Biological characteristic analysis system based on multi-modal data and residual image attention network

Abstract

The invention discloses a biological characteristic analysis system based on multi-modal data and a residual image attention network. The biological characteristic analysis system comprises a multi-modal data processing module, an individual similarity network construction module, a biological characteristic extraction module and a biological characteristic analysis module. Clinical image data and multi-modal data of gene expression are collected, on one hand, a comprehensive biological individual similarity network is fused and constructed, on the other hand, feature extraction is carried out based on a multi-modal auto-encoder fusion method, and finally, a constructed residual image attention network model is input for feature processing. Therefore, accurate classification and analysis of various biological characteristics and samples can be realized. The biological potential feature analysis accuracy and reliability are effectively improved, powerful reference is provided for biological scientific research and health management, and early discovery of potential changes of biological features and implementation of personalized health management are facilitated.

Description

Technical Field

[0001] The present invention belongs to the field of biological science research and health management, and relates to biometric analysis technology. Specifically, it relates to a biometric analysis system based on multimodal data and residual graph attention network, which is suitable for accurately analyzing various types of biological data and samples and predicting potential changes in biometrics. Background Art

[0002] Biometric analysis is a crucial component of biological science research and health management, playing a crucial role in improving the efficiency of biological research, optimizing health management programs, and reducing medical costs. Therefore, research on biometric analysis technology has become a key topic in both the scientific and medical fields.

[0003] Traditional methods for biometric analysis primarily include traditional physiological indicator testing, genetic testing, and imaging studies. Traditional physiological indicator testing uses samples such as blood and urine to obtain individual physiological indicator information for biometric analysis. Genetic testing assists in biometric analysis by analyzing information such as mutations and expression levels of specific genes. Imaging studies utilize technologies such as X-rays, CT scans, and MRIs to capture images of the human body and identify potential biometric changes.

[0004] However, these methods have many limitations in practical application. Physiological indicator testing relies on the accuracy of the testing equipment and is difficult to fully reflect the complexity of biological characteristics. Some early changes or minor lesions may be difficult to accurately identify. While genetic testing can provide information at the molecular level, relying solely on genetic data often struggles to fully reflect the complexity of biological characteristics. Testing is also costly and places stringent demands on sample quality and analytical techniques. While imaging examinations can intuitively display structural information within the human body, they have low sensitivity to early, subtle changes and are unable to directly reflect the biological behavior and molecular characteristics of biological characteristics.

[0005] These issues limit the accuracy and reliability of biometric analysis, leading to frequent analysis delays, high misjudgment rates, and difficulty meeting personalized needs in complex biological research and health management scenarios. With the rise of machine learning, a large number of methods based on support vector machines, random forests, and logistic regression have emerged in the field of biometric analysis. These methods analyze the biometric characteristics of samples by learning from biological data. However, traditional machine learning methods have high requirements for data quality, and noise and outliers can affect analysis results.

[0006] To overcome these challenges, biometric analysis techniques based on multimodal data fusion using deep learning methods have become a research hotspot in recent years. By fusing information from multiple modalities, such as imaging and genomic data, and leveraging deep learning techniques to extract feature data from massive, high-dimensional data, this method is currently the most advanced in the field of biometric analysis. However, existing methods often struggle with missing or noisy modalities. Most deep learning methods assume complete and noise-free data, which is difficult to achieve in real-world scenarios and limits model performance. Biometric analysis systems based on multimodal data and residual graph attention networks can fully leverage the morphological information of imaging data and the molecular features of genomic data to deeply mine biometric information across multiple data sets. However, existing fusion strategies still have limitations when processing multimodal data, making it difficult to capture long-range dependencies within the data. Weight distribution in the model's hidden layers is often imbalanced, making it difficult to automatically adjust the weights of feature information. Summary of the Invention

[0007] This invention addresses the shortcomings of existing biometric analysis methods, such as low accuracy, poor reliability, and low universality, by providing a biometric analysis system based on multimodal data and residual graph attention networks. By integrating multimodal data from clinical imaging and gene expression, a comprehensive individual similarity network is constructed. This is then processed using a residual graph attention network model to achieve precise classification and analysis of various biometric features and samples, effectively improving analysis accuracy and reliability. This provides a powerful reference for biological science research and health management, facilitating the early detection of potential changes in biometrics and the implementation of personalized health management.

