Methods, systems, apparatus, and media for analysis of cancer cross-modality data

Abstract

The application discloses an analysis method, system and device based on cancer cross-modal data and a medium, and comprises the following steps: in response to receiving an analysis request input by a user, inputting the analysis request to a first large model, determining an analysis task and an analysis tool for executing the analysis task; in the case where the analysis task comprises a target analysis task, extracting information from clinical data input by the user by using an information extraction engine to obtain object basic data; obtaining a gene expression matrix formed by cancer transcriptome data after dimension reduction based on the object basic data by using a wisdom algorithm contraction engine; performing survival analysis on the gene expression matrix after dimension reduction by using a risk mapping engine to obtain a survival analysis result and a target gene set affecting survival; and checking the target gene set based on reference data by using an arbitration evaluation engine in the analysis tool to obtain a checking result.

Description

Technical Field

[0001] This invention relates to the field of cross-modal analysis technology, specifically to analysis methods, systems, devices, and media based on cancer cross-modal data. Background Technology

[0002] Cancer survival risk prediction is currently of great significance in clinical decision-making and disease management, but existing technologies still have significant shortcomings. On the one hand, traditional statistical models often rely on single-modal data, making it difficult to fully utilize cross-modal information such as clinical texts and transcriptome data. Furthermore, they are easily affected by noise when processing high-dimensional features, resulting in insufficient model stability and prediction accuracy.

[0003] On the other hand, although deep learning methods have strong feature extraction capabilities, they generally lack adaptability to clinical scenarios, have weak model interpretability, and their training process is highly dependent on manual operation, making it impossible to achieve automated and reproducible survival analysis.

[0004] Furthermore, existing methods generally lack external biological validation mechanisms for model outputs, making it difficult to determine whether the screened genes have a real disease association, resulting in limited reliability of the analysis results. As clinical and molecular heterogeneity among cancer patients continues to increase, traditional analytical frameworks are unable to dynamically adjust analytical strategies based on patient characteristics, failing to meet the demands for high-precision, interpretable, and verifiable intelligent survival analysis. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Firstly, a method for analyzing cancer cross-modal data, comprising:

[0007] Upon receiving an analysis request from a user, the analysis request is input to the instruction distillation engine to obtain the user intent represented by the analysis request, and to match the analysis task with the analysis tools used to execute the analysis task; the analysis tools include an information extraction engine, an intelligent computing contraction engine, a risk mapping engine, and an arbitration evaluation engine;

[0008] The information extraction engine in the analysis tool is used to extract information from the user's clinical data to obtain the object's basic data;

[0009] Using the intelligent shrinkage engine in the analysis tool, based on the basic data of the object, key genes are identified in the initial gene expression matrix formed by cancer transcriptome data to obtain the dimensionality-reduced gene expression matrix.

[0010] The risk mapping engine in the analysis tool is used to perform survival analysis on the dimensionality-reduced gene expression matrix to obtain the survival analysis results and the set of target genes affecting survival.

[0011] Using the arbitration evaluation engine in the analysis tool, the target gene set is validated based on reference data to obtain validation results, which are then used to update the target gene set. The user is then fed back the survival analysis results and the updated target gene set.

[0012] As a preferred embodiment of the cancer cross-modal data analysis method described in this invention, the instruction distillation engine has a built-in first large language model to generate a first prompt message with low information density and a second prompt message with high information density.

[0013] The information extraction engine includes inputting clinical data and first prompt information into a second large language model to extract information and obtain initial object basic data and confidence level; if the confidence level is lower than the confidence level threshold, generating second prompt information based on the initial object basic data, and inputting the second prompt information and clinical data into the second large language model to obtain object basic data.

[0014] As a preferred embodiment of the cancer cross-modal data analysis method described in this invention, the key gene identification includes determining normal transcriptome data that matches the basic data of the object from a set of normal transcriptome datasets characterizing health status.

[0015] Covariance processing was performed on the normal transcriptome data and the cancer transcriptome data respectively to obtain the covariance matrix of the normal group and the covariance matrix of the cancer group;

[0016] Based on the covariance matrix of the normal group and the covariance matrix of the cancer group, a comparison matrix is ​​obtained;

[0017] The comparison matrix was analyzed using principal component analysis to obtain the principal component matrix;

[0018] Based on the principal component matrix, the contribution of each gene is evaluated to obtain the contribution degree of each gene;

[0019] Based on the contribution of each gene, the dimensionality-reduced gene expression matrix is ​​determined from the initial gene expression matrix.

[0020] As a preferred embodiment of the cancer cross-modal data analysis method described in this invention, the survival analysis includes determining a target survival analysis model that matches the basic data of the object from an initial survival analysis model;

[0021] The gene expression matrix is ​​input into the target survival analysis model to obtain the set of target genes that characterize survival and the survival analysis results.

[0022] As a preferred embodiment of the cancer cross-modal data analysis method described in this invention, the target survival analysis model includes generating an initial survival analysis model using a multilayer perceptron.

[0023] The target training samples that match the basic data of the object are input into the initial survival analysis model to obtain the sample gene set and sample survival analysis results;

[0024] The sample gene set and the current feature selection parameter set are processed using a dynamic penalty loss function to obtain a first loss value; wherein, the current feature selection parameter set includes the parameters used in the initial survival analysis model to select genes that affect survival risk;

[0025] The survival analysis results of the samples and the survival analysis results of the remaining training samples in the risk sample set within the predetermined observation period are input into the negative log-partial likelihood loss function to obtain the second loss value; the risk sample set includes multiple training samples with the same basic data of the same object; the remaining training samples include other training samples in the risk sample set except for the target training sample.

[0026] The current survival analysis parameter set in the initial survival analysis model is processed using a regularized loss function to obtain a third loss value; wherein, the current survival analysis parameter set includes all parameters in the initial survival analysis model except for the current feature selection parameter;

[0027] The target loss value is obtained by adding the first loss value, the second loss value, and the third loss value.

[0028] Based on the target loss value, the parameters of the initial survival analysis model are adjusted to obtain the target survival analysis model.

