A deep learning-based gene medical big data integration analysis system

Abstract

This invention discloses a deep learning-based gene-medical big data integration and analysis system, comprising: a data acquisition module for collecting multi-source gene-medical data of tested subjects and generating a unified object identifier; a data processing module for generating structured multi-source data; a standardization mapping module for generating standardized multi-source gene-medical data; a data fusion module for generating multi-modal fusion data records; a feature construction module for constructing a multi-modal feature vector set; a cross-modal analysis module for performing cross-modal representation learning using an improved CHRep model and outputting multi-modal association features; a result generation module for generating analysis results; and a report output module for outputting a gene-medical big data integration and analysis report, thereby realizing the unified integration of multi-source gene-medical data, cross-modal association analysis, and intelligent result output.

Description

Technical Field

[0001] This invention relates to the field of medical big data processing, and in particular to a gene medical big data integration and analysis system based on deep learning. Background Technology

[0002] With the development of high-throughput sequencing technology, clinical information systems, and medical data platforms, medical institutions and research units are able to collect a large amount of genetic medical data related to the subjects being examined. This type of data typically includes gene sequencing data, gene variant annotation data, transcriptome expression data, clinical diagnosis and treatment data, laboratory test data, medication records, and disease outcome data. This data can describe the subjects at the molecular, clinical phenotype, diagnosis and treatment process, and disease outcome levels, providing a data foundation for disease risk assessment, pathogenic variant identification, drug response analysis, and patient stratification. Therefore, how to effectively collect, desensitize, preprocess, standardize, map, correlate, time-align, and fuse fields of multi-source genetic medical data has become a crucial technical issue in genetic medical big data analysis systems.

[0003] In existing technologies, gene sequencing data, gene variant annotation data, and transcriptome expression data are typically generated separately by sequencing platforms or bioinformatics analysis workflows. Clinical diagnosis and treatment data, laboratory test data, medication records, and disease outcome data are usually stored in hospital information systems, laboratory systems, electronic medical record systems, or follow-up systems. Due to differences in data sources, field structures, data formats, encoding methods, and time recording methods across different systems, existing systems often can only analyze data from a single source when processing multi-source genetic medical data, or they rely on manual methods to organize and match partial data, making it difficult to establish a unified data foundation for the same examinee. Especially in different data sources, the same examinee may correspond to different names, hospital numbers, examination numbers, sample numbers, or internal system numbers. Without a unified object identifier, it is difficult to ensure that data from different sources can be accurately attributed to the same examinee, thus affecting the reliability of subsequent data association and analysis results. Summary of the Invention

[0004] One objective of this invention is to propose a gene medical big data integration and analysis system based on deep learning. This invention fully utilizes multi-source gene medical data integration, standard terminology mapping, multimodal feature construction, and improved CHRep model cross-modal representation learning technology. It has the advantages of high data integration consistency, strong cross-modal correlation analysis capability, more comprehensive prediction and analysis results, strong privacy protection capability for the tested subjects, and high efficiency of clinical auxiliary analysis.

[0005] According to an embodiment of the present invention, a gene-based medical big data integration and analysis system based on deep learning includes: The data acquisition module is used to collect multi-source genetic medical data of the tested subjects and generate a unified object identifier; The data processing module is used to desensitize multi-source genetic medical data, establish a mapping relationship between desensitized object identifiers and unified object identifiers, and perform preprocessing to generate structured multi-source data. The standardization mapping module is used to map structured multi-source data to standard terms to generate standardized multi-source genomic medical data. The data fusion module is used to perform sample association, time alignment, and field fusion of standardized multi-source medical gene data using a unified object identifier as the primary key, generating multimodal fusion data records. The feature construction module is used to construct a multimodal feature vector set based on multimodal fusion data records; The cross-modal analysis module is used to input multimodal feature vector sets into the improved CHRep model to perform cross-modal representation learning and output multimodal correlation features; The results generation module is used to perform disease risk prediction, pathogenic variant identification, drug response prediction, and patient stratification analysis using multimodal association features, and generate analysis results. The report output module is used to write the analysis results and corresponding gene feature contribution information and clinical feature contribution information into the results database, and output the integrated analysis report of gene medical big data through a visualization interface.

[0006] Optionally, gene sequencing data, gene variation annotation data, transcriptome expression data, as well as relevant clinical diagnosis and treatment data, laboratory test data, medication record data, and disease outcome data for each subject can be collected through medical institutions, scientific research databases, or laboratory sequencing platforms. During the data collection process, a unique and unified object identifier is generated for each subject.

[0007] Optionally, the data processing module includes: Desensitizing multi-source genetic medical data, establishing a mapping relationship between desensitized object identifiers and unified object identifiers, and generating desensitized multi-source genetic medical data; Perform format parsing and quality control on gene sequencing data in desensitized multi-source gene medical data to generate standardized sequencing data; Preprocessing is performed on the clinical diagnosis and treatment data, laboratory test data, medication record data and disease outcome data in the desensitized multi-source gene medical data to generate structured clinical data; Standardized sequencing data is spliced ​​with structured clinical data to form unified structured multi-source data.

[0008] Optionally, the normalization mapping module includes: Each type of data field in the structured multi-source data is identified, and a corresponding data type and unit mapping relationship is established for each field to form a field set; For each field in the field set, standard terminology mapping is performed, uniformly mapping the original field names and codes to a predefined standard vocabulary to form a standardized field set; For each inspected record in the structured multi-source data, the data is rearranged and encoded according to the standardized field set to generate a standardized data record; The standardized data records of all subjects are integrated to form a unified, standardized multi-source genetic medical data.

[0009] Optionally, the data fusion module includes: Using the unified object identifier in standardized multi-source genetic medical data as the primary key, various data records of the same subject are grouped to form a subset of subject data. Perform time alignment on a subset of the inspected object's data to form a time-consistent multimodal data sequence; The multimodal data sequences are correlated at the sample level to generate a preliminary multimodal fusion data subset for each tested object; The preliminary multimodal fusion data subset of each inspected object is uniformly arranged and integrated at the field level to form a multimodal fusion data record.

[0010] Optionally, the feature construction module includes: Based on multimodal fusion data records, extract the multimodal data fields corresponding to each inspected object to generate inspected object feature construction records; The fields in the feature construction record of the inspected object are classified into numerical fields, categorical fields and time fields. The numerical fields are normalized, the categorical fields are encoded and the time fields are encoded by time intervals to generate the standard feature record of the inspected object. Missing fields in the standard feature record of the inspected object are marked for missing information, and the missing mark values ​​are written into the corresponding fields to generate a complete feature record of the inspected object. Modal segmentation is performed on the complete feature records of the subjects to generate gene feature vectors, expression feature vectors, clinical feature vectors, test feature vectors, medication feature vectors, and outcome feature vectors; The gene feature vector, expression feature vector, clinical feature vector, test feature vector, medication feature vector, and outcome feature vector are concatenated to generate a multimodal feature vector for each subject. All multimodal feature vectors corresponding to the inspected objects are indexed according to a unified object identifier to form a multimodal feature vector set.

