A gastric cancer immunotherapy individualized survival prediction model based on a GBM-Cox model and a construction method thereof

Abstract

This invention belongs to the field of bioinformatics data analysis technology, and relates to a personalized survival prediction model for gastric cancer immunotherapy based on the GBM-Cox model and its construction method. Baseline patient data is collected, including follow-up time, event status, etc. ARID1A Data preprocessing was performed on gene mutation, TMB, age, metastasis status, and sex. A GBM-Cox model was constructed. The negative logarithm of the partial likelihood function of the Cox model was selected as the loss function. A gradient boosting iterative algorithm was used to train the risk scoring function, a weighted regression tree was fitted, and the optimal output constant of the leaf nodes was calculated. The risk scoring function was updated. The iteration was repeated to optimize the risk scoring function. This invention uses the GBM-Cox survival prediction model to improve the model's applicability to small sample data and enhance prediction accuracy. Simultaneously, it integrates biomarkers. ARID1A Mutation status, along with other key clinical features, provides a more accurate method for predicting immunotherapy.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics data analysis technology, and relates to a method for fusing genomic features (such as...) ARID1A A gradient booster (GBM) survival prediction model based on mutation status and clinical characteristic variables is used to guide individualized treatment of patients with advanced gastric cancer with immune checkpoint inhibitors (ICIs) and optimize treatment decisions; specifically, it involves an individualized survival prediction model for gastric cancer immunotherapy based on the GBM-Cox model and its construction method. Background Technology

[0002] In the medical field, especially in cancer treatment, survival prediction models have become a key tool for personalized treatment decisions. Immunotherapy has shown potential in improving the prognosis of patients with gastric malignancies; however, the response to immunotherapy varies greatly among gastric cancer patients, making the accurate prediction of the efficacy and survival of immunotherapy a significant clinical challenge.

[0003] Currently, most common gastric cancer survival prediction models rely on traditional statistical methods, such as the Cox proportional hazards model. However, these methods often overlook the complexity of nonlinear relationships and high-dimensional data, and are prone to overfitting in small sample sizes. Furthermore, current gastric cancer immunotherapy prediction models typically depend on a single biomarker or on macroscopic clinical characteristics such as age, TNM stage, and pathological type, and are generally constructed using the Cox proportional hazards regression model. Biomarkers such as programmed death-ligand-1 (PD-L1), tumor mutational burden (TMB), and mismatch repair (MMR) are widely used in predicting immunotherapy, but their predictive ability in gastric cancer is limited, leading to poor prediction of clinical immunotherapy efficacy and an inability to accurately identify patient groups that respond well to ICIs. While the Cox proportional hazards regression model is widely used clinically, it has two inherent limitations: First, it fundamentally ignores the molecular heterogeneity of tumors. Gastric cancer exhibits high diversity at the molecular level, including gene mutations and signaling pathway regulation. This is the core reason why patients with similar clinical characteristics show vastly different treatment responses and survival outcomes. Analysis based solely on macroscopic characteristics cannot reveal the molecular essence, limiting the model's predictive accuracy. Second, its methodological approach is limited by model assumptions. The Cox model presupposes that the effect of each covariate on risk is linear and additive (i.e., each factor acts independently and uniformly). However, many biological or clinical variables have complex and non-linear modes of action, contradicting the assumptions of the Cox model and further restricting predictive accuracy.

[0004] Existing gastric cancer survival prediction models (such as Cox regression) are mainly based on clinical characteristics (age, TNM stage, etc.), neglecting the fact that... ARID1AWhile key gene mutation states significantly impact the efficacy and prognosis of immunotherapy, the model has failed to systematically integrate traditional biomarkers (such as TMB) with clinical characteristics, resulting in incomplete information dimensions and limitations in predictive accuracy. Furthermore, the model algorithms are limited; for example, the Cox regression model assumes a linear relationship between variables and survival risk, failing to capture the complex nonlinear interactions between variables in the real world. Introducing molecular markers (such as gene mutations) leads to high-dimensional and sparse data (increased features, but potentially small sample sizes for specific mutation types). Traditional statistical models (such as Cox regression) are prone to overfitting in such data environments, meaning the model over-memorizes noise from the training data rather than universal patterns, resulting in poor generalization ability.