[0008] To achieve the above object, the present invention adopts the following technical solutions:

[0009] In a first aspect, embodiments of the present application provide a biometric feature analysis system based on multimodal data and residual graph attention network, comprising the following modules:

[0010] Multimodal data processing module: pre-processes the collected raw clinical data and raw genomic data to construct a data set.

[0011] Individual similarity network construction module: An individual similarity network is constructed using a similarity network fusion method based on a hierarchical spatial fusion strategy.

[0012] Biometric feature extraction module: Use a method based on multimodal autoencoder fusion to extract features and obtain a multimodal data feature matrix.

[0013] Biometric analysis module: The residual graph attention network model is used to perform biometric analysis based on the obtained individual similarity network feature matrix and multimodal data feature matrix.

[0014] In one possible implementation, the multimodal data processing module cleans the raw data through outlier detection and missing value processing methods, performs feature selection on the genomic data using the L1 regularized linear support vector machine algorithm, and constructs a valid data set for subsequent work.

[0015] In one possible implementation, the individual similarity network construction module integrates the processed clinical data and genomic data, introduces a hierarchical spatial fusion strategy, and generates an individual similarity network through iteration and integration that fully retains the similarity between biological individuals and the correlation characteristics between data.

[0016] In one possible implementation, the biometric feature extraction module constructs independent autoencoders for clinical data and genomic data, maps the original data into a low-dimensional latent space, and then integrates the features of the two modalities through an integration layer to obtain an integration matrix; then, the decoder maps the integration matrix back to the original input space to obtain a dispersed feature representation, and finally, these feature representations are merged through matrix splicing to obtain a more representative and robust feature representation, namely, a multimodal data feature matrix.

[0017] In one possible implementation, the biometric analysis module is specifically implemented as follows:

[0018] First, the obtained individual similarity network feature matrix and multimodal data feature matrix are used as model inputs. By designing a two-layer graph convolutional neural network model, the self-learning ability of the neural network is utilized to perform gradient descent on the input, and the final prediction result is obtained through iterative correction. In order to improve the accuracy of feature aggregation, the attention mechanism is introduced to dynamically adjust the contribution of neighboring nodes and perform weighted processing on the input features. The residual connection is designed and applied to the two-layer graph convolutional neural network model that incorporates the attention mechanism to enhance the feature expression ability of the original sample and form the final residual graph attention network model. Finally, the cross-entropy loss function is used to quantify the difference between the predicted and true labels to optimize the model performance. After completing the construction and optimization of the residual graph attention network model, the model is used to conduct in-depth analysis of the characteristics of the biological samples and finally output the predicted feature labels of the biological samples.

[0019] In a second aspect, an embodiment of the present application provides a biometric feature analysis method based on multimodal data and a residual graph attention network, comprising the following steps:

[0020] Multimodal data processing: Preprocess the collected raw clinical and genomic data to construct a dataset. The raw data is cleaned using outlier detection and missing value processing. A linear support vector machine algorithm with L1 regularization is used for feature selection on the genomic data to construct a valid dataset for subsequent work.

[0021] Individual Similarity Network Construction: We use a similarity network fusion method based on a hierarchical spatial fusion strategy to construct an individual similarity network. We integrate processed clinical and genomic data, introduce a hierarchical spatial fusion strategy, and through iteration and integration, generate an individual similarity network that fully preserves the similarities between biological individuals and the correlation characteristics between data.

[0022] Biometric feature extraction: A method based on multimodal autoencoder fusion is used for feature extraction to obtain a multimodal data feature matrix.

[0023] Biometric analysis: The residual graph attention network model is used to perform biometric analysis based on the obtained individual similarity network feature matrix and multimodal data feature matrix.