[0011] Optionally, the cross-modal analysis module includes: The improved CHRep model inputs a multimodal feature vector set to perform cross-modal representation learning. The improved CHRep model includes a heteromodal anchoring encoding module, an object consistency interaction representation module, a temporal constraint association learning module, and a posterior offset correction output module. The improvement of the improved CHRep model is that it extends the bimodal representation method of image features and gene expression features in the traditional CHRep model to a cross-modal representation method for multimodal feature vector sets, and introduces heteromodal anchoring, object-level hypergraph high-order interaction, hyperbolic prototype alignment, time interval constraints, and posterior offset correction mechanisms to generate multimodal association features. In the heterogeneous modality anchoring coding module, based on the multimodal feature vector set, modality identification embedding, field source anchoring and hyperbolic space projection processing are performed on the multimodal feature vector corresponding to each inspected object to generate heterogeneous modality anchoring features; In the object consistency interaction representation module, the heterogeneous modal anchoring features corresponding to the same tested object are aggregated at the object level based on the unified object identifier, an object-level hypergraph structure is constructed, high-order relations are transferred to the object-level hypergraph structure, and the heterogeneous modal anchoring features of the same tested object are mapped to the same object prototype neighborhood through the hyperbolic prototype alignment mechanism to generate object consistency interaction features. In the temporal constraint association learning module, time information is read from the multimodal fusion data records, time interval constraints are applied to the object consistency interaction features, and temporal constraint association features are generated. In the posterior offset correction output module, modality center features corresponding to each modality are generated based on temporal constraint association features. The representation offset between temporal constraint association features and modality center features is calculated, and posterior offset correction is performed on temporal constraint association features to generate multimodal association features.

[0012] Optionally, the result generation module includes: Disease risk prediction is performed based on multimodal association features, generating disease risk prediction results for each subject. Based on multimodal association features, pathogenic variants are identified, and pathogenic variant identification results are generated for each tested object. Drug response prediction is performed based on multimodal association features, generating drug response prediction results for each subject. Patient stratification analysis is performed based on multimodal association features to generate patient stratification results for each subject. The disease risk prediction results, pathogenic variant identification results, drug response prediction results, and patient stratification results are linked and integrated according to a unified object identifier to generate analysis results corresponding to each subject.

[0013] Optionally, the report output module includes: Based on the correspondence between multimodal association features and analysis results, gene feature contribution information and clinical feature contribution information corresponding to the analysis results are generated; The analysis results, along with the corresponding gene feature contribution information and clinical feature contribution information, are recorded in the results database. In the results database, report data is generated according to the order of the report fields; The report data is converted into a display format through a visualization interface to generate an integrated analysis report of gene and medical big data.

[0014] The beneficial effects of this invention are: This invention collects, desensitizes, preprocesses, and maps standard terminology to multi-source genetic medical data of tested subjects. This enables gene sequencing data, gene variant annotation data, transcriptome expression data, clinical diagnosis and treatment data, laboratory test data, medication records, and disease outcome data to be standardized into multi-source genetic medical data under a unified object identifier. Compared with existing technologies that suffer from inconsistent data formats, inconsistent field meanings, and difficulty in data correlation among different sources, this invention can preserve the correlation between different data sources while protecting the privacy of tested subjects, providing a unified data foundation for subsequent sample correlation, time alignment, and field fusion.

[0015] This invention uses a unified object identifier as the primary key to perform sample association, time alignment, and field fusion on standardized multi-source genetic medical data, generating multimodal fused data records and further constructing a multimodal feature vector set. This processing method avoids the problems of unclear data correspondence, missing temporal relationships, and chaotic field structures caused by simply splicing genetic and clinical data in existing technologies. It enables data from the same subject at different time points and in different modalities to form a clearly structured, consistent, and time-corresponding feature expression.

[0016] This invention inputs a multimodal feature vector set into an improved CHRep model to perform cross-modal representation learning. Through heterogeneous modality anchoring encoding, object-consistent interactive representation, temporal constraint association learning, and posterior bias correction, it generates multimodal association features. Compared to the traditional bimodal CHRep model, which only targets image and gene expression features, this invention extends it to multi-source genomic medical data scenarios. This enables the association modeling of multiple types of information, such as genes, expression, clinical data, laboratory tests, medication, and outcomes, within the same subject dimension, thereby improving the expressive power of cross-modal associations in complex genomic medical data.

[0017] This invention utilizes multimodal association features for disease risk prediction, pathogenic variant identification, drug response prediction, and patient stratification analysis, generating analysis results along with corresponding gene feature contribution information and clinical feature contribution information. In this way, the system can not only output analysis results specific to the tested subjects but also provide the basis for the feature contribution of the analysis results, expanding the gene medical big data analysis process from a single predictive output to an analysis output combining predictive results and feature contribution information.

[0018] This invention writes the analysis results, gene feature contribution information, and clinical feature contribution information into a results database, and generates a gene-medical big data integrated analysis report through a visualization interface, forming a complete closed loop of multi-source data processing, deep learning analysis, and results display. This system can improve the automation level and data utilization efficiency of gene-medical big data integrated analysis, providing unified, standardized, and traceable data analysis results for disease risk assessment, pathogenic variant analysis, drug response judgment, and patient stratification management. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a deep learning-based gene medical big data integration and analysis system proposed in this invention. Figure 2 This is a schematic diagram illustrating the construction of a multimodal feature vector set for a deep learning-based gene medical big data integration and analysis system proposed in this invention. Figure 3 This is a schematic diagram of the structure of an improved CHRep model for a deep learning-based gene medical big data integration and analysis system proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figures 1-3 A deep learning-based gene-based medical big data integration and analysis system includes: The data acquisition module is used to collect multi-source genetic medical data of the tested subjects and generate a unified object identifier; The data processing module is used to desensitize multi-source genetic medical data, establish a mapping relationship between desensitized object identifiers and unified object identifiers, and perform preprocessing to generate structured multi-source data. The standardization mapping module is used to map structured multi-source data to standard terms to generate standardized multi-source genomic medical data. The data fusion module is used to perform sample association, time alignment, and field fusion of standardized multi-source medical gene data using a unified object identifier as the primary key, generating multimodal fusion data records. The feature construction module is used to construct a multimodal feature vector set based on multimodal fusion data records; The cross-modal analysis module is used to input multimodal feature vector sets into the improved CHRep model to perform cross-modal representation learning and output multimodal correlation features; The results generation module is used to perform disease risk prediction, pathogenic variant identification, drug response prediction, and patient stratification analysis using multimodal association features, and generate analysis results. The report output module is used to write the analysis results and corresponding gene feature contribution information and clinical feature contribution information into the results database, and output the integrated analysis report of gene medical big data through a visualization interface.