[0029] As a preferred embodiment of the cancer cross-modal data analysis method described in this invention, the verification result includes determining the reference data from a knowledge base based on the basic data of the object and the type of cancer suffered by the target object;

[0030] If the reference data matches the target gene set, the verification result of the characterization analysis is determined to be normal.

[0031] If the reference data does not match the target gene set, the verification result of the characterization analysis is determined to be abnormal.

[0032] In a preferred embodiment of the cancer cross-modal data analysis method described in this invention, if the verification result characterizes an anomaly, the target gene set is updated based on the verification result to obtain an updated target gene set.

[0033] Secondly, an analysis system based on cancer cross-modal data using the method described in this invention is characterized by: a task analysis module that receives an analysis request input by a user, inputs the analysis request to an instruction distillation engine, obtains the user intent represented by the analysis request, matches the analysis task with the analysis tools used to perform the analysis task; the analysis tools include an information extraction engine, an intelligent computation shrinkage engine, a risk mapping engine, and an arbitration assessment engine;

[0034] The extraction module uses the information extraction engine in the analysis tool to extract information from the user's clinical data to obtain the object's basic data;

[0035] The intelligent shrinkage module utilizes the intelligent shrinkage engine in the analysis tool to identify key genes in the initial gene expression matrix formed by cancer transcriptome data based on the object's basic data, thereby obtaining a dimension-reduced gene expression matrix.

[0036] The survival analysis module uses the risk mapping engine in the analysis tool to perform survival analysis on the dimensionality-reduced gene expression matrix, and obtains the survival analysis results and the set of target genes that affect survival.

[0037] The verification module uses the arbitration evaluation engine in the analysis tool to verify the target gene set based on reference data, obtain verification results, update the target gene set based on the verification results, and feed back the user the survival analysis results and the updated target gene set.

[0038] Thirdly, a computer device includes: a memory and a processor; the memory stores a computer program, wherein: when the processor executes the computer program, it implements the steps of the method described in any one of the present invention.

[0039] Fourthly, a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described in any one of the present invention.

[0040] The beneficial effects of this invention are as follows: By constructing a large-scale model task parsing mechanism driven by prompting engineering, and combining a multi-module collaborative system of information extraction engine, intelligent shrinkage engine, risk mapping engine, and arbitration assessment engine, this invention achieves an automated, biologically consistent, and interpretable survival analysis process for cross-modal cancer data. Compared with existing technologies, this invention can not only automatically extract reliable basic data of subjects from unstructured clinical texts, but also achieve accurate key gene identification based on the covariance difference between the normal group and the cancer group, effectively reducing noise interference from high-dimensional omics data and improving the stability of feature selection. Furthermore, this invention achieves individualized survival risk prediction through an adaptive elastic sparse deep neural network, and constructs a dual verification system through internal indicator verification and external literature evidence analysis, significantly improving the reliability, biological interpretability, and clinical acceptability of the analysis results. Simultaneously, this invention supports automatic triggering of iterative optimization when verification anomalies occur, enabling the output results to have self-correcting capabilities, thereby forming an end-to-end, reproducible, and scalable intelligent survival analysis process, greatly improving the efficiency and intelligence level of cancer data analysis. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is an overall flowchart of the analysis method based on cancer cross-modal data;

[0043] Figure 2 This diagram illustrates application scenarios for analytical methods based on cross-modal cancer data.

[0044] Figure 3 A schematic diagram of an analytical tool for a cross-modal cancer data analysis method;

[0045] Figure 4 A schematic diagram of an instruction distillation engine for an analysis method based on cancer cross-modal data;

[0046] Figure 5 This is a schematic diagram illustrating the execution of analysis tasks in a cross-modal cancer data analysis method.

[0047] Figure 6 This is a block diagram of an analysis system based on cancer cross-modal data. Detailed Implementation

[0048] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0049] Reference Figures 1-6 As an embodiment of the present invention, an analysis method based on cancer cross-modal data is provided, comprising:

[0050] Figure 2 The diagram illustrates an application scenario of the cancer cross-modal data analysis method according to an embodiment of the present method.

[0051] like Figure 2 As shown, application scenario 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0052] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0053] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0054] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0055] It should be noted that the cancer cross-modal data analysis method provided in this embodiment can generally be executed by server 105. Correspondingly, the cancer cross-modal data analysis device provided in this embodiment can generally be located in server 105. The cancer cross-modal data analysis method provided in this embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the cancer cross-modal data analysis device provided in this embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0056] It should be understood that Figure 2 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0057] The following will be based on Figure 2 The described scene, through Figure 1 and Figures 3-5 The analysis method based on cancer cross-modal data according to the disclosed embodiments is described in detail.

[0058] like Figure 1 As shown, the analysis method based on cancer cross-modal data in this embodiment includes operations S210 to S250. This analysis and processing method based on cancer cross-modal data can be executed by a server or a terminal device.

[0059] In operation S210, in response to receiving an analysis request from the user, the analysis request is input to the first large language model (instruction distillation engine) to determine the analysis task and the analysis tool used to perform the analysis task based on the user intent represented by the analysis request.

[0060] When operating S220, if the analysis task includes a target analysis task, the information extraction engine in the analysis tool is used to extract information from the clinical data input by the user to obtain the basic data of the object.

[0061] When operating the S230, the intelligent shrinkage engine in the analysis tool is used to identify key genes in the initial gene expression matrix formed by cancer transcriptome data in cancer cross-modal data based on the object's basic data, and the dimensionality-reduced gene expression matrix is ​​obtained.

[0062] When operating S240, the risk mapping engine in the analysis tool is used to perform survival analysis on the dimensionality-reduced gene expression matrix to obtain the survival analysis results and the set of target genes affecting survival.

[0063] When operating the S250, the arbitration evaluation engine in the analysis tool is used to verify the target gene set based on reference data, and the verification results are obtained. Based on the verification results, the target gene set is optimized, and the user's survival analysis results and the optimized target gene set are fed back.

