Cancer survival prediction system, electronic device, and computer program product

Abstract

This application relates to the field of medical technology, providing a cancer survival prediction system, electronic device, and computer program product. The system includes: a feature extraction module for extracting global gene features, pathological text features, and gene-level pathway features based on the patient's gene expression profile, pathological text data, and biological pathway data; a fusion module for fusing the global gene features, pathological text features, and gene-level pathway features to obtain multimodal fused features; an effect value calculation module for calculating the total effect value and the natural direct effect value based on the global gene features, pathological text features, and multimodal fused features, and obtaining the total indirect effect value by subtracting the total effect value from the natural direct effect value; and a prediction module for inputting the total indirect effect value into a target survival prediction model to obtain the cancer survival prediction information output by the target survival prediction model. This system can remove text bias and improve the accuracy of cancer survival prediction.

Description

Technical Field

[0001] This application belongs to the field of medical technology, and in particular relates to a cancer survival prediction system, electronic device and computer program product. Background Technology

[0002] In the field of clinical oncology, cancer survival prediction refers to the professional medical analysis process in which medical personnel and researchers combine multi-dimensional biomedical data of patients and use statistical models, artificial intelligence algorithms, and other technologies to quantitatively assess the probability of survival, survival duration, and risk of tumor recurrence and progression in cancer patients over a specific future period. Multi-dimensional biomedical data, such as gene expression profiles, histopathological images, and pathological texts (e.g., pathology reports), contain key biological information and clinical characteristics related to the occurrence and development of cancer in patients and are the core data for achieving accurate cancer survival prediction.

[0003] However, existing cancer survival prediction technologies struggle to effectively remove textual biases caused by spurious associations in pathological text data when integrating gene-related features with pathological text data. This results in the fused features carrying interfering data, affecting the accuracy of cancer survival prediction results. Summary of the Invention

[0004] This application provides a cancer survival prediction system, electronic device, and computer program product to solve the problem in the prior art that it is difficult to effectively remove text bias caused by false associations in pathological text data, resulting in the fused features carrying interfering data and affecting the accuracy of cancer survival prediction results.

[0005] A first aspect of this application provides a cancer survival prediction system, comprising: The feature extraction module is used to extract global gene features from the patient's gene expression profile, extract pathological text features from the patient's pathological text data, and extract gene-level pathway features from the patient's biological pathway data based on the gene expression profile; the gene expression profile contains multiple genes, and the biological pathway data contains multiple biological pathways. The fusion module is used to fuse the global gene features, the pathological text features, and the gene-level pathway features to obtain multimodal fused features; The effect value calculation module is used to calculate the total effect value and the natural direct effect value based on the global gene features, the pathological text features, and the multimodal fusion features, and to obtain the total indirect effect value by subtracting the total effect value and the natural direct effect value; the total effect value includes the true biological effect value and the text bias effect value, the natural direct effect value includes the text bias effect value, and the total indirect effect value includes the true biological effect value; The prediction module is used to input the total indirect effect value into the target survival prediction model to obtain the cancer survival prediction information output by the target survival prediction model.

[0006] A second aspect of this application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described in the first aspect.

[0007] A third aspect of this application provides a computer program product comprising a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0008] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0009] As can be seen from the above, this application extracts global gene features, pathological text features, and gene-level pathway features based on the patient's gene expression profile, pathological text data, and biological pathway data, respectively. Multimodal fusion features are obtained by fusing these features. Subsequently, based on the global gene features, pathological text features, and multimodal fusion features, the total effect value, including the true biological effect value and the text bias effect value, as well as the natural direct effect value, including the text bias effect value, are calculated. The difference between the total effect value and the natural direct effect value yields the total indirect effect value after removing the text bias. This text-bias-free total indirect effect value is then input into the target survival prediction model, resulting in significantly improved accuracy in cancer survival prediction information. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a structural diagram of a cancer survival prediction system provided in an embodiment of this application; Figure 2 This is a structural diagram of a feature extraction module in a cancer survival prediction system provided in an embodiment of this application; Figure 3 This is a structural diagram of a fusion module in a cancer survival prediction system provided in an embodiment of this application; Figure 4 This is a structural diagram of an effect value calculation module in a cancer survival prediction system provided in an embodiment of this application; Figure 5 This is a structural diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0012] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0013] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0014] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0015] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0016] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [the described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [the described condition or event] is detected," or "in response to detection of [the described condition or event]."

[0017] In specific implementations, the terminals described in the embodiments of this application include, but are not limited to, other portable devices such as mobile phones, laptop computers, or tablet computers with touch-sensitive surfaces (e.g., touchscreen displays and / or touchpads). It should also be understood that in some embodiments, the device is not a portable communication device, but a desktop computer with touch-sensitive surfaces (e.g., touchscreen displays and / or touchpads).

[0018] The following discussion describes terminals that include displays and touch-sensitive surfaces. However, it should be understood that terminals may include one or more other physical user interface devices such as physical keyboards, mice, and / or joysticks.

[0019] The terminal supports a variety of applications, such as one or more of the following: drawing applications, presentation applications, word processing applications, website creation applications, disc burning applications, spreadsheet applications, game applications, telephone applications, video conferencing applications, email applications, instant messaging applications, exercise support applications, photo management applications, digital camera applications, digital camcorder applications, web browsing applications, digital music player applications, and / or digital video player applications.

[0020] Various applications that can run on the terminal can use at least one common physical user interface device, such as a touch-sensitive surface. One or more functions of the touch-sensitive surface and the corresponding information displayed on the terminal can be adjusted and / or changed between and / or within applications. In this way, the terminal's common physical architecture (e.g., the touch-sensitive surface) can support various applications with user interfaces that are intuitive and transparent to the user.

[0021] It should be understood that the sequence number of each step in this embodiment does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application embodiment.

[0022] To illustrate the technical solution described in this application, specific embodiments are provided below.

[0023] See Figure 1 , Figure 1 This is a structural diagram of a cancer survival prediction system provided in an embodiment of this application. Figure 1 As shown, a cancer survival prediction system 100 includes: a feature extraction module 101, a fusion module 102, an effect value calculation module 103, and a prediction module 104.

[0024] The feature extraction module 101 is used to extract global gene features from the patient's gene expression profile, extract pathological text features from the patient's pathological text data, and extract gene-level pathway features from the patient's biological pathway data based on the gene expression profile; the gene expression profile contains multiple genes, and the biological pathway data contains multiple biological pathways.

[0025] The fusion module 102 is used to perform feature fusion on the global gene features, the pathological text features and the gene-level pathway features to obtain multimodal fusion features.

[0026] The effect value calculation module 103 is used to calculate the total effect value and the natural direct effect value based on the global gene features, the pathological text features and the multimodal fusion features, and to obtain the total indirect effect value by subtracting the total effect value and the natural direct effect value; the total effect value includes the true biological effect value and the text bias effect value, the natural direct effect value includes the text bias effect value, and the total indirect effect value includes the true biological effect value.

[0027] The prediction module 104 is used to input the total indirect effect value into the target survival prediction model to obtain the cancer survival prediction information output by the target survival prediction model.

[0028] Gene expression profiling refers to a dataset containing the expression values ​​of multiple genes in a patient's body, denoted as . , , which includes The gene expression value is the gene .

