAP method is used to analyze the direction and magnitude of the contribution of core features to the model output, making the diagnostic criteria of the machine learning model more transparent and solving the problem of the lack of mechanism explanation for the model results.

[0017] This invention also provides a breast cancer diagnostic system, comprising: a data input interface for receiving breast cancer transcriptome sample data; a preprocessing module for performing logarithmic transformation and standardization on the expression levels of the transcriptome sample data; a feature selection module for selecting a core feature set from the transcriptome sample data according to the above method; a model inference module for using the core feature set as input to call a pre-trained diagnostic classification model to output breast cancer discrimination results; and an interpretability output module for calculating and outputting the SHAP contribution value of each feature to the current sample discrimination result.

[0018] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method. Attached Figure Description

[0019] Figure 1 This is a flowchart of the multi-stage feature screening and diagnostic modeling of the present invention;

[0020] Figure 2 A schematic diagram illustrating consensus gene screening using three feature selectors: LASSO, SVM-RFE, and XGBoost. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to illustrate the present invention and should not be regarded as limiting the scope of the present invention.

[0023] First Embodiment

[0024] This embodiment provides a multi-stage feature screening and construction method for a breast cancer diagnostic model.

[0025] Obtain transcriptome sample data for breast cancer.

[0026] The breast cancer transcriptome sample data obtained in this embodiment comes from the RNA-Seq data of the TCGA-BRCA project, including 1,118 breast cancer samples and 113 normal control samples, totaling 1,231 samples. The original expression profile contains 60,659 gene features.

[0027] Expression data preprocessing.

[0028] The expression levels of the acquired transcriptome sample data were transformed using a log2(x+1) transformation to compress the dynamic range of the data and reduce the impact of skewed distributions and extreme values. Subsequently, Z-score normalization was performed using the statistics of the training set to ensure that the gene features were at comparable levels and to avoid information leakage from the test set.

[0029] The training and test sets are divided into hierarchical groups.

[0030] Based on the health / breast cancer label, the sample data was randomly divided into a training set and an independent test set in a 7:3 ratio to maintain the class ratio of the two sets consistent with the original data. All standardized parameter calculations, differential expression analysis, and feature selection operations were strictly limited to the training set.

[0031] Initial screening for differential expression.

[0032] 7. After performing the Shapiro-Wilk normality test and the Kolmogorov-Smirnov normality test, considering that gene expression data generally do not conform to a normal distribution, the Mann-Whitney U nonparametric test was used for initial screening of differentially expressed genes. Differentially expressed genes were screened by combining the fold change threshold (|log2 Fold Change|>1) and the significance threshold (adjusted P-value<0.05). In this example, 4715 significantly differentially expressed genes were identified, including 1329 upregulated genes and 3386 downregulated genes.

[0033] LASSO feature selection.

[0034] Using differentially expressed genes from the training set as input, a logistic regression model with L1 regularization was employed for feature selection. The regularization strength parameter C was optimized using hierarchical 5-fold cross-validation and the ROC-AUC objective function to obtain the optimal regularization parameter. The sparsity property of LASSO was utilized to compress the coefficients of unimportant gene features to zero, forming the first candidate feature set. In this embodiment, the optimal C value was 0.0464, resulting in 36 candidate feature genes.

[0035] SVM-RFE feature selection.

[0036] Using differentially expressed genes from the training set as input, feature selection is performed based on a linear kernel support vector machine and a recursive feature elimination mechanism. The penalty coefficient C is optimized by combining 5-fold cross-validation and grid search, and class weights are set to accommodate the class imbalance between healthy and disease samples. A second candidate feature set is obtained by iteratively removing gene features that contribute little to the support vector margin.

[0037] XGBoost Feature Selection.

[0038] Using differentially expressed genes from the training set as input, an XGBoost model is constructed for feature selection. A stochastic hyperparameter search is employed to optimize hyperparameters including n_estimators, max_depth, learning_rate, subsample, colsample_bytree, gamma, reg_alpha, and reg_lambda. The importance of each gene feature is evaluated using the built-in gain importance metric of the XGBoost model, resulting in a third candidate feature set. This embodiment yields 50 candidate features.

[0039] Consensus gene identification and consensus index assessment.

[0040] The intersection of the first candidate feature set obtained by LASSO feature selection, the second candidate feature set obtained by SVM-RFE feature selection, and the third candidate feature set obtained by XGBoost feature selection is used to obtain the consensus gene that simultaneously satisfies linear sparsity, support vector margin contribution, and nonlinear tree model importance.

[0041] Consensus genes are evaluated using a consistency index. The consistency index is defined as the ratio of the number of times a consensus gene is selected among the three feature selection methods to the total number of selections using the three methods. For example, if a gene is selected in all three selectors (LASSO, SVM-RFE, and XGBoost), its consistency index is 1, indicating that the gene is a strong consensus gene; if it is selected only in two selectors, its consistency index is two-thirds, indicating that the gene is a moderate consensus gene.

