Cancer prognosis prediction methods, devices, equipment and storage media

By constructing a cancer prognostic model based on visual neural networks and fully connected neural networks, the problem of data acquisition difficulties in existing technologies has been solved, enabling efficient prognostic risk assessment for cancer patients and enhancing the interpretability of the model.

CN117854722BActive Publication Date: 2026-11-10GUANGZHOU KINGMED CENTER FOR CLINICAL LABORATORY CO LTD +1
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202311747452.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-12-18
Publication Date
2026-11-10
Estimated Expiration
2043-12-18

AI Technical Summary

Technical Problem

Most existing cancer prognostic models rely on multi-omics data from the publicly available TCGA dataset, making data acquisition difficult, and there is a lack of corresponding prognostic models for cancer types with limited data in TCGA.

Method used

A pre-defined cancer prognostic model is constructed using visual neural networks and fully connected neural networks. By collecting genetic information and current clinical information of target cancer patients, the genetic information is processed using visual neural networks, and the output information of visual neural networks and current clinical information are integrated through fully connected neural networks to generate prognostic risks.

Benefits of technology

It improves the effectiveness of cancer prognostic risk assessment, enabling accurate prediction of prognostic risk for cancer patients based on limited genetic and current clinical information, and enhances the interpretability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117854722B_ABST
    Figure CN117854722B_ABST
Patent Text Reader

Abstract

This invention discloses a method, apparatus, device, and storage medium for predicting cancer prognosis. The method includes: collecting genetic information and current clinical information of a target cancer patient; inputting the genetic information and current clinical information into a preset cancer prognosis model to obtain prognostic risk. The preset cancer prognosis model is constructed based on a visual neural network and a fully connected neural network. The visual neural network is used to process and encode the genetic information, and the fully connected neural network is used to further integrate the output information of the visual neural network and the current clinical information to generate the corresponding prognostic risk. Because this invention constructs a high-performance preset cancer prognosis model using visual neural networks and fully connected neural networks, it can predict the prognostic risk of a target cancer patient with the assistance of the model based on relatively little genetic information and current clinical information, thereby improving the effectiveness of cancer prognostic risk assessment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of cancer prognostic analysis technology, and in particular to a method, apparatus, device, and storage medium for predicting cancer prognosis. Background Technology

[0002] The high incidence of cancer in recent years has spurred the development of medical aids, with prognostic risk analysis being a key one. This technology helps select appropriate treatment plans based on the potential prognostic risks of different patients. Prognostic prediction primarily studies the relationship between the final clinical outcome and closely related influencing factors, using biomedical data such as molecular and clinical data to infer the time of death or disease recurrence. Modern medical imaging and high-throughput sequencing technologies can be used to rapidly collect large amounts of biomedical data from cancer patients, such as genomic and transcriptomic data. Prognostic prediction techniques can then be used to estimate tumor progression or disease recurrence time, effectively assisting physicians in making decisions regarding cancer patient care and treatment.

[0003] However, most existing pre-defined cancer prognostic models rely on multi-omics data from the publicly available TCGA dataset, making it difficult to obtain the necessary data. Moreover, most of these models target cancer types with large datasets in TCGA, while some cancers are overlooked due to the limited amount of data in the TCGA dataset, resulting in a lack of corresponding prognostic models.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a method, apparatus, device, and storage medium for predicting cancer prognosis, aiming to solve the technical problems that most existing pre-defined cancer prognosis models rely on multi-omics data from the publicly available TCGA dataset, making it difficult to obtain the required data; moreover, most of them are for cancer types with large amounts of data in TCGA, while some cancers are overlooked because of the small amount of data in the TCGA dataset, and corresponding prognosis models are still lacking.

[0006] To achieve the above objectives, the present invention provides a method for predicting cancer prognosis, the method comprising the following steps:

[0007] Collect genetic and current clinical information from patients with the target cancer.

[0008] The genetic information and the current clinical information are input into a preset cancer prognosis model to obtain the prognostic risk of the target cancer patient. The preset cancer prognosis model is constructed based on a visual neural network and a fully connected neural network. The visual neural network is used to process and encode the genetic information, and the fully connected neural network is used to further integrate the output information of the visual neural network and the current clinical information to generate the corresponding prognostic risk.

[0009] Optionally, before collecting the genetic information and current clinical information of the target cancer patient, the process may further include:

[0010] A visual neural network is obtained by constructing an initial pathway network using a preset set of gene variation features and preset gene pathway relationships.

[0011] The visual neural network is connected to the fully connected neural network to obtain an initial prognostic model;

[0012] The initial prognostic model is trained to obtain a preset cancer prognostic model.