[0029] Pathological text data, also known as histopathological reports, are structured diagnostic documents containing highly semantically abstract and clinical information, denoted as... It contains textual information such as the anatomical location, pathological features, and diagnostic conclusions related to the tumor.

[0030] Biological pathway data refers to the set of gene functional pathways defined by biomedical databases, denoted as... , , which includes biological pathways Each biological pathway is composed of a set of genes that have synergistic biological functions (tumor development and progression).

[0031] Global gene features are feature vectors that characterize the global distribution characteristics of gene expression profiles, denoted as... It can reflect the overall expression pattern of genes.

[0032] Pathological text features are computer-recognizable feature vectors obtained by semantically encoding unstructured pathological text data, denoted as . .

[0033] Gene-level pathway features refer to features that integrate higher-order gene synergistic effects and interactions between biological pathways, denoted as... It characterizes the structured association between genes and biological pathways.

[0034] The total effect (TE) is the effect size calculated when all information is available, denoted as . It includes the true biological effect value (causal contribution) and the text bias effect value (spurious association).

[0035] The true biological effect value is the core contribution effect based on cancer pathological mechanisms and real clinical causal relationships, and serves as an effective basis for target survival prediction models to predict cancer survival. The true biological effect value consists of two parts of truly effective information: first, core biological association information such as higher-order synergistic effects and inter-pathway interactions between gene expression profiles and biological pathway data, reflecting the intrinsic molecular mechanisms of cancer development; second, clinical semantic information in pathological texts that is truly relevant to patient survival prognosis (such as tumor invasion degree and pathological grade), providing prognostic value based on the actual pathological characteristics of the disease.

[0036] Text bias effect values ​​are the biased contribution effects caused by spurious statistical associations in pathological texts, and are ineffective components that need to be identified and suppressed. Text bias effect values ​​originate from "linguistic shortcuts" in pathological texts, where non-causal terms in pathology reports (such as anatomical location terms like "lymph node" or purely descriptive terms) form spurious statistical associations with survival prognosis due to frequency distribution biases in the training dataset. Traditional survival prediction models are prone to overlearning these associations during training, making predictions based on these spurious statistical associations. These effect values ​​lack actual disease pathology mechanism support and generally do not have real clinical prognostic guidance value. In complex clinical scenarios, they reduce the decision reliability and generalization ability of survival prediction models.

[0037] The Natural Direct Effect (NDE) is data that includes only text bias effects, denoted as... This characterizes spurious statistical associations in pathological texts.

[0038] The total indirect effect (TIE) is the effect value after removing text bias, denoted as . It only includes real biological effect values.

[0039] In some embodiments, the feature extraction module 101 designs feature extraction logic for three types of data: genes, pathological texts, and biological pathways. This enables refined and targeted feature extraction of multi-source heterogeneous cancer-related biomedical data, achieving characteristic representation of multi-source heterogeneous biomedical data and avoiding the simple homogenization of data from different modalities.

[0040] Among them, the global gene features extracted based on gene expression profiles preserve the global regularity of gene expression, the pathological text features extracted based on pathological text data capture clinical semantic information, and the gene-level pathway features extracted based on gene expression profiles and biological pathway data characterize the structured biological association between genes and biological pathways, providing a high-quality feature foundation for subsequent multimodal feature fusion.

[0041] like Figure 2 As shown, Figure 2 This is a structural diagram of a feature extraction module in a cancer survival prediction system provided in this application embodiment. The feature extraction module 101 includes: a gene feature extraction unit 1011, a text feature extraction unit 1012, and a pathway feature extraction unit 1013.

[0042] The gene feature extraction unit 1011 is used to extract the global gene features from the gene expression profile.

[0043] The text feature extraction unit 1012 is used to extract the pathological text features from the pathological text data.

[0044] The pathway feature extraction unit 1013 is used to construct a gene pathway hypergraph based on the membership relationship between multiple genes in the gene expression profile and multiple biological pathways in the biological pathway data, and to perform feature aggregation based on the gene pathway hypergraph to obtain the gene-level pathway features.

[0045] The gene feature extraction unit is a functional unit specifically designed to process gene expression profiles and extract global gene features. Its core is a global feature extraction model such as Self-Normalizing Neural Networks (SNN).

[0046] The text feature extraction unit is a functional unit specifically designed to process pathological text data and extract semantic features from it. Its core is a pre-trained language model in the clinical or biomedical field, such as the Clinical Bidirectional Encoder Representations from Transformers (ClinicalBERT).

[0047] The pathway feature extraction unit is a functional unit specifically designed to model the structured relationship between genes and pathways and extract gene-level pathway features. Its core is the Hypergraph Neural Network (HGNN) architecture, which is the core unit for realizing high-order biological feature representation.

[0048] In some embodiments, the gene feature extraction unit performs global feature extraction on the gene expression profile through continuous linear layers and nonlinear activation layers in a pre-trained SNN, thereby obtaining global gene features, i.e. This unit does not consider local associations between genes and has no complex computational overhead for local associations. Through the gene feature extraction unit, it can efficiently capture the global distribution characteristics of gene expression, providing comprehensive global genome characterization data for subsequent multimodal fusion.

[0049] In some embodiments, the text feature extraction unit performs semantic encoding processing on the preprocessed pathological text data (such as word segmentation and stop word removal) using a text encoder such as ClinicalBERT, converting the text sequence into low-dimensional, dense pathological text features, i.e. By using a text feature extraction unit, clinical semantic information in pathological text data is accurately captured, providing high-quality text feature representation data for subsequent multimodal fusion and causal bias removal.

[0050] In some embodiments, depending on the application scenario, pre-trained models in the biomedical field (such as Biomedical Bidirectional Encoder Representations from Transformers (BioBERT) or PubMed Bidirectional Encoder Representations from Transformers (PubMedBERT)) or large-scale generative pre-trained models (such as Large Language Models (LLMs)) can be selected to implement semantic encoding in order to obtain text semantic representation data of different dimensions.

[0051] In some embodiments, the pathway feature extraction unit obtains the affiliation between genes and biological pathways through a biomedical database, constructs a gene pathway hypergraph, performs feature aggregation of hypergraph nodes and hyperedges based on the gene pathway hypergraph, captures high-order synergistic effects between genes and complex interactions between pathways, and obtains gene-level pathway features accordingly.

[0052] Among them, biomedical databases include Reactome, HallmarkGene Sets, Kyoto Encyclopedia of Genes and Genomes (KEGG), WikiPathways, and Gene Ontology (GO), etc. Appropriate biomedical databases can be selected according to different biological pathway research needs.

[0053] The feature extraction module comprises three independent functional units, enabling specialized processing of data from different modalities and avoiding interference between feature extraction modes. Among them, the pathway feature extraction unit models the gene-pathway membership relationship through a hypergraph structure, overcoming the limitation of traditional neural networks that treat genes as independent features, and facilitating the effective capture of higher-order biological interactions.

