Methods, systems, devices, and media for classifying cancer gene expression data based on multi-feature fusion and deep learning models.

Abstract

This invention belongs to the field of cancer assessment technology and discloses a method, system, device, and medium for classifying cancer gene expression data based on multi-feature fusion and a deep learning model. The method includes: acquiring gene expression data to be classified and converting it into a gene expression matrix; inputting the gene expression matrix into a parallel feature extraction network for feature extraction to obtain multiple feature subsets; inputting each feature subset into a corresponding deep learning classification model based on preset matching rules for cancer type classification prediction to obtain multiple classification results; wherein the deep learning classification model includes a multilayer perceptron model, a Transformer model, and a long short-term memory network model; and selecting the result with the highest cancer type prediction probability from the classification results as the cancer classification label. This invention enables effective feature mining and accurate classification of high-dimensional gene expression data.

Description

Technical Field

[0001] This invention belongs to the field of cancer assessment technology, and in particular relates to a method, system, device and medium for classifying cancer gene expression data based on multi-feature fusion and deep learning models. Background Technology

[0002] Accurate cancer classification is a core prerequisite for tumor diagnosis and personalized treatment. Gene expression data contains key molecular information about tumor development and progression, and is an important basis for cancer classification. However, gene expression data is inherently characterized by high dimensionality, low sample size, and significant noise interference. Directly using it for cancer classification can easily lead to the curse of dimensionality, resulting in overfitting of the classification model and a significant decrease in its generalization ability.

[0003] In existing technologies, cancer classification is mostly accomplished using a single feature extraction method or a single machine learning model. Single feature extraction methods (such as using only PCA dimensionality reduction or only LASSO regression for feature selection) cannot effectively mine information from gene expression data across multiple dimensions, easily leading to feature loss. Meanwhile, traditional single machine learning classification models have limited fitting capabilities for complex biological data and struggle to adapt to high-dimensional gene feature data after dimensionality reduction. These shortcomings combine to result in insufficient accuracy and poor stability in existing cancer classification methods, failing to meet the technical requirements for precise cancer classification in clinical practice. Therefore, developing a cancer classification method that integrates the information mining advantages of multi-feature extraction with the efficient fitting advantages of deep learning models to achieve effective feature mining and accurate classification of high-dimensional gene expression data has become an urgent technical problem to be solved in this field. Summary of the Invention

[0004] The purpose of this invention is to provide a method, system, device and medium for classifying cancer gene expression data based on multi-feature fusion and deep learning models, so as to solve the technical defects of the existing cancer classification methods, such as insufficient feature mining, low classification accuracy and poor model generalization ability.

[0005] In a first aspect, to achieve the above objectives, this invention provides a method for classifying cancer gene expression data based on multi-feature fusion and a deep learning model, comprising:

[0006] Accurate cancer classification is crucial for tumor diagnosis and personalized treatment, but the high dimensionality and low sample size of gene expression data pose challenges to this task. This invention proposes a cancer classification method based on the fusion of multi-feature extraction and deep learning, aiming to address the problems of insufficient information mining and inadequate classification accuracy in existing single-method approaches. This method first acquires gene expression data for five cancer types—BRCA (Breast Invasive Carcinoma), COAD (Colon Adenocarcinoma), KIRC (Kidney Renal Clear Cell Carcinoma), LUAD (Lung Adenocarcinoma), and PRAD (Prostate Adenocarcinoma)—from the TCGA database (The Cancer Genome Atlas). After obtaining a standardized gene expression matrix and undergoing standardized preprocessing, three parallel feature extraction paths are constructed: LASSO regression is used to screen 293 key gene features, PCA is used to extract 433 principal components with a cumulative variance explanation rate of 95.02%, and random forest is used to screen the top 200 important features. Subsequently, the three feature subsets are input into adapted MLP, Transformer, and LSTM deep learning models, respectively, under unified training parameters (50 epochs, batch size 32, Adam). Training and prediction are performed under an optimizer; finally, the highest predicted probability among the three sets of results is taken as the final classification result. Experimental verification shows that the LASSO+MLP path achieves 100% accuracy in both training and validation, the PCA+Transformer path achieves 0.9895 accuracy in validation, and the RF+LSTM path achieves 98.5% accuracy on the test set. The overall classification accuracy and stability are significantly better than traditional methods. The specific steps include:

[0007] S1: Obtain gene expression data to be classified;

[0008] S2: Convert the gene expression data to be classified into a gene expression matrix;

[0009] S3: Input the gene expression matrix into a parallel feature extraction network for feature extraction to obtain multiple feature subsets; wherein, the parallel feature extraction network is used to execute the LASSO regression algorithm, PCA principal component analysis algorithm and random forest RF algorithm in parallel;

[0010] S4: Based on preset matching rules, each feature subset is input into the corresponding deep learning classification model for cancer type classification prediction, resulting in multiple classification results; wherein, the deep learning classification model includes a multilayer perceptron model, a Transformer model, and a long short-term memory network model; the result with the highest cancer type prediction probability among the classification results is selected as the cancer classification label.

[0011] Optionally, step S2 specifically includes:

[0012] Missing values ​​were imputed and the gene expression data to be classified were standardized to obtain preprocessed gene expression data. A gene expression matrix was then constructed based on the preprocessed gene expression data.

[0013] Optionally, the training process of each deep learning classification model in step S4 specifically includes:

[0014] Missing values ​​were filled and standardized in the gene expression data of different cancer types in the Cancer Genome Atlas (TCGA) database. After standardization, a gene expression matrix was constructed as training data.

[0015] The training data is input into a parallel feature extraction network. The LASSO regression algorithm is used to select key gene features with non-zero coefficients in the training data to form the first feature set. The PCA principal component analysis algorithm is used to extract principal component features in the training data to form the second feature set. The Random Forest (RF) algorithm is used to extract the top 200 gene features in the training data by importance to form the third feature set.

[0016] Based on preset matching rules, the first feature set, the second feature set, and the third feature set are respectively input into the corresponding deep learning classification model for cancer type classification prediction. The model is then trained based on the target loss function to obtain the trained deep learning classification model. The cancer type classification prediction task is then performed based on the trained deep learning classification model.

[0017] Optionally, step S4 specifically includes:

[0018] The first feature set is input into the multilayer perceptron model, and the first classification result is output.

[0019] The second feature set is input into the Transformer model, which outputs the second classification result.

[0020] The third feature set is input into the Long Short-Term Memory network model, and the third classification result is output.

[0021] The result with the highest predicted probability among the three classification results is taken as the cancer classification label, which is used to assist doctors in making classification judgments.

[0022] Secondly, to achieve the above objectives, this invention provides a cancer gene expression data classification system based on multi-feature fusion and a deep learning model, comprising:

[0023] The data acquisition module is used to acquire gene expression data to be classified.

[0024] The preprocessing module is used to convert the gene expression data to be classified into a gene expression matrix;

[0025] The feature extraction module is used to input the gene expression matrix into a parallel feature extraction network for feature extraction, obtaining multiple feature subsets; wherein, the parallel feature extraction network is used to execute the LASSO regression algorithm, the PCA principal component analysis algorithm, and the random forest RF algorithm in parallel;

[0026] The classification prediction module is used to input each feature subset into the corresponding deep learning classification model according to the preset matching rules to perform cancer type classification prediction and obtain multiple classification results; wherein, the deep learning classification model includes a multilayer perceptron model, a Transformer model and a long short-term memory network model; the result with the highest cancer type prediction probability among the classification results is selected as the cancer classification label.

[0027] Thirdly, to achieve the above objectives, the present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the cancer gene expression data classification method based on multi-feature fusion and deep learning model described in the first aspect.

[0028] Fourthly, to achieve the above objectives, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the cancer gene expression data classification method based on multi-feature fusion and deep learning model described in the first aspect.