[0064] Applications can be built based on cancer cross-modal data analysis methods to enable human-computer interaction. In this interaction, users can input analysis requests through input boxes and upload clinical data as well as cancer transcriptome data using a document upload control.

[0065] User authentication and session management features can be configured for the application. After authentication, users can log in and perform analysis, ultimately obtaining a structured survival analysis report based on the analysis results and an optimized target gene set. This structured survival analysis report is then displayed to the user through an interactive interface.

[0066] The instruction distillation engine can utilize the primary language model to perform user intent analysis on analysis requests and, based on this intent, determine the analysis task and tools. The model structure of the primary language model is not limited; for example, it can include pre-trained large language models. The key is that it can take the analysis request as input and perform tasks such as user intent recognition, task analysis, and tool determination.

[0067] The type of analysis task is not limited. For example, if it is a single task such as "performing multiple student survival analysis based on existing data", then "bioinformatics analysis tools" will be directly assigned. If it includes multiple tasks with different fields or different thinking patterns, such as "extracting pathological features from clinical data such as cases" and "reducing dimensionality of cancer transcriptome data", then analysis tools will be assigned to each analysis task.

[0068] Optionally, when the analysis task includes a target analysis task, the analysis tools may include an information extraction engine, an intelligent shrinkage engine, a risk mapping engine, and an arbitration assessment engine. Correspondingly, the target analysis task can be understood as a task of performing survival analysis based on cancer cross-modal data.

[0069] By leveraging the first major language model to flexibly adjust the analysis tasks and tools used to perform them based on the user intent represented by the analysis request, the analysis becomes flexible and targeted, thereby improving the effectiveness of the analysis.

[0070] Based on basic object data, such as user age, occupation, and cancer type, key genes are identified in the initial gene expression matrix. This improves the accuracy of key gene identification, thereby reducing the amount of data processing and noise content in the dimensionality-reduced gene expression matrix, and improving the accuracy of subsequent processing.

[0071] After performing survival analysis using the dimensionality-reduced gene expression matrix, an arbitration evaluation engine is used to verify the obtained survival analysis results and the set of target genes affecting survival based on reference data, thereby improving the reliability and effectiveness of the final results fed back to the user.

[0072] Figure 3 A schematic diagram of the analysis tool according to an embodiment of the method is shown.

[0073] like Figure 3 As shown, the analysis tools may include an information extraction engine, an intelligent computing contraction engine, a risk mapping engine, and an arbitration assessment engine.

[0074] like Figure 3 As shown, the instruction distillation engine, with its built-in first major language model, is used to parse user-input analysis requests. Based on prompting engineering and thought chain derivation processes, it optimizes, reconstructs, and intelligently enhances high-level instructions for cancer survival analysis tasks. The instruction distillation engine, with its built-in first major language model, generates low-information-density first prompts and high-information-density second prompts. The first prompt, with low information density, is generated by semantically simplifying user commands, retaining the core task objective, and removing redundant context. It is used to trigger basic semantic parsing operations in downstream modules such as the information extraction engine to ensure that the model can complete the initial judgment under minimal interference. The second prompt, with high information density, is generated by the command distillation engine when the confidence of the initial execution result is insufficient, the task recognition is ambiguous, or there is potential ambiguity. It combines knowledge base content, scenario inference results, and the most similar examples in the few-shot example library to the current task. The prompt is then extended and inferred based on the task semantics, parameter constraints, data format requirements, and professional background information. This makes the prompt more targeted, has a richer medical context, and clearer analytical boundaries, thereby guiding the downstream engine to obtain a higher confidence output in the second processing.

[0075] Large language models are deployed separately for multi-tool collaborative scheduling of complex tasks and execution of specific subtasks. Based on the completed instruction distillation engine, high-level system prompts are generated for the construction of the adaptive scheduling engine and downstream execution engine.

[0076] The high-order instructions output by the instruction distillation engine are preprocessed by the job routing unit, which decomposes the complex task of cancer survival analysis into multiple sub-tasks with well-defined execution boundaries and matches them to downstream execution engine jobs. This scheduling process can be completed automatically by the system without user intervention.

[0077] like Figure 4 As shown, the instruction distillation engine may include: an intent perception unit, an expert dispatch unit, a scenario simulation unit, and a standardized operation unit.

[0078] The intent-aware unit is used to deeply analyze the intent of user commands, accurately extract the task scenario (survival analysis), core objectives (predicting risks and identifying key genes), and expected output (a structured report containing risk scores and a gene list), and pass the basic information to other optimization units.

[0079] The expert dispatch unit is used to construct prompts, such as including a virtual expert identity and clearly defining their area of ​​expertise and behavioral norms, thus building a precise cognitive framework for the downstream execution engine. For example, the prompts could include:

Expert Identity

Behavioral Norms

[0080] The scenario simulation unit is used to simulate the medical background knowledge that may be involved in the task execution process. This includes highly relevant professional contextual information, such as introducing specific cancer subtypes, common prognostic factors, and related biomarkers to help the model understand biological heterogeneity. It also includes key details such as data sources, analysis considerations, and data constraints, integrating medical background knowledge into higher-level instructions to enrich the decision-making basis of the downstream execution engine. Data constraints mainly clarify the structural and quality constraints that data must meet, such as: gene expression data must be in matrix format, specifying rows and columns and using standard naming conventions; clinical text must be parsable plain text or JSON format. It also proposes strategies for handling data quality issues.

[0081] Standardized work units are used to transform users' explicit or implicit demands into clear, unambiguous, and quantifiable work requirements.

[0082] The format standardization unit is used to strictly define the data structure and interaction format of the task output, ensuring that the output results of the downstream execution engine can be directly used in the automated analysis process.

[0083] When assigning roles to tasks based on expert dispatch units, the engine initiates a local knowledge base retrieval and imports the results into higher-level instructions. The locally vectorized knowledge base contains high-quality medical instruction templates and statistical method descriptions that have been manually selected, ensuring the professionalism and standardization of the interaction.