[0022] In this embodiment, gene sequencing data, gene variation annotation data, transcriptome expression data, and related clinical diagnosis and treatment data, test data, medication record data, and disease outcome data of each subject are collected through medical institutions, scientific research databases, or laboratory sequencing platforms. During the data collection process, a unique unified object identifier is generated for each subject. The unified object identifier is generated by hashing or randomly generating the subject's name, date of birth, and hospital number.

[0023] In this embodiment, the data processing module includes: Desensitizing multi-source genetic medical data, establishing a mapping relationship between desensitized object identifiers and unified object identifiers, and generating desensitized multi-source genetic medical data; The specific process for generating desensitized multi-source genetic medical data is as follows: Identifier analysis is performed on the multi-source genetic medical data, including sensitive information such as the examinee's name, date of birth, and hospital number; encryption encoding or hash algorithms are used to map the examinee's name, date of birth, and hospital number to unique desensitized object identifiers, and a mapping relationship is established between the desensitized object identifiers and their corresponding unified object identifiers; direct identity information fields in the multi-source genetic medical data are deleted or replaced, and the gene sequencing data, gene variation annotation data, transcriptome expression data, clinical diagnosis and treatment data, laboratory test data, medication record data, and disease outcome data associated with the examinee are uniformly mapped to their corresponding desensitized object identifiers, forming complete desensitized multi-source genetic medical data with consistent field correspondences; Perform format parsing and quality control on gene sequencing data in desensitized multi-source gene medical data to generate standardized sequencing data; The process of generating standardized sequencing data is as follows: Based on the sequence information and sequencing quality information contained in the gene sequencing data of the desensitized multi-source gene medical data, the data is organized according to the system's preset table structure or database format to form a unified data representation, ensuring consistency in field names, data types, and storage formats; the quality score corresponding to each sequence in the gene sequencing data is checked, and sequences below the preset quality threshold are marked as abnormal and removed from the dataset; duplicate records are deduplicated to ensure that each gene variant site has a unique corresponding entry in the data; after completing the above processing, all sequence information and quality information are integrated to generate standardized sequencing data; Preprocessing is performed on the clinical diagnosis and treatment data, laboratory test data, medication record data and disease outcome data in the desensitized multi-source gene medical data to generate structured clinical data; The process of generating structured clinical data specifically involves: uniformly identifying and mapping fields in diagnostic and treatment data, laboratory test data, medication records, and disease outcome data to ensure consistency in field names, data types, and units for each record; marking missing data and employing supplementation or imputation strategies based on data type to ensure data integrity; identifying and correcting outliers in the data, including checking numerical ranges, time series consistency, and logical constraints, and deleting or correcting records that do not conform to the rules; standardizing time information by arranging various events according to a unified time format and time order, and organizing data from different sources according to a unified field order to ensure accurate correspondence among multiple types of data for the same subject; and integrating the preprocessed data according to a unified data structure to form structured clinical data. Standardized sequencing data is spliced ​​with structured clinical data to form unified structured multi-source data.

[0024] In this embodiment, the standardized mapping module includes: Each type of data field in the structured multi-source data is identified, and a corresponding data type and unit mapping relationship is established for each field to form a field set; The formation process of the field set is as follows: traverse various types of data in the structured multi-source data, including gene sequencing data, gene mutation site information, transcriptome expression data, clinical diagnosis and treatment data, laboratory test data, medication record data, and disease outcome data; analyze and record the name, data type, and unit of each field, and establish a corresponding mapping relationship for each field according to predefined data type and unit specifications; organize all fields and their corresponding data type and unit mapping relationships into a unified data structure set, forming the field set; For each field in the field set, standard terminology mapping is performed, uniformly mapping the original field names and codes to a predefined standard vocabulary to form a standardized field set; The process of forming a standardized field set is as follows: each field in the field set and its corresponding data type and unit information are read sequentially; the original name and code of each field are compared with a predefined standard vocabulary, and the original fields are replaced and mapped according to the unified field names and codes in the standard vocabulary; and the data is organized according to a unified data structure to form a standardized field set. For each inspected record in the structured multi-source data, the data is rearranged and encoded according to the standardized field set to generate a standardized data record; The process of generating standardized data records is as follows: Data from each field of each inspected record is read; the fields of each record are rearranged according to the standardized field set to ensure that the field order, field name, and encoding are completely consistent with the standardized field set; the data of each field is encoded and the units are standardized, mapping the data values ​​in the original record to the standardized format to ensure that the data type, numerical range, and units of each record meet the standardization requirements; and all processed field data are integrated to form a single complete standardized data record. The standardized data records of all subjects are integrated to form a unified, standardized multi-source genetic medical data. The formation of standardized multi-source genetic medical data is specifically as follows: Standardized data records for each subject are read one by one according to the data storage structure, and each field in the record is uniformly arranged and coded according to a standardized field set; each record is classified and indexed according to the unique identifier of the subject, so that each standardized data record can accurately correspond to each subject in the integrated data set; the standardized data records of all subjects are sequentially added to a unified data set and stored using a unified data structure to generate standardized multi-source genetic medical data.