[0013] Optionally, the step of constructing an initial pathway network using a preset gene variation feature set and preset gene pathway relationships to obtain a visual neural network includes:

[0014] Obtain pathway gene sets and pathway hierarchical relationships through a biomolecular pathway knowledge database;

[0015] The pathway gene set and the pathway hierarchical relationship are correlated and extracted to obtain a preset gene pathway relationship, which includes the gene-pathway relationship and the relationship between pathways.

[0016] Based on the relationships between genes and pathways, the relationships between pathways, and a preset set of gene variation features, an initial pathway network is constructed to obtain a visual neural network.

[0017] Optionally, training the initial prognostic model to obtain a preset cancer prognostic model includes:

[0018] The preset gene mutation feature set and preset clinical information are input into the initial prognostic model to obtain the prognostic risk score of cancer patients. The preset clinical information includes information such as age at diagnosis, treatment plan, cancer type and cancer status.

[0019] Based on a preset loss function, the current loss of the model is calculated according to the prognostic risk score, survival time, and survival status.

[0020] The initial prognostic model is trained based on the current loss of the model to obtain a preset cancer prognostic model.

[0021] Optionally, the step of calculating the current loss of the model based on the pre-defined loss function, according to the prognostic risk score, survival time, and survival status, includes:

[0022] Obtain the survival time and survival status of the cancer patients;

[0023] Based on the survival time and the survival status, obtain the event set and survival set of the cancer patient;

[0024] Based on a preset loss function, the current loss of the model is calculated according to the prognostic risk score, the event set, and the survival set.

[0025] Optionally, the preset loss function includes a nonlinear proportional risk loss function, specifically:

[0026]

[0027] Where θ represents the current parameters of the initial prognostic model; and R(t) represents the prognostic risk score for the i-th and j-th cancer patients under the current parameters; D is the event set, representing the set of cancer patients who have been observed to die; i Let be the survival set, representing the time t when the i-th cancer patient dies. i A collection of cancer patients who are still alive.

[0028] Optionally, training the initial prognostic model based on the current model loss to obtain a preset cancer prognostic model includes:

[0029] The initial prognostic model is trained based on the current loss of the model to obtain the current parameters;

[0030] Determine whether the current parameters meet the fitting conditions;

[0031] When the current parameters do not meet the fitting conditions, the initial prognostic model is validated using ten-fold cross-validation until the current parameters meet the fitting conditions. The current parameters that meet the fitting conditions are then used as hyperparameters, and a preset cancer prognostic model is obtained based on the hyperparameters.

[0032] Furthermore, to achieve the above objectives, the present invention also proposes a cancer prognosis prediction device, the device comprising:

[0033] The information collection module is used to collect genetic information and current clinical information of the target cancer patients;

[0034] The risk prediction module is used to input the genetic information and the current clinical information into a preset cancer prognosis model to obtain the prognostic risk of the target cancer patient. The preset cancer prognosis model is constructed based on a visual neural network and a fully connected neural network. The visual neural network is used to process and encode the genetic information, and the fully connected neural network is used to further integrate the output information of the visual neural network and the current clinical information to generate the corresponding prognostic risk.

[0035] Furthermore, to achieve the above objectives, the present invention also proposes a cancer prognosis prediction device, the device comprising: a memory, a processor, and a cancer prognosis prediction program stored in the memory and executable on the processor, the cancer prognosis prediction program being configured to implement the steps of the cancer prognosis prediction method as described above.

[0036] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing a cancer prognosis prediction program, which, when executed by a processor, implements the steps of the cancer prognosis prediction method as described above.

[0037] This invention collects the genetic information and current clinical information of a target cancer patient; it then inputs this genetic information and current clinical information into a preset cancer prognostic model to obtain the prognostic risk of the target cancer patient. The preset cancer prognostic model is constructed based on a visual neural network and a fully connected neural network. The visual neural network processes and encodes the genetic information, while the fully connected neural network further integrates the output information of the visual neural network and the current clinical information to generate the corresponding prognostic risk. Because this invention constructs a high-performance preset cancer prognostic model using visual neural networks and fully connected neural networks, it can predict the prognostic risk of a target cancer patient with the assistance of a relatively small amount of genetic and current clinical information, thereby improving the effectiveness of cancer prognostic risk assessment. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the structure of a cancer prognosis prediction device in the hardware operating environment involved in the embodiments of the present invention;

[0039] Figure 2 This is a flowchart illustrating the first embodiment of the cancer prognosis prediction method of the present invention;

[0040] Figure 3 This is a flowchart illustrating the second embodiment of the cancer prognosis prediction method of the present invention;

[0041] Figure 4 This is a schematic diagram of the model structure and application framework in the second embodiment of the cancer prognosis prediction method of the present invention;

[0042] Figure 5 This is a structural block diagram of the first embodiment of the cancer prognosis prediction device of the present invention.