[0084] like Figure 4 As shown, a local vector knowledge base and a few-shot sample library are also configured for the instruction distillation engine to improve the professionalism and standardization of high-level instructions.

[0085] The data sources in the vector knowledge base include, but are not limited to, clinical practice guidelines, medical and statistics textbooks, authoritative academic journals, and high-quality databases reviewed by experts.

[0086] The few-shot example library can include manually designed and written few-shot sample pairs, containing original user instructions, optimized inference processes, and higher-order instruction outputs. It uses a large language model to expand the number of sample pairs, covering a variety of typical task scenarios.

[0087] The original data is uniformly text-encoded and transformed into a high-dimensional vector representation. Independent index sets are built for the local vector knowledge base and the few-shot example library, and different retrieval strategies and return quantity thresholds are configured.

[0088] When enriching the task context information based on the scenario inference unit, the locally stored few-shot example library is called to select the case with the highest similarity to the current task as a learning sample and concatenate it with higher-order instructions to generate prompt information.

[0089] The adaptive scheduling engine will be explained below. Figure 5 As shown, the adaptive scheduling engine receives high-order instructions output from upstream instruction distillation instructions, performs task parsing and planning on them, and coordinates downstream execution engines to complete the cancer survival analysis task.

[0090] An adaptive scheduling engine may include: a data storage unit, a job routing unit, a status management unit, etc.

[0091] The data storage unit is used to receive and store the raw data input by the user, while encapsulating and maintaining all intermediate data and final results generated by downstream execution engines, and establishing a unified, dynamically updated task data pool.

[0092] The job routing unit is used to parse the core objectives and operation processes of high-order instructions. Based on task type and data characteristics, it breaks down complex tasks into clearly defined sub-tasks and accurately matches them to the corresponding downstream execution engines.

[0093] The status management unit is used to establish one-to-one status monitoring for each downstream execution engine and dynamically adjust the subtask execution strategy based on task progress and feedback results.

[0094] Clinical data features are extracted to obtain basic data about the objects. The information extraction engine can sequentially perform context-aware word segmentation, syntactic analysis, entity recognition, and semantic judgment on the clinical data to identify key basic data such as age information, determine the semantic relationship between age and cancer, and finally output the extracted clinical features in a preset JSON format and return them to the data storage unit.

[0095] like Figure 3 As shown, the intelligent computational shrinkage engine is used to perform dimensionality reduction processing on transcriptomics data based on basic object data to identify key genes and obtain a dimensionality-reduced gene expression matrix. The engine can generate Python code from its built-in large language model, such as the third largest language model. Based on the code execution results, the status management unit automatically identifies the task completion status. If the code encounters errors, it collaborates with the execution engine to perform cause feedback analysis and provide adjustment strategies until the dimensionality-reduced gene expression matrix stabilizes and is then stored.

[0096] In this embodiment, the transcriptomics data still exhibits high dimensionality and sparsity, requiring preliminary dimensionality reduction preprocessing before it can be used for subsequent survival analysis modeling. If the job routing unit identifies fused data and dimensionality reduction instructions, it invokes the intelligent shrinkage engine to execute them.

[0097] like Figure 3 As shown, if the job routing unit identifies low-dimensional preprocessed data and analysis instructions, it calls the risk mapping engine to execute them. The risk mapping engine is designed with a multi-group adaptive elastic sparse deep neural network architecture, which drives the engine to perform modeling through high-order instructions.

[0098] Specifically, this architecture replaces the linear estimation part in the traditional Cox proportional hazards model with a multiple perceptron neural network. Multiple groups of Cox neural networks are designed based on the heterogeneity stratification of age information. Each group of neural networks has an independent parameter set after training, and each gene has a one-to-one feature connection weight in each neural network. By co-training multiple groups of adaptive elastic sparse deep neural networks, accurate prediction of survival risk for cancer patients and screening of target gene sets under age heterogeneity stratification are simultaneously achieved.

[0099] like Figure 3 As shown, the arbitration evaluation engine is driven by high-level instructions to perform internal performance verification and external biological verification. If the functional unit recognizes the analysis results and evaluation instructions, it invokes the arbitration evaluation engine to execute them.

[0100] The arbitration evaluation engine includes internal performance validation and external biological validation. Internal performance validation employs a five-fold cross-validation method, calculating the consistency index, true positive rate, and false discovery rate corresponding to the risk scores predicted by a multi-group adaptive elastic sparse deep neural network, and comparing these scores with set thresholds to evaluate model performance. External biological validation, based on a selected set of characteristic genes, calls the built-in large language model to generate query strings conforming to PubMed (a database) retrieval rules, automatically retrieving abstract texts of relevant literature from the database through a secure application programming interface. The built-in large language model is then restarted to perform deep semantic analysis of the literature content, identifying gene-cancer associations and grading them as low, medium, or high evidence strength. The evidence strength grading primarily depends on the nature of the literature itself and the specific description of the results; for example, direct validation from basic experiments is stronger than case-control studies, and terms like "significant" are stronger than "preliminary findings." The large language model analyzes this information automatically, ultimately generating a structured report containing dual validation, which is returned to the user as the result of a cancer cross-modal data survival analysis task. The output report is in a preset JSON format, with key fields including: cancer type, age stratification, associated genes, individual risk score set, consistency index, true positive rate, and false discovery rate.

[0101] The literature search content is used to biologically verify the final results and enhance their interpretability, which will be directly reflected in the output report; the knowledge base content is the definition of specific statistical methods or working standards relied upon by the downstream execution engine during the coding process, or a supplementary example of optimizing the generation of high-order task instructions, mainly to improve its professional adaptability.

[0102] The following sections will explain each analytical tool in turn.

[0103] like Figure 5 As shown, when the analysis task is a target analysis task, the target analysis task can be completed through the information extraction engine, intelligent computing contraction engine, risk mapping engine and arbitration evaluation engine, resulting in a structured analysis report, which is then fed back to the user through an interactive interface.

[0104] The following section will explain each analysis tool, starting with a detailed explanation of the information extraction engine.