[0054] In some embodiments, the pathway feature extraction unit 1013 includes: The hypergraph construction unit 10131 is used to construct each hypergraph node corresponding to each of the genes, and based on the membership relationship between the multiple genes and the multiple biological pathways, construct a hyperedge connecting the multiple hypergraph nodes to obtain the gene pathway hypergraph containing multiple hypergraph nodes and multiple hyperedges. The hypergraph processing unit 10132 is used to perform bidirectional feature aggregation of the hypergraph nodes and hyperedges on the gene pathway hypergraph through hypergraph convolution operations to obtain the target interaction features corresponding to each hypergraph node. The pathway aggregation unit 10133 is used to aggregate the target interaction features of all the hypergraph nodes associated with each hyperedge in the gene pathway hypergraph to obtain initial pathway features, and then to perform weighted fusion of the initial pathway features through a pathway self-attention mechanism to obtain the gene-level pathway features.

[0055] Hypergraph building units are functional units used to construct gene pathway hypergraphs.

[0056] Gene pathway hypergraph is noted as ,in, For a hypergraph node set, It is a super-edge set. This is a gene pathway association matrix.

[0057] A hypergraph node corresponds one-to-one with a gene, with each hypergraph node representing a gene. The set of hypergraph nodes is represented as follows: .

[0058] Each hyperedge corresponds one-to-one with a biological pathway. Each hyperedge represents a biological pathway and connects the hypergraph nodes corresponding to all genes in that pathway. The set of hyperedges is represented as follows: .

[0059] The gene pathway association matrix is ​​a matrix that describes the connection relationships between nodes and hyperedges in a hypergraph. It is constructed based on gene expression profiles, biological pathway data, and the membership relationships between genes and biological pathways. , Indicates the first The hypergraph node corresponding to the gene belongs to the _th ... The hyperedge corresponding to each biological pathway This indicates that the two are not related.

[0060] Membership relationship specifically refers to the affiliation relationship between genes and biological pathways, that is, whether a gene is a component of a certain biological pathway: if a gene is a gene of that biological pathway, the two have a membership relationship; if a gene has no functional connection with that biological pathway and is not a component of it, then the two do not have a membership relationship.

[0061] Bidirectional feature aggregation refers to the two-stage message passing in the gene pathway hypergraph from hypergraph nodes to hyperedges and from hyperedges to hypergraph nodes.

[0062] The target interaction feature is a gene fusion feature obtained after multiple rounds of bidirectional feature aggregation of a hypergraph node, which integrates the biological information of the pathway to which the hypergraph node belongs and the associated pathways.

[0063] Inter-pathway self-attention mechanisms are attention mechanisms that weight initial pathway features, dynamically weigh the importance of each biological pathway, and capture the synergistic relationships between pathways.

[0064] In some embodiments, the hypergraph construction unit maps each gene to a hypergraph node and each biological pathway to a hyperedge. Based on the gene pathway association matrix, the hyperedge corresponding to each biological pathway is connected to all the hypergraph nodes it contains, ultimately constructing a gene pathway hypergraph. By constructing a hypergraph structure that fits the natural biological topology of genes and pathways using the hypergraph construction unit, the limitations of traditional binary interaction modeling are overcome, laying a structured foundation for subsequent capture of high-order synergistic effects of multiple genes and complex interactions between pathways.

[0065] In some embodiments, when constructing a gene pathway hypergraph, the hypergraph nodes are calculated. Degree and hyper-edge The degree of a hypergraph node is denoted as [degree] to facilitate subsequent bidirectional feature aggregation. , representing the number of hyperedges connected to the hypergraph node, and the degrees of multiple hypergraph nodes constitute the node degree matrix. (The diagonal elements in the matrix represent the number of hyperedges connecting each hypergraph node). The degree of a hyperedge is denoted as... The degree of a hyperedge represents the number of hypergraph nodes it contains. Since there is a one-to-one correspondence between hypergraph nodes and genes, and a one-to-one correspondence between hyperedges and biological pathways, it is also equivalent to the number of genes contained in the biological pathway corresponding to that hyperedge. The degrees of multiple hyperedges constitute a hyperedge degree matrix. (The diagonal elements in the matrix represent the number of hypergraph nodes contained in each hyperedge).

[0066] In some embodiments, the hypergraph processing unit performs hypergraph convolution operations on the constructed gene pathway hypergraph, and captures high-order synergistic effects between genes and multi-scale interaction information between biological pathways through multi-round bidirectional feature aggregation of nodes and hyperedges, thereby obtaining the target interaction features of each hypergraph node.

[0067] In some embodiments, the hypergraph processing unit is configured to: project the gene expression profile onto a set projection space to obtain a node feature matrix corresponding to the gene pathway hypergraph; the node feature matrix contains node features that correspond one-to-one with multiple hypergraph nodes; based on the node feature matrix, perform feature aggregation on all hypergraph nodes associated with each hyperedge to obtain a hyperedge feature matrix corresponding to multiple hyperedges; based on the hyperedge feature matrix, perform feature aggregation on all hyperedges associated with each hypergraph node to obtain a new node feature matrix corresponding to multiple hypergraph nodes; return to execute the step of performing feature aggregation on all hypergraph nodes associated with each hyperedge based on the node feature matrix to obtain a hyperedge feature matrix corresponding to multiple hyperedges, until a set number of bidirectional feature aggregations are completed, and use the node features corresponding to each hypergraph node in the node feature matrix obtained when the set number of aggregations is reached as the target interaction features corresponding to each hypergraph node.

[0068] The projection space is defined as a pre-defined low-dimensional feature space, whose feature dimensions are represented as follows: It is used to map high-dimensional gene expression profiles into low-dimensional node features.

[0069] The node feature matrix is ​​a matrix containing multiple node features, i.e., gene features, denoted as . , Each row in the node feature matrix corresponds to a feature of a hypergraph node.

[0070] The hyperedge feature matrix is ​​a matrix containing multiple hyperedge features, denoted as . Each row of the hyperedge feature matrix corresponds to a feature of a hyperedge, representing the feature representation of the hyperedge by aggregating node features.

[0071] Set number of times The number of layers in the hypergraph convolution is a hyperparameter that can be adjusted. More layers result in a richer array of gene pathway interactions captured. This application implements hypergraph convolution operations through multi-layer hypergraph convolution.

[0072] In some embodiments, the core of the hypergraph processing unit is the message passing mechanism of the hypergraph neural network, and the specific implementation steps are as follows: (1) Node feature initialization: Gene expression profile Projecting onto a defined projection space, the initial node feature matrix is ​​obtained by defining the linear layer of the projection space and the ReLU (Rectified Linear Unit) activation function. Specifically =ReLU(Linear(G)) ,in, ; (2) Bidirectional feature aggregation: Set the number of layers in the hypergraph convolution to be... For each round of bidirectional feature aggregation, feature aggregation from hypergraph nodes to hyperedges and from hyperedges to hypergraph nodes are performed sequentially; where the feature aggregation from hypergraph nodes to hyperedges is based on the node feature matrix of the current round. For each hyperedge, feature aggregation is performed on all hypergraph nodes associated with it to obtain the hyperedge feature matrix. , , Features from hyperedges to hypergraph nodes are aggregated into a feature matrix based on the hyperedge. For each hypergraph node, feature aggregation is performed on all hyperedges associated with it, combined with a learnable weight matrix. and nonlinear activation functions A new node feature matrix is ​​obtained. , .