[0005] Based on this, the present invention constructs a framework combining the GBM model and the Cox model, and integrates new biomarkers ( ARID1A (Mutation status) and other key clinical features are used to more effectively predict the immunotherapy efficacy in gastric cancer patients. Summary of the Invention

[0006] This invention addresses the problems of existing models having limited prediction accuracy due to their single prediction dimension and failure to integrate predictive molecular markers, as well as the tendency for models to overfit in small sample data. It proposes a personalized survival prediction model for gastric cancer immunotherapy based on the GBM-Cox model and its construction method.

[0007] This invention employs the GBM-Cox survival prediction model, aiming to improve the model's applicability to small sample data and enhance prediction accuracy; simultaneously, it integrates novel biomarkers. ARID1A By combining mutation status with other key clinical features and the characteristics of the immune microenvironment associated with gastric cancer, a more accurate method for predicting immunotherapy has been constructed. This method will help improve the immunotherapy response rate of gastric cancer patients, accurately identify patient groups sensitive to ICIs, comprehensively assess the impact of immunotherapy on patient survival, thereby increasing the benefit of immunotherapy and promoting the personalized development of immunotherapy for gastric cancer.

[0008] The technical solution of this invention is: This invention provides a method for constructing a personalized survival prediction model for gastric cancer immunotherapy based on the GBM-Cox model, comprising the following steps: (1) Collect baseline data of patients receiving immunotherapy for gastric cancer, including follow-up time. T Event Status δ , and eigenvectors X The feature vector includes ARID1A Gene mutation, TMB, age, metastasis status, and sex; After preprocessing the feature vectors, the sample data of a single patient is represented as follows: The entire patient dataset is represented as ;in T i For patients i Follow-up time, δ i For patients i The status of death events, X i For patients i The preprocessed feature vector data.

[0009] ARID1A Gene mutations are closely related to the regulation of the tumor immune microenvironment in gastric cancer. ARID1A As an important gene for chromatin remodeling, its mutations may lead to increased TMB, upregulated PD-L1 expression, and altered immune cell infiltration patterns. These factors combined result in... ARID1A Mutation status has become a highly promising biomarker for predicting the response to immunotherapy in gastric cancer.

[0010] (2) Construct the GBM-Cox model, which is defined as follows: Patient samples are represented as The risk function is ; in, Reflecting on the patient's time t The instantaneous risk of an event occurring. As the benchmark risk function, It is a risk scoring function;

[0011] in, M This represents the total number of iterations. v Indicates the learning rate; L m For the first m The total number of leaf nodes in the tree; For the first m The trees in the The predicted constant values ​​at each leaf node ; This is an indicator function.

[0012] GBM is an ensemble learning method that constructs a strong learner by iteratively training multiple weak learners (usually decision trees). It possesses powerful nonlinear modeling capabilities, good feature selection capabilities, and strong resistance to overfitting. Especially with small sample data, the GBM model can effectively avoid overfitting through mechanisms such as regularization, ensemble learning, and feature selection, thereby providing more accurate prediction results.

[0013] (3) The negative logarithm of the partial likelihood function of the Cox proportional hazards model is selected as the loss function. The risk scoring function is trained using a gradient boosting iterative algorithm. First, initialize the risk function, then calculate the gradient. and second derivative With pseudo residuals− As a goal, with As weights, a weighted regression tree is fitted, which incorporates the patient feature vectors. X Divided into several leaf node regions In each Internally calculate the optimal output value for this node. ; Based on the weighted regression tree obtained in the current iteration, update the risk scoring function using the following formula:

[0014] Update the model. Let m be the risk scoring function after the m-th iteration. This is the risk scoring function after the (m-1)th iteration; Iterate gradually until completion. M Rounds of iteration to optimize the risk scoring function .

[0015] Furthermore, in step (1), the feature vector X It consists of the patient's genetic and clinical characteristics, among which the genetic characteristics include TMB, ARID1A Mutation status and clinical characteristics include age, metastasis status, and gender; Preprocessing methods include: Z-score standardization of age and TMB, and analysis of migration patterns. ARID1A Mutation state and sex are binary-classified and coded, with the specific coding rules as follows: Transfer status: 0 = not transferred, 1 = transferred; ARID1A Mutation state: 0 = no mutation, 1 = mutation; Gender: 0 = female, 1 = male.

[0016] Furthermore, in step (2), when the patient feature vector X Falling into the corresponding feature space region The value is 1 if it is true, and 0 otherwise.

