Gastric cancer postoperative MRD recurrence early warning system based on multi-source data fusion

Abstract

The invention relates to the technical field of tumor data fusion, and particularly discloses a stomach cancer postoperative MRD recurrence early warning system based on multi-source data fusion, which is used for solving the problems that in stomach cancer postoperative MRD monitoring, the multi-modal data dimension and structure difference is large, missing values and noise are common, especially high-dimensional genome data features are few, and the MRD recurrence early warning effect is poor. The system solves the problem that data fusion and key feature extraction face a large challenge, and comprises a data acquisition module, a data preprocessing module, a machine learning module, a risk early warning module, a multi-mode auto-encoder module, a feature weighted fusion module, a Banach space embedding module and a Fisher-Rao distance measurement module, the data acquisition module is connected with the data preprocessing module, and the data preprocessing module is connected with the Fisher-Rao distance measurement module. The data preprocessing module is connected with the machine learning module, the machine learning module is connected with the risk early warning module, key features are extracted through a multi-modal self-encoder, and under optimization of feature weighted fusion, Banach space embedding and Fisher-Rao distance measurement, modal difference and information loss are reduced, and the MRD recurrence risk is accurately evaluated.

Description

Technical Field

[0001] The present invention relates to the technical field of tumor data fusion, and more specifically, to a postoperative MRD recurrence warning system for gastric cancer based on multi-source data fusion. Background Art

[0002] In the monitoring of MRD (minimal residual disease) after gastric cancer surgery, multimodal data are derived from genomic, transcriptomic, proteomic, metabolomic and clinical information. These data have different dimensions and structures, and there are significant differences between the modalities (e.g., genomic data is in sequence form, while proteomic data is concentration expression). Therefore, direct comparison and integration are challenging. At the same time, missing values ​​and noise will appear during the data collection process. In particular, genomic data has extremely high dimensions and few effective features, which further increases the difficulty of data fusion and calculation. Therefore, how to process missing data, remove noise and effectively extract key features from multi-source data has become a core technical issue to improve the performance of monitoring models. Summary of the Invention

[0003] In order to overcome the above-mentioned defects of the prior art, the present invention provides a MRD recurrence warning system for gastric cancer surgery based on multi-source data fusion. Key features are extracted through a multimodal autoencoder, and modality differences and information loss are reduced through feature weighted fusion, Banach space embedding and optimization of Fisher-Rao distance metric. Ultimately, a machine learning model is used to accurately monitor MRD status and assess recurrence risk, thereby improving prediction accuracy.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] A gastric cancer postoperative MRD recurrence warning system based on multi-source data fusion includes a data acquisition module, a data preprocessing module, a machine learning module and a risk warning module, and also includes a multimodal autoencoder module, a feature weighted fusion module, a Banach space embedding module and a Fisher-Rao distance measurement module. The data acquisition module is connected to the data preprocessing module, the data preprocessing module is connected to the machine learning module, and the machine learning module is connected to the risk warning module. The multimodal autoencoder module extracts key features from the multimodal data and maps them to a unified low-dimensional space to form a feature representation. The feature weighted fusion module extracts key features from the multimodal data and maps them to a unified low-dimensional space to form a feature representation. The importance of modal features is weighted and fused to generate a unified feature representation. The Banach space embedding module embeds the weighted fused features into the Banach space to handle the complexity and high dimensionality of different modal data and ensure the smoothness of the features. The Fisher-Rao distance measurement module uses the Fisher-Rao distance measurement to measure the difference between the fused modalities, optimize the fused feature representation, and reduce the difference between the modalities by distributing the data of different modalities. The machine learning module inputs the fused data into the deep neural network and conducts MRD monitoring and risk assessment by learning the key features of different modalities.

[0006] As a further solution of the present invention, the execution process of the data acquisition module collecting multi-source data includes:

[0007] Step 11, initialization phase: When the system is initially started, it first loads the predetermined configuration file and connects to the required database and medical equipment according to the system requirements. After the configuration is loaded, the data acquisition module is activated and ready to communicate with different data sources, waiting for the start of the data acquisition task;

[0008] Step 12, data source connection and verification: The data acquisition module connects multiple data sources of genomics, transcriptomics, proteomics, metabolomics, and patient clinical information, and verifies the connection stability of the data sources;

[0009] Step 13, execute data collection task: according to the collection frequency configured by the system, activate the data collection task regularly, automatically capture data from different data sources through predefined APIs and data interfaces, and cache them in temporary storage space.