[0073] In some embodiments, after obtaining a new node feature matrix Subsequently, the feature matrix of this node can be optimized using layer normalization and the ReLU activation function. This enhances the ability to represent features.

[0074] In some embodiments, when completed After round-trip bidirectional feature aggregation, the final node feature matrix is ​​obtained. The node features corresponding to each hypergraph node are used as the target interaction features of the hypergraph nodes. These target interaction features are fused with... Layer-scale gene-pathway and pathway-pathway interaction information.

[0075] In some embodiments, the message passing mechanism of the hypergraph neural network can be adjusted according to the actual application requirements to a hyperedge aggregation operator based on self-attention mechanism, a multi-scale hypergraph convolution operator, or a hybrid deep architecture that combines graph convolutional network (GCN) and hypergraph convolution (HGNN).

[0076] By employing a feature initialization-multi-round message passing-target feature output process, multi-scale high-order representation of gene features is achieved. Each round of hypergraph node-hyperedge bidirectional feature aggregation incorporates normalization of both the node degree matrix and the hyperedge degree matrix, avoiding issues of excessively large or small feature values. Furthermore, the introduction of layer normalization and the ReLU activation function further enhances convolutional stability and non-linear feature expression capabilities, effectively capturing high-order synergistic effects within biological pathways and cascading effects between pathways.

[0077] In some embodiments, the pathway aggregation unit determines gene-level pathway features through gene-to-pathway pooling operations and inter-pathway self-attention weighting operations.

[0078] In the gene-to-pathway pooling stage, the target interaction features of all hypergraph nodes associated with each hyperedge are aggregated to obtain the initial pathway features. ,in, In this stage, high-order gene interaction features, i.e. target interaction features, are aggregated into pathway features, thereby integrating the collaborative interaction information of all genes within a single biological pathway.

[0079] During the inter-path self-attention weighting stage, the initial path features are... A self-attention layer between input pathways calculates the weight coefficients of each biological pathway through attention scoring, and then performs weighted fusion of the initial pathway features to finally obtain gene-level pathway features. Dynamically learn the weight coefficients of each biological pathway, accurately weigh the importance of biological pathways, precisely capture the complex interactions between biological pathways, and improve the effectiveness of biological associations of pathway features.

[0080] This application achieves refined processing of the entire process of gene pathway hypergraph construction, feature aggregation, and pathway weighting through hypergraph construction unit, hypergraph processing unit, and pathway aggregation unit. It breaks through the limitation of traditional graph neural networks that can only model binary pairwise relationships, captures high-order synergistic effects between multiple genes and complex interaction relationships between biological pathways, and improves the biological interpretability of gene-level pathway features.

[0081] In some embodiments, after the feature extraction module 101 obtains global gene features, pathological text features and gene-level pathway features, it sends these features to the fusion module 102, which integrates the global gene features, pathological text features and gene-level pathway features to obtain multimodal fusion features.

[0082] like Figure 3 As shown, Figure 3 This is a structural diagram of a fusion module in a cancer survival prediction system provided in an embodiment of this application. The fusion module 102 includes: a feature splicing unit 1021, a cross-modal processing unit 1022, and a multi-granularity fusion unit 1023.

[0083] The feature splicing unit 1021 is used to splice the global gene features and the pathological text features to obtain spliced ​​features.

[0084] The cross-modal processing unit 1022 is used to perform cross-modal feature interaction on the spliced ​​features and the gene-level pathway features using a cross-modal attention mechanism to obtain cross-modal features.

[0085] The multi-granularity fusion unit 1023 is used to perform feature fusion on the cross-modal features, the global gene features and the pathological text features through residual aggregation operation to obtain the multi-modal fused features.

[0086] The feature concatenation unit is used to perform concatenation operations on different feature vectors.

[0087] The concatenated feature is a vector formed by concatenating global gene features and pathological text features, denoted as... .

[0088] The CrossAttention mechanism uses concatenated features as query vectors and gene-level pathway features as key and value vectors to achieve cross-modal feature interaction between genes, text, and pathways.

[0089] Cross-modal features are the output of cross-modal attention mechanisms, representing high-order interaction information that integrates pathological text, global genes, and biological pathways. They are denoted as... .

[0090] Multimodal fusion features are the final fused features that combine single-modal fine-grained features and cross-modal high-order interaction features, denoted as... .

[0091] In some embodiments, the feature splicing unit will combine global gene features and pathological text features Perform feature concatenation to obtain concatenated features. , It can quickly integrate global gene features and pathological text features to construct a cross-modal unified context vector.

[0092] In some embodiments, query vectors adapted to cross-modal attention mechanisms can be flexibly constructed. This concatenated feature... as query vector Alternatively, project the concatenated features and use the resulting features as the query vector. , ,in, For learnable weight matrix, This is the bias term. The projection can be a linear layer projection, which enables feature dimension matching and semantic representation enhancement, laying a high-quality feature foundation for subsequent accurate retrieval of associated pathway information.

[0093] In some embodiments, the cross-modal processing unit employs a cross-modal attention mechanism to query vectors. As input, gene-level pathway features as a key vector Sum value vector Perform cross-modal attention computation to obtain cross-modal features. By leveraging a cross-modal attention mechanism, we can achieve precise interaction between text-gene context and pathway features, efficiently retrieve biological pathway information related to clinical texts, uncover higher-order complementary associations across modalities, and generate cross-modal features that integrate multimodal semantics.

[0094] In some embodiments, the multi-granularity fusion unit performs residual aggregation operations on cross-modal features. Global genetic characteristics and pathological text features The features are then fused to obtain the final multimodal fusion features. , ,in, This represents the residual aggregation operator. By leveraging residual aggregation, the information decay problem during the fusion process is mitigated, enabling deep integration of multi-granularity and multi-dimensional features.

[0095] The fusion module achieves hierarchical fusion of multimodal features through three sub-units, breaking through the limitations of traditional fusion methods such as vector concatenation and cross-attention: the cross-modal attention mechanism realizes the interaction of high-order features of the three modalities of genes, text, and pathways, and captures complementary information across modalities; the residual aggregation operation effectively reduces information attenuation during the fusion process, preserving both the single-modal fine-grained features of genes and text and integrating high-order collaborative signals across modalities.

[0096] The fusion module balances the independence of single-modal features with the interactivity of cross-modal features, providing high-quality multimodal fusion features for subsequent causal debiasing and survival prediction.

[0097] The effect value calculation module 103 is the core quantitative execution module for realizing causal debiasing in pathological texts. By decomposing causal effects, it accurately quantifies the contribution of different inference paths to cancer survival prediction, providing clear numerical basis for subsequent adaptive debiasing calculations.

[0098] The design of the dual calculation paths for total effect value and natural direct effect value in the effect value calculation module 103, as well as the definition of each effect value, are all centered around solving the core technical problems of spurious associations (text bias) in pathological texts and model-dependent language shortcuts.

[0099] The effect value calculation module includes two paths: total effect value calculation and natural direct effect value calculation. In essence, it is an engineered reproduction of the two core inference paths in the Structural Causal Model (SCM), corresponding to the causal inference path and text bias path in cancer survival prediction.