[0029] The technical effects of this invention are as follows:

[0030] This invention achieves multi-dimensional and effective feature mining of gene expression data by constructing three parallel feature extraction paths. By combining the adaptation design of features and deep learning models with the integration strategy of multi-classification results, it achieves accurate dimensionality reduction and efficient classification of high-dimensional gene expression data, solving the technical drawbacks of existing single methods. Ultimately, it achieves accurate classification and identification of five types of cancer: BRCA, COAD, KIRC, LUAD, and PRAD, providing reliable and efficient technical support for the clinical auxiliary diagnosis of cancer.

[0031] This invention effectively mitigates the risk of overfitting in high-dimensional gene data by integrating multi-path feature complementarity mining with deep learning model adaptation, and improves classification accuracy, providing precise and reliable technical support for cancer auxiliary diagnosis. It is applicable to tumor gene data analysis scenarios in the field of bioinformatics. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0034] Figure 1 This is a flowchart of the cancer gene expression data classification process using multi-feature fusion and deep models in an embodiment of the present invention.

[0035] Figure 2 This is a LASSO+MLP classification path analysis report from an embodiment of the present invention;

[0036] Figure 3 This is a PCA+Transformer classification path analysis report from an embodiment of the present invention;

[0037] Figure 4 This is a PF+LSTM classification path analysis report in an embodiment of the present invention. Detailed Implementation

[0038] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as a limitation of the present invention, but rather as a more detailed description of certain aspects, features, and embodiments of the present invention.

[0039] It should be understood that the terminology used in this invention is merely for describing particular embodiments and is not intended to limit the invention. Furthermore, with respect to numerical ranges in this invention, it should be understood that each intermediate value between the upper and lower limits of the range is also specifically disclosed. Every smaller range between any stated value or intermediate value within a stated range, and any other stated value or intermediate value within said range, is also included in this invention. The upper and lower limits of these smaller ranges may be independently included or excluded from the range.

[0040] Various modifications and variations can be made to the specific embodiments described in this specification without departing from the scope or spirit of the invention, as will be apparent to those skilled in the art. Other embodiments derived from this specification will also be obvious to those skilled in the art. This application specification and embodiments are merely exemplary.

[0041] The terms “include,” “including,” “have,” “contain,” etc., used in this article are all open-ended terms, meaning that they include but are not limited to.

[0042] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0043] Example 1

[0044] like Figure 1 - Figure 4 As shown, this embodiment provides a cancer gene expression data classification method based on multi-feature fusion and deep learning models, including:

[0045] (1) Obtain and preprocess gene expression datasets of five types of cancer in the TCGA database, and perform preprocessing operations on the datasets to obtain standardized gene expression matrices;

[0046] (2) Lasso regression (LASSO), principal component analysis (PCA), and random forest (RF) algorithms were used to extract features from the standardized dataset to obtain three feature subsets.

[0047] (3) The three feature subsets are respectively input into the multilayer perceptron (MLP), Transformer, and long short-term memory (LSTM) classification models for training and prediction, and three sets of classification results are obtained.

[0048] (4) Integrate the classification results of the three groups, output the final classification result, and complete the classification and identification of the five types of cancer.

[0049] In practice, step (2) yields three feature sets through three parallel feature extraction paths, specifically:

[0050] Path 1: The LASSO regression algorithm is used to screen features of the standardized gene expression matrix, and 293 key gene features with non-zero coefficients are selected to form the first feature set;

[0051] Path 2: Principal Component Analysis (PCA) algorithm is used to perform feature dimensionality reduction on the standardized gene expression matrix, and the first 433 principal components are extracted. The cumulative variance explained by the 433 principal components is 95.02%, and the variance explained by the first 3 principal components are 0.11, 0.09 and 0.08, respectively, which constitute the second feature set.

[0052] Path 3: Use the Random Forest (RF) algorithm to calculate the importance of each gene feature in the standardized gene expression matrix, and select the top 200 gene features by importance to form the third feature set.