[0017] Furthermore, in step (3), the loss function for:

[0018] in, Indicates the patient i The state of the death event, time T i The state of death events at time δ i =1 indicates that a death event has occurred, δ i =0 indicates deletion. Indicates the patient i Risk score; Indicates time T i This refers to all patients who are still under follow-up and have not yet experienced any deaths. Indicates the patient j The risk value; express j Take all risk sets Includes patients i All patients, including those in the group; This represents the total risk value for all patients within the risk set.

[0019] Furthermore, in step (3), the risk scoring function is trained using a gradient boosting iterative algorithm. The model parameters are gradually optimized through multiple iterations. The specific iteration process is as follows: Initialize risk function: Set the initialization risk function. =0; Calculating gradients and second derivatives: Predicted values ​​for each patient Calculate the loss function right gradient and second derivative The formula is as follows: ; ; in, Indicates the patient i The deviation between predictions and actual results under the current model; δ represents the curvature of the loss function. i Indicates the patient i The state of death events, δ j Indicates the patient j The status of the death event; Indicates the patient i In patients j The relative risk proportion of risk concentration, among which Indicates in risk set The total risk value for all patients in the group. Indicates the patient i The risk value, Indicates the first k Feature vectors of individual patients.

[0020] Furthermore, in step (3), an iterative update is performed to set the number of iteration trees. M =2000, interaction depth is 4, learning rate v It is 0.01.

[0021] The present invention also provides a GBM-Cox survival prediction model obtained by constructing a GBM-Cox model-based individualized survival prediction model for gastric cancer immunotherapy as described in any of the above claims.

[0022] This invention further provides a method for predicting individualized survival probability of gastric cancer immunotherapy based on the GBM-Cox model. The method uses the GBM-Cox survival prediction model obtained by any of the above-described construction methods to predict individualized survival probability, and includes the following steps: (1) Input the preprocessed feature vector of each patient into the trained risk scoring function. In the process of calculating individualized risk scores ; (2) Using the Breslow method, estimate the baseline cumulative risk function based on the training set data. The calculation formula is:

[0023] in, u Indicates the time point in the training set where a death event occurs, satisfying the following conditions: u ≤ t ; d ( u ) indicates at a point in time u The number of patients who died; Indicates at a point in time u The risk set consists of all patients who are still under follow-up and have not yet experienced a death. This represents the total risk value for all patients in the risk cluster; (3) At any follow-up time point t Accumulated risk function based on benchmark and individualized risk scoring functions for patients Calculate the individualized survival probability of this patient. The calculation formula is: .

[0024] The beneficial effects of this invention are: (1) The present invention innovatively uses ARID1A Gene mutation status was included as an independent predictor in the model for predicting the efficacy of ICIs in advanced gastric cancer. ARID1AMutations serve as core genomic biomarkers for predicting immunotherapy in gastric cancer. By integrating novel biomarkers, we can provide a more accurate method for predicting immunotherapy.

[0025] This invention employs a multimodal data fusion architecture to break down data barriers, simultaneously integrating multi-dimensional features. While incorporating traditional biomarkers (such as TMB), it retains key clinical characteristic variables (such as age, gender, and metastasis status) and adds... ARID1A Mutations serve as novel biomarkers. By constructing a joint prediction model that integrates clinical and molecular features, we can overcome the shortcomings of traditional models that rely solely on baseline features and break through the limitations of traditional models that depend on a single clinical variable, thereby improving the timeliness of predictions.

[0026] (2) This invention directly predicts the patient’s survival probability at a specific time point (such as the 1-year OS rate) based on the risk score, transforming the prediction model into a clinical decision-making tool to directly guide the selection of individualized treatment pathways.

[0027] (3) This invention selects GBM as the core modeling framework. By integrating a large number of decision trees, GBM can naturally capture the nonlinear relationship between variables and outcomes as well as the high-order interaction between variables without any prior assumptions. This enables the model to better understand and fit the complex biological mechanisms in cancer prognosis, effectively avoid overfitting, and ensure that even in sample data containing rare mutation events, it can learn stable and reliable patterns, significantly enhancing the reliability, stability and generalization ability of the model in actual clinical applications, thereby making more accurate predictions. Attached Figure Description

[0028] Figure 1 ROC curves for the GBM-Cox survival prediction model and the Cox model; Figure 2 The calibration curve for the GBM-Cox survival prediction model. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] To further understand the present invention, it will be further described in conjunction with the accompanying drawings and embodiments.