[0025] In this embodiment, the data fusion module includes: Using the unified object identifier in standardized multi-source genetic medical data as the primary key, various data records of the same subject are grouped to form a subset of subject data. The specific process for forming the subset of the tested object data is as follows: traverse each record in the standardized multi-source gene medical data, read the corresponding unified object identifier in the record, group all data records with the same unified object identifier, and classify the gene sequencing data, gene variation annotation data, transcriptome expression data, clinical diagnosis and treatment data, laboratory test data, medication record data, and disease outcome data associated with the tested object into the same set according to the original field structure; organize the records in each set according to the field order to ensure that each record maintains a consistent field arrangement and data type in the set, thus forming the subset of the tested object data; Perform time alignment on a subset of the inspected object's data to form a time-consistent multimodal data sequence; The process of forming a multimodal data sequence is as follows: read the time information of various types of data in each subset of the subject's data, sort the different types of data records according to a unified time standard, and mark or interpolate time points with inconsistent or missing time intervals so that the corresponding records of various types of data on the same time axis can be accurately aligned. Combine the various types of data after time standardization and sorting in chronological order to form a multimodal data sequence organized by time step. Each time step contains gene sequencing data, gene variation annotation data, transcriptome expression data and clinically relevant data corresponding to the time point. The multimodal data sequences are correlated at the sample level to generate a preliminary multimodal fusion data subset for each tested object; The process of generating the preliminary multimodal fusion data subset is as follows: read the record of each time step in the multimodal data sequence of each tested object; match the data of the corresponding time point or the closest time point in each time step; match the data of all modalities at the sample level; and organize the matched data according to the time order and field order to form a preliminary multimodal data subset containing all modalities. The preliminary multimodal fusion data subset of each inspected object is uniformly arranged and integrated at the field level to form a multimodal fusion data record; The process of forming multimodal fusion data records is as follows: read the records of each time step in each preliminary multimodal fusion data subset, arrange the fields in the records in a unified order according to the standardized field set, integrate the data of the same field type in each record, including combining the time-aligned multimodal information according to the field index, and ensure that the field name, code and data type of each record are consistent with the standardized field set; classify and organize the integrated records according to the unified object identifier of the inspected object to form multimodal fusion data records.

[0026] In this embodiment, the feature construction module includes: Based on multimodal fusion data records, extract the multimodal data fields corresponding to each inspected object to generate inspected object feature construction records; The specific process for generating the feature construction record of the inspected object is as follows: Read the multimodal fusion data record, locate the record range corresponding to each inspected object using a unified object identifier, extract the multimodal data fields used for feature construction within the record range, and organize the extracted multimodal data fields according to the field name, field code, field order, and field source in the multimodal fusion data record; for cases where there are multiple records for the same inspected object, the system will group multimodal data fields with the same unified object identifier into the same record unit, and retain the time information and field values ​​corresponding to each field; after completing field extraction and record aggregation, an inspected object feature construction record corresponding to a single inspected object is formed. The fields in the feature construction record of the inspected object are classified into numerical fields, categorical fields and time fields. The numerical fields are normalized, the categorical fields are encoded and the time fields are encoded by time intervals to generate the standard feature record of the inspected object. The specific process for generating the standard feature record of the inspected object is as follows: Based on the characteristics of the inspected object, each field in the record is constructed. The fields are then divided into numerical fields, categorical fields, and time fields according to their value format and meaning. For numerical fields, the system extracts the field values ​​and normalizes them according to the maximum, minimum, mean, or standard deviation of the field across all inspected object records. For categorical fields, the system performs discrete encoding according to the category name or category code corresponding to the field value, converting the categorical fields into coded features. For time fields, the system calculates the time interval between the corresponding time of each time field and the baseline time, using the baseline time for the same inspected object as the starting point, and converts the time interval into time interval coded features. The normalized numerical features, coded categorical features, and time interval coded features are arranged according to the field order in the record constructed from the characteristics of the inspected object, generating the standard feature record of the inspected object corresponding to the inspected object. Missing fields in the standard feature record of the inspected object are marked for missing information, and the missing mark values ​​are written into the corresponding fields to generate a complete feature record of the inspected object. Modal segmentation is performed on the complete feature records of the subjects to generate gene feature vectors, expression feature vectors, clinical feature vectors, test feature vectors, medication feature vectors, and outcome feature vectors; The modality segmentation process is as follows: Each field in the complete feature record of the tested subject is read, and each field is assigned to its corresponding modality based on its source, meaning, and encoding. Fields from gene sequencing data and gene variation annotation data are assigned to the gene modality, and their values ​​are extracted in order to generate gene feature vectors. Fields from transcriptome expression data are assigned to the expression modality, and their values ​​are extracted in order to generate expression feature vectors. Fields from clinical diagnosis and treatment data are assigned to the clinical modality, and their values ​​are extracted in order to generate clinical feature vectors. Fields from laboratory test data are assigned to the laboratory modality, and their values ​​are extracted in order to generate laboratory feature vectors. Fields from medication records are assigned to the medication modality, and their values ​​are extracted in order to generate medication feature vectors. Fields from disease outcome data are assigned to the outcome modality, and their values ​​are extracted in order to generate outcome feature vectors. The gene feature vector, expression feature vector, clinical feature vector, test feature vector, medication feature vector, and outcome feature vector are concatenated to generate a multimodal feature vector for each subject. All multimodal feature vectors corresponding to the inspected objects are indexed according to a unified object identifier to form a multimodal feature vector set; The formation process of the multimodal feature vector set is as follows: Based on the multimodal feature vectors corresponding to all tested objects, obtain the unified object identifier corresponding to each multimodal feature vector; using the unified object identifier as the index key, write the multimodal feature vector of each tested object into the corresponding index position, and check whether there are duplicate unified object identifiers; when there are multiple multimodal feature vectors corresponding to the same unified object identifier, arrange the multiple multimodal feature vectors according to time information or record generation order, and classify them under the same unified object identifier; perform a dimensionality consistency check on all indexed multimodal feature vectors to ensure that each multimodal feature vector contains gene feature vectors, expression feature vectors, clinical feature vectors, test feature vectors, medication feature vectors, and outcome feature vectors in the same order; after completing the indexing and dimensionality check, a multimodal feature vector set is formed with the unified object identifier as the index and the multimodal feature vectors as the data content.