[0043] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0044] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0045] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of a cancer prognosis prediction device in the hardware operating environment of an embodiment of the present invention.

[0046] like Figure 1 As shown, the cancer prognosis prediction device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0047] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on cancer prognosis prediction devices and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0048] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a cancer prognosis prediction program.

[0049] exist Figure 1In the cancer prognosis prediction device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the cancer prognosis prediction device of the present invention can be set in the cancer prognosis prediction device, and the cancer prognosis prediction device calls the cancer prognosis prediction program stored in the memory 1005 through the processor 1001 and executes the cancer prognosis prediction method provided in the embodiment of the present invention.

[0050] This invention provides a method for predicting cancer prognosis, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the cancer prognosis prediction method of the present invention.

[0051] Cancer, a major global public health problem, causes millions of deaths annually. In recent years, cancer research has made significant progress in early diagnosis and treatment. Advances in genomics technologies have improved our understanding of the molecular mechanisms of cancer, enabling targeted therapy. However, the high heterogeneity of cancer remains a major obstacle limiting the efficacy of targeted therapies and impairing treatment outcomes. Therefore, prognostic assessment or subtype classification of cancer patients is crucial for precision medicine in clinical practice.

[0052] With the widespread application of artificial intelligence in the biomedical field, an increasing number of machine learning methods based on genetic information are being used for cancer risk prediction and survival analysis. However, most existing pre-defined cancer prognostic models are developed based on multi-omics data from the publicly available TCGA dataset, which has several limitations. First, considering the data dependency of machine learning models, these prognostic models are mostly designed for cancer types with large datasets in TCGA. Therefore, for cancer types with small datasets or even those not included in TCGA, corresponding prognostic models are still lacking. Second, although models developed based on multi-omics data often achieve better predictive performance than single-modality models, the time and financial costs required to acquire multi-modality data limit the clinical application potential of these models.

[0053] Therefore, this invention provides a new approach to the mechanistic explanation of biomedical problems based on visual machine learning, such as visual neural networks (VNNs). In a visualized structure, the hidden layer neurons in a neural network are mapped to biological entities such as genes, proteins, pathways, and cellular subsystems, and the connections between neurons represent biological relationships. Therefore, VNNs offer higher model interpretability than standard black-box neural network structures. This embodiment utilizes the relatively limited genetic information and highly interpretable VNN neural network model to construct a high-performance, pre-defined cancer prognostic model, thereby improving the effectiveness of cancer prognostic risk assessment.

[0054] In this embodiment, the cancer prognosis prediction method includes the following steps:

[0055] Step S10: Collect genetic information and current clinical information of the target cancer patient.

[0056] It should be noted that the execution subject of the method in this embodiment can be an electronic device with functions of genetic information collection, gene screening, and gene identification, such as a gene sequencer or gene detector. It can also be other electronic devices capable of achieving the same or similar functions, such as the aforementioned cancer prognosis prediction device. This embodiment does not limit this. Here, the aforementioned cancer prognosis prediction device (hereinafter referred to as the prediction device) will be used to specifically describe this embodiment and the following embodiments.

[0057] Understandably, genetic information refers to the genetic information contained in the DNA sequence carried by the target cancer patient. Genes are specific regions on DNA molecules that carry genetic information that determines the characteristics and functions of an organism. Applying VNNs to pre-defined cancer prognostic models allows for accurate assessment of prognostic risk based on limited genetic information and current clinical data.

[0058] It should be noted that the current clinical information refers to the clinical treatment information of the target cancer patient, such as the age at diagnosis, cancer type and status, diagnosis and staging information, and treatment methods. Specifically, it may include whether the age at diagnosis is over 60 years old, whether R-CHOP treatment has been received, and whether it is a newly diagnosed cancer, etc. It can also include other information, which is not limited in this embodiment. In summary, inputting the current clinical information and a small amount of genetic information into a preset cancer prognostic model can ultimately yield the patient's prognostic risk.

[0059] Step S20: Input the genetic information and the current clinical information into a preset cancer prognosis model to obtain the prognostic risk of the target cancer patient. The preset cancer prognosis model is constructed based on a visual neural network and a fully connected neural network. The visual neural network is used to process and encode the genetic information, and the fully connected neural network is used to further integrate the output information of the visual neural network and the current clinical information to generate the corresponding prognostic risk.

[0060] Understandably, the pre-defined cancer prognostic model is a model built and trained based on a visible neural network (VNN) and a fully connected neural network. In this visualized structure, input neurons are mapped to gene variation features, and hidden layer neurons are mapped to biological pathways. Based on VNN, the interpretability of the pre-defined cancer prognostic model can be improved, thereby assisting in the discovery of genetic drivers.