[0031] Example 1 This embodiment provides a method for constructing a personalized survival prediction model for gastric cancer immunotherapy based on the GBM-Cox model, including the following steps: Step 1: Data Preprocessing Baseline data of patients undergoing immunotherapy for gastric cancer were collected, and triplet variables were defined to describe the survival and characteristic information of each patient, as follows: Time variable: Follow-up time (unit: months) denoted as T It is used to reflect the length of time from when a patient receives immunotherapy until the occurrence of the endpoint event (death) or the end of follow-up; Event variable: Death event state, denoted as δ A binary classification coding system is used, in which δ =1 indicates that a death event has occurred. δ =0 indicates censoring (the patient did not die at the end of the follow-up period). Feature vector: Composed of patient genetic and clinical features, with a dimension of 5, denoted as Among them, the genetic characteristic is tumor mutation burden (TMB). ARID1A Mutation status; clinical characteristics include age, metastasis status, and gender. Regarding the above feature vectors... X Preprocessing is performed to eliminate interference from dimensional differences and categorical variable encoding, ensuring the stability and accuracy of the model. The specific processing methods are as follows: Continuous variable handling: Age and TMB were standardized using Z-scores to avoid the impact of differences in dimensions on model stability; Categorical variable coding: for transition situations, ARID1A Mutation status and gender are binary coded, with the following coding rules: Transmission status: 0 = no transmission, 1 = transmission; ARID1A Mutation state: 0 = no mutation, 1 = mutation; Gender: 0 = female, 1 = male.

[0032] After preprocessing, the sample data of a single patient are represented as follows: The entire patient dataset is represented as .in T i For patients i Follow-up time, δ i For patients i The status of death events, X i For patients i The preprocessed feature vector data.

[0033] in T i For patients i Follow-up time, δ i For patients i The status of death events,X i For patients i The preprocessed feature vector data.

[0034] A survival prediction model for gastric cancer immunotherapy was constructed using a framework combining gradient boosting machine (GBM) and Cox proportional hazards model, aiming to capture the complex nonlinear relationship between patient genetic characteristics, clinical characteristics, and survival outcomes. The model is defined as follows: Patient samples are represented as The risk function is

[0035] in The risk scoring function learned from GBM is used to reflect the individual risk of patients after receiving immunotherapy; Reflecting on the patient's time t The instantaneous risk of an event (death); The baseline risk function is used to ensure that the model reflects the overall impact of immunotherapy on survival while providing personalized risk assessments based on individual characteristics.

[0036]

[0037] middle M This represents the total number of iterations (the total number of decision trees). v This represents the learning rate, with a value of 0.01. L m For the first m The total number of leaf nodes in the tree; For the first m The trees in the The predicted constant values ​​at each leaf node ,in It is the first derivative (gradient). It is the second derivative; This is an indicator function. When the patient feature vector... X Falling into the corresponding feature space region The value is 1 if it is true, and 0 otherwise.

[0038] Step 3: Training the GBM-Cox model Step 3.1 Loss Function The negative logarithm of the partial likelihood function of the Cox proportional hazards model is selected as the loss function to measure the deviation between the model's predictions and actual survival data. By minimizing this loss function, patients who experience death events are given a higher probability in the risk distribution. The specific loss function is as follows:

[0039] Indicates the patient i The state of the death event, time T i The state of death events at time δ i =1 indicates that a death event has occurred, δ i =0 indicates deletion. Indicates the patient i Risk score; Indicates time T i This refers to all patients who are still under follow-up and have not yet experienced any deaths. Indicates the patient j The risk value; express j Take all risk sets All patients (including patients) i ); This represents the total risk value for all patients within the risk set.