[0024] In one possible implementation, biometric feature extraction is specifically as follows: independent autoencoders are constructed for clinical data and genomic data, the raw data is mapped into a low-dimensional latent space, and the features of the two modalities are integrated through an integration layer to obtain an integration matrix; then, the decoder maps the integration matrix back to the original input space to obtain a dispersed feature representation, and finally, these feature representations are merged through matrix splicing to obtain a more representative and robust feature representation, namely, a multimodal data feature matrix.

[0025] In one possible implementation, the biometric analysis is specifically as follows:

[0026] First, the obtained individual similarity network feature matrix and multimodal data feature matrix are used as model inputs. By designing a two-layer graph convolutional neural network model, the self-learning ability of the neural network is utilized to perform gradient descent on the input, and the final prediction result is obtained through iterative correction. In order to improve the accuracy of feature aggregation, the attention mechanism is introduced to dynamically adjust the contribution of neighboring nodes and perform weighted processing on the input features. The residual connection is designed and applied to the two-layer graph convolutional neural network model that incorporates the attention mechanism to enhance the feature expression ability of the original sample and form the final residual graph attention network model. Finally, the cross-entropy loss function is used to quantify the difference between the predicted and true labels to optimize the model performance. After completing the construction and optimization of the residual graph attention network model, the model is used to conduct in-depth analysis of the characteristics of the biological samples and finally output the predicted feature labels of the biological samples.

[0027] In a third aspect, an embodiment of the present application provides an electronic device, comprising:

[0028] A processor; a memory for storing processor-executable instructions; wherein the processor is configured to call the memory to store instructions to execute any one of the methods described in the second aspect.

[0029] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the storage medium stores computer program instructions, and the computer program instructions are executed by a processor to implement any one of the methods described in the second aspect.

[0030] The beneficial effects of the present invention are as follows:

[0031] The method of the present invention can retain more sufficient characteristic information of clinical data by constructing an individual similarity network and innovatively introducing a hierarchical spatial fusion strategy; it designs a method of autoencoder fusion for feature extraction of multimodal data to obtain better genomic feature information; it constructs a residual attention network model, integrates the attention mechanism and residual connection module into the structure of the double-layer graph convolutional network, and has a more accurate prediction effect than existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flowchart of a biometric analysis system based on multimodal data and residual graph attention network according to an embodiment of the present invention.

[0033] Figure 2 A schematic diagram of constructing an individual similarity network according to an embodiment of the present invention.

[0034] Figure 3 Schematic diagram of feature extraction of a multimodal autoencoder according to an embodiment of the present invention.

[0035] Figure 4 This is a flow chart of the residual graph attention network model according to an embodiment of the present invention.

[0036] Figure 5 Schematic diagram of the attention mechanism according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The present invention will be further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0038] Biometric analysis is crucial for understanding the pathogenesis of health problems and for health management. This invention, by improving the accuracy and robustness of assessments, can better predict the progression of an individual's health status, develop appropriate health management plans, and promote advancements in health management. By integrating multimodal encoders with deep learning technology and designing a feature mining method, this invention can more accurately predict an individual's health status, avoiding the low precision, poor reliability, and strong dependence on specific datasets inherent in traditional methods.

[0039] The present invention provides a biometric feature analysis system based on multimodal data and residual graph attention network, including the following modules:

[0040] Multimodal Data Processing Module: Preprocesses the collected raw clinical and genomic data to construct a dataset. Raw data is cleaned using outlier detection and missing value processing. Feature selection is performed on the genomic data using a linear support vector machine algorithm with L1 regularization to construct a valid dataset for subsequent work.

[0041] Individual Similarity Network Construction Module: This module uses a similarity network fusion method based on a hierarchical spatial fusion strategy to construct an individual similarity network. Processed clinical and genomic data are integrated, and a hierarchical spatial fusion strategy is introduced. Through iteration and integration, an individual similarity network is generated that fully preserves the similarities between individuals and the correlation characteristics between data. This network serves as the input to the residual graph attention network model of the biometric analysis module for biometric analysis.