[0010] As a further solution of the present invention, the workflow of the data preprocessing module receiving the multi-source data collected by the data acquisition module and performing preprocessing includes:

[0011] Step 21, data format standardization: Use one-hot encoding to convert each base in the genomic data into a corresponding numerical vector, convert the gene information in sequence form into a numerical matrix, and directly input the transcriptome, proteome, and metabolome data into a numerical matrix;

[0012] Step 22, data cleaning: noise monitoring and outlier identification of the data, using statistical methods to detect extreme values ​​and correct or delete them, and using filters to remove noise that occurs during the collection process;

[0013] Step 23, missing data processing: linear interpolation is used to fill missing data points for continuous data, and low-rank matrix decomposition completion method is used to fill missing values ​​for high-dimensional data;

[0014] Step 24, data standardization and normalization: standardize the data to a mean of 0 and a standard deviation of 1 to eliminate the dimensional differences of data of different modalities and scale the data to [0, 1];

[0015] Step 25, data dimensionality reduction: use linear discriminant analysis to reduce the dimensionality of high-dimensional data and retain key features;

[0016] Step 26, data balancing: oversample the less common categories in the data to increase the number of samples of that category, and undersample the more common categories in the data to reduce the number of samples of that type.

[0017] As a further solution of the present invention, the multimodal autoencoder module processes genomic data with spatial structure through convolutional autoencoders, extracts local patterns in sequences, uses variational autoencoders to process transcriptome and metabolome data, learns latent variable distributions to capture complex relationships in the data, uses denoising autoencoders to process patients' clinical information, helps remove missing values ​​of noise in measurements, uses sparse autoencoders to process high-dimensional sparse genomic data, and extracts key features through sparse regularization.

[0018] As a further solution of the present invention, in the multimodal autoencoder module, the autoencoder networks for processing genomic and transcriptomic data share the convolution kernel parameters of the convolution layer to extract local information in the sequence, and the autoencoder networks for processing proteomic and metabolomic data share the fully connected layers and sparse layers in the encoder layer to capture biologically relevant metabolic pathways.

[0019] As a further solution of the present invention, in the multimodal autoencoder module, the variational autoencoders of genomic data and transcriptomic data share the mean and variance parameters in the latent space, generating a unified latent variable distribution, predicting the expression of the transcriptome from the genomic data, and predicting the abundance of the proteome through the metabolomic data.

[0020] As a further solution of the present invention, in the multimodal autoencoder module, the variational autoencoder network of clinical information and metabolomics data shares convolutional layers and fully connected layers to extract common information in patient data, while the coefficient regularization layer and denoising layer retain the modality-specific feature learning ability and capture the special information in the metabolomics.

[0021] As a further solution of the present invention, the execution process of the multimodal autoencoder module to generate feature representations for genomic and transcriptomic data is as follows:

[0022] Using Convolutional Autoencoders to process genomic data:

[0023] Step 311, input data: input genome data;

[0024] Step 312, convolutional layer parameter sharing: Use shared convolution kernels to extract local features from genomic data. Multiple convolution layers extract high-level sequence patterns layer by layer. For genomic and transcriptomic data, convolution kernel parameters are shared in the convolution layers to extract similar sequence information.

[0025] Step 313, generating feature representation: the feature map output by the convolutional layer is further compressed through the pooling layer to generate a low-dimensional feature representation that represents the potential features of the genome;

[0026] Using variational autoencoders to process transcriptome data:

[0027] Step 321, input data: input transcriptome data;

[0028] Step 322, sharing latent variables: Through shared convolutional layers and fully connected layers, the features of transcriptome data and genomic data are mapped to the same latent space. The shared encoder generates joint mean and variance parameters to form a cross-modal latent variable distribution.

[0029] Step 323, generating feature representation: generating a low-dimensional feature representation of the transcriptome data by sampling latent variables, which shares a similar distribution with the genomic data in the latent space for cross-modal reasoning;

[0030] As a further solution of the present invention, the execution process of the multimodal autoencoder module for generating feature representations for multiple data sources of proteome and metabolome, as well as clinical information data of patients is as follows:

[0031] Using sparse encoders on proteomic data:

[0032] Step 331, input data: input proteomic data;

[0033] Step 332, parameter sharing: the shared encoding layer extracts key features in the protein data through sparse regularization, and the fully connected layers and sparse layers shared by the proteomic and metabolomic data in the encoder capture related biological pathways;

[0034] Step 333, generating feature representation: generating feature representation of the proteome through sparse coding to capture key nodes in protein expression;

[0035] Using variational autoencoders to process metabolomics data:

[0036] Step 341, input data: input metabolite abundance data of the metabolome;

[0037] Step 342, parameter sharing and regularization: The metabolomics data and the proteomics data share the mean and variance of the latent space to generate the same latent variable distribution. A regularization layer is used in the encoder to extract common information between the modalities while maintaining the specificity of the metabolomics data.