[0061] It's important to note that a fully connected neural network (MLP), also known as a multilayer perceptron, is a common artificial neural network architecture. It consists of multiple neurons, each connected to all neurons in the previous layer. Each connection has an associated weight used to adjust the importance of the input signal. Neurons receive weighted inputs from neurons in the previous layer, process them through an activation function, and then produce an output that is passed to the next layer's neurons. During training, the weights are updated using backpropagation, allowing the fully connected neural network to learn and gradually optimize the model's predictive ability.

[0062] In its implementation, the predictive device first collects the genetic information and current clinical information of the target cancer patient. This genetic and clinical information is then input into a pre-defined cancer prognostic model. In a visualized structure, input neurons are mapped to gene mutation features, and hidden layer neurons are mapped to biological pathways, improving the interpretability of the pre-defined cancer prognostic model. Finally, based on a fully connected neural network, the output information of the visualized neural network and the current clinical information are used to predict the prognostic risk of the target cancer patient.

[0063] This embodiment first constructs a preset cancer prognostic model based on visual neural networks and fully connected neural networks. Then, a prediction device collects the genetic information and current clinical information of the target cancer patient. Finally, the genetic information and current clinical information are input into the preset cancer prognostic model. In a visualized structure, input neurons are mapped to gene variation features, and hidden layer neurons are mapped to biological pathways, achieving interpretability not found in traditional black-box neural networks. Then, the output information of the visual neural network and the current clinical information are used to predict the prognostic risk of the target cancer patient based on the fully connected neural network. Because this invention constructs a high-performance preset cancer prognostic model using visual neural networks and fully connected neural networks, it can predict the prognostic risk of the target cancer patient with the assistance of the model, based on relatively little genetic information and current clinical information, thereby improving the effectiveness of cancer prognostic risk assessment.

[0064] refer to Figure 3 , Figure 3 This is a flowchart illustrating the second embodiment of the cancer prognosis prediction method of the present invention.

[0065] Based on the first embodiment described above, in this embodiment, considering the construction of a preset cancer prognosis model, before step S10, the method further includes:

[0066] Step S01: Construct the initial pathway network using a preset gene variation feature set and preset gene pathway relationships to obtain a visual neural network.

[0067] It should be noted that the preset gene variant feature set is a set of gene variant features that are pre-selected or defined and are related to the prognostic impact of cancer. These features can include different types of gene variants such as single nucleotide variants (SNV), insertions / deletions (Indel), chromosomal structural variants, gene rearrangements, and copy number variants (CNV).

[0068] Understandably, the pre-defined gene pathway relationships are a set of gene pathway relationships that are pre-selected or defined and are related to cancer prognosis. The initial pathway network is a series of neurons and connections that construct a visual neural network. In the visualized structure, input neurons are mapped to gene variation features, hidden layer neurons are mapped to biological pathways, and the connections between neurons represent the relationships between the biological entities they refer to.

[0069] Step S02: Connect the visual neural network with the fully connected neural network to obtain the initial prognostic model.

[0070] Step S03: Train the initial prognostic model to obtain a cancer prognostic model.

[0071] In the specific implementation, for ease of understanding, please refer to... Figure 4 , Figure 4 This is a schematic diagram illustrating the model structure and application framework in the second embodiment of the cancer prognosis prediction method of the present invention. The network structure of this model mainly includes two parts (see reference). Figure 4 These are, respectively, a VNN structure for characterizing gene variation profiles and an FCNN structure for integrating gene variation information with key clinical information. To design a VNN structure that simulates cell biological processes, pathway gene sets and pathway hierarchy relationship files can first be downloaded from Reactome (https: / / reactome.org / ) to obtain gene-pathway and pathway-pathway relationship information. Then, a pathway-guided visual neural network (VNN) structure is constructed based on the input gene variation feature set. Figure 4 ).

[0072] As a sparse neural network structure, each neuron in the hidden layer of a VNN is mapped to a different pathway, and there are connections between related neurons (pathways). These sparse connections between neurons can be achieved through weight masking.

[0073] W ′ =W⊙M;

[0074] Where W is the original weight matrix, and M is a binary masking matrix with the same dimension as the weight matrix, taking the value 0 or 1. The weight matrix is ​​modified by multiplication between elements so that the model can selectively ignore some weights.

[0075] Specifically, such as Figure 4 As shown, the input to the VNN structure consists of 117 gene variation features. Neurons in the lowest-level pathways are selectively connected to relevant gene features. Then, the neurons in the sub-pathways transmit the integrated information to the neurons in the corresponding parent pathways. For the pathway relationship P1→P2, P1 is denoted as the sub-pathway (child node), and P2 as the parent pathway (parent node). Information from the sub-pathway is continuously transmitted to higher-level parent pathways until the current pathway no longer has a parent pathway. At this point, the neurons in this pathway are directly connected to the root node, and the output of the root node is used as the feature embedding of the gene mutation spectrum.