[0105] According to this embodiment, the information extraction engine includes a second large language model. The second large language model can be a large language model.

[0106] For example Figure 1The operation S220 shown utilizes the information extraction engine in the analysis tool to extract information from the clinical data input by the user, obtaining basic object data. This may include: inputting the clinical data and the first prompt information into the second language model for information extraction, obtaining initial basic object data and confidence level. If the confidence level is lower than the confidence level threshold, generating a second prompt information based on the initial basic object data, and inputting the second prompt information and the clinical data into the second language model to obtain the basic object data.

[0107] The information extraction engine includes a second language model, which is optimized and trained using data from multiple professional medical terminology databases. This data is used to improve the medical standardization of the second language model in extracting cancer diagnosis and treatment information.

[0108] For the entity recognition results based on the initial object basic data, a corresponding confidence score can be output simultaneously. If the confidence score is lower than the confidence score threshold, the information extraction engine is triggered to perform a secondary semantic enhancement loop until the object basic data with a qualified confidence score is output. The object basic data can be output in a preset JSON format, and key fields can include: name, age, and cancer type.

[0109] Optionally, the first prompt can be generated using an instruction distillation engine to prompt the second language model to perform the information extraction task. The second prompt can also be generated by the instruction distillation engine to prompt the second language model to perform a secondary semantic reinforcement loop task. Specifically, when the confidence level of the initial object's basic data is low, it will not directly accept this unreliable result, but will initiate a more complex "thinking" process. This process can be understood as "letting the model think again carefully." For example, internal re-inference: "The above recognition confidence is insufficient. Please perform a secondary analysis. Consider the following context: '65' may represent age, blood pressure, or other values. Please combine the overall semantics of the sentence 'patient's age' and common medical expressions to re-judge and give a higher confidence answer."

[0110] By using the second language model to extract information from large amounts of clinical data, the extracted data can be made accurate and effective through semantic relevance, avoiding the problem of extracting incorrect information due to the fact that the basic data of the object includes simple numbers.

[0111] The intelligent computing shrinkage engine will be explained in detail below.

[0112] According to this embodiment, for example, Figure 1Operation S230, as shown, utilizes the intelligent shrinkage engine in the analysis tool to identify key genes in the initial gene expression matrix formed from cancer transcriptome data based on the object's basic data, obtaining a dimensionality-reduced gene expression matrix. This may include: identifying normal transcriptome data matching the object's basic data from a set of normal transcriptome data representing healthy states; performing covariance processing on the normal and cancer transcriptome data respectively to obtain normal group covariance matrices and cancer group covariance matrices; obtaining a contrast matrix based on the normal group covariance matrix and cancer group covariance matrix containing biological reference data; analyzing the contrast matrix using principal component analysis to obtain principal component matrices; evaluating the contribution of each gene based on the principal component matrices to obtain the contribution degree of each gene; and determining the dimensionality-reduced gene expression matrix from the initial gene expression matrix based on the contribution degrees of each gene.

[0113] Given that a large number of genes in transcriptomics data are either under-transcribed or not transcribed, traditional principal component analysis is difficult to extract the nonlinear relationships between genes, which can easily lead to difficulties in convergence or overfitting of the prediction model, severely reducing the model's generalization ability and the reliability of the analysis results.

[0114] Therefore, cancer transcriptome data can be dimensionality reduced to identify genes with high contribution rates, resulting in a gene expression matrix with low noise and low data volume after dimensionality reduction.

[0115] Considering the instability of estimating empirical covariance matrices in high-dimensional data, and the fact that focusing only on the internal variance structure of the target without a control reference makes it difficult for the original data to highlight cancer-specific gene signals after dimensionality reduction, thus weakening the biological interpretive value of the results.

[0116] Normal transcriptome data matching the basic data of the subjects can be identified from normal transcriptome data characterizing healthy states. Using the normal group covariance matrix formed from the normal transcriptome data as biological reference data, covariance processing is performed on the cancer transcriptome data to obtain a comparison matrix. Therefore, by using normal transcriptome data as a control reference, cancer-specific gene signals are highlighted, thereby improving the biological interpretability of the results.

[0117] Alternatively, the contrast matrix can be represented by the following formula (1).

[0118] (1)

[0119] This indicates the degree of difference between the two sets of covariance matrices; This represents the covariance matrix obtained after shrinking the Lediot-Wolf estimation formula for cancer transcriptome data; This represents the covariance matrix obtained after shrinking normal transcriptome data according to the Lediot-Wolf estimation formula. The representation matrix showing the difference in covariance between the cancer group and the normal group. This represents the expression matrix corresponding to the cancer transcriptome data. Y is the expression matrix of the normal group.

[0120] right Perform eigenvalue decomposition to obtain eigenvectors and corresponding eigenvalues.

[0121]

[0122] in, , representing the unit eigenvector matrix along the principal component direction; , where represents the eigenvalue matrix in the corresponding direction. This indicates transpose. express The p-th eigenvector. express In the eigenvector Eigenvalues ​​in the direction.

[0123] Based on the preset variance explanation ratio choose The directions of the principal components are denoted as follows: The principal component matrix is ​​obtained by projecting the original high-dimensional feature matrix onto a low-dimensional space. ,in, Represents the centralized To enable interpretable analysis of cancer transcriptome data, the contribution of each gene was assessed based on its loading value in the principal components, and the top genes whose contribution ratio reached a set threshold were selected. These are individual genes. In cancer transcriptome data, only the transcriptome data corresponding to these genes are retained to form a dimensionality-reduced gene expression matrix. .

[0124] The intelligent shrinkage engine outputs the dimensionality-reduced gene expression matrix to the data storage unit for subsequent risk prediction and analysis.

[0125] The dimensionality reduction method provided in this embodiment can highlight cancer-specific gene signals with high contribution in the dimensionality-reduced gene expression matrix, while reducing noise content and data processing volume, thereby improving subsequent processing efficiency and accuracy.