[0027] In this embodiment, the cross-modal analysis module includes: The improved CHRep model inputs a multimodal feature vector set to perform cross-modal representation learning. The improved CHRep model includes a heteromodal anchoring encoding module, an object consistency interaction representation module, a temporal constraint association learning module, and a posterior offset correction output module. The improvement of the improved CHRep model is that it extends the bimodal representation method of image features and gene expression features in the traditional CHRep model to a cross-modal representation method for multimodal feature vector sets, and introduces heteromodal anchoring, object-level hypergraph high-order interaction, hyperbolic prototype alignment, time interval constraints, and posterior offset correction mechanisms to generate multimodal association features. The improved CHRep model's specific structure includes a heterogeneous modality anchoring encoding module, an object consistency interaction representation module, a temporal constraint association learning module, and a posterior offset correction output module, connected sequentially. During data transmission, the multimodal feature vector set uses the tested object as the basic input unit, with each tested object corresponding to a multimodal feature vector. Its data format is a one-dimensional vector formed by concatenating gene feature vectors, expression feature vectors, clinical feature vectors, test feature vectors, medication feature vectors, and outcome feature vectors in a fixed modal order. After receiving the multimodal feature vectors, the heterogeneous modality anchoring encoding module converts them into heterogeneous modality anchored features with modality identifiers and field source anchors. Its data format is a two-dimensional feature matrix containing a unified object identifier, modality identifier, field source anchor, and feature values. The object consistency interaction... The mutual representation module receives heterogeneous modal anchoring features and constructs an object-level hypergraph structure according to a unified object identifier. It uses each modal feature as a node and multiple related modal features within the same examined object as hyperedges. After high-order relation propagation and hyperbolic prototype alignment, it outputs object-consistent interaction features, which are in the form of an object-level feature matrix indexed by a unified object identifier. The temporal constraint association learning module receives the object-consistent interaction features and, combined with the temporal information in the multimodal fusion data records, converts the object-level feature matrix into a temporal feature matrix arranged in chronological order. After time interval constraints, it outputs temporal constraint association features. The posterior offset correction output module receives the temporal constraint association features, generates modality center features corresponding to each modality, calculates the representation offset, performs posterior offset correction, and finally outputs multimodal association features. The training data for the improved CHRep model comes from a multimodal feature vector set. Each subject is associated with a unified object identifier, along with corresponding disease risk labels, pathogenic variant labels, drug response labels, and patient stratification labels. Disease risk labels are generated from disease outcome data, pathogenic variant labels from gene variant annotation data, drug response labels from medication records and disease outcome data, and patient stratification labels from clinical diagnosis and treatment data, laboratory test data, and disease outcome data. During training, the multimodal feature vector set is input into the improved CHRep model, sequentially generating heterogeneous modality anchoring features, object consistency interaction features, temporal constraint association features, and multimodal relationship features. The model incorporates features and updates parameters based on a multi-task loss function, which includes disease risk prediction loss, pathogenic variant identification loss, drug response prediction loss, patient stratification loss, object consistency constraint loss, time interval constraint loss, and posterior bias correction loss. Training parameters include learning rate, batch size, number of training epochs, weight decay coefficient, feature dimension, time window length, and weights for each loss term. During training, the model parameters are iteratively updated according to batch input. Training stops and the improved CHRep model is updated when the total loss on the validation set fails to decrease for several consecutive training epochs, the evaluation metric of the analysis results reaches a set threshold, or the number of training epochs reaches a set upper limit. In the heterogeneous modality anchoring coding module, based on the multimodal feature vector set, modality identification embedding, field source anchoring and hyperbolic space projection processing are performed on the multimodal feature vector corresponding to each inspected object to generate heterogeneous modality anchoring features; The specific process of generating heterogeneous modality anchoring features is as follows: Based on the multimodal feature vector set, extract the multimodal feature vector corresponding to each tested object; according to the modality type to which each field in the multimodal feature vector belongs, assign corresponding modality identifiers to the data of different modalities, and write the modality identifiers into the feature positions of the corresponding fields to complete the modality identifier embedding; based on the field source of each field in the multimodal fusion data record, establish a field source anchor point for each field, and bind the field values ​​to the corresponding field source anchor points to complete the field source anchoring; map the multimodal feature vectors after modality identifier embedding and field source anchoring to hyperbolic space, and update the spatial position of the feature values ​​according to the attribution relationship between different modalities and the correspondence between field sources and field values; output the spatially updated features as heterogeneous modality anchoring features. In the object consistency interaction representation module, the heterogeneous modal anchoring features corresponding to the same tested object are aggregated at the object level based on the unified object identifier, an object-level hypergraph structure is constructed, high-order relations are transferred to the object-level hypergraph structure, and the heterogeneous modal anchoring features of the same tested object are mapped to the same object prototype neighborhood through the hyperbolic prototype alignment mechanism to generate object consistency interaction features. The specific process for generating object-consistent interaction features is as follows: Heterogeneous modal anchoring features are read, and those features belonging to the same tested object are grouped into the same object feature unit according to a unified object identifier. Each modal feature in the object feature unit is used as a node, and multiple related modal features within the same tested object are used as a hyperedge to construct an object-level hypergraph structure. In the object-level hypergraph structure, higher-order relations are passed between multiple modal features connected by the same hyperedge, enabling each node to receive information from other related modal features within the same tested object and update the feature values ​​of the corresponding node. An object prototype is generated for each tested object through a hyperbolic prototype alignment mechanism, and each heterogeneous modal anchoring feature corresponding to the tested object is mapped to the neighborhood range corresponding to the object prototype, so that different modal features of the same tested object form an object-level correspondence in the hyperbolic space. The modal features after completing the higher-order relation passing and hyperbolic prototype alignment are summarized according to a unified object identifier to generate object-consistent interaction features. In the temporal constraint association learning module, time information is read from the multimodal fusion data records, time interval constraints are applied to the object consistency interaction features, and temporal constraint association features are generated. The specific process for generating time-constrained association features is as follows: Read the time information from the multimodal fusion data records, extract the object consistency interaction features corresponding to the same inspected object according to the unified object identifier, determine the temporal sequence relationship between different records of the same inspected object based on the time information, and calculate the time interval between adjacent records; write the time interval as a constraint factor into the corresponding object consistency interaction features, so that the object consistency interaction features at different time points are arranged according to their temporal sequence; update the association of object consistency interaction features at adjacent time points according to the time interval constraint factor, so that the object consistency interaction features at each time point contain the association information of its previous and next time points; summarize the object consistency interaction features after time arrangement and association update to generate time-constrained association features. In the posterior offset correction output module, modality center features corresponding to each modality are generated based on temporal constraint association features. The representation offset between temporal constraint association features and modality center features is calculated, and posterior offset correction is performed on temporal constraint association features to generate multimodal association features. The process of generating multimodal association features is as follows: Temporal constraint association features are divided into feature groups corresponding to each modality according to field source and modality type; the feature mean is calculated for each feature group of each modality to generate the modality center feature corresponding to each modality; the difference between the temporal constraint association feature of each modality and the corresponding modality center feature is calculated, and the difference is used as the representation offset; the temporal constraint association feature of the corresponding modality is reverse-corrected according to the representation offset to align the temporal constraint association feature of each modality to the corresponding modality center feature; the modal features after post-abstract offset correction are combined according to a unified object identifier and field order to generate multimodal association features.