[0076] It should be noted that due to the inherent characteristics of pathway interactions in Reactome, the hidden neurons in the VNN structure are not strictly hierarchical. Therefore, there are some neurons with skip connections in the VNN structure.

[0077] At the same time, such as Figure 4 As shown, to more accurately predict patient risk, clinical information that significantly impacts patient prognosis can be introduced based on prior knowledge, such as whether the age of diagnosis is over 60 years, whether the patient received R-CHOP treatment, and whether the cancer is newly diagnosed. Figure 4 As shown, the output of the root node of the VNN structure is concatenated with supplementary clinical information and then fed into a two-layer fully connected neural network (FCNN) to finally output the patient's accurate prognostic risk, thus improving the effectiveness of cancer prognostic assessment.

[0078] Furthermore, considering the construction of a visual neural network, step S01 in this embodiment includes: obtaining a pathway gene set and pathway hierarchical relationships through a biomolecular pathway knowledge database; performing association extraction on the pathway gene set and the pathway hierarchical relationships to obtain a preset gene pathway relationship, wherein the preset gene pathway relationship includes gene-pathway relationship and pathway-to-pathway relationship; constructing an initial pathway network based on the gene-pathway relationship, the pathway-to-pathway relationship, and a preset gene variation feature set to obtain a visual neural network.

[0079] It should be noted that a biomolecular pathway knowledge database is a database used to store, integrate, and provide information about biomolecular pathways. It contains information on the interactions and regulation between genes, proteins, and other biomolecules, and organizes them into a hierarchical pathway network, such as the Reactome database.

[0080] In practical implementation, the pathway gene set and pathway hierarchy relationship files can first be downloaded from Reactome to obtain gene-pathway and pathway-pathway relationship information. Then, a pathway-guided visual neural network (VNN) structure can be constructed for the input preset gene variation feature set. In the visualization structure, input neurons are mapped to gene variation features, hidden layer neurons are mapped to biological pathways, and the connections between neurons represent the relationships between the biological entities they refer to, thus achieving interpretability that traditional black-box neural networks do not possess.

[0081] Furthermore, considering model training, step S03 in this embodiment includes: inputting a preset gene mutation feature set and preset clinical information into the initial prognostic model to obtain a prognostic risk score for cancer patients, wherein the preset clinical information includes information such as age at diagnosis, treatment plan, cancer type, and cancer status; calculating the current loss of the model based on a preset loss function, according to the prognostic risk score, survival time, and survival status; and training the initial prognostic model based on the current loss of the model to obtain a preset cancer prognostic model.

[0082] Understandably, the preset loss function is a function used to update the model parameters, which can improve the model's prediction performance.

[0083] Furthermore, the step of calculating the current loss of the model based on the pre-defined loss function, according to the prognostic risk score, survival time, and survival status, includes: obtaining the survival time and survival status of the cancer patient; obtaining the event set and survival set of the cancer patient based on the survival time and survival status; and calculating the current loss of the model based on the pre-defined loss function, according to the prognostic risk score, the event set, and the survival set.

[0084] It should be noted that survival time is the time during which the event (death) was observed in cancer patients. The event set is the set of patients for whom the event (death) was observed.

[0085] Furthermore, the preset loss function in this embodiment includes a non-linear proportional risk loss function, specifically:

[0086]

[0087] Where θ represents the current parameters of the initial prognostic model; and R(t) represents the prognostic risk score for the i-th and j-th cancer patients under the current parameters; D is the event set, representing the set of cancer patients who have been observed to die; iLet be the survival set, representing the time t when the i-th cancer patient dies. i A collection of cancer patients who are still alive.

[0088] Furthermore, in this embodiment, training the initial prognostic model based on the current model loss to obtain a preset cancer prognostic model includes: training the initial prognostic model based on the current model loss to obtain current parameters; determining whether the current parameters meet the fitting condition; when the current parameters do not meet the fitting condition, validating the initial prognostic model using ten-fold cross-validation until the current parameters meet the fitting condition, using the current parameters that meet the fitting condition as hyperparameters, and obtaining the preset cancer prognostic model based on the hyperparameters.

[0089] Understandably, 10-fold cross-validation is a commonly used cross-validation method for evaluating model performance and selecting model hyperparameters. Specifically, it divides the dataset into 10 equal-sized subsets, with 9 subsets used as training data and the remaining subset used as validation data. By using 10-fold cross-validation, model performance can be evaluated more accurately, avoiding bias caused by the selection of specific training and test sets, and better assessing the model's generalization ability to unknown data.