[0053] The regularization parameter of the Lasso Regression (LASSO) algorithm described in Path 1 is set to 0.01; the number of trees in the Random Forest (RF) model described in Path 3 is set to 100.

[0054] The three feature extraction paths described in step (3) correspond one-to-one with the deep learning classification models, specifically:

[0055] The first feature set of path 1 is input into the multilayer perceptron (MLP) model, and the first classification result is output.

[0056] The second feature set of path 2 is input into the Transformer model, and the second classification result is output.

[0057] The third feature set of path 3 is input into the Long Short-Term Memory (LSTM) network model, and the third classification result is output.

[0058] The feasible structure of the MLP model is as follows: input layer (293 neurons) → hidden layer 1 (128 neurons, activation function is ReLU) → hidden layer 2 (64 neurons, activation function is ReLU) → output layer (5 neurons, activation function is Softmax).

[0059] The Transformer model contains three encoder layers, each of which includes a multi-head attention mechanism and a feedforward neural network; the LSTM model contains two LSTM layers and one fully connected output layer, with the activation function being Softmax.

[0060] The training parameters for the MLP model, Transformer model, and LSTM model are uniformly set as follows: 50 training epochs, batch size 32, Adam optimizer, and cross-entropy loss function.

[0061] The feasible method for integrating the classification results of each path in step (4) is to take the result with the highest prediction probability among the three classification results as the final cancer classification prediction result.

[0062] The five cancer types mentioned are BRCA, COAD, KIRC, LUAD, and PRAD. The classification accuracy of the method meets the following requirements: the training and validation accuracy of the LASSO+MLP path reaches 100%, the training accuracy of the PCA+Transformer path reaches 0.9982, the validation accuracy reaches 0.9895, and the test set accuracy of the RF+LSTM path reaches 98.5%.

[0063] This embodiment also provides a cancer gene expression data classification system based on multi-feature fusion and a deep learning model, including:

[0064] The data acquisition module is used to acquire gene expression data to be classified.

[0065] The preprocessing module is used to convert the gene expression data to be classified into a gene expression matrix;

[0066] The feature extraction module is used to input the gene expression matrix into a parallel feature extraction network for feature extraction, obtaining multiple feature subsets; wherein, the parallel feature extraction network is used to execute the LASSO regression algorithm, the PCA principal component analysis algorithm, and the random forest RF algorithm in parallel;

[0067] The classification prediction module is used to input each feature subset into the corresponding deep learning classification model according to the preset matching rules to perform cancer type classification prediction and obtain multiple classification results; wherein, the deep learning classification model includes a multilayer perceptron model, a Transformer model and a long short-term memory network model; the result with the highest cancer type prediction probability among the classification results is selected as the cancer classification label.

[0068] Compared with existing technologies, this embodiment effectively solves the technical defects of existing cancer classification methods, such as insufficient feature mining, poor model generalization ability, and insufficient classification accuracy, through innovative designs of multi-path feature extraction, precise feature-model adaptation, and multi-result integration. It has the following significant beneficial effects:

[0069] 1. Multi-dimensional feature mining to alleviate technical bottlenecks in high-dimensional data: This embodiment integrates three complementary feature extraction methods: LASSO, PCA, and RF. It performs differentiated mining on high-dimensional gene expression data from three dimensions: key gene screening, core information dimensionality reduction, and feature importance ranking. This not only preserves key biological features that are highly correlated with cancer classification, but also achieves efficient dimensionality reduction. It effectively alleviates the dimensionality curse and overfitting risk caused by high dimensionality and low sample size of gene expression data, while avoiding the information loss problem of single feature extraction methods and fully mining the effective classification information in the data.

[0070] 2. Precise feature-model matching enhances model classification ability: Based on the structural characteristics and data attributes of the three feature sets, matching deep learning models are used—MLP model is good at handling nonlinear mapping of high-dimensional selected features, Transformer model is good at capturing sequence association information of principal component features, and LSTM model is good at mining temporal dependencies of highly important features. This achieves a deep combination of feature attributes and model advantages, significantly improving the classification model's fitting ability and generalization of complex biological data, and solving the problem of poor adaptability of traditional single models.