[0100] While multimodal fusion modules achieve deep integration of global gene features, pathological text features, and gene-level pathway features, the fused full-information features inevitably contain spurious statistical associations from the pathological text (such as non-causal associations of anatomical locations and descriptive terms due to dataset distribution bias). If the fused features are used directly for survival prediction, the survival prediction model will still rely on "linguistic shortcuts" in the text, failing to fundamentally address the problem of decreased prediction reliability and generalization ability caused by text bias. Existing technologies lack quantitative methods to distinguish between "true causal contributions" and "spurious bias contributions," and simple feature filtering or weight adjustment cannot accurately isolate text bias and its resulting spurious bias contributions.

[0101] Therefore, this application sets up a dedicated effect value calculation module 103, which decomposes the predicted contribution into different quantifiable effect values ​​based on the intervention idea of ​​structural causal model, so that the subsequent debiasing operation changes from "unfounded empirical adjustment" to "precise calculation with numerical value", ensuring the effectiveness and controllability of debiasing.

[0102] In structural causal models, there exist indirect causal paths ( Gene expression profile ( ), pathological text data ( Mediation variables were obtained through multimodal fusion. Then by Determine the survival outcome This path represents the true causal collaborative contribution of multimodal features. Structural causal models also exhibit direct bias paths (…). ): Bypassing the multimodal fusion process, directly using pathological text data ( Derive the survival outcome This path represents the pure bias contribution from text bias. The total effect value and the natural direct effect value are calculated along two paths corresponding one-to-one with the above paths. By calculating through these two paths separately, the causal contribution and the bias contribution are completely separated. If only a single path is used for calculation, it is impossible to distinguish which features in the fused features are genuine multimodal collaborative information and which are spurious textual associations. Subsequent debiasing operations lack precise quantitative basis, easily leading to incomplete or excessive debiasing and loss of effective textual information.

[0103] The total effect value is a quantitative result of the normal inference path of the whole information. It is obtained by directly inputting global genetic features, pathological text features, and multimodal fusion features into the target effect value calculation model by the effect value calculation unit. Its core connotation is the total predictive value that includes both true causal contributions and spurious bias contributions. It integrates the true biological and clinical semantic collaborative information of genomics and pathological texts, but inevitably also contains spurious statistical associations formed by the distribution bias of the dataset in the pathological text.

[0104] The natural direct effect value is the quantification result of the pure text bias path. It is obtained by replacing global gene features and multimodal fusion features with preset empty features in the effect value calculation unit, and then inputting them together with pathological text features into the target effect value calculation model. This operation blocks the indirect causal path in the structural causal model through causal intervention (do-intervention), preventing the model from obtaining biological information from the genome and collaborative information from multiple modalities. It can only rely on pathological text features for prediction. Therefore, its core content is the pure spurious association contribution brought by the shortcut of pathological text, which is the bias quantification value that needs to be removed in the subsequent debiasing calculation.

[0105] The effect value calculation module obtains the total effect value and the natural direct effect value through dual-path calculation, which respectively quantifies the "total contribution of all information" and the "contribution of pure text bias". This provides a clear and calculable numerical basis for the weighted bias removal operation of the subsequent bias removal calculation unit, so as to accurately remove false associations in the text, while retaining the causal information with real prognostic value in the pathological text.

[0106] like Figure 4 As shown, Figure 4 This is a structural diagram of an effect value calculation module in a cancer survival prediction system provided in this application embodiment. The effect value calculation module 103 includes: an effect value calculation unit 1031 and a biased calculation unit 1032.

[0107] The effect value calculation unit 1031 is used to input the global gene features, the pathological text features and the multimodal fusion features into the target effect value calculation model to obtain the total effect value output by the target effect value calculation model; and to replace the global gene features and the multimodal fusion features with preset empty features, and then input them together with the pathological text features into the target effect value calculation model to obtain the natural direct effect value output by the target effect value calculation model.

[0108] The debiasing calculation unit 1032 is used to perform weighted calculation on the natural direct effect value and the debiasing weight to obtain the debiased natural effect value, and to subtract the total effect value from the debiased natural effect value to obtain the total indirect effect value.

[0109] The effect value calculation unit is a functional unit that calculates effect values ​​based on the structural causal model.

[0110] The target effect value calculation model is a trained neural network model that can output effect values ​​under different intervention scenarios, and it is the core carrier for effect value calculation.

[0111] The pre-defined empty features are feature vectors that simulate missing information. They can be represented by zero vectors, constant vectors, trainable vectors learned from end-to-end learning, random noise vectors that conform to a specific distribution, or adversarial samples generated by adversarial training.

[0112] The bias removal calculation unit is a functional unit that realizes text bias removal. Its core is to perform a difference operation on the effect value of the adaptive bias removal weight.

[0113] The bias removal weights are learnable parameters, denoted as . Used for dynamic balancing of depolarization intensity, with a value range of [value range missing]. .

[0114] The total indirect effect value is the weighted and subtracted effect value, denoted as . .

[0115] The core of the effect size calculation module is counterfactual debiasing, which separates the causal contribution and textual bias in the total effect by constructing factual and counterfactual scenarios. The specific implementation steps are as follows: (1) Total effect size (TE) calculation: The effect size calculation unit will calculate the global gene characteristics. Pathological text features and multimodal fusion features Input into the target effect calculation model, at this time the structural causal model and All were activated, and the total effect value was obtained. ,in, The mapping function for calculating the target effect size model is... This indicates that all paths are active.

[0116] (2) Calculation of Natural Direct Effect Value (NDE): The effect value calculation unit constructs a counterfactual scenario through do-intervention, and incorporates global gene characteristics. and multimodal fusion features Replace them with preset empty features respectively and , to feature pathological texts and preset empty features and Input the target effect value calculation model, at this time the blockade Activate only ( 0), to obtain the natural direct effect value .

[0117] (3) Adaptive bias removal calculation: The bias removal calculation unit introduces learnable bias removal weights. First, the direct natural effect values ​​are weighted to obtain the debiased natural effect values. Then the total effect value Subtracting the natural effect value after removing bias yields the total indirect effect value after removing text bias. ,in, To completely eliminate bias, This indicates that the retained portion of the text has prognostic value and is a direct signal.

[0118] In some embodiments, the total indirect effect value can also be expressed as , , equivalent to Completely eliminate bias.

[0119] The effect value calculation module achieves precise removal of spurious associations in pathological texts through counterfactual debiasing inference, overcoming the limitations of traditional multimodal models that rely on linguistic shortcuts for prediction. Specifically, the dual-context design based on a structural causal model can deconstruct the true biological effect and textual bias effect in the total effect from a technical perspective, ensuring that the model's decisions are based on the real disease pathology mechanism, significantly improving the model's decision reliability and generalization ability in complex clinical scenarios; the learnable debiasing weights achieve a dynamic adaptive balance of debiasing intensity, avoiding the loss of prognostic information caused by excessive debiasing, while effectively suppressing non-causal spurious statistical associations in the text.

[0120] In some embodiments, the target survival prediction model in the prediction module 104 can be a classic model in the field of survival analysis (such as the Cox Proportional Hazards Model, random survival forest, discrete-time survival model, etc.), or it can be a custom neural network model.