[0090] In practice, a non-linear proportional risk loss function can be used to update model parameters during model training.

[0091]

[0092] By employing 10-fold cross-validation to determine the model's hyperparameters, it is ensured that the model can be trained and validated on various data subsets, effectively avoiding overfitting and selecting the optimal hyperparameters across a broad data distribution. This allows for more accurate evaluation of model performance, avoids bias caused by specific training and test set selections, and better assesses the model's generalization ability to unknown data.

[0093] Furthermore, after the model is trained, its performance can be evaluated using the Concordance Index (C-index), which is widely used in pre-defined cancer prognostic models. The C-index value is between 0 and 1, representing the consistency between the survival time predicted by the model and the actual observed survival time for two randomly selected samples.

[0094]

[0095] Where N effect This refers to a valid patient match. and Let I be the prognostic risk predicted by the model for the i-th and j-th patients, and let I be the indicative function. The function value is 1 when it is true, and 0 otherwise.

[0096] In addition, several external validation cohorts can be used to validate the model's performance. These external validation sets come from studies completely independent of the training cohorts, thus objectively reflecting the model's generalization ability on unseen data. However, it is worth noting that due to differences in the detection methods and gene detection ranges used in each study, some features may be missing in the independent validation cohorts. For missing features, a default value (0 or 1) can be uniformly assigned to all cases in the current cohort.

[0097] Furthermore, the interpretability of the pre-defined cancer prognostic model obtained through the above model training mainly includes two parts: post-model interpretation based on SHAP (SHapley Additive exPlanations) and visual interpretation based on VNN structure. SHAP-based interpretability can clarify the degree and direction of a gene's influence on patient prognosis within a gene set, thereby improving the efficiency of genetic driver discovery. The Python package SHAP can be used to calculate importance scores for each input gene feature and the gene pathways that act as hidden layer neurons. These scores accurately reflect the feature's impact on patient prognosis. Therefore, key genes that contribute most to model prediction can be identified based on SHAP values, and the risk trends of genes can be distinguished based on the positive or negative SHAP values. Simultaneously, visual interpretation based on VNN structure can help clarify the pathways through which genes influence patient prognosis, which is of significant value for understanding the biological mechanisms of disease and guiding future experimental research.

[0098] To facilitate understanding, specific experimental data are used to demonstrate the effectiveness of the embodiments of the present invention, but this does not limit the scope of the embodiments. For example, for diffuse large B-cell lymphoma (DLBCL), a dataset of 928 cases (referred to as HMRN) published by Stuart et al. in their DLBCL-related research was used. This dataset includes variant profiles (variation information) of 117 genes and other clinical information such as survival time, survival status, age at diagnosis, and treatment status. To verify the model's performance and generalization ability, two external validation sets were introduced: the DLBC cohort (48 cases) and the BCCA cohort (329 cases) in TCGA. The final model achieved an average C-index of 0.72 on the HMRN dataset with 10-fold cross-validation, and C-indexes of 0.66 and 0.62 on the two external validation sets TCGA and BCCA, respectively. This allows for the prediction of prognostic risk in target cancer patients with relatively little genetic and current clinical information, improving the effectiveness of cancer prognostic risk assessment.

[0099] This embodiment downloads pathway gene sets and pathway hierarchy relationship files from Reactome to obtain gene-pathway and pathway-pathway relationship information. Then, a pathway-guided visual neural network (VNN) structure can be constructed for the input preset gene variation feature set. In this visualized structure, input neurons are mapped to gene variation features, hidden layer neurons are mapped to biological pathways, and the connections between neurons represent relationships between the biological entities they represent, achieving interpretability not found in traditional black-box neural networks. Furthermore, ten-fold cross-validation can be used to determine the model's hyperparameters, ensuring the model can be trained and validated on various data subsets, effectively avoiding overfitting and selecting the optimal hyperparameters across a wide data distribution. This allows for more accurate evaluation of model performance, avoids bias caused by specific training and test set selections, and better assesses the model's generalization ability to unknown data.

[0100] Furthermore, embodiments of the present invention also propose a storage medium storing a cancer prognosis prediction program, which, when executed by a processor, implements the steps of the cancer prognosis prediction method described above.

[0101] Reference Figure 5 , Figure 5 This is a structural block diagram of the first embodiment of the cancer prognosis prediction device of the present invention.

[0102] like Figure 5 As shown, the cancer prognosis prediction device proposed in this embodiment of the invention includes:

[0103] Information acquisition module 501 is used to collect genetic information and current clinical information of the target cancer patient;

[0104] The risk prediction module 502 is used to input the genetic information and the current clinical information into a preset cancer prognosis model to obtain the prognostic risk of the target cancer patient. The cancer prognosis model is constructed based on a visual neural network and a fully connected neural network. The visual neural network is used to process and encode the genetic information, and the fully connected neural network is used to further integrate the output information of the visual neural network and the current clinical information to generate the corresponding prognostic risk.