[0038] Step 343, generating feature representation: generating feature representation of metabolomics data by variational autoencoding, and capturing common features with proteomics in the latent space;

[0039] Using Denoising Autoencoders to Process Clinical Information:

[0040] Step 351, input data: input the patient's clinical information;

[0041] Step 352, denoising and feature extraction: The clinical information is noised during the encoding phase, and the denoising autoencoder generates a noise-free latent representation by learning the denoising capability. After the denoising layer, the core features of the patient's clinical information are extracted for integration with the biological data.

[0042] Step 353, generating feature representation: generating a denoised feature representation of the patient's clinical information.

[0043] As a further solution of the present invention, the feature weighted fusion module, the Banach space embedding module, the Fisher-Rao distance measurement module, the machine learning module, and the risk warning module perform feature fusion based on the multi-source data feature representation generated by the multimodal autoencoder module, and perform risk warning through the machine learning model. The specific implementation process includes:

[0044] Step 41, weighted feature fusion: accepting the genomic feature representation, transcriptomic feature representation, proteomic feature representation, metabolomic feature representation, and clinical information group feature representation generated by the multimodal autoencoder for weighted fusion, assigning weights to each modality in a predefined manner, performing weighted fusion, and generating a unified feature representation;

[0045] Step 42, Banach space embedding: mapping the fused feature representation into the Banach space to generate a feature representation embedded in the Banach space;

[0046] Step 43, Fisher-Rao distance metric: Receive the feature representation embedded in the Banach space obtained from step 42, use the Fisher-Rao distance metric to measure the difference in feature distribution between different modalities, optimize the feature representation embedded in the Banach space, and obtain the final feature representation of multi-source data fusion;

[0047] Step 44, machine learning prediction: The feature representation of the final multi-source data fusion obtained in step 43 is input into the trained deep neural network to obtain the predicted risk probability of MRD recurrence after gastric cancer surgery;

[0048] Step 45, MRD recurrence warning after gastric cancer surgery: When the predicted risk probability of MRD recurrence after gastric cancer surgery exceeds a predetermined threshold, a recurrence warning is issued and relevant medical personnel are notified.

[0049] Compared with the existing technology, the technical effect of the present invention is as follows: the present invention can extract key features from different data modalities of genome, transcriptome, proteome, metabolome and patient medical information through a multimodal autoencoder module, and map them to a unified low-dimensional space, thereby reducing modal differences. The weights are adjusted through a feature weighted fusion module to achieve optimized fusion of different modal features, making the key features more prominent. The Banach space embedding module further ensures the smoothness and consistency of high-dimensional data. The Fisher-Rao distance measurement module optimizes the distribution differences between modalities and reduces the information loss generated during the fusion process. Finally, the machine learning module accurately monitors the MRD status and evaluates the risk of recurrence, significantly improving the performance and prediction accuracy of the model, and can effectively solve the problems of large differences between multimodal data, complex data dimensions, missing values ​​and noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a system block diagram of the present invention. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the accompanying drawings. Obviously, the technical solutions described are only part of the present invention, not the entire invention. Based on the technical solutions of the present invention, all other technical solutions obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0052] like Figure 1As shown, the present invention proposes an MRD recurrence warning system for gastric cancer surgery based on multi-source data fusion, which includes a data acquisition module, a data preprocessing module, a machine learning module and a risk warning module, and also includes a multimodal autoencoder module, a feature weighted fusion module, a Banach space embedding module and a Fisher-Rao distance measurement module. The data acquisition module is connected to the data preprocessing module, the data preprocessing module is connected to the machine learning module, the machine learning module is connected to the risk warning module, the multimodal autoencoder module extracts key features from the multimodal data, maps them to a unified low-dimensional space to form a feature representation, and the feature weighted fusion module is connected to the machine learning module. The block strengthens the fusion of each modal feature according to its importance to generate a unified feature representation. The Banach space embedding module embeds the weighted fused features into the Banach space to handle the complexity and high dimensionality of different modal data and ensure the smoothness of the features. The Fisher-Rao distance measurement module uses the Fisher-Rao distance measurement to measure the difference between the fused modalities, optimize the fused feature representation, and reduce the difference between the modalities by distributing the data of different modalities. The machine learning module inputs the fused data into the deep neural network and performs MRD monitoring and risk assessment by learning the key features of different modalities.

[0053] Through the multimodal autoencoder module, the present invention can extract key features from different data modalities of genome, transcriptome, proteome, metabolome and patient medical information, and map them to a unified low-dimensional space, reducing modal differences. The weights are adjusted through the feature weighted fusion module to achieve optimized fusion of different modal features, making the key features more prominent. The Banach space embedding module further ensures the smoothness and consistency of high-dimensional data. The Fisher-Rao distance measurement module optimizes the distribution differences between modalities and reduces the information loss generated during the fusion process. Finally, the machine learning module accurately monitors MRD status and evaluates recurrence risk, significantly improving the performance and prediction accuracy of the model, and can effectively solve the problems of large differences between multimodal data, complex data dimensions, missing values ​​and noise.