[0042] Consensus genes whose consistency index reaches a preset threshold are selected as the core feature set. In this embodiment, the preset threshold is two-thirds, meaning that consensus genes selected in at least two feature selection methods are selected as the core feature set. This screening mechanism ensures that the selected gene features are robust across methods, reducing the risk of accidental omissions or misselections due to parameter settings of a single selector or data noise. In this embodiment, COL10A1 was ultimately determined as the core consensus gene.

[0043] Training and evaluation of diagnostic classification models.

[0044] Using the core feature set as input, four classification models were trained respectively: XGBoost, Logistic Regression (with L2 regularization), Random Forest, and Support Vector Machine (RBF kernel).

[0045] The preferred hyperparameters for the XGBoost model include: subsample=0.9, reg_lambda=2, reg_alpha=0, n_estimators=50; other tree structure parameters and learning rate, etc., can be determined based on cross-validation of the training set or random search results.

[0046] Accuracy, precision, recall, F1 score, Matthews correlation coefficient, and ROC-AUC of each model were calculated using an independent test set. The model with the best overall performance was selected as the final diagnostic classification model. In this embodiment, the XGBoost model performed best, with an accuracy of 0.9703, precision of 1, recall of 0.9673, F1 score of 0.9834, Matthews correlation coefficient of 0.8549, and ROC-AUC of 0.9915 on the independent test set. The confusion matrix showed that all 34 healthy samples were correctly classified, and 325 out of 336 disease samples were correctly classified.

[0047] Model interpretability analysis.

[0048] The SHAP method was used to analyze the contribution of each feature in the core feature set to diagnostic discrimination. The SHAP value of each core gene feature was calculated for each sample in the training set, quantifying the direction and magnitude of each feature's contribution to the model output. SHAP analysis showed that high expression of COL10A1 promoted the model's output of a positive breast cancer diagnosis, while low expression promoted a negative healthy diagnosis. The relationship between core genes and breast cancer development, tumor-specific expression, and potential clinical translational value was explained by combining literature and functional annotations.

[0049] The output of this embodiment includes: breast cancer / health classification results; core consensus gene COL10A1; SHAP contribution value of COL10A1 to the classification results of a single sample; model performance evaluation metrics, including accuracy, precision, recall, F1 score, Matthews correlation coefficient, and ROC-AUC.

[0050] Figure 1 This is a flowchart of the multi-stage feature selection and diagnostic modeling process in this embodiment. Figure 1 As shown, the entire process starts with the input of transcriptome data, and proceeds sequentially through data preprocessing, training / testing hierarchical partitioning, differential expression screening, multi-algorithm parallel feature selection (LASSO, SVM-RFE, XGBoost), consensus gene identification and consistency index evaluation, diagnostic classification model training and evaluation, and finally outputs diagnostic discrimination results and interpretability analysis.

[0051] Figure 2 This diagram illustrates consensus gene screening using three feature selectors: LASSO, SVM-RFE, and XGBoost. Figure 2 As shown, the three feature selectors each select their own feature subsets from the differentially expressed gene set. The intersection of the three feature subsets is used to obtain the consensus gene, and then the core feature set is determined by the consistency index evaluation.

[0052] Second Embodiment

[0053] This embodiment provides a breast cancer diagnostic system, which is constructed based on the method described in the first embodiment and is intended for deployment and use in actual clinical or research scenarios.

[0054] The system includes a data input interface, a preprocessing module, a feature filtering module, a model inference module, and an interpretable output module.

[0055] The data input interface is used to receive transcriptome sample data from breast cancer. Supported data formats include expression matrix files in CSV, TSV, or Excel formats. The interface also supports batch data import, allowing simultaneous processing of data from multiple samples.

[0056] The preprocessing module communicates with the data input interface to perform log2(x+1) transformation and Z-score normalization on the expression levels of the input transcriptome sample data. The preprocessing module also includes a stratification unit to divide the processed sample data into a training set and an independent test set in a 7:3 ratio based on the healthy / breast cancer label. All normalization parameter calculations, differential expression analysis, and feature selection are performed within the training set.

[0057] The feature selection module is communicatively connected to the preprocessing module and is used to select a core feature set from the preprocessed data according to the method described in the first embodiment. Specifically, the feature selection module performs the following operations in sequence: initial screening of differential expression is performed using the Mann-Whitney U nonparametric test within the training set; three feature selection methods—LASSO, SVM-RFE, and XGBoost—are executed in parallel; the intersection of the three feature selection results is taken to obtain consensus genes; a consistency index is evaluated on the consensus genes, where the consistency index is the ratio of the number of times the consensus gene is selected in the three feature selection results to the total number of times the three feature selection methods are used; and consensus genes with a consistency index of two-thirds are selected as the core feature set. The feature selection module outputs a list of genes from the core feature set.