[0121] The total indirect effect value output by the effect value calculation module is input into the target survival prediction model. The model outputs the corresponding cancer survival prediction information through a nonlinear transformation of the total indirect effect value.

[0122] In some embodiments, cancer survival prediction information includes patient risk scores, survival probability curves, risk stratification results, and predicted event occurrence times.

[0123] Among them, the patient risk score is used to quantify the patient's cancer prognostic risk. The higher the score, the higher the patient's risk of death and the shorter the survival time. The survival probability curve is used to predict the patient's survival probability at different follow-up time points, which intuitively reflects the patient's survival trend. The risk stratification result refers to dividing patients into high, medium and low risk groups, which provides a basis for the formulation of clinical personalized treatment plans. The event occurrence prediction time refers to predicting the time when the patient will experience the target clinical event (such as cancer recurrence or death).

[0124] The prediction module uses the total indirect effect value after removing text bias for survival prediction, ensuring that the model's prediction results are based on real biological causal relationships rather than spurious statistical associations in pathological texts. Combined with the professional survival analysis capabilities of the target survival prediction model, it outputs accurate and comprehensive cancer survival prediction information, providing reliable decision-making basis for clinicians to develop personalized treatment plans and improve patient prognosis.

[0125] In some embodiments, the system further includes a model training module, configured to: input global gene sample features, pathological text sample features, and multimodal fusion sample features into an effect value calculation model to obtain the total sample effect value output by the effect value calculation model; replace the global gene sample features and the multimodal fusion sample features with the preset empty features, and input them together with the pathological text sample features into the effect value calculation model to obtain the sample natural direct effect value output by the effect value calculation model; perform probability normalization processing on the total sample effect value and the sample natural direct effect value respectively to obtain a first probability distribution corresponding to the total sample effect value and a second probability distribution corresponding to the sample natural direct effect value; calculate a first survival loss value corresponding to the total sample effect value and a second survival loss value corresponding to the sample natural direct effect value based on the first probability distribution and the second probability distribution respectively; calculate the total model loss based on the first survival loss value, the second survival loss value, and a distribution divergence loss mechanism; iteratively optimize the model parameters of the effect value calculation model based on the total model loss until the total model loss reaches a set loss value, and use the effect value calculation model corresponding to the total model loss that reaches the set loss value as the target effect value calculation model.

[0126] The total model loss refers to the joint loss obtained by fusing the first survival loss and the second survival loss through the distribution divergence loss mechanism, which is used to iteratively optimize the model parameters for calculating the effect value.

[0127] The loss value is set as a pre-defined loss threshold, which is the loss value at which the model converges. Reaching this value indicates that the model training is complete.

[0128] In some embodiments, the model training module trains the effect value calculation model based on discrete survival prediction and joint loss optimization. The specific implementation steps are as follows: (1) Data preprocessing: The continuous survival time of all patients in the training set was used to preprocess the data. Discretize into A time interval [0, ), ,[ , ), for each patient Label the corresponding survival information, i.e., the observation time. (Falling within a certain discrete time interval mentioned above) and event indicator ( This indicates that the target event has occurred, and the data is not censored. (This indicates that the target event was not observed, which is right-censored data). At the same time, global gene sample features, pathological text sample features, and multimodal fusion sample features of patients are extracted from the training set. (2) Calculation of sample effect value: Based on the factual and counterfactual scenarios, the global gene sample features, pathological text sample features and multimodal fusion sample features of the patient are input into the effect value calculation model to obtain the total effect value of the sample. At the same time, the global gene sample features and multimodal fusion sample features of the patient are set to preset empty features and input into the effect value calculation model together with the pathological text sample features to obtain the natural direct effect value of the sample. (3) Probability normalization: The total effect value of the patient sample is normalized using methods such as the softmax function. and sample natural direct effect value Probability normalization is performed to map the unbounded effect value to... The probability distribution of events in dimension 1 yields the first probability distribution. Second probability distribution , , In the probability distribution, each dimension corresponds to the probability of a target event occurring within a time interval, and the sum of the probabilities of all dimensions in the probability distribution is 1. (4) Survival loss calculation: Define a general survival loss function based on negative log-likelihood. Calculate the survival loss values ​​corresponding to the first probability distribution and the second probability distribution respectively, adapt the error penalty for censored and uncensored data, and calculate the first survival loss based on the first probability distribution. Calculate the second survival loss based on the second probability distribution Specifically, for patients The survival loss function is ,in, For patients Observation time The first term of the discrete time interval to which it belongs penalizes the prediction error of a single interval for non-censored events, and the second term penalizes the prediction bias of the total probability of subsequent intervals for right-censored data. (5) Divergence loss calculation: Based on the Kullback-Leibler (KL) divergence loss mechanism, This aligns the sharpness of the second probability distribution with that of the first probability distribution, preventing unimodal branch degradation. For the total number of patients, To stop the gradient, the first probability distribution of the total effect branch is used as the fixed objective to avoid gradient backpropagation to the total effect branch, and only the natural direct effect branch is optimized. (6) Joint total loss calculation: The three losses are weighted and fused to obtain the total model loss. ,in, and To balance the hyperparameters of each loss term and adapt them to the training requirements of different cancer datasets; (7) Model parameter optimization: Gradient descent, adaptive moment estimation (Adam), stochastic gradient descent (SGD) and other optimization algorithms are used to iteratively update the model parameters of the effect value calculation model in combination with the total model loss until the total model loss reaches the set loss value. The model at this time is used as the target effect value calculation model.

[0129] The model training module, based on a joint loss optimization strategy, completed end-to-end training of the effect value calculation model. Its core technological achievements are reflected in multi-dimensional training optimization and capability enhancement: Probability normalization maps the predicted effect value to an event probability distribution over a discrete time interval, converting unbounded predicted values ​​to compliant probability distributions, thus laying the foundation for loss calculation. The survival loss function, designed based on negative log-likelihood, is specifically adapted to the characteristics of clinical data in cancer survival prediction, effectively handling right-censored data in survival analysis and ensuring the rationality and clinical relevance of loss calculation. The introduced KL divergence loss mechanism effectively avoids the training degradation problem of single-modal branches in the text, ensuring the model's accurate capture and quantification of textual bias information. Through joint optimization of multiple loss terms, the effect value calculation model can simultaneously learn the true causal relationships between multimodal features and the bias features in the text. The final trained target effect value calculation model possesses superior effect value prediction accuracy and causal inference ability, providing a reliable basis for subsequent adaptive causal debiasing operations.

[0130] In some embodiments, the target effect value calculation model is obtained by the effect value calculation model training unit in the model training module through joint loss iterative optimization training. This unit provides an accurate effect value quantification basis for subsequent causal correction. The model training module also integrates a survival prediction model training unit, which is used to complete end-to-end training of the survival prediction model based on the total indirect effect value of the sample and the cancer survival information of the sample, to obtain the target survival prediction model.

[0131] In some embodiments, the survival prediction model training unit and the effect value calculation model training unit share the underlying parameters of the multimodal feature extraction network. The parameter sharing mechanism reduces the computational overhead of model training, while ensuring the consistency and continuity of multimodal feature representation, thereby further improving the overall model training efficiency and prediction performance.