[0054] It should be noted that the network layer structure of the convolutional autoencoder in the multimodal autoencoder module is:

[0055] Input layer: receives sequence data (genomic DNA sequence or transcriptome expression matrix).

[0056] Shared convolutional layer: Uses shared convolution kernels to extract local sequence features (genomic and transcriptomic data share these convolution kernel parameters), including convolution layer 1 and convolution layer 2. Convolution layer 1: extracts low-level local features (convolution kernel size is 3×3); convolution layer 2: extracts high-level local features (convolution kernel size is 5×5), and further combines the convolution outputs;

[0057] Pooling layer: After the shared convolution layer, the feature dimension is reduced through the maximum pooling operation.

[0058] Fully connected layer: Performs a fully connected operation on the features output by the convolutional layer and further compresses them into low-dimensional feature representation.

[0059] Decoding layer: decodes the compressed features and reconstructs the input data.

[0060] Output layer: Generates low-dimensional feature representations of the genome and transcriptome.

[0061] Genomic and transcriptomic data share similar sequence characteristics. Therefore, by sharing the convolutional kernel parameters in the convolutional layers, common local information can be extracted from the sequences, such as conserved regions of gene sequences or local patterns of gene expression. Sharing these convolutional layers reduces the number of model parameters and improves the efficiency of information fusion between different modalities.

[0062] The variational autoencoder is used to process proteomic and metabolomic data, learn the latent variable distribution and generate latent space representation. Its network structure is:

[0063] Input layer: receives proteomic or metabolomic data (numeric type, indicating expression abundance or metabolic concentration);

[0064] The shared encoder layer includes fully connected layer 1, fully connected layer 2, and sparse layer: Fully connected layer 1: linearly transforms the input data through the shared fully connected layer to generate a preliminary potential representation; fully connected layer 2: further extracts features to capture relevant read and write pathway features; sparse layer: applies sparse regularization to extract key features in the data, which is particularly suitable for high-dimensional and sparse proteomic data;

[0065] Latent space layer: Shares the latent variable distribution parameters of mean and variance, which are used to generate the potential representation of metabolome and proteome, including mean layer, variance layer and sampling layer: the mean layer calculates the mean of the latent variable, the variance layer calculates the variance of the latent variable, and the sampling layer samples the latent variable through reparameterization techniques;

[0066] Decoding layer: reconstructs proteomic or metabolomic data through the decoder;

[0067] Output layer: Generates a low-dimensional latent representation of the proteomic or metabolomic data.

[0068] Mechanism of sharing fully connected layers and sparse layers: There is a biological correlation between proteomic and metabolomic data (such as the correlation between metabolic pathways and protein expression). By sharing fully connected layers and sparse layers, these common features can be captured, thereby improving the model's understanding of multimodal data. Sparse layers extract key metabolite or protein expression features by applying regularization to high-dimensional data. Mechanism of sharing latent space: Genomic and transcriptomic data share the mean and variance parameters in the variational autoencoder, which means that the two data modalities are mapped to the same distribution in the latent space, so that the expression of the transcriptome can be predicted from the genomic data, and vice versa.

[0069] Sparse autoencoders are used to process proteomic data, especially high-dimensional sparse data. Their network layer structure is as follows:

[0070] Input layer: receives proteomic data (high-dimensional sparse data);

[0071] Fully connected layer: performs linear transformation on the input data and extracts sparse features;

[0072] Sparse regularization layer: Maintain feature sparsity through L1 regularization to extract key expressions in proteomic data;

[0073] Decoding layer: decodes sparse features and reconstructs the original input data;

[0074] Output layer: Generates low-dimensional feature representation of proteomic data.

[0075] The denoising autoencoder is used to process the patient's clinical information data, remove noise and complete missing data. Its network layer structure includes:

[0076] Input layer: receives clinical information;

[0077] Adding a noise layer: Artificially adding noise to simulate incomplete or erroneous data.

[0078] Shared convolutional layer: a convolutional layer shared with metabolomics data to extract local features from clinical information;

[0079] Fully connected layer: The shared fully connected layer extracts global features from clinical data and helps capture correlations with metabolomics data;

[0080] Denoising layer: removes noise, restores original data, and extracts clear features;

[0081] Decoding layer: reconstructs the denoised data;

[0082] Output layer: Generates denoised clinical information feature representation.

[0083] Mechanism of sharing convolutional and fully connected layers: Clinical information and metabolomics data may be biologically related. By sharing convolutional and fully connected layers, common features between the two can be captured. For example, the concentration of metabolites in the metabolome may be closely related to the patient's clinical status.

[0084] Denoising and Regularization Layers: The denoising layer is used to process noise in clinical data, ensuring that the generated feature representation is clearer and more accurate. The coefficient regularization layer preserves specific features in the data, ensuring that the differential characteristics of data from different modalities are retained and preventing the loss of important features due to over-fusion.