[0071] 3. Unified training parameters and probability fusion strategy to ensure classification accuracy and stability: This embodiment sets unified training hyperparameters for the three types of deep learning models, eliminating the interference of parameter differences on classification results and ensuring the fairness and comparability of model training; at the same time, it adopts the result integration strategy of the highest predicted probability value to integrate the classification advantages of multiple models, further improving the accuracy and stability of cancer classification. Experimental verification shows that each classification path in this embodiment achieves high accuracy classification, and the overall classification effect is significantly better than the existing cancer classification methods with single feature extraction and single model.

[0072] 4. The method is highly feasible and has high clinical application value: The technical solution of this embodiment is based on the TCGA public database, with reliable data sources. The algorithm parameters and model structure of each step are clearly and specifically defined, and the method has strong repeatability and feasibility. At the same time, this embodiment achieves accurate classification of five common cancer types: BRCA, COAD, KIRC, LUAD, and PRAD, providing reliable and efficient technical support for the clinical auxiliary diagnosis of cancer. It can be directly applied to tumor gene data analysis scenarios in the field of bioinformatics and has broad prospects for clinical translation and application.

[0073] This embodiment specifically includes the following steps:

[0074] 1. Data preprocessing: Gene expression datasets for five cancer types (BRCA, COAD, KIRC, LUAD, and PRAD) were obtained from the TCGA database. The datasets contained 801 samples and 16,382 gene features. Missing values ​​were filled and the raw data were standardized in sequence to eliminate differences in data dimensions and noise interference, resulting in a standardized gene expression matrix that can be used for subsequent analysis.

[0075] 2. Multi-path parallel feature extraction: Three parallel feature extraction paths are constructed to perform differential feature mining on the standardized gene expression matrix, resulting in three feature sets, specifically:

[0076] Path 1: The LASSO algorithm of lasso regression is used to screen features of the standardized gene expression matrix. The regularization parameter is set to 0.01, and 293 key gene features with non-zero coefficients are selected to form the first feature set.

[0077] Path 2: Principal Component Analysis (PCA) algorithm is used to reduce the dimensionality of the standardized gene expression matrix and extract the top 433 principal components. The cumulative variance explained by these 433 principal components reaches 95.02%, effectively preserving the core information of the original data and forming the second feature set. Path 3: Random Forest (RF) algorithm is used to rank the features of the standardized gene expression matrix by importance. The number of trees is set to 100, and the top 200 gene features by importance are selected to form the third feature set.

[0078] 3. Deep Learning Model Adaptation Training and Prediction: The three feature sets mentioned above are input one-to-one into the adapted deep learning classification model. Model training and cancer classification prediction are performed under unified training hyperparameters, resulting in three independent classification results, specifically:

[0079] (1) Input the first feature set into the multilayer perceptron (MLP) model. The network structure of the MLP model is: input layer (293 neurons) → first hidden layer (128 neurons, activation function is ReLU) → second hidden layer (64 neurons, activation function is ReLU) → output layer (5 neurons, activation function is Softmax).

[0080] (2) Input the second feature set into the Transformer model, which contains three encoder layers. Each encoder layer integrates a multi-head attention mechanism and a feedforward neural network, and the activation function of the feedforward neural network is GELU.

[0081] (3) Input the third feature set into the Long Short-Term Memory (LSTM) network model. The LSTM model contains two LSTM layers and one fully connected output layer. The activation function of the LSTM layer is ReLU, and the activation function of the fully connected output layer is Softmax. The training hyperparameters of the MLP model, Transformer model and LSTM model are uniformly set as follows: 50 training rounds, batch size 32, optimizer is Adam, and loss function is cross-entropy.

[0082] 4. Integration and output of multiple results: The classification results output by the above three sets of deep learning models are integrated, and the result with the highest cancer type prediction probability among the three classification results is taken as the final cancer classification prediction result, thus completing the accurate classification and identification of five types of cancer.