[0132] In some embodiments, the system further includes a parameter optimization module for: optimizing and adjusting the debiasing weights based on the cancer survival prediction information and the actual cancer survival information.

[0133] Accurate information on cancer survival is the patient's actual survival outcome, including information such as actual follow-up time and actual events.

[0134] The parameter optimization module uses error feedback between cancer survival prediction information and actual information to dynamically adjust the debiasing weights, ensuring that while removing spurious correlations in the text, it retains causal information with real prognostic value in the pathological text to the greatest extent possible. The specific implementation process is as follows: (1) Obtain cancer survival prediction information and corresponding real cancer survival information; (2) Calculate the survival difference value between the actual cancer survival information and the cancer survival prediction information. This survival difference value quantifies the deviation between the prediction result and the actual survival result under the current debiasing weight. (3) Using optimization algorithms such as Adam and SGD, with the goal of minimizing survival differences, biased weights are adjusted. Perform backpropagation and iterative updates, continuously adjusting The possible values ​​of ; (4) Continue to repeat the above steps until the survival difference value tends to stabilize. At this point, the value corresponding to the stable survival difference value is... As the optimal debiasing weight, it is applied to subsequent causal debiasing to complete the optimization and adjustment of the debiasing weight.

[0135] In some embodiments, this application demonstrates the feasibility of its cancer survival prediction system in cancer survival prediction tasks through systematic experimental verification on large-scale real clinical datasets.

[0136] Extensive testing was conducted on real-world datasets from five major cancer cohorts, including Bladder Urothelial Carcinoma (BLCA), Breast Invasive Carcinoma (BRCA), Head and Neck Squamous Cell Carcinoma (HNSC), Stomach Adenocarcinoma (STAD), and Colon Adenocarcinoma and Rectum Adenocarcinoma (COADREAD). The test results are shown in Table 1 below. Table 1 presents the predictive performance of various survival prediction models provided in the embodiments of this application.

[0137] Table 1. Predictive performance of various survival prediction models

[0138] Table 1 uses the consistency index (C-Index) as the core evaluation indicator. All data were obtained through 5-fold cross-validation. The comparison objects include single-modal prediction models based on whole-slide images (WSI), pathological text data (Text), or omics data (Omics), i.e. gene expression profiles, as well as multimodal prediction models based on multimodal data.

[0139] The cancer survival prediction system proposed in this application is referred to as the CaDe-HG model.

[0140] As shown in Table 1, the CaDe-HG model performed best, achieving an average C-index of 0.706 across the five cancer types. It was the only predictive model among all participating models to exceed 0.7 in average C-index, demonstrating excellent performance across all cancer cohorts, particularly outstanding in the COADREAD cohort with a C-index of 0.776 ± 0.043. In contrast, other predictive models exhibited significant performance limitations. The average C-index of the pathology text unimodal predictive model was only 0.611, while the average C-index of the omics unimodal predictive models generally ranged from 0.65 to 0.67. Although traditional multimodal predictive models outperformed unimodal models, their average C-index did not exceed 0.7. This is mainly because traditional multimodal predictive models lacked high-order collaborative modeling capabilities for gene pathways and failed to address the text bias problem caused by spurious associations in pathology texts.

[0141] The experimental results in Table 1 not only verify the synergistic effect of high-order gene modeling and counterfactual text debiasing in the CaDe-HG model, but also demonstrate the feasibility and advantages of fusing pathological texts with omics data. At the same time, they solve the industry pain points of difficulty in acquiring whole-slice image data and high computational resource consumption.

[0142] To deeply analyze the contributions of each core module in the cancer survival prediction system, namely the CaDe-HG model, this application conducted systematic ablation experiments. By adopting strategies of removing or replacing specific modules, the impact of hypergraph neural networks and counterfactual text debiasing on the final performance of the CaDe-HG model was quantitatively evaluated. Detailed results are shown in Table 2, which is an ablation experiment data table provided by an embodiment of this application.

[0143] Table 2 Ablation Experiment Data

[0144] Among them, Full Model represents the complete CaDe-HG model, w / o Debiase is a variant model with the counterfactual text debiening module removed, Debiase (w / o G) is a variant model with only the counterfactual text debiening module retained but without higher-order gene modeling, Gene-Only (w / o G) is a pure gene monomodal model without higher-order gene modeling, Gene-Only (w / G) is a pure gene monomodal model with higher-order gene modeling, and Text-Only is a pure pathological text monomodal model.

[0145] Experimental results show that the complete CaDe-HG model performs best, with an overall C-index of 0.706. It also achieves the highest predictive performance across all cancer cohorts, with the COADREAD cohort showing particularly outstanding performance, achieving a C-index of 0.776±0.043. After removing the counterfactual text debiase module, the overall C-index of the model drops to 0.675, and the predictive performance of all cancer types declines to varying degrees. In particular, the COADREAD cohort experiences a sharp drop to 0.715±0.022, confirming the important role of this operation in suppressing spurious text associations and improving predictive robustness.

[0146] The performance of all other simplified variant models was significantly lower than that of the complete CaDe-HG model. The overall C-Index of the text-only model was only 0.624, and the overall C-Index of the single-gene modality model was only 0.658~0.675. This fully demonstrates the importance of the high-order co-modeling of genes in the hypergraph neural network of this application in improving feature representation ability and the counterfactual text debiasing in effectively suppressing non-causal spurious statistical associations in the text.

[0147] This application overcomes the limitations of traditional binary interaction modeling by constructing a global pathway-aware hypergraph structure, accurately capturing high-order synergistic effects between multiple genes and complex cross-pathway interaction information, significantly enhancing the ability to represent the underlying biological logic of genomic data; at the same time, relying on a unified hypergraph neural network architecture, it effectively overcomes the resource redundancy problem caused by the independent calculation of hundreds of biological pathways in existing technologies, and achieves synergistic optimization of gene modeling capabilities and computational efficiency.

[0148] This application introduces a counterfactual text debiasing technique to address the bias interference problem in medical texts. By precisely deconstructing the natural direct effects (NDEs) and total indirect effects (TIEs), it effectively identifies and suppresses spurious statistical associations in pathological texts, eliminates the influence of dataset distribution bias, and ensures that the survival prediction decisions of the survival prediction model are entirely based on the true pathological causal characteristics.

[0149] In this embodiment, global gene features, pathological text features, and gene-level pathway features are extracted based on the patient's gene expression profile, pathological text data, and biological pathway data, respectively. Multimodal fusion features are obtained by fusing these features. Then, based on the global gene features, pathological text features, and multimodal fusion features, a total effect value including the true biological effect value and the text bias effect value, as well as a natural direct effect value including the text bias effect value, are calculated. The difference between the total effect value and the natural direct effect value yields a total indirect effect value removed from the text bias. This total indirect effect value removed from the text bias is input into the target survival prediction model, resulting in significantly improved cancer survival prediction information.

[0150] Figure 5 This is a structural diagram of an electronic device provided in an embodiment of this application. As shown in the figure, the electronic device 4 of this embodiment includes: at least one processor 40 ( Figure 5 (Only one is shown in the diagram), memory 41, and computer program 42 stored in said memory 41 and executable on said at least one processor 40, which, when executed, implements the steps in any of the above method embodiments.