[0085] By sharing parameters between different layers (convolutional layer, fully connected layer, latent space) and modules, the system has the following technical effects: shared convolutional layers extract local information in the sequence and generate feature representations. The variational autoencoders of genomic and transcriptomic data share the mean and variance parameters in the latent space, and predict the expression of the transcriptome from the genomic data; by sharing fully connected layers and sparse layers, biologically relevant metabolic pathways are captured, and the abundance of the proteome is predicted through metabolomic data; by sharing convolutional layers and fully connected layers, common information in patient data is extracted, and the denoising layer retains the patient-specific feature learning ability and captures special information in the metabolome.

[0086] This structure improves the efficiency and performance of the model, ensuring the depth of data feature extraction and effectively realizing the fusion and reasoning of data from different modalities.

[0087] It should be noted that the execution process of the data acquisition module to collect multi-source data includes:

[0088] Step 11, initialization phase: When the system is initially started, it first loads the predetermined configuration file and connects to the required database and medical equipment according to the system requirements. After the configuration is loaded, the data acquisition module is activated and ready to communicate with different data sources, waiting for the start of the data acquisition task;

[0089] Step 12, data source connection and verification: The data acquisition module connects multiple data sources of genomics, transcriptomics, proteomics, metabolomics, and patient clinical information, and verifies the connection stability of the data sources;

[0090] Step 13, execute data collection task: according to the collection frequency configured by the system, activate the data collection task regularly, automatically capture data from different data sources through predefined APIs and data interfaces, and cache them in temporary storage space.

[0091] It should be noted that the workflow of the data preprocessing module receiving the multi-source data collected by the data acquisition module and performing preprocessing includes:

[0092] Step 21, data format standardization: Use one-hot encoding to convert each base in the genomic data into a corresponding numerical vector, convert the gene information in sequence form into a numerical matrix, and directly input the transcriptome, proteome, and metabolome data into a numerical matrix;

[0093] Step 22, data cleaning: noise monitoring and outlier identification of the data, using statistical methods to detect extreme values ​​and correct or delete them, and using filters to remove noise that occurs during the collection process;

[0094] Step 23, missing data processing: linear interpolation is used to fill missing data points for continuous data, and low-rank matrix decomposition completion method is used to fill missing values ​​for high-dimensional data;

[0095] Step 24, data standardization and normalization: standardize the data to a mean of 0 and a standard deviation of 1 to eliminate the dimensional differences of data of different modalities and scale the data to [0, 1];

[0096] Step 25, data dimensionality reduction: use linear discriminant analysis to reduce the dimensionality of high-dimensional data and retain key features;

[0097] Step 26, data balancing: oversample the less common categories in the data to increase the number of samples of that category, and undersample the more common categories in the data to reduce the number of samples of that type.

[0098] It should be noted that the multimodal autoencoder module processes spatially structured genomic data using convolutional autoencoders to extract local patterns in sequences. It also uses variational autoencoders for transcriptomic and metabolomic data, learning latent variable distributions to capture complex relationships in the data. A denoising autoencoder processes patient clinical information to help remove missing values ​​from measurement noise. A sparse autoencoder processes high-dimensional, sparse genomic data, extracting key features through sparsity regularization. The autoencoder networks for genomic and transcriptomic data share convolutional kernel parameters to extract local information from sequences. The autoencoder networks for proteomic and metabolomic data share fully connected and sparse layers in the encoder layer to capture biologically relevant metabolic pathways. The variational autoencoders for genomic and transcriptomic data share the mean and variance parameters in the latent space, generating a unified latent variable distribution. This allows for prediction of transcriptomic expression from genomic data and proteomic abundance from metabolomic data. The variational autoencoder network of clinical information and metabolomics data shares convolutional layers and fully connected layers to extract common information in patient data. The coefficient regularization layer and denoising layer retain the modality-specific feature learning ability and capture the special information in the metabolomics.

[0099] Convolutional autoencoders process spatially structured genomic data, extracting local patterns within sequences and capturing important biological information within the data. This enables the model to accurately identify specific local features in genomic and transcriptomic data. Variational autoencoders, by sharing a latent space, learn the distribution of latent variables across the genome, transcriptome, and metabolome, effectively capturing the complex relationships between data modalities. This allows for prediction of transcriptomic expression from genomic data and inference of proteomic abundance from metabolomic data, improving the correlation and reasoning capabilities between data modalities. Denoising autoencoders remove noise and fill in missing values ​​when processing clinical data, ensuring data integrity and accuracy, and improving the performance of subsequent models for recurrence risk assessment. Sparse autoencoders extract key features from high-dimensional sparse data through sparse regularization, addressing the challenges of high dimensionality and sparse information in genomic and proteomic data and effectively reducing data redundancy. The genome and transcriptome share convolutional kernel parameters, while the proteome and metabolome share fully connected and sparse layers, reducing model redundancy between modalities and improving model learning efficiency and generalization. This sharing mechanism also facilitates information exchange and feature synergy between modalities, enabling multimodal data to collectively support disease prediction and decision-making. By sharing latent spaces and feature layers, data from different modalities can achieve cross-modal reasoning, particularly within biological pathways, enabling the prediction of proteomic features from metabolomic data. This cross-modal prediction improves the effectiveness of correlation analysis between data, enabling the system to more accurately capture key biomarkers in disease onset and progression.