[0028] In this embodiment, the result generation module includes: Disease risk prediction is performed based on multimodal association features, generating disease risk prediction results for each subject. The specific process for generating disease risk prediction results is as follows: Based on multimodal association features, extract the multimodal association feature records corresponding to each subject according to a unified object identifier; perform weighted calculation on the feature values ​​related to disease risk prediction in the multimodal association feature records to obtain the risk output value corresponding to each disease category; normalize the risk output values ​​corresponding to each disease category to obtain the disease risk score corresponding to each disease category; determine the disease category with the highest disease risk score as the disease risk category of the corresponding subject; associate the unified object identifier, disease risk category, and disease risk score to generate the disease risk prediction result corresponding to each subject. Based on multimodal association features, pathogenic variants are identified, and pathogenic variant identification results are generated for each tested object. The specific process for generating pathogenic variant identification results is as follows: Based on multimodal association features, multimodal association feature records corresponding to each tested object are extracted according to a unified object identifier; feature values ​​related to gene variants are extracted from the multimodal association feature records, and the feature values ​​are correlated with gene variant site information in standardized multi-source gene medical data to obtain variant association features corresponding to each gene variant site; pathogenicity is calculated for the variant association features corresponding to each gene variant site to obtain a pathogenic variant score corresponding to each gene variant site; the pathogenic variant type corresponding to each gene variant site is determined according to the gene variant site information; the unified object identifier, pathogenic variant site, pathogenic variant type, and pathogenic variant score are correlated and recorded to generate pathogenic variant identification results corresponding to each tested object. Drug response prediction is performed based on multimodal association features, generating drug response prediction results for each subject. The specific process for generating drug response prediction results is as follows: Based on multimodal association features, multimodal association feature records corresponding to each tested subject are extracted according to a unified object identifier; feature values ​​related to drug response are extracted from the multimodal association feature records, and the feature values ​​are correlated with drug record data in standardized multi-source gene medical data to obtain drug association features corresponding to each drug; drug response is calculated for the drug association features corresponding to each drug to obtain the drug sensitivity score and adverse drug reaction risk score corresponding to each drug; the drug response category corresponding to each drug is determined based on the drug sensitivity score and adverse drug reaction risk score; the unified object identifier, drug response category, drug sensitivity score, and adverse drug reaction risk score are correlated and recorded to generate drug response prediction results corresponding to each tested subject. Patient stratification analysis is performed based on multimodal association features to generate patient stratification results for each subject. The specific process for generating patient stratification results is as follows: Based on multimodal association features, multimodal association feature records corresponding to each subject are extracted according to a unified object identifier; feature values ​​related to patient stratification analysis are extracted from the multimodal association feature records, and similarity calculations are performed on the feature values ​​to obtain the matching degree between the subject and different patient stratification categories; the matching degree corresponding to each patient stratification category is normalized to obtain the stratification confidence score corresponding to each patient stratification category; the patient stratification category with the highest stratification confidence score is determined as the patient stratification category of the corresponding subject; the unified object identifier, patient stratification category, and stratification confidence score are associated and recorded to generate the patient stratification result corresponding to each subject. The disease risk prediction results, pathogenic variant identification results, drug response prediction results, and patient stratification results are linked and integrated according to a unified object identifier to generate analysis results corresponding to each subject.

[0029] In this embodiment, the report output module includes: Based on the correspondence between multimodal association features and analysis results, gene feature contribution information and clinical feature contribution information corresponding to the analysis results are generated; The specific process for generating gene feature contribution information and clinical feature contribution information is as follows: Based on the multimodal association features and analysis results, and according to a unified object identifier, establish the correspondence between the multimodal association features and analysis results of the same subject; based on the source of each field in the multimodal association features, feature fields from gene sequencing data, gene variant annotation data, and transcriptome expression data are classified into gene feature fields, and feature fields from clinical diagnosis and treatment data, laboratory test data, medication record data, and disease outcome data are classified into clinical feature fields; calculate the contribution weights of gene feature fields and clinical feature fields to the disease risk prediction results, pathogenic variant identification results, drug response prediction results, and patient stratification results in the analysis results, and sort them according to the contribution weight from largest to smallest; record the gene feature field name, corresponding contribution weight, and sorting result as gene feature contribution information, and record the clinical feature field name, corresponding contribution weight, and sorting result as clinical feature contribution information; The analysis results, along with the corresponding gene feature contribution information and clinical feature contribution information, are recorded in the results database. In the results database, report data is generated according to the order of the report fields. The report data includes disease risk prediction information, pathogenic variant identification information, drug response prediction information, patient stratification information, genetic feature contribution information, and clinical feature contribution information. The report data can be converted into a display format through a visualization interface to generate an integrated analysis report of gene medical big data. The specific process of format conversion is as follows: Report data is read through a visual interface, and the data is categorized according to the report display fields, including disease risk prediction information, pathogenic variant identification information, drug response prediction information, patient stratification information, gene feature contribution information, and clinical feature contribution information. The categorized report data is then converted into the data structures corresponding to the display components. Disease risk prediction information is converted into risk level display data and risk score display data; pathogenic variant identification information is converted into variant site display data, variant type display data, and pathogenic variant score display data; drug response prediction information is converted into drug response category display data, drug sensitivity score display data, and adverse drug reaction risk score display data; patient stratification information is converted into patient stratification category display data and stratification confidence level display data; and gene feature contribution information and clinical feature contribution information are converted into feature contribution ranking display data. The data structures corresponding to each display component are arranged according to the report page order, and corresponding tables, graphs, and text display fields are generated. The arranged tables, graphs, and text display fields are combined to form a gene medical big data integrated analysis report.

[0030] Example 1: A gene-based medical big data analysis scenario in a precision oncology center of a tertiary hospital. In its daily clinical practice, the center conducts gene testing on subjects suspected of having genetically predisposed tumors, recurrent tumors, and those with significant differences in medication response. It also collects clinical data, laboratory test data, medication records, and disease outcome data from these subjects. Since this data originates from sequencing platforms, electronic medical record systems, laboratory systems, medication management systems, and follow-up systems, there are differences in data field naming, time formats, unit standards, and subject identification methods between these systems. For example, the same subject is represented by a sample number in the sequencing platform, a hospital number in the electronic medical record system, and a follow-up number in the follow-up system. This makes it difficult for existing systems to accurately correlate multiple types of data from the same subject. Furthermore, some clinical fields suffer from inconsistent units, inconsistent time information, missing fields, and abnormal records, making it difficult to form a unified data foundation for deep learning analysis between gene and clinical data.

[0031] In this scenario, the hospital first collects multi-source genetic medical data from the subjects, including gene sequencing data, gene variant annotation data, transcriptome expression data, clinical diagnosis and treatment data, laboratory test data, medication records, and disease outcome data. During the data collection process, the system generates a unique identifier for each subject and anonymizes identity information such as name, date of birth, and hospital number, establishing a mapping relationship between the anonymized identifier and the unified identifier. This allows data from different systems to be accurately linked without exposing direct identity information. Subsequently, the system preprocesses the anonymized multi-source genetic medical data. Gene sequencing data undergoes format parsing and quality control to generate standardized sequencing data. Clinical diagnosis and treatment data, laboratory test data, medication records, and disease outcome data undergo field cleaning, missing value marking, outlier handling, timestamp standardization, and unit standardization to generate structured clinical data. Finally, the standardized sequencing data and structured clinical data are concatenated to form structured multi-source data.