[0083] The following implementation examples provide a detailed explanation of the specific implementation process of the three feature extraction-deep learning classification paths. The gene expression datasets used in each example are standardized gene expression matrices of five cancer types (BRCA, COAD, KIRC, LUAD, and PRAD) in the TCGA database. These matrices are obtained from the original dataset after missing value imputation and standardization preprocessing, and contain 801 samples and 16382 gene features.

[0084] Example 1: Lasso Regression (LASSO) Feature Extraction + Multilayer Perceptron (MLP) Classification Path

[0085] 1. Feature extraction: The preprocessed standardized gene expression matrix was input into the LASSO regression model. The regularization parameter of the model was set to 0.01. After model training, 293 key gene features with non-zero coefficients were selected to form the first feature set. Among them, the gene features with the top 20 absolute values ​​of feature coefficients contributed the most to cancer classification and were the core features of classification.

[0086] 2. Model Training: The first feature set mentioned above is input into a multilayer perceptron (MLP) classification model for training. The network structure of the MLP model is as follows: input layer (293 neurons) → first hidden layer (128 neurons, activation function is ReLU) → second hidden layer (64 neurons, activation function is ReLU) → output layer (5 neurons, activation function is Softmax). The model training hyperparameters are set as follows: 50 training epochs, batch size 32, optimizer is Adam, and loss function is cross-entropy.

[0087] 3. Result Validation: During training, both training loss and validation loss continuously decreased and tended to stabilize, with the final training loss at 0.0635 and the validation loss at 0.0352; both training accuracy and validation accuracy reached 1.0000. The confusion matrix showed that there were no erroneous samples in the classification results of the five cancer types (BRCA, COAD, KIRC, LUAD, PRAD), achieving accurate classification of the five cancer types.

[0088] Example 2: Principal Component Analysis (PCA) Feature Extraction + Transformer Classification Path

[0089] 1. Feature Extraction: The preprocessed standardized gene expression matrix is ​​input into the principal component analysis (PCA) model for feature dimensionality reduction. The first 433 principal components are extracted to form the second feature set. The cumulative variance explained by the 433 principal components reaches 95.02%, which can effectively preserve the core information in the original gene expression data and achieve efficient dimensionality reduction of high-dimensional data.

[0090] 2. Model Training: The second feature set mentioned above is input into the Transformer classification model for training. The Transformer model contains three encoder layers. Each encoder layer integrates a multi-head attention mechanism and a feedforward neural network. The activation function of the feedforward neural network is GELU. The model training hyperparameters are consistent with those in Example 1, namely, 50 training rounds, batch size of 32, Adam optimizer, and cross-entropy loss function.

[0091] 3. Results Validation: During training, both training loss and validation loss continuously decreased, ultimately reaching 0.1607 for training loss and 0.1910 for validation loss; training accuracy reached 0.9982 and validation accuracy reached 0.9895. The confusion matrix showed that only a few samples exhibited classification bias, and overall, high-precision classification of the five cancer types was achieved, demonstrating excellent stability of the classification results.

[0092] Example 3: Random Forest (RF) feature extraction + Long Short-Term Memory (LSTM) classification path

[0093] 1. Feature extraction: The preprocessed standardized gene expression matrix is ​​input into the random forest (RF) model. The number of trees in the model is set to 100. The importance score of each gene feature is calculated by the model. Based on the score, the top 200 gene features in terms of importance are selected to form the third feature set.

[0094] 2. Model Training: The third feature set mentioned above is input into a Long Short-Term Memory (LSTM) classification model for training. The LSTM model contains two LSTM layers and one fully connected output layer. The activation function of the LSTM layers is ReLU, and the activation function of the fully connected output layer is Softmax. The model training hyperparameters are kept the same as in Examples 1 and 2, namely, 50 training epochs, batch size of 32, Adam optimizer, and cross-entropy loss function.