[0105] This embodiment first constructs a preset cancer prognostic model based on visual neural networks and fully connected neural networks. Then, a prediction device collects the genetic information and current clinical information of the target cancer patient. Finally, the genetic information and current clinical information are input into the preset cancer prognostic model. In a visualized structure, input neurons are mapped to gene variation features, hidden layer neurons are mapped to biological pathways, and the connections between neurons represent the relationships between the biological entities they represent, thus achieving interpretability that traditional black-box neural networks lack. Then, the output information of the visual neural network and the current clinical information are used to predict the prognostic risk of the target cancer patient based on the fully connected neural network. Because this invention constructs a high-performance preset cancer prognostic model using visual neural networks and fully connected neural networks, it can predict the prognostic risk of the target cancer patient with the assistance of the model based on relatively little genetic information and current clinical information, thereby improving the effectiveness of cancer prognostic risk assessment.

[0106] Based on the first embodiment of the cancer prognosis prediction device of the present invention, a second embodiment of the cancer prognosis prediction device of the present invention is proposed.

[0107] In this embodiment, the cancer prognosis prediction device further includes a model training module 503, which is used to construct an initial pathway network by using a preset gene mutation feature set and a preset gene pathway relationship to obtain a visual neural network; connect the visual neural network with a fully connected neural network to obtain an initial prognosis model; and train the initial prognosis model to obtain a cancer prognosis model.

[0108] Furthermore, the model training module 503 is also used to obtain pathway gene sets and pathway hierarchical relationships through a biomolecular pathway knowledge database; to perform association extraction on the pathway gene sets and pathway hierarchical relationships to obtain preset gene pathway relationships, the preset gene pathway relationships including gene-pathway relationships and relationships between pathways; and to construct an initial pathway network based on the gene-pathway relationships, the relationships between pathways, and a preset gene variation feature set to obtain a visual neural network.

[0109] Furthermore, the model training module 503 is also used to input a preset gene mutation feature set and preset clinical information into the initial prognostic model to obtain the prognostic risk score of the cancer patient, wherein the preset clinical information includes information such as age at diagnosis, treatment plan, cancer type and cancer status; calculate the current loss of the model based on the preset loss function according to the prognostic risk score, survival time and survival status; and train the initial prognostic model according to the current loss of the model to obtain a preset cancer prognostic model.

[0110] Furthermore, the model training module 503 is also used to obtain the survival time and survival status of the cancer patient; obtain the event set and survival set of the cancer patient based on the survival time and survival status; and calculate the current loss of the model based on the prognostic risk score, the event set and the survival set based on a preset loss function.

[0111] Furthermore, in the model training module 503, the preset loss function includes a non-linear proportional risk loss function, specifically:

[0112]

[0113] Where θ represents the current parameters of the initial prognostic model; and R(t) represents the prognostic risk score for the i-th and j-th cancer patients under the current parameters; D is the event set, representing the set of cancer patients who have been observed to die; i Let be the survival set, representing the time t when the i-th cancer patient dies. i A collection of cancer patients who are still alive.

[0114] Furthermore, the model training module 503 is also used to train the initial prognostic model based on the current loss of the model to obtain current parameters; determine whether the current parameters meet the fitting conditions; when the current parameters do not meet the fitting conditions, verify the initial prognostic model through ten-fold cross-validation until the current parameters meet the fitting conditions, take the current parameters that meet the fitting conditions as hyperparameters, and obtain a preset cancer prognostic model based on the hyperparameters.

[0115] Other embodiments or specific implementations of the cancer prognosis prediction device of the present invention can be referred to the above-described method embodiments, and will not be repeated here.

[0116] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0117] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0118] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0119] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for predicting cancer prognosis, characterized in that, The cancer prognosis prediction method includes: A visual neural network is obtained by constructing an initial pathway network using a preset set of gene mutation features and preset gene pathway relationships. The input neurons of the visual neural network are mapped to preset gene mutation features, and the hidden layer neurons are mapped to biological pathways. The connections between neurons represent the relationships between biological entities. The connections between the hidden layer neurons are constructed using preset gene pathway relationships to form a sparse connection structure. The sparse connections between neurons are achieved through weight masking. The visual neural network is connected to the fully connected neural network to obtain an initial prognostic model; The initial prognostic model is trained to obtain a preset cancer prognostic model; Collect genetic and current clinical information from patients with the target cancer. The genetic information and the current clinical information are input into a preset cancer prognosis model to obtain the prognostic risk of the target cancer patient. The visual neural network is used to process and encode the genetic information. The root node output of the visual neural network is used as a feature embedding of the gene mutation spectrum, which is concatenated with the current clinical information and then connected to two layers of the fully connected neural network. The fully connected neural network is used to further integrate the output information of the visual neural network and the current clinical information to generate the corresponding prognostic risk.