[0032] After generating structured multi-source data, the system further maps the structured multi-source data to standard terms, unifying fields with the same meaning but different names in different systems into a standardized field set. For example, fields representing white blood cell counts, tumor markers, and liver and kidney function indicators in different testing systems are unified into corresponding standard fields; drug names in different medication records are unified into the same drug code; and mutation sites, mutation types, and related annotation information in gene mutation annotation data are unified into a standardized field structure, thereby generating standardized multi-source genetic medical data. Then, using a unified object identifier as the primary key, the system performs sample association, time alignment, and field fusion on the standardized multi-source genetic medical data of the same examinee, forming a multimodal fusion data record. This multimodal fusion data record can organize gene, expression, clinical, testing, medication, and outcome information of the same examinee at different time points into a unified data structure, avoiding data mismatches and temporal order disorder caused by simply splicing tables manually.

[0033] In the deep learning analysis phase, the system constructs a multimodal feature vector set based on multimodal fusion data records and inputs this set into an improved CHRep model to perform cross-modal representation learning. The improved CHRep model generates heteromodal anchoring features by performing modality identification embedding, field source anchoring, and hyperbolic space projection processing on data from different modalities through a heteromodal anchoring encoding module. It then generates object-consistent interaction features by performing object-level aggregation of heteromodal anchoring features for the same subject based on a unified object identifier through an object-level hypergraph structure, performing high-order relation transfer and hyperbolic prototype alignment. Finally, it generates temporal-constrained association features by reading time information from the multimodal fusion data records through a temporal constraint association learning module, applying time interval constraints to the object-consistent interaction features. Finally, it generates multimodal association features through a posterior offset correction output module. The system utilizes these multimodal association features for disease risk prediction, pathogenic variant identification, drug response prediction, and patient stratification analysis, generating analysis results and further generating gene feature contribution information and clinical feature contribution information. Ultimately, it outputs a gene-medical big data integrated analysis report through a visualization interface.

[0034] To verify the application effect of this embodiment, the hospital selected 850 tumor-related subjects who completed gene testing and had complete follow-up records between January 2023 and June 2024 as test samples. Of these, 680 were training samples, 85 were validation samples, and 85 were test samples. During the test, the system of this invention was compared with a traditional manual data processing and conventional machine learning analysis process. In the traditional process, data processing mainly relies on manual field matching and table splicing, and model analysis uses single gene variant features or partial clinical features for prediction. The system of this invention, however, employs unified object identification, desensitized object identification mapping, structured multi-source data generation, standard terminology mapping, multimodal fusion data record construction, and improved CHRep model cross-modal representation learning. Test results show that the system of this invention exhibits significant improvements in data association completeness, field standardization consistency, time alignment accuracy, disease risk prediction accuracy, drug response prediction accuracy, and report generation efficiency.

[0035] Table 1. Comparison of Test Results between the Gene Therapy Big Data Integration and Analysis System and Traditional Processes

[0036] As shown in Table 1, with the same number of test samples, the traditional manual data processing plus conventional machine learning process only achieves a multi-source data association completeness rate of 82.4% due to the lack of a unified object identifier, a mapping relationship between the desensitized object identifier and the unified object identifier, and a standardized multi-source genetic medical data generation mechanism. The system of this invention associates data of the same examinee using a unified object identifier and retains the mapping relationship between the desensitized object identifier and the unified object identifier after desensitization, achieving a multi-source data association completeness rate of 96.5%. This result demonstrates that this invention can solve the problem of difficulty in mapping data from different sources to the same examinee, reducing data omissions and mismatches caused by inconsistencies in sample numbers, hospital numbers, and follow-up numbers.

[0037] In terms of field standardization consistency rate and time alignment accuracy, the traditional process achieved 78.8% and 75.3% respectively. This is mainly due to inconsistent field naming, inconsistent units, and the lack of a unified timeline between diagnosis / treatment time, testing time, medication time, and outcome time across different systems. The system of this invention, through structured multi-source data generation and standard terminology mapping, transforms different fields into a standardized field set. During the construction of multimodal fusion data records, it performs sample association, time alignment, and field fusion, achieving a field standardization consistency rate of 95.2% and a time alignment accuracy of 93.8%. This data demonstrates that this invention does not simply splice data, but rather organizes multiple types of data from the same examined object into multimodal fusion data records suitable for model analysis, based on a unified object identifier and time information.

[0038] In disease risk prediction, pathogenic variant identification, drug response prediction, and patient stratification analysis, the system of this invention achieves improvements of 9.4, 8.6, 10.6, and 9.4 percentage points respectively compared to traditional processes. This improvement stems from the invention's use of multimodal feature vector sets and an improved CHRep model for cross-modal representation learning, enabling gene features, expression features, clinical features, laboratory features, medication features, and outcome features to form multimodal association features under a unified object identifier. Particularly in drug response prediction, traditional processes typically rely solely on medication records or limited gene variant information for judgment, while the system of this invention simultaneously combines medication record data, disease outcome data, and multimodal association features, thus increasing the accuracy of drug response prediction from 76.5% to 87.1%.

[0039] Regarding report generation efficiency, traditional processes require manual organization of analysis results, manual supplementation of feature interpretations, and manual report generation, with an average report generation time of 18.6 minutes per case. The system of this invention writes analysis results, gene feature contribution information, and clinical feature contribution information into the results database and converts the display format through a visualization interface, reducing the average report generation time to 3.8 minutes per case. Simultaneously, the proportion of data records requiring manual correction in the traditional process is 21.2%, while the system of this invention reduces this to 6.7%, indicating that this invention reduces manual intervention in data standardization, field correspondence, result correlation, and report generation. The above embodiments demonstrate that this invention can address the problems of difficulty in unifying and correlating multi-source genomic medical data, insufficient field standardization, missing temporal relationships, weak cross-modal analysis capabilities, and low report generation efficiency in existing technologies, forming a complete processing flow from data acquisition to analysis report output.

[0040] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A gene-based medical big data integration and analysis system based on deep learning, characterized in that, include: The data acquisition module is used to collect multi-source genetic medical data of the tested subjects and generate a unified object identifier; The data processing module is used to desensitize multi-source genetic medical data, establish a mapping relationship between desensitized object identifiers and unified object identifiers, and perform preprocessing to generate structured multi-source data. The standardization mapping module is used to map structured multi-source data to standard terms to generate standardized multi-source genomic medical data. The data fusion module is used to perform sample association, time alignment, and field fusion of standardized multi-source medical gene data using a unified object identifier as the primary key, generating multimodal fusion data records. The feature construction module is used to construct a multimodal feature vector set based on multimodal fusion data records; The cross-modal analysis module is used to input multimodal feature vector sets into the improved CHRep model to perform cross-modal representation learning and output multimodal correlation features; The results generation module is used to perform disease risk prediction, pathogenic variant identification, drug response prediction, and patient stratification analysis using multimodal association features, and generate analysis results. The report output module is used to write the analysis results and corresponding gene feature contribution information and clinical feature contribution information into the results database, and output the integrated analysis report of gene medical big data through a visualization interface.