[0100] It should be noted that the execution process of the multimodal autoencoder module to generate feature representations for genomic and transcriptomic data is as follows:

[0101] Using Convolutional Autoencoders to process genomic data:

[0102] Step 311, input data: input genome data;

[0103] Step 312, convolutional layer parameter sharing: Use shared convolution kernels to extract local features from genomic data. Multiple convolution layers extract high-level sequence patterns layer by layer. For genomic and transcriptomic data, convolution kernel parameters are shared in the convolution layers to extract similar sequence information.

[0104] Step 313, generating feature representation: the feature map output by the convolutional layer is further compressed through the pooling layer to generate a low-dimensional feature representation that represents the potential features of the genome;

[0105] Using variational autoencoders to process transcriptome data:

[0106] Step 321, input data: input transcriptome data;

[0107] Step 322, sharing latent variables: Through shared convolutional layers and fully connected layers, the features of transcriptome data and genomic data are mapped to the same latent space. The shared encoder generates joint mean and variance parameters to form a cross-modal latent variable distribution.

[0108] Step 323, generating feature representation: generating a low-dimensional feature representation of the transcriptome data by sampling latent variables, which shares a similar distribution with the genomic data in the latent space for cross-modal reasoning;

[0109] It should be noted that the execution process of the multimodal autoencoder module for generating feature representations for multiple data sources of proteomics and metabolomics, as well as patient clinical information data, is as follows:

[0110] Using sparse encoders on proteomic data:

[0111] Step 331, input data: input proteomic data;

[0112] Step 332, parameter sharing: the shared encoding layer extracts key features in the protein data through sparse regularization, and the fully connected layers and sparse layers shared by the proteomic and metabolomic data in the encoder capture related biological pathways;

[0113] Step 333, generating feature representation: generating feature representation of the proteome through sparse coding to capture key nodes in protein expression;

[0114] Using variational autoencoders to process metabolomics data:

[0115] Step 341, input data: input metabolite abundance data of the metabolome;

[0116] Step 342, parameter sharing and regularization: The metabolomics data and the proteomics data share the mean and variance of the latent space to generate the same latent variable distribution. A regularization layer is used in the encoder to extract common information between the modalities while maintaining the specificity of the metabolomics data.

[0117] Step 343, generating feature representation: generating feature representation of metabolomics data by variational autoencoding, and capturing common features with proteomics in the latent space;

[0118] Using Denoising Autoencoders to Process Clinical Information:

[0119] Step 351, input data: input the patient's clinical information;

[0120] Step 352, denoising and feature extraction: The clinical information is noised during the encoding phase, and the denoising autoencoder generates a noise-free latent representation by learning the denoising capability. After the denoising layer, the core features of the patient's clinical information are extracted for integration with the biological data.

[0121] Step 353, generating feature representation: generating a denoised feature representation of the patient's clinical information.

[0122] It should be noted that the feature weighted fusion module, Banach space embedding module, Fisher-Rao distance measurement module, machine learning module, and risk warning module perform feature fusion based on the multi-source data feature representation generated by the multimodal autoencoder module, and the specific implementation process of risk warning through the machine learning model includes:

[0123] Step 41, weighted feature fusion: accepting the genomic feature representation, transcriptomic feature representation, proteomic feature representation, metabolomic feature representation, and clinical information group feature representation generated by the multimodal autoencoder for weighted fusion, assigning weights to each modality in a predefined manner, performing weighted fusion, and generating a unified feature representation;

[0124] Step 42, Banach space embedding: mapping the fused feature representation into the Banach space to generate a feature representation embedded in the Banach space;

[0125] Step 43, Fisher-Rao distance metric: Receive the feature representation embedded in the Banach space obtained from step 42, use the Fisher-Rao distance metric to measure the difference in feature distribution between different modalities, optimize the feature representation embedded in the Banach space, and obtain the final feature representation of multi-source data fusion;

[0126] Step 44, machine learning prediction: The feature representation of the final multi-source data fusion obtained in step 43 is input into the trained deep neural network to obtain the predicted risk probability of MRD recurrence after gastric cancer surgery;

[0127] Step 45, MRD recurrence warning after gastric cancer surgery: When the predicted risk probability of MRD recurrence after gastric cancer surgery exceeds a predetermined threshold, a recurrence warning is issued and relevant medical personnel are notified.