[0151] The electronic device 4 can be a desktop computer, laptop, handheld computer, or cloud server, etc. The electronic device 4 may include, but is not limited to, a processor 40 and a memory 41. Those skilled in the art will understand that... Figure 5 This is merely an example of electronic device 4 and does not constitute a limitation on electronic device 4. It may include more or fewer components than shown, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0152] The processor 40 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0153] The memory 41 can be an internal storage unit of the electronic device 4, such as a hard disk or memory. The memory 41 can also be an external storage device of the electronic device 4, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory 41 can include both internal and external storage units of the electronic device 4. The memory 41 is used to store the computer program and other programs and data required by the electronic device. The memory 41 can also be used to temporarily store data that has been output or will be output.

[0154] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0155] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0156] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0157] In the embodiments provided in this application, it should be understood that the disclosed systems / electronic devices and methods can be implemented in other ways. For example, the system / electronic device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of systems or units may be electrical, mechanical, or other forms.

[0158] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0159] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0160] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0161] The processes in the above-described embodiments can be implemented by a computer program product. When the computer program product is run on an electronic device, the electronic device executes the steps in the above-described method embodiments.

[0162] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A cancer survival prediction system, characterized in that, include: The feature extraction module is used to extract global gene features from the patient's gene expression profile, extract pathological text features from the patient's pathological text data, and extract gene-level pathway features from the patient's biological pathway data based on the gene expression profile. The gene expression profile contains multiple genes, and the biological pathway data contains multiple biological pathways; The fusion module is used to fuse the global gene features, the pathological text features, and the gene-level pathway features to obtain multimodal fused features; The effect value calculation module is used to calculate the total effect value and the natural direct effect value based on the global gene features, the pathological text features, and the multimodal fusion features, and to obtain the total indirect effect value by subtracting the total effect value and the natural direct effect value; the total effect value includes the true biological effect value and the text bias effect value, the natural direct effect value includes the text bias effect value, and the total indirect effect value includes the true biological effect value; The prediction module is used to input the total indirect effect value into the target survival prediction model to obtain the cancer survival prediction information output by the target survival prediction model.

2. The system according to claim 1, characterized in that, The feature extraction module includes: A gene feature extraction unit is used to extract the global gene features from the gene expression profile; A text feature extraction unit is used to extract the pathological text features from the pathological text data; The pathway feature extraction unit is used to construct a gene pathway hypergraph based on the membership relationship between multiple genes in the gene expression profile and multiple biological pathways in the biological pathway data, and to perform feature aggregation based on the gene pathway hypergraph to obtain the gene-level pathway features.

3. The system according to claim 2, characterized in that, The pathway feature extraction unit includes: A hypergraph construction unit is used to construct each hypergraph node corresponding to each of the genes, and based on the membership relationship between the multiple genes and the multiple biological pathways, construct a hyperedge connecting the multiple hypergraph nodes to obtain the gene pathway hypergraph containing multiple hypergraph nodes and multiple hyperedges. The hypergraph processing unit is used to perform bidirectional feature aggregation of the hypergraph nodes and hyperedges on the gene pathway hypergraph through hypergraph convolution operations to obtain the target interaction features corresponding to each hypergraph node. The pathway aggregation unit is used to aggregate the target interaction features of all hypergraph nodes associated with each hyperedge in the gene pathway hypergraph to obtain initial pathway features, and then to perform weighted fusion of the initial pathway features through a pathway self-attention mechanism to obtain the gene-level pathway features.

4. The system according to claim 3, characterized in that, The hypergraph processing unit is used for: The gene expression profile is projected onto a set projection space to obtain the node feature matrix corresponding to the gene pathway hypergraph; the node feature matrix contains node features that correspond one-to-one with multiple nodes of the hypergraph. Based on the node feature matrix, feature aggregation is performed on all hypergraph nodes associated with each hyperedge to obtain hyperedge feature matrices corresponding to multiple hyperedges; Based on the hyperedge feature matrix, feature aggregation is performed on all the hyperedges associated with each hypergraph node to obtain a new node feature matrix corresponding to multiple hypergraph nodes; Return to the step of performing feature aggregation on all hypergraph nodes associated with each hyperedge based on the node feature matrix to obtain a hyperedge feature matrix corresponding to multiple hyperedges, until the bidirectional feature aggregation is completed a set number of times, and the node features corresponding to each hypergraph node in the node feature matrix obtained when the set number of times is reached are used as the target interaction features corresponding to each hypergraph node.

5. The system according to claim 1, characterized in that, The fusion module includes: The feature splicing unit is used to splice the global gene features and the pathological text features to obtain spliced ​​features; A cross-modal processing unit is used to perform cross-modal feature interaction on the spliced ​​features and the gene-level pathway features using a cross-modal attention mechanism to obtain cross-modal features; The multi-granularity fusion unit is used to perform feature fusion on the cross-modal features, the global gene features, and the pathological text features through residual aggregation operations to obtain the multi-modal fused features.

6. The system according to claim 1, characterized in that, The effect value calculation module includes: The effect value calculation unit is used to input the global gene features, the pathological text features, and the multimodal fusion features into the target effect value calculation model to obtain the total effect value output by the target effect value calculation model; and to replace the global gene features and the multimodal fusion features with preset empty features, and then input them together with the pathological text features into the target effect value calculation model to obtain the natural direct effect value output by the target effect value calculation model. The debiasing calculation unit is used to perform weighted calculation on the natural direct effect value and the debiasing weight to obtain the debiased natural effect value, and to subtract the total effect value from the debiased natural effect value to obtain the total indirect effect value.

7. The system according to claim 6, characterized in that, The system also includes a model training module for: The global gene sample features, pathological text sample features, and multimodal fusion sample features are input into the effect value calculation model to obtain the total effect value of the sample output by the effect value calculation model. After replacing the global gene sample features and the multimodal fusion sample features with the preset empty features, they are input together with the pathological text sample features into the effect value calculation model to obtain the sample natural direct effect value output by the effect value calculation model. The total effect value and the natural direct effect value of the sample are respectively subjected to probability normalization to obtain the first probability distribution corresponding to the total effect value and the second probability distribution corresponding to the natural direct effect value of the sample. Based on the first probability distribution and the second probability distribution, calculate the first survival loss value corresponding to the total effect value of the sample and the second survival loss value corresponding to the natural direct effect value of the sample, respectively. Based on the first survival loss value, the second survival loss value, and the distribution divergence loss mechanism, the total model loss is calculated; The model parameters of the effect value calculation model are iteratively optimized based on the total loss of the model until the total loss of the model reaches a set loss value, and the effect value calculation model corresponding to the total loss of the model that has reached the set loss value is taken as the target effect value calculation model.

8. The system according to claim 6, characterized in that, The system also includes a parameter optimization module, used for: Based on the cancer survival prediction information and the actual cancer survival information, the bias removal weights are optimized and adjusted.

9. An electronic device, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the electronic device performs the steps performed by the system as described in any one of claims 1 to 8.

10. A computer program product, characterized in that, Includes a computer program, which, when run, causes the steps performed by the system as described in any one of claims 1 to 8 to be performed.

Images