[0095] 3. Result Validation: After the model training was completed, the classification effect was validated on the test set. The results showed that the model's classification accuracy on the test set reached 98.5%, the training accuracy was 0.9871, and the validation accuracy was 0.9688, indicating that the classification path has good generalization ability and can be effectively applied to the actual classification and identification scenarios of five types of cancer.

[0096] The above description is merely a preferred embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for classifying cancer gene expression data based on multi-feature fusion and a deep learning model, characterized in that, include: S1: Obtain gene expression data to be classified; S2: Convert the gene expression data to be classified into a gene expression matrix; S3: Input the gene expression matrix into a parallel feature extraction network for feature extraction to obtain multiple feature subsets; wherein, the parallel feature extraction network is used to execute the LASSO regression algorithm, PCA principal component analysis algorithm and random forest RF algorithm in parallel; S4: Based on preset matching rules, each feature subset is input into the corresponding deep learning classification model for cancer type classification prediction, resulting in multiple classification results; wherein, the deep learning classification model includes a multilayer perceptron model, a Transformer model, and a long short-term memory network model; the result with the highest cancer type prediction probability among the classification results is selected as the cancer classification label.

2. The cancer gene expression data classification method based on multi-feature fusion and deep learning model according to claim 1, characterized in that, Step S2 specifically includes: Missing values ​​were imputed and the gene expression data to be classified were standardized to obtain preprocessed gene expression data. A gene expression matrix was then constructed based on the preprocessed gene expression data.

3. The cancer gene expression data classification method based on multi-feature fusion and deep learning model according to claim 1, characterized in that, The training process of each deep learning classification model in step S4 specifically includes: Missing values ​​were filled and standardized in the gene expression data of different cancer types in the Cancer Genome Atlas (TCGA) database. After standardization, a gene expression matrix was constructed as training data. The training data is input into a parallel feature extraction network. The LASSO regression algorithm is used to select key gene features with non-zero coefficients in the training data to form the first feature set. The PCA principal component analysis algorithm is used to extract principal component features in the training data to form the second feature set. The Random Forest (RF) algorithm is used to extract the top 200 gene features in the training data by importance to form the third feature set. Based on preset matching rules, the first feature set, the second feature set, and the third feature set are respectively input into the corresponding deep learning classification model for cancer type classification prediction. The model is then trained based on the target loss function to obtain the trained deep learning classification model. The cancer type classification prediction task is then performed based on the trained deep learning classification model.

4. The cancer gene expression data classification method based on multi-feature fusion and deep learning model according to claim 3, characterized in that, Step S4 specifically includes: The first feature set is input into the multilayer perceptron model, and the first classification result is output. The second feature set is input into the Transformer model, which outputs the second classification result. The third feature set is input into the Long Short-Term Memory network model, and the third classification result is output. The result with the highest predicted probability among the three classification results is taken as the cancer classification label, which is used to assist doctors in making classification judgments.

5. A cancer gene expression data classification system based on multi-feature fusion and deep learning model, characterized in that, include: The data acquisition module is used to acquire gene expression data to be classified. The preprocessing module is used to convert the gene expression data to be classified into a gene expression matrix; The feature extraction module is used to input the gene expression matrix into a parallel feature extraction network for feature extraction, obtaining multiple feature subsets; wherein, the parallel feature extraction network is used to execute the LASSO regression algorithm, the PCA principal component analysis algorithm, and the random forest RF algorithm in parallel; The classification prediction module is used to input each feature subset into the corresponding deep learning classification model according to the preset matching rules to perform cancer type classification prediction and obtain multiple classification results; wherein, the deep learning classification model includes a multilayer perceptron model, a Transformer model and a long short-term memory network model; the result with the highest cancer type prediction probability among the classification results is selected as the cancer classification label.

6. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to enable the electronic device to perform a cancer gene expression data classification method based on multi-feature fusion and deep learning model according to any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements a cancer gene expression data classification method based on multi-feature fusion and a deep learning model as described in any one of claims 1-4.

Images