[0128] In order to clearly illustrate the specific implementation process of multi-source data fusion of this technical solution, the following examples are analyzed and explained.

[0129] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

[0130] Finally: The above description is only an illustration of the implementation mode of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A gastric cancer postoperative MRD recurrence early warning system based on multi-source data fusion, comprising a data acquisition module, a data preprocessing module, a machine learning module and a risk early warning module, characterized in that: It also includes a multimodal autoencoder module, a feature weighted fusion module, a Banach space embedding module and a Fisher-Rao distance measurement module. The data acquisition module is connected to the data preprocessing module, the data preprocessing module is connected to the machine learning module, and the machine learning module is connected to the risk warning module. The multimodal autoencoder module extracts key features from the multimodal data, maps them to a unified low-dimensional space, and forms a feature representation. The feature weighted fusion module enhances the fusion of each modal feature according to its importance to generate a unified feature representation. The Banach space embedding module embeds the weighted fused features into the Banach space to handle the complexity and high dimensionality of different modal data and ensure the smoothness of the features. The Fisher-Rao distance measurement module uses the Fisher-Rao distance to measure the difference between the fused modalities, optimizes the fused feature representation, and reduces the difference between the modalities by distributing the data of different modalities. The machine learning module inputs the fused data into a deep neural network, and performs MRD monitoring and risk assessment by learning the key features of different modalities.

2. The MRD recurrence warning system after gastric cancer surgery based on multi-source data fusion according to claim 1, characterized in that: The execution process of the data acquisition module to collect multi-source data includes: Step 11, initialization phase: when the system is initially started, first load the predetermined configuration file, connect the required database and medical equipment according to the system requirements, and after the configuration is loaded, activate the data acquisition module, prepare to communicate with different data sources, and wait for the start of the data acquisition task; Step 12, data source connection and verification: The data acquisition module connects multiple data sources of genomics, transcriptomics, proteomics, metabolomics, and the patient's clinical information, and verifies the connection stability of the data source; Step 13, execute data collection tasks: according to the collection frequency configured by the system, activate the data collection tasks regularly, automatically capture data from different data sources through predefined APIs and data interfaces, and cache them in temporary storage space.

3. The MRD recurrence warning system after gastric cancer surgery based on multi-source data fusion according to claim 2, characterized in that: The workflow of the data preprocessing module receiving the multi-source data collected by the data acquisition module and preprocessing includes: Step 21, data format standardization: use one-hot encoding to convert each base in the genomic data into a corresponding numerical vector, convert the gene information in sequence form into a numerical matrix, and directly input the data of the transcriptome, proteome, and metabolome into a numerical matrix; Step 22, data cleaning: noise monitoring and outlier identification of data, using statistical methods to detect extreme values ​​and correct or delete them, and using filters to remove noise that occurs during the collection process; Step 23, missing data processing: linear interpolation is used to fill missing data points for continuous data, and low-rank matrix decomposition completion method is used to fill missing values ​​for high-dimensional data; Step 24, data standardization and normalization: standardize the data with a mean of 0 and a standard deviation of 1, eliminate the dimensional differences of data of different modalities, and scale the data to [0,1]; Step 25, data dimensionality reduction: use linear discriminant analysis to reduce the dimension of high-dimensional data and retain key features; Step 26, data balancing: oversample the less common categories in the data to increase the number of samples of that category, and undersample the more common categories in the data to reduce the number of samples of that type.

4. The MRD recurrence warning system after gastric cancer surgery based on multi-source data fusion according to claim 3, characterized in that: The multimodal autoencoder module processes genomic data with spatial structure through convolutional autoencoders, extracts local patterns in sequences, uses variational autoencoders to process transcriptome and metabolome data, learns latent variable distributions to capture complex relationships in the data, uses denoising autoencoders to process patients' clinical information, helps remove missing values ​​of noise in measurements, uses sparse autoencoders to process high-dimensional and sparse genomic data, and extracts key features through sparse regularization.

5. The MRD recurrence warning system after gastric cancer surgery based on multi-source data fusion according to claim 4, characterized in that: In the multimodal autoencoder module, the autoencoder networks for processing genomic and transcriptomic data share the convolution kernel parameters of the convolutional layer to extract local information in the sequence, and the autoencoder networks for proteomic and metabolomic data share the fully connected layers and sparse layers in the encoder layer to capture biologically relevant metabolic pathways.

6. The MRD recurrence warning system after gastric cancer surgery based on multi-source data fusion according to claim 5, characterized in that: In the multimodal autoencoder module, the variational autoencoders of genomic data and transcriptomic data share the mean and variance parameters in the latent space, generating a unified latent variable distribution, predicting the expression of the transcriptome from the genomic data, and predicting the abundance of the proteome through the metabolomic data.