2. The cancer prognosis prediction method as described in claim 1, characterized in that, The process of constructing a visual neural network by using a preset gene variation feature set and preset gene pathway relationships to construct an initial pathway network includes: Obtain pathway gene sets and pathway hierarchical relationships through a biomolecular pathway knowledge database; The pathway gene set and the pathway hierarchical relationship are correlated and extracted to obtain a preset gene pathway relationship, which includes the gene-pathway relationship and the relationship between pathways. Based on the relationships between genes and pathways, the relationships between pathways, and a preset set of gene variation features, an initial pathway network is constructed to obtain a visual neural network.

3. The cancer prognosis prediction method as described in claim 1, characterized in that, The step of training the initial prognostic model to obtain a preset cancer prognostic model includes: The preset gene mutation feature set and preset clinical information are input into the initial prognostic model to obtain the prognostic risk score of cancer patients. The preset clinical information includes information such as age at diagnosis, treatment plan, cancer type and cancer status. Based on a preset loss function, the current loss of the model is calculated according to the prognostic risk score, survival time, and survival status. The initial prognostic model is trained based on the current loss of the model to obtain a preset cancer prognostic model.

4. The cancer prognosis prediction method as described in claim 3, characterized in that, The calculation of the model's current loss based on a preset loss function, according to the prognostic risk score, survival time, and survival status, includes: Obtain the survival time and survival status of the cancer patients; Based on the survival time and the survival status, obtain the event set and survival set of the cancer patient; Based on a preset loss function, the current loss of the model is calculated according to the prognostic risk score, the event set, and the survival set.

5. The cancer prognosis prediction method as described in claim 4, characterized in that, The preset loss function includes a nonlinear proportional risk loss function, specifically: ; in, These are the current parameters of the initial prognostic model; and Let be the prognostic risk score for the i-th and j-th cancer patients under the current parameters; D is the event set, representing the set of cancer patients observed to have died; Let be the survival set, representing the time of death of the i-th cancer patient. A collection of cancer patients who are still alive.

6. The cancer prognosis prediction method as described in claim 5, characterized in that, The step of training the initial prognostic model based on the current loss of the model to obtain a preset cancer prognostic model includes: The initial prognostic model is trained based on the current loss of the model to obtain the current parameters; Determine whether the current parameters meet the fitting conditions; When the current parameters do not meet the fitting conditions, the initial prognostic model is validated using ten-fold cross-validation until the current parameters meet the fitting conditions. The current parameters that meet the fitting conditions are then used as hyperparameters, and a preset cancer prognostic model is obtained based on the hyperparameters.

7. A cancer prognosis prediction device, characterized in that, The device includes: The model training module is used to construct an initial pathway network using a preset set of gene mutation features and preset gene pathway relationships to obtain a visual neural network. The input neurons of the visual neural network are mapped to preset gene mutation features, and the hidden layer neurons are mapped to biological pathways. The connections between neurons represent relationships between biological entities. The connections between the hidden layer neurons are constructed using preset gene pathway relationships to form a sparse connection structure, and the sparse connections between neurons are achieved through weight masking. The visual neural network is then connected to a fully connected neural network to obtain an initial prognostic model. The initial prognostic model is trained to obtain a preset cancer prognostic model. The information collection module is used to collect genetic information and current clinical information of the target cancer patients; The risk prediction module is used to input the genetic information and the current clinical information into a preset cancer prognosis model to obtain the prognostic risk of the target cancer patient. The visual neural network is used to process and encode the genetic information. The root node output of the visual neural network is used as a feature embedding of the gene mutation spectrum, which is concatenated with the current clinical information and then connected to two layers of the fully connected neural network. The fully connected neural network is used to further integrate the output information of the visual neural network and the current clinical information to generate the corresponding prognostic risk.

8. A cancer prognosis prediction device, characterized in that, The device includes: a memory, a processor, and a cancer prognosis prediction program stored in the memory and executable on the processor, the cancer prognosis prediction program being configured to implement the steps of the cancer prognosis prediction method as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium stores a cancer prognosis prediction program, which, when executed by a processor, implements the steps of the cancer prognosis prediction method as described in any one of claims 1 to 6.

Citation Information

Patent Citations

  • Cancer survival prediction method based on graph neural network and multi-modal data

    CN116092669A

  • Cancer prognosis prediction model training device, medium and electronic equipment

    CN116881725A

  • Prognostic risk prediction system for colorectal cancer

    CN117079801A