[0058] The model inference module is communicatively connected to the feature selection module. It uses the core feature set output by the feature selection module as input to call the pre-trained diagnostic classification model for inference calculations, outputting the breast cancer discrimination result. The diagnostic classification model is the XGBoost model trained in the first embodiment, with hyperparameters including subsample=0.9, reg_lambda=2, reg_alpha=0, and n_estimators=50.

[0059] The interpretability output module communicates with the model inference module and is used to calculate the contribution value of each feature in the core feature set to the current sample discrimination result using the SHAP method. It outputs the SHAP contribution value of each feature to the discrimination result of a single sample and generates a feature contribution report containing a SHAP force map. The SHAP force map visually displays the positive or negative contribution direction and magnitude of each feature to the model output.

[0060] The system also includes a model update module, which is connected in communication with the model inference module. This module is used to retrain or fine-tune the diagnostic classification model using incremental data after obtaining new labeled data, so as to continuously improve the model performance.

[0061] The system is deployed on a computing device, which includes a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements the functions of the aforementioned modules. The system can also be deployed on a cloud server to provide breast cancer diagnostic services to terminal devices via an application programming interface (API).

[0062] This embodiment, by solidifying the method of the first embodiment into a specific system module, achieves full automation from data input to diagnostic report output. The above descriptions are merely embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of this specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this invention.

Claims

1. A method for multi-stage feature screening and construction of a breast cancer diagnostic model, characterized in that, Includes the following steps: Obtain transcriptome sample data for breast cancer; Logarithmic transformation and standardization were performed on the expression levels of the transcriptome sample data; The processed sample data is divided into training set and independent test set according to a preset ratio; The Mann-Whitney U nonparametric test was used to initially screen for differential expression in the training set to obtain the differentially expressed gene set. Using the differentially expressed gene set as input, LASSO feature selection, SVM-RFE feature selection and XGBoost feature selection are performed in parallel to obtain the first candidate feature set, the second candidate feature set and the third candidate feature set, respectively. The consensus gene is obtained by taking the intersection of the first candidate feature set, the second candidate feature set, and the third candidate feature set. The consensus gene is evaluated using a consistency index, which is the ratio of the number of times the consensus gene is selected in the three feature selection results to the total number of the three feature selection methods. Consensus genes whose consistency index reaches a preset threshold are selected as the core feature set; The diagnostic classification model is trained using the core feature set as input, and its performance is evaluated using the independent test set to output the breast cancer discrimination result.

2. The method for multi-stage feature screening and construction of a breast cancer diagnostic model according to claim 1, characterized in that, The LASSO feature selection employs a logistic regression model with L1 regularization, and optimizes the regularization strength through hierarchical 5-fold cross-validation and ROC-AUC objective to obtain the first candidate feature set.

3. The method for multi-stage feature screening and construction of a breast cancer diagnostic model according to claim 1, characterized in that, The SVM-RFE feature selection is based on linear kernel support vector machine and recursive feature elimination mechanism, combined with 5-fold cross-validation and grid search to optimize the penalty coefficient C, and class weights are set to adapt to sample imbalance, thus obtaining a second candidate feature set.

4. The method for multi-stage feature screening and construction of a breast cancer diagnostic model according to claim 1, characterized in that, The XGBoost feature selection is achieved by constructing an XGBoost model and using random hyperparameter search to optimize hyperparameters including n_estimators, max_depth, learning_rate, subsample, colsample_bytree, gamma, reg_alpha, and reg_lambda, and then using the model's built-in feature importance to select a third candidate feature set.

5. The method for multi-stage feature screening and construction of a breast cancer diagnostic model according to claim 1, characterized in that, The preset threshold for the consistency index is two-thirds.

6. The method for multi-stage feature screening and construction of a breast cancer diagnostic model according to claim 1, characterized in that, The diagnostic classification model is an XGBoost model, whose hyperparameters include: subsample=0.9, reg_lambda=2, reg_alpha=0, n_estimators=50.

7. The method for multi-stage feature screening and construction of a breast cancer diagnostic model according to claim 1, characterized in that, It also includes using the SHAP method to analyze the contribution of each feature in the core feature set to the model output, and outputting the SHAP contribution value of each feature to the discrimination result of a single sample.

8. The method for multi-stage feature screening and construction of a breast cancer diagnostic model according to claim 1, characterized in that, The preset ratio is 7:3, and the data is randomly divided into stratified groups based on the health / breast cancer label, so that the sample category ratio of the training set and the independent test set is consistent with the original data; all standardized parameter calculations, differential expression analysis and feature selection are limited to the training set.

9. A breast cancer diagnostic system, characterized in that, include: Data input interface for receiving breast cancer transcriptome sample data; The preprocessing module is used to perform logarithmic transformation and standardization on the expression levels of the transcriptome sample data; A feature screening module is used to screen the transcriptome sample data to obtain a core feature set according to the method of any one of claims 1 to 8; The model inference module is used to take the core feature set as input and call a pre-trained diagnostic classification model to output the breast cancer discrimination result; The interpretability output module is used to calculate and output the SHAP contribution value of each feature to the current sample discrimination result.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images