2. The gene medical big data integration and analysis system based on deep learning according to claim 1, characterized in that, Gene sequencing data, gene variation annotation data, transcriptome expression data, and related clinical diagnosis and treatment data, laboratory test data, medication records and disease outcome data are collected for each subject through medical institutions, scientific research databases or laboratory sequencing platforms. During the data collection process, a unique and unified object identifier is generated for each subject.

3. The gene medical big data integration and analysis system based on deep learning according to claim 1, characterized in that, The data processing module includes: Desensitizing multi-source genetic medical data, establishing a mapping relationship between desensitized object identifiers and unified object identifiers, and generating desensitized multi-source genetic medical data; Perform format parsing and quality control on gene sequencing data in desensitized multi-source gene medical data to generate standardized sequencing data; Preprocessing is performed on the clinical diagnosis and treatment data, laboratory test data, medication record data and disease outcome data in the desensitized multi-source gene medical data to generate structured clinical data; Standardized sequencing data is spliced ​​with structured clinical data to form unified structured multi-source data.

4. The gene medical big data integration and analysis system based on deep learning according to claim 1, characterized in that, The standardized mapping module includes: Each type of data field in the structured multi-source data is identified, and a corresponding data type and unit mapping relationship is established for each field to form a field set; For each field in the field set, standard terminology mapping is performed, uniformly mapping the original field names and codes to a predefined standard vocabulary to form a standardized field set; For each inspected record in the structured multi-source data, the data is rearranged and encoded according to the standardized field set to generate a standardized data record; The standardized data records of all subjects are integrated to form a unified, standardized multi-source genetic medical data.

5. The gene medical big data integration and analysis system based on deep learning according to claim 1, characterized in that, The data fusion module includes: Using the unified object identifier in standardized multi-source genetic medical data as the primary key, various data records of the same subject are grouped to form a subset of subject data. Perform time alignment on a subset of the inspected object's data to form a time-consistent multimodal data sequence; The multimodal data sequences are correlated at the sample level to generate a preliminary multimodal fusion data subset for each tested object; The preliminary multimodal fusion data subset of each inspected object is uniformly arranged and integrated at the field level to form a multimodal fusion data record.

6. The gene medical big data integration and analysis system based on deep learning according to claim 1, characterized in that, The feature construction module includes: Based on multimodal fusion data records, extract the multimodal data fields corresponding to each inspected object to generate inspected object feature construction records; The fields in the feature construction record of the inspected object are classified into numerical fields, categorical fields and time fields. The numerical fields are normalized, the categorical fields are encoded and the time fields are encoded by time intervals to generate the standard feature record of the inspected object. Missing fields in the standard feature record of the inspected object are marked for missing information, and the missing mark values ​​are written into the corresponding fields to generate a complete feature record of the inspected object. Modal segmentation is performed on the complete feature records of the subjects to generate gene feature vectors, expression feature vectors, clinical feature vectors, test feature vectors, medication feature vectors, and outcome feature vectors; The gene feature vector, expression feature vector, clinical feature vector, test feature vector, medication feature vector, and outcome feature vector are concatenated to generate a multimodal feature vector for each subject. All multimodal feature vectors corresponding to the inspected objects are indexed according to a unified object identifier to form a multimodal feature vector set.

7. The gene medical big data integration and analysis system based on deep learning according to claim 1, characterized in that, The cross-modal analysis module includes: The improved CHRep model inputs a multimodal feature vector set to perform cross-modal representation learning. The improved CHRep model includes a heteromodal anchoring encoding module, an object consistency interaction representation module, a temporal constraint association learning module, and a posterior offset correction output module. The improvement of the improved CHRep model is that it extends the bimodal representation method of image features and gene expression features in the traditional CHRep model to a cross-modal representation method for multimodal feature vector sets, and introduces heteromodal anchoring, object-level hypergraph high-order interaction, hyperbolic prototype alignment, time interval constraints, and posterior offset correction mechanisms to generate multimodal association features. In the heterogeneous modality anchoring coding module, based on the multimodal feature vector set, modality identification embedding, field source anchoring and hyperbolic space projection processing are performed on the multimodal feature vector corresponding to each inspected object to generate heterogeneous modality anchoring features; In the object consistency interaction representation module, the heterogeneous modal anchoring features corresponding to the same tested object are aggregated at the object level based on the unified object identifier, an object-level hypergraph structure is constructed, high-order relations are transferred to the object-level hypergraph structure, and the heterogeneous modal anchoring features of the same tested object are mapped to the same object prototype neighborhood through the hyperbolic prototype alignment mechanism to generate object consistency interaction features. In the temporal constraint association learning module, time information is read from the multimodal fusion data records, time interval constraints are applied to the object consistency interaction features, and temporal constraint association features are generated. In the posterior offset correction output module, modality center features corresponding to each modality are generated based on temporal constraint association features. The representation offset between temporal constraint association features and modality center features is calculated, and posterior offset correction is performed on temporal constraint association features to generate multimodal association features.

8. The gene medical big data integration and analysis system based on deep learning according to claim 1, characterized in that, The result generation module includes: Disease risk prediction is performed based on multimodal association features, generating disease risk prediction results for each subject. Based on multimodal association features, pathogenic variants are identified, and pathogenic variant identification results are generated for each tested object. Drug response prediction is performed based on multimodal association features, generating drug response prediction results for each subject. Patient stratification analysis is performed based on multimodal association features to generate patient stratification results for each subject. The disease risk prediction results, pathogenic variant identification results, drug response prediction results, and patient stratification results are linked and integrated according to a unified object identifier to generate analysis results corresponding to each subject.

9. The gene medical big data integration and analysis system based on deep learning according to claim 1, characterized in that, The report output module includes: Based on the correspondence between multimodal association features and analysis results, gene feature contribution information and clinical feature contribution information corresponding to the analysis results are generated; The analysis results, along with the corresponding gene feature contribution information and clinical feature contribution information, are recorded in the results database. In the results database, report data is generated according to the order of the report fields; The report data is converted into a display format through a visualization interface to generate an integrated analysis report of gene and medical big data.

Images