[0042] Biometric Feature Extraction Module: This module uses a multimodal autoencoder fusion-based method for feature extraction. Separate autoencoders are constructed for clinical and genomic data, mapping the raw data into a low-dimensional latent space. The integration layer then integrates the features of the two modalities to produce an integration matrix. A decoder then maps the integration matrix back into the original input space to obtain a decentralized feature representation. Finally, these feature representations are combined through matrix concatenation to produce a more representative and robust feature representation, namely the multimodal data feature matrix.

[0043] Biometric Analysis Module: Utilizes the residual graph attention network model for biometric analysis. First, the obtained individual similarity network feature matrix and multimodal data feature matrix are used as model inputs. By designing a two-layer graph convolutional neural network model, the self-learning ability of the neural network is utilized to perform gradient descent on the input, and the final prediction result is obtained through iterative correction. To improve the accuracy of feature aggregation, an attention mechanism is introduced to dynamically adjust the contribution of neighboring nodes and perform weighted processing on the input features. Residual connections are designed and applied to the two-layer graph convolutional neural network model that incorporates the attention mechanism to enhance the feature expression ability of the original sample and form the final residual graph attention network model. Finally, the cross-entropy loss function is used to quantify the difference between the predicted and true labels to optimize model performance. After completing the construction and optimization of the residual graph attention network model, the model is used to conduct in-depth analysis of the characteristics of the biological sample, and finally output the predicted feature label of the biological sample.

[0044] The embodiment of the present application also provides a biometric feature analysis method based on multimodal data and residual graph attention network, which specifically includes the following steps:

[0045] Step (1) Data preprocessing is performed through the multimodal data processing module.

[0046] Data preprocessing ensures data quality and provides effective information for the model. For raw clinical data acquired based on the application scenario, we first use methods such as median filtering, wavelet transform, or convolutional neural networks to remove noise from the image data. We then perform image registration on data from different devices or time points. Next, we perform image cropping and normalization on the areas of key interest for prediction. Furthermore, when the raw clinical data lacks sufficient data samples, we apply data augmentation techniques to rotate, translate, and scale the images to expand the training dataset.

[0047] For the original genomic data obtained, which contains information such as gene expression and DNA methylation, we first remove missing values, standardize and normalize them, and interpolate the missing values. Then, this example uses the L1 regularized linear support vector machine algorithm for feature selection. Assuming the dataset is The CCP has Data examples:

[0048]

[0049] In the formula For the data instances with Gene features and a label value , Specifically expressed as:

[0050]

[0051] In the formula For the The value of each feature. The sparse solution of the feature value is used for feature selection. The mathematical expression is:

[0052]

[0053] Where, is the weight vector, is the bias term, is the error penalty parameter. The algorithm uses To control the sparsity of genomic data, The larger the value, the more features are selected. Finally, one-hot encoding is used to encode the original categorical data and convert it into numerical data as the input of the subsequent model.

[0054] Step (2) constructs the individual similarity network through the individual similarity network construction module.

[0055] After completing the processing of multimodal data, this example uses the similarity network fusion method to construct an individual similarity network. To accomplish this task, the original data is first converted into a graph data structure, and a hierarchical spatial fusion strategy is used to effectively integrate the correlation between multimodal data. The process is as follows Figure 2 The specific steps are as follows:

[0056] Graph data structures can generally be used To express, is a node set, which in this instance represents an individual set. Representative Name individual. is the set of edges, Indicates connection and The edge is used to express the similarity between nodes. Its calculation formula is expressed as:

[0057]

[0058] Where, and Represents individual and The range of cosine similarity is , where 1 means completely similar and -1 means completely dissimilar. Total individuals, then we can determine the edge set matrix , and define the normalized similarity matrix of all individuals and the local similarity matrix for:

[0059]

[0060]

[0061] Where, For nodes The set of neighbor nodes of . Indicates the In the iteration Normalized similarity matrix of type data. To fully utilize the complementary information of multimodal data, this example introduces a hierarchical spatial fusion strategy. Data of different modalities are fused at different levels. For clinical data and genomic data, basic features (such as clinical indicators and gene expression levels) are extracted at the low level, and semantic features (such as disease classification and gene function annotation) are extracted at the high level. Cosine similarity is calculated at each level to form multiple similarity matrices. Through cross-validation, the similarity matrices at different levels are weighted and fused. The iterative fusion formula is:

[0062]

[0063]

[0064] Where, Representative Local similarity matrix of type data, Represents the total data type, Indicates the In the iteration Layer local similarity matrix, It is The weights of the layers, The similarity network fusion method iteratively fuses local similarity matrices at different levels according to this formula. Each iteration performs a normalization operation to ensure that the similarity between an individual and itself is always higher than the similarity with other individuals. Finally, an individual similarity network is formed, and the individual similarity network feature matrix is ​​obtained. , which contains the association information between different types of data. In this example, this method is used to utilize the individual similarity network of clinical data and genomic data as the input of the subsequent model for prediction.

[0065] Step (3) performs feature extraction through the multimodal autoencoder of the biometric feature extraction module.

[0066] After completing the processing of multimodal data, this example completes the feature extraction work by constructing a multimodal autoencoder. Multimodal Autoencoder (MAE) is an unsupervised learning neural network model whose goal is to learn an effective representation of input data, thereby achieving dimensionality reduction and feature extraction of high-dimensional data. The multimodal autoencoder MAE constructed by the present invention consists of multiple autoencoders and a decoder. Each encoder is responsible for processing data of one modality, and the decoder integrates the encoded features of all modalities and reconstructs the original data. The flow chart is as follows Figure 3 shown.

[0067] First, each autoencoder maps the input data of different modalities into a latent space, capturing important features in the data and generating low-dimensional feature representations. In this example, the encoder structure is designed based on the characteristics of different data modalities. For clinical image data, a convolutional neural network is used as the autoencoder; for genomic data with serialized characteristics, a recurrent neural network architecture is used as the autoencoder. The encoding process of the autoencoder can be expressed as:

[0068]

[0069] in, Indicates the The weight matrix of the encoder, is the encoder bias term, Is the activation function, generally used 、 and To express.

[0070] The decoder maps the integrated matrix back to the original input space to obtain the decentralized feature representation.

[0071] Next, the integration layer integrates the compressed feature representations generated by the autoencoder to obtain an integration matrix. The decoder maps this integration matrix to the original input data, restoring the original multimodal data as much as possible. The fully connected layer of the decoder is constructed to decode the output features of each encoder into the original dimension. The decoding process can be expressed as:

[0072]

[0073] Where, is the weight matrix of the decoder, is the decoder bias term, is the activation function, Is the reconstructed feature matrix. In order to minimize the error between the input data and the reconstructed output data, the mean square error is used to optimize the objective function:

[0074]

[0075] Where, is the number of features of the sample, The actual value is Features, is the corresponding reconstructed output. The feature matrix after processing by the multimodal autoencoder , which will be used as input for the next step of biometric analysis.

[0076] Step (4) performs biometric analysis through the residual graph attention network model of the biometric analysis module.

[0077] This example designs a residual graph attention network model. First, a basic two-layer graph convolutional neural network is designed to perform gradient descent on the input. In order to improve the accuracy of feature aggregation, an attention mechanism is introduced to dynamically adjust the contribution of neighboring nodes and perform weighted processing on the input features. Residual connections are designed and applied to the two-layer graph convolutional neural network model that incorporates the attention mechanism to form the final residual graph attention network model. It can combine the individual similarity network and the multimodal data feature matrix to learn the potential representation and finally obtain the predicted label result. The flowchart is as follows: Figure 4 shown.

[0078] First, for the individual similarity network feature matrix input in step 2 And the multimodal data feature matrix obtained in step 3 The individual similarity network feature matrix contains the similarity information between individuals, and the multimodal data feature matrix contains the clinical data and genomic data features after fusion by the multimodal autoencoder. The node features learned by the two-layer graph convolutional neural network model are:

[0079]

[0080] Where, is the weight parameter matrix of the graph convolutional neural network, is the activation function, Embed the output features of the previous layer of graph convolutional neural network; extract features through a multi-head self-attention mechanism, where each head focuses on a different part of the feature subspace, assigns weights to each neighbor node, and dynamically adjusts the contribution of each neighbor by calculating the relationship between the node and its neighbors to improve model performance; at this point, the update formula becomes:

[0081]

[0082] Where, is a node and The attention coefficient between is calculated as follows:

[0083]

[0084] in, It is Nodes in the head and The energy value between and bias Calculation yields:

[0085]

[0086] Where, is the input feature vector, Is a learnable vector; by weighted summing the features of neighboring nodes, the updated node features are obtained ; Then, the node feature representation after multi-head self-attention processing is concatenated with the multimodal feature matrix, and the activation function is used Complete feature propagation, the formula is:

[0087]

[0088] On this basis, this example introduces residual connections to further enhance the feature expression ability of the original sample and form the final residual graph attention network model. Through an independent fully connected layer, the dimension of the input feature is ensured to be consistent with the node feature dimension of the network, and a residual connection is made with the output:

[0089]

[0090] Where, is the feature representation after residual connection, Represents the multimodal feature matrix Passed to two fully connected layers. Then, all processed nodes are connected through the fully connected layer for prediction:

[0091]

[0092] in, is the final prediction output of the residual graph attention network model.

[0093] During the training process, the cross entropy loss function is used to quantify the prediction loss of the model. It can measure the difference between the model's prediction results and the true value, expressed as:

[0094]

[0095] Where, The number of categories for classifying specific symptoms depends on the specific application scenario; Is the sign function, set to 0 or 1, when the sample The true category is equal to 1 when it is, otherwise 0; Representation sample Belong to category The predicted probability of .

[0096] After completing the construction and optimization of the above residual graph attention network model, the trained residual graph attention network model is used to perform in-depth analysis of the clinical images and genomic data features of the biological samples to be analyzed, and finally output the predicted feature labels of the biological samples .

[0097] The present application also provides a computer-readable storage medium having computer program instructions stored thereon, wherein the computer program instructions implement the above method when executed by a processor. The computer-readable storage medium may be a non-volatile computer-readable storage medium.

[0098] An embodiment of the present application also provides an electronic device, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to perform the above method.

[0099] The electronic device may be provided as a terminal, a server, or other forms of devices.

[0100] The present invention may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present invention.

[0101] A computer-readable storage medium can be a tangible device that can hold and store instructions for use by an instruction execution device. A computer-readable storage medium can be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination thereof. As used herein, a computer-readable storage medium is not to be construed as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through an electrical wire.

[0102] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions to be stored in the computer-readable storage medium in each computing / processing device.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Ordinary technicians in the relevant field can still modify or replace the specific implementation methods of the present invention with equivalents by referring to the above embodiments. Any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention are within the scope of protection of the claims of the present invention to be approved.

Claims

1. A biometric analysis system based on multimodal data and residual graph attention network, characterized by: Includes the following modules: Multimodal data processing module: pre-processes the collected raw clinical data and raw genomic data to construct a data set; Individual similarity network construction module: Use the similarity network fusion method based on the hierarchical space fusion strategy to construct the individual similarity network; Biometric feature extraction module: uses a method based on multimodal autoencoder fusion to extract features and obtain a multimodal data feature matrix; Biometric analysis module: The residual graph attention network model is used to perform biometric analysis based on the obtained individual similarity network feature matrix and multimodal data feature matrix.

2. The biometric feature analysis system based on multimodal data and residual graph attention network according to claim 1, characterized in that The multimodal data processing module cleans the raw data through outlier detection and missing value processing methods, performs feature selection on genomic data using the L1 regularized linear support vector machine algorithm, and constructs an effective data set for subsequent work.

3. The biometric feature analysis system based on multimodal data and residual graph attention network according to claim 1, characterized in that The individual similarity network construction module integrates the processed clinical data and genomic data, introduces a hierarchical spatial fusion strategy, and through iteration and integration, generates an individual similarity network that fully retains the similarity between biological individuals and the correlation characteristics between data. It serves as the input of the residual graph attention network model of the biometric analysis module for biometric analysis.