[0126] According to this embodiment, the intelligent contraction engine is constructed as follows: Formula and description information for performing the key gene identification task are determined from a knowledge base. Parameter information of the data to be processed is determined. The formula, description, and parameter information of the data to be processed are input into a third language model to obtain the execution code for performing the key gene identification task. Based on the execution code, the intelligent contraction engine is generated.

[0127] The intelligent computing shrinkage engine also includes an external interpreter module, used to execute built-in large language models, such as the third largest language model, to generate runtime code. The dimensionality reduction task of the intelligent computing shrinkage engine is completed based on the runtime code. The status management unit automatically identifies the task completion status based on the execution results. If the runtime code encounters errors, it performs cause feedback analysis and adjusts the strategy, repeating the process until a stable result is successfully output.

[0128] The code can be written in Python.

[0129] The knowledge base stores text descriptions and key formulas related to dimensionality reduction methods, which form the basis for code generation. Before the third language model generates the running code, it automatically obtains the specific dimensions of the data to be processed, as well as user instructions or system default settings. The above content is then integrated and input into the third language model to generate the running code for the specific dimensionality reduction task.

[0130] Stable results include, but are not limited to: no errors in the code execution, and no issues such as dimension mismatch, calculation errors, or undefined variables. The successfully obtained dimensionality-reduced gene expression matrix satisfies the dimensionality requirement. ,and Greater than 0 and less than the dimension of the initial gene expression matrix The dimensionality-reduced gene expression matrix does not contain invalid values ​​such as non-numeric or infinity. The obtained dimensionality-reduced gene expression matrix successfully reaches the preset variance explanation ratio or overall contribution ratio threshold. The eigenvalues ​​on the obtained principal components exhibit a skewed distribution rather than a uniform distribution. Downstream validation results indirectly verify the high quality of dimensionality reduction.

[0131] According to this embodiment, the use of the third language model to automatically generate running code can achieve the effect of automated analysis by automated generation tools, thereby reducing R&D costs and time and improving processing efficiency.

[0132] According to this embodiment, for example, Figure 1Operation S240, as shown, utilizes the risk mapping engine in the analysis tool to perform survival analysis on the dimensionality-reduced gene expression matrix, obtaining the survival analysis results and the set of target genes affecting survival. This may include: determining a target survival analysis model that matches the basic data of the object from multiple survival analysis models; inputting the gene expression matrix into the target survival analysis model to obtain the set of target genes representing the impact on survival and the survival analysis results.

[0133] The target survival analysis model is constructed using a deep neural network structure. By leveraging the powerful nonlinear mapping capability and multi-level feature extraction capability of deep neural networks, it provides important technical support for capturing the complex interactions between genes in cancer patients.

[0134] However, the "black box" nature of deep neural networks limits their interpretability in clinical practice. Furthermore, most existing cancer survival prediction models are based on single-modality omics data, focusing only on effective information at the gene level and neglecting the differential characteristics exhibited by patients during diagnosis and treatment. These non-genetic characteristics contain rich clinical information and are closely related to survival risk.

[0135] In this embodiment, multiple survival analysis models that are mapped to the basic data of the target can be pre-established, allowing different survival analysis models to be used to process data from different users. This distinguishes cancer patients with similar genetic characteristics but vastly different clinical situations from different dimensions, further improving the accuracy and interpretability of survival prediction results for clinical decision-making.

[0136] According to this embodiment, the target survival analysis model is trained in the following manner.

[0137] An initial survival analysis model is generated using a multilayer perceptron. Target training samples matching the basic data of the object are input into the initial survival analysis model to obtain a sample gene set and sample survival analysis results. A dynamic penalty loss function is used to process the sample gene set and the current feature selection parameter weight set to obtain a first loss value. The current feature selection parameter set includes parameters in the initial survival analysis model used to select genes affecting survival risk. The sample survival analysis results and the survival analysis results of the remaining training samples in the risk sample set within a predetermined observation period are input into a negative log-partial likelihood loss function to obtain a second loss value. The risk sample set includes multiple training samples with the same basic data of the object, and the remaining training samples include other training samples in the risk sample set besides the target training samples. A regularized loss function is used to process the current survival analysis parameter set in the initial survival analysis model to obtain a third loss value. The current survival analysis parameter set includes all parameters in the initial survival analysis model except for the current feature selection parameter set. Based on the first, second, and third loss values, a target loss value is obtained.

[0138] The "target training samples matched with the target individual's basic data" refers to the system's matching and screening of training samples based on the target individual's basic data before risk prediction, thereby improving the model's adaptability to individual differences. Specifically, the target individual's basic data includes age, gender, tumor type, pathological stage, molecular subtype, and other key clinical parameters that can affect survival risk. Based on this basic data, the system selects samples from a pre-constructed training dataset that are identical to or within a comparable range of the target individual in terms of the aforementioned clinical characteristics, forming a target training sample set that matches the target individual's features. By executing the initial survival analysis model on this homogeneous sample set, the model's feature learning can better align with the target individual's biological background and clinical status, thereby improving the accuracy, stability, and individualization of risk prediction.

[0139] The risk mapping engine, based on deep learning, uses a multi-group adaptive elastic sparse deep neural network architecture and a composite loss function that includes a dynamic penalty loss function and a negative log-partial likelihood loss function to simultaneously predict the survival risk score of cancer patients and select feature genes under heterogeneous stratification.

[0140] The gradient descent algorithm can be used to adjust the parameters of the initial survival analysis model based on the target loss value, with the goal of maximizing the likelihood of survival data under the model, and ultimately achieving accurate prediction of individual survival risk and screening of heterogeneous stratified features.

[0141] However, this is not the only option. The negative log-partial likelihood loss function can also be used to ensure the model's goodness of fit to the survival risk score. Based on a dynamic penalty loss function for group sparsity, two hyperparameters are introduced for the feature connection weights to control the overall sparsity intensity and the degree of heterogeneity between groups, allowing specific genes to be associated only with the basic data of the target population. Based on the model generalization fitting loss, L2 regularization is applied to the non-feature sparse weight parameters to prevent overfitting.

[0142] Specifically, the target loss value can be determined by the loss function of the following formula (2).