7. The MRD recurrence early warning system for gastric cancer after surgery based on multi-source data fusion according to claim 1, characterized in that: In the multimodal autoencoder module, the variational autoencoder network of clinical information and metabolomics data shares convolutional layers and fully connected layers to extract common information in patient data. The coefficient regularization layer and denoising layer retain the modality-specific feature learning ability and capture the special information in the metabolome.

8. The MRD recurrence early warning system for gastric cancer after surgery based on multi-source data fusion according to claim 7, characterized in that: The execution process of the multimodal autoencoder module to generate feature representations for genomic and transcriptomic data is as follows: Using Convolutional Autoencoders to process genomic data: Step 311, input data: input genome data; Step 312, convolution layer parameter sharing: using a shared convolution kernel to extract local features from genomic data, multiple convolution layers extract high-level sequence patterns layer by layer, and for genomic and transcriptomic data, convolution kernel parameters are shared in the convolution layer to extract similar sequence information; Step 313, generating feature representation: the feature map output by the convolutional layer is further compressed through the pooling layer to generate a low-dimensional feature representation, which represents the potential features of the genome; Using variational autoencoders to process transcriptome data: Step 321, input data: input transcriptome data; Step 322, sharing latent variables: Mapping the features of transcriptome data and genomic data to the same latent space through shared convolutional layers and fully connected layers, and the shared encoder generates joint mean and variance parameters to form a cross-modal latent variable distribution; Step 323, generate feature representation: generate a low-dimensional feature representation of the transcriptome data by sampling latent variables, which shares a similar distribution with the genomic data in the latent space for cross-modal reasoning.

9. The MRD recurrence warning system after gastric cancer surgery based on multi-source data fusion according to claim 7, characterized in that: The execution process of the multimodal autoencoder module to generate feature representation for multi-source data sources of proteomics and metabolomics, as well as clinical information data of patients is as follows: Use sparse encoders to process proteomic data: Step 331, input data: input proteomic data; Step 332, parameter sharing: the shared encoding layer extracts key features in the protein data through sparse regularization, and the fully connected layers and sparse layers shared by the proteome and metabolome data in the encoder capture related biological pathways; Step 333, generate feature representation: Generate feature representation of proteome through sparse coding to capture key nodes in protein expression; Using variational autoencoders to process metabolomics data: Step 341, input data: input metabolite abundance data of the metabolome; Step 342, parameter sharing and regularization: the metabolomics data and the proteomics data share the mean and variance of the latent space, generate the same latent variable distribution, and use a regularization layer in the encoder to extract the common information between the modalities while maintaining the specificity of the metabolomics data; Step 343, generating feature representation: generating feature representation of metabolome data by variational autoencoding, and capturing common features with proteome in latent space; Using Denoising Autoencoders to Process Clinical Information: Step 351, input data: input the patient's clinical information; Step 352, denoising and feature extraction: the clinical information is added with noise in the encoding stage, and the denoising autoencoder generates a noise-free latent representation by learning the denoising ability. After the denoising layer, the core features of the patient's clinical information are extracted for integration with the biological data; Step 353, generate feature representation: generate a denoised feature representation of the patient's clinical information.

10. The MRD recurrence warning system after gastric cancer surgery based on multi-source data fusion according to claim 1, characterized in that: The feature weighted fusion module, Banach space embedding module, Fisher-Rao distance measurement module, machine learning module and risk warning module perform feature fusion based on the multi-source data feature representation generated by the multimodal autoencoder module, and the specific implementation process of risk warning through the machine learning model includes: Step 41, feature weighted fusion: accept the genome feature representation, transcriptome feature representation, proteome feature representation, metabolome feature representation and clinical information group feature representation generated by the multimodal autoencoder for weighted fusion, assign weights to each modality in a predefined manner, perform weighted fusion, and generate a unified feature representation; Step 42, Banach space embedding: mapping the fused feature representation into the Banach space to generate a feature representation embedded in the Banach space; Step 43, Fisher-Rao distance measurement: receiving the feature representation embedded in the Banach space obtained from step 42, using the Fisher-Rao distance measurement to measure the difference in feature distribution between different modalities, optimizing the feature representation embedded in the Banach space, and obtaining the feature representation of the final multi-source data fusion; Step 44, machine learning prediction: input the feature representation of the final multi-source data fusion obtained in step 43 into the trained deep neural network to obtain the predicted risk probability of MRD recurrence after gastric cancer surgery; Step 45, MRD recurrence warning after gastric cancer surgery: When the predicted risk probability of MRD recurrence after gastric cancer surgery exceeds a predetermined threshold, a recurrence warning is issued and relevant medical personnel are notified.

Images