4. The biometric feature analysis system based on multimodal data and residual graph attention network according to claim 1, characterized in that The biometric feature extraction module constructs independent autoencoders for clinical data and genomic data, maps the original data into a low-dimensional latent space, and then integrates the features of the two modalities through an integration layer to obtain an integration matrix; then, the decoder maps the integration matrix back to the original input space to obtain a dispersed feature representation, and finally, these feature representations are merged through matrix splicing to obtain a more representative and robust feature representation, namely the multimodal data feature matrix.

5. The biometric feature analysis system based on multimodal data and residual graph attention network according to claim 1, characterized in that The biometric analysis module is specifically implemented as follows: First, the obtained individual similarity network feature matrix and multimodal data feature matrix are used as model inputs. A two-layer graph convolutional neural network model is designed. The self-learning ability of the neural network is utilized to perform gradient descent on the input, and the final prediction result is obtained through iterative correction. To improve the accuracy of feature aggregation, an attention mechanism is introduced to dynamically adjust the contribution of neighboring nodes and perform weighted processing on the input features. A residual connection is designed and applied to a two-layer graph convolutional neural network model that incorporates an attention mechanism to enhance the feature expression ability of the original sample and form the final residual graph attention network model; finally, the difference between the predicted and true labels is quantified through the cross-entropy loss function to optimize the model performance; after completing the construction and optimization of the residual graph attention network model, the model is used to conduct in-depth analysis of the characteristics of biological samples and finally output the predicted feature labels of the biological samples.

6. A biometric feature analysis method based on multimodal data and residual graph attention network, characterized in that: The following steps are involved: Multimodal data processing: Preprocess the collected raw clinical data and raw genomic data to construct a data set; clean the raw data through outlier detection and missing value processing methods, and perform feature selection on the genomic data using the linear support vector machine algorithm with L1 regularization to construct a valid data set for subsequent work; Individual similarity network construction: A similarity network fusion method based on a hierarchical spatial fusion strategy is used to construct an individual similarity network. The processed clinical data and genomic data are integrated, and a hierarchical spatial fusion strategy is introduced. Through iteration and integration, an individual similarity network is generated that fully preserves the similarities between biological individuals and the correlation characteristics between data. Biometric feature extraction: A multimodal autoencoder fusion method is used to extract features and obtain a multimodal data feature matrix. Biometric analysis: The residual graph attention network model is used to perform biometric analysis based on the obtained individual similarity network feature matrix and multimodal data feature matrix.

7. The biometric feature analysis method based on multimodal data and residual graph attention network according to claim 6, characterized in that The specific biological feature extraction is as follows: independent autoencoders are constructed for clinical data and genomic data, the original data is mapped into a low-dimensional latent space, and the features of the two modalities are integrated through the integration layer to obtain an integrated matrix; then, the decoder maps the integrated matrix back to the original input space to obtain a dispersed feature representation, and finally, these feature representations are merged through matrix splicing to obtain a more representative and robust feature representation, namely the multimodal data feature matrix.

8. The biometric feature analysis method based on multimodal data and residual graph attention network according to claim 6, characterized in that The biometric analysis is as follows: First, the obtained individual similarity network feature matrix and multimodal data feature matrix are used as model inputs. A two-layer graph convolutional neural network model is designed. The self-learning ability of the neural network is utilized to perform gradient descent on the input, and the final prediction result is obtained through iterative correction. To improve the accuracy of feature aggregation, an attention mechanism is introduced to dynamically adjust the contribution of neighboring nodes and perform weighted processing on the input features. A residual connection is designed and applied to a two-layer graph convolutional neural network model that incorporates an attention mechanism to enhance the feature expression ability of the original sample and form the final residual graph attention network model; finally, the difference between the predicted and true labels is quantified through the cross-entropy loss function to optimize the model performance; after completing the construction and optimization of the residual graph attention network model, the model is used to conduct in-depth analysis of the characteristics of biological samples and finally output the predicted feature labels of the biological samples.

9. An electronic device, characterized in that: It includes: processor; A memory for storing processor-executable instructions; wherein the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 6 to 8.

10. A computer-readable storage medium, characterized in that The storage medium stores computer program instructions, and the computer program instructions are executed by a processor to implement the method according to any one of claims 6 to 8.

Images