[0143] (2)

[0144] in, This represents the regularization loss function; Represents the dynamic penalty loss function; This represents the negative log-partial likelihood loss function.

[0145] (3)

[0146] in, This represents the weight parameters of the j-th input feature (gene) in the deep neural network of the risk mapping engine; The inter-group sparsity coefficient represents the sparsity strength of the control inputs to the initial survival analysis model sample input data. The coefficients represent the degree of heterogeneity of the dimensionality-reduced variables in the initial survival analysis model. The dimension of the gene expression matrix after dimensionality reduction is represented by ; t represents the t-th sample group, where the basic data of objects in each sample group are the same; T represents the total number of sample groups. This represents the feature selection parameter corresponding to the j-th gene in the t-th initial survival analysis model; for all genes j... The set formed constitutes the set of feature selection parameters for the t-th initial survival analysis model used to select genes that affect survival risk.

[0147] (4)

[0148] in, This represents the negative log-partial likelihood loss function. This represents the set of samples in the t-th sample group where a survival event (such as death) occurred. q represents the sample index, which, unlike i, is used to represent all samples in the risk set. Let represent the risk prediction function corresponding to the initial survival analysis model of the t-th sample group. Let represent the input feature vector of sample i in the t-th hierarchical model. This represents the sample survival analysis results of the target training samples output by the initial survival analysis model within the predetermined observation period; Indicates the first The risk set consists of the uncensored samples that survive the expected event within a finite observation period. This represents the survival analysis results of the remaining training samples in the risk sample set within the predetermined observation period. The risk sample set includes multiple training samples with the same basic data of the same object, and the remaining training samples include other training samples in the risk sample set besides the target training sample.

[0149] (5)

[0150] in, This indicates control over the severity of punishment. This represents the set of parameters in the initial survival analysis model, excluding the feature selection parameter set. It is a regularization penalty term used to penalize "non-feature selection parameters" in the model, preventing the model from overfitting.

[0151] According to this embodiment, the target survival analysis model trained in the above manner can learn knowledge from different dimensions, thereby making the knowledge and capabilities learned by the target survival analysis model comprehensive and effective, and improving the accuracy and effectiveness of model training.

[0152] The arbitration evaluation engine will be explained in detail below.

[0153] According to this embodiment, for example, Figure 1 The operation S250 shown utilizes the arbitration evaluation engine in the analysis tool to validate the target gene set based on reference data, obtaining a validation result. This may include: determining reference data from a knowledge base based on the object's basic data; determining a validation result indicating normal characterization analysis if the reference data matches the target gene set; and determining a validation result indicating abnormal characterization analysis if the reference data does not match the target gene set.

[0154] The arbitration evaluation engine can be used to perform internal and external biological validation of survival analysis results and generate structured reports.

[0155] Internal validation employs a multiple sampling statistical validation method to calculate performance metrics, which are then compared to set thresholds to determine the final performance. Optionally, the reference data in the knowledge base can refer to set thresholds that match the performance metrics. If the performance metrics exceed the set thresholds, the validation result is considered normal. Conversely, if they fall below the thresholds, the validation result is considered abnormal.

[0156] External biological validation, based on a selected target gene set, uses a built-in large language model to generate query strings that conform to medical database retrieval rules. It connects to an external database via a secure application programming interface (API) and automatically retrieves relevant literature abstracts as reference data. The built-in large language model is then used again to perform deep semantic analysis of the literature content, identifying gene-gene relationships and grading the strength of evidence. If a relationship exists between the reference data and genes in the target gene set, a match is determined, indicating a normal validation result. Otherwise, the validation result is considered abnormal.

[0157] A structured report containing the double-validation results can be generated based on the validation results and returned to the user as the execution result of the cross-modal data cancer survival analysis task.

[0158] The results analyzed using the above methods, such as the target gene set and survival analysis results, can be evaluated through validation to ensure the accuracy and effectiveness of the analysis results. This avoids providing users with inaccurate results due to failure to pass validation, thereby improving the user experience of human-computer interaction.

[0159] The following section provides a detailed explanation of how to couple and iterate the arbitration assessment engine and the risk mapping engine.

[0160] According to this embodiment, when performing such Figure 1 Following operation S250, the method further includes: in the case of abnormal characterization analysis of the verification results, optimizing the target gene set based on the verification results to obtain an optimized target gene set.

[0161] Preferably, the risk mapping engine and the arbitration assessment engine can be jointly constructed to form a coupled iterative mechanism.

[0162] When the status management unit detects that internal validation metrics have not reached a set threshold or that external biological assessment support is insufficient, it triggers the model to run again. In this loop, based on reasoning about the causes of analysis failure, the status management unit dynamically adjusts the model parameters and retrieval strategy of the risk mapping engine, initiating a new round of data analysis and validation. This optimizes the target gene set, resulting in an optimized target gene set. Finally, it generates an output report that meets the preset requirements.

[0163] By leveraging the coupled iterative mechanism of the risk mapping engine and the arbitration assessment engine, it is possible to ensure that when anomalies are found in the verification result characterization analysis, existing analysis tools can be used in a timely and effective manner to optimize the results and improve the optimization effect.

[0164] On the other hand, the cancer cross-modal data analysis system 600 of this embodiment includes a task analysis module 610, an extraction module 620, an intelligent calculation shrinkage module 630, a survival analysis module 640, and a verification module 650.

[0165] The task analysis module 610 responds to received user input analysis requests by inputting the requests into a first language model. The first language model then uses this model to determine the analysis task and the analysis tools used to execute it, based on the user intent represented by the analysis request. The extraction module 620, when the analysis task includes a target analysis task, uses the information extraction engine within the analysis tool to extract information from the user-input clinical data, obtaining basic object data. The intelligent shrinkage module 630 uses the intelligent shrinkage engine within the analysis tool to identify key genes in the initial gene expression matrix formed from cancer transcriptome data, based on the basic object data, resulting in a dimensionality-reduced gene expression matrix. The survival analysis module 640 uses the risk mapping engine within the analysis tool to perform survival analysis on the dimensionality-reduced gene expression matrix, obtaining survival analysis results and a set of target genes affecting survival. The verification module 650 uses the arbitration evaluation engine within the analysis tool to verify the target gene set based on reference data, obtaining verification results. These results are then used to optimize the target gene set, and the user is fed back the survival analysis results and the optimized target gene set.

[0166] If the above functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0167] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0168] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0169] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0170] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An analysis method based on cancer cross-modal data, characterized in that, include: Upon receiving an analysis request from a user, the analysis request is input into the instruction distillation engine to obtain the user intent represented by the analysis request, and to match the analysis task and the analysis tool used to perform the analysis task. The analytical tools include an information extraction engine, an intelligent computing contraction engine, a risk mapping engine, and an arbitration assessment engine. The information extraction engine in the analysis tool is used to extract information from the user's clinical data to obtain the object's basic data; Using the intelligent shrinkage engine in the analysis tool, based on the basic data of the object, key genes are identified in the initial gene expression matrix formed by cancer transcriptome data to obtain the dimensionality-reduced gene expression matrix. The risk mapping engine in the analysis tool is used to perform survival analysis on the dimensionality-reduced gene expression matrix to obtain the survival analysis results and the set of target genes affecting survival. Using the arbitration evaluation engine in the analysis tool, the target gene set is validated based on reference data to obtain validation results, so as to update the target gene set based on the validation results; It also provides feedback on the survival analysis results provided by the user and the updated target gene set.

2. The analysis method based on cancer cross-modal data as described in claim 1, characterized in that: The instruction distillation engine has a built-in first language model to generate a first prompt message with low information density and a second prompt message with high information density. The information extraction engine includes inputting clinical data and first prompt information into a second language model to extract information and obtain initial object basic data and confidence level. If the confidence level is lower than the confidence threshold, a second prompt message is generated based on the initial basic object data. The second prompt message and clinical data are then input into the second large language model to obtain the basic object data.

3. The analysis method based on cancer cross-modal data as described in claim 2, characterized in that: The key gene identification includes determining normal transcriptome data that matches the basic data of the object from a set of normal transcriptome datasets characterizing health status; Covariance processing was performed on the normal transcriptome data and the cancer transcriptome data respectively to obtain the covariance matrix of the normal group and the covariance matrix of the cancer group; Based on the covariance matrix of the normal group and the covariance matrix of the cancer group, a comparison matrix is ​​obtained; The comparison matrix was analyzed using principal component analysis to obtain the principal component matrix; Based on the principal component matrix, the contribution of each gene is evaluated to obtain the contribution degree of each gene; Based on the contribution of each gene, the dimensionality-reduced gene expression matrix is ​​determined from the initial gene expression matrix.

4. The analysis method based on cancer cross-modal data as described in claim 3, characterized in that: The survival analysis includes determining a target survival analysis model that matches the basic data of the object from an initial survival analysis model; The gene expression matrix is ​​input into the target survival analysis model to obtain the set of target genes that characterize survival and the survival analysis results.

5. The analysis method based on cancer cross-modal data as described in claim 4, characterized in that: The target survival analysis model includes generating an initial survival analysis model using a multilayer perceptron. The target training samples that match the basic data of the object are input into the initial survival analysis model to obtain the sample gene set and the sample survival analysis results. The sample gene set and the current feature selection parameter set are processed using a dynamic penalty loss function to obtain a first loss value; wherein, the current feature selection parameter set includes the parameters used in the initial survival analysis model to select genes that affect survival risk; The survival analysis results of the samples and the survival analysis results of the remaining training samples in the risk sample set within the predetermined observation period are input into the negative log-partial likelihood loss function to obtain the second loss value; the risk sample set includes multiple training samples with the same basic data of the same object; the remaining training samples include other training samples in the risk sample set except for the target training sample. The current survival analysis parameter set in the initial survival analysis model is processed using a regularized loss function to obtain a third loss value; wherein, the current survival analysis parameter set includes all parameters in the initial survival analysis model except for the current feature selection parameter; The target loss value is obtained by adding the first loss value, the second loss value, and the third loss value. Based on the target loss value, the parameters of the initial survival analysis model are adjusted to obtain the target survival analysis model.

6. The analysis method based on cancer cross-modal data as described in claim 5, characterized in that: The verification result includes determining the reference data from the knowledge base based on the object's basic data and the type of cancer the target object suffers from; If the reference data matches the target gene set, the verification result is determined to be normal in the characterization analysis. If the reference data does not match the target gene set, the verification result of the characterization analysis is determined to be abnormal.

7. The analysis method based on cancer cross-modal data as described in claim 6, characterized in that: If the verification result characterization analysis is abnormal, the target gene set is updated based on the verification result to obtain an updated target gene set.

8. A cancer cross-modal data analysis system employing the method described in any one of claims 1-7, characterized in that: The task analysis module receives an analysis request from the user, inputs the analysis request into the instruction distillation engine, obtains the user intent represented by the analysis request, and matches the analysis task with the analysis tool used to execute the analysis task. The analytical tools include an information extraction engine, an intelligent computing contraction engine, a risk mapping engine, and an arbitration assessment engine. The extraction module uses the information extraction engine in the analysis tool to extract information from the user's clinical data to obtain the object's basic data; The intelligent shrinkage module utilizes the intelligent shrinkage engine in the analysis tool to identify key genes in the initial gene expression matrix formed by cancer transcriptome data based on the object's basic data, thereby obtaining a dimension-reduced gene expression matrix. The survival analysis module uses the risk mapping engine in the analysis tool to perform survival analysis on the dimensionality-reduced gene expression matrix, and obtains the survival analysis results and the set of target genes that affect survival. The verification module uses the arbitration evaluation engine in the analysis tool to verify the target gene set based on reference data and obtain the verification result, so as to update the target gene set based on the verification result; It also provides feedback on the survival analysis results provided by the user and the updated target gene set.

9. A computer device, comprising: A memory and a processor; the memory stores a computer program, characterized in that: when the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images