[0040] Step 3.2 Iterative optimization of the risk function based on gradient boosting The risk scoring function is trained using a gradient boosting iterative algorithm. The model parameters are gradually optimized through multiple iterations. The specific iteration process is as follows: (1) Initialize the risk function: Set the initial risk function. =0, meaning that the risk score of all patients is 0 in the initial state; (2) Calculate the gradient and second derivative: the predicted value for each patient Calculate the loss function right First derivative (gradient) ) and second derivative (Heisenberg) ),in Used to measure the deviation between the current model's predicted value and the actual value. The curvature of the loss function is used to reflect the stable model update process. The specific calculation formula is as follows: ; ; Indicates the patient i The deviation between predictions and actual results under the current model; δ represents the curvature of the loss function, used for stable updates. i Indicates the patient i The state of death events, δ j Indicates the patient j The status of the death event; Indicates the patienti In patients j The relative risk proportion of risk concentration, among which Indicates in risk set The total risk value for all patients in the group. Indicates the patient i The risk value, Indicates the first k Feature vectors of individual patients; i , j , k Indexing patients, i This represents the target patient index for calculating the gradient and second derivative. j An index representing patients who experienced death events. express A group of patients who are still at risk; Indicates all that satisfy index of conditions j The set, Indicates time The group of patients who were still at risk.

[0041] During model training, negative gradients are used. As an optimization direction, a regression tree is fitted to approximate this direction, thereby gradually correcting the risk scoring function; simultaneously, the second derivative is combined with... The update magnitude is weighted and adjusted to improve the stability and convergence efficiency of the model update. Through the gradient-direction-based iterative update method described above, each round of model updates proceeds along the direction of loss function descent, thereby gradually reducing the loss function value and optimizing the model parameters.

[0042] Step 3.3 Fitting a weighted regression tree With pseudo residuals The gradient is used as the objective, with As weights, a weighted regression tree is fitted, which incorporates the patient feature vectors. X Divided into several leaf node regions , ( m This represents the current iteration round number. (This is the leaf node number of the current decision tree).

[0043] Step 3.4 Calculate the optimal output constant for the leaf nodes. In each leaf node region Within, based on all patients within that node gradient and weight Calculate the optimal output value for this node. The calculation formula is as follows:

[0044] Step 3.5 Update the risk scoring function Based on the weighted regression tree obtained in the current iteration, update the risk scoring function using the following formula:

[0045] Update the model. Let m be the risk scoring function after the m-th iteration. Let m be the risk scoring function after the (m-1)th iteration. v This represents the learning rate, with a value of 0.01, which helps avoid overfitting on small samples. As an indicator function, when the patient feature vector X Belonging to the feature space region of leaf nodes The value is 1 if the condition is met, and 0 otherwise.

[0046] Step 3.6 Repeat the iteration Iterate gradually until completion. M Rounds of iteration to optimize the risk scoring function This allows the model to effectively differentiate survival risk differences among patients receiving immunotherapy. To balance robustness with small sample data and model expressive power, the number of iteration trees is set. M =2000, interaction depth is 4, learning rate is 0.01.

[0047] Example 2

[0048] This embodiment provides a method for predicting individualized survival probability in gastric cancer immunotherapy based on the GBM-Cox model, using patient follow-up time. T The event state δ is used as survival data, with patient genetic and clinical characteristics as input variables. X Risk scoring function is learned iteratively through gradient boosting. Individualized survival probabilities are calculated based on the Cox proportional hazards framework.

[0049] Specifically, the following steps are included: Based on the GBM-Cox model trained in Example 1, an individualized risk score is calculated for each patient according to the input genetic and clinical characteristics. By combining the Breslow method to estimate the baseline cumulative hazard function, the personalized survival probability of each gastric cancer patient at a specified time point after immunotherapy is calculated. The specific steps are as follows: (1) Calculate individualized risk scores: The feature vector of each patient after preprocessing Input to the risk scoring function after training is complete In this process, the patient's individualized risk score was obtained. ; (2) Evaluation of the baseline cumulative risk function: The baseline cumulative risk function is estimated based on the training set data using the Breslow method. The calculation formula is:

[0050] u Indicates the time point in the training set where a death event occurs, satisfying the following conditions: u ≤ t ; d ( u ) indicates at a point in time u The number of patients who died; Indicates at a point in time u The risk set consists of all patients who are still under follow-up and have not yet experienced a death. This represents the total risk value of all patients in the risk cluster.

[0051] (3) Calculate individualized survival probability: at any follow-up time point t Accumulated risk function based on benchmark and individualized risk scores of patients Calculate the individualized survival probability of this patient. The calculation formula is: .

[0052] Example 3 GBM-Cox model evaluation To verify the effectiveness of the method of the present invention in predicting immunotherapy for gastric cancer, the consistency index (C-index), receiver operating characteristic (ROC) curve and area under the ROC curve (AUC) were used to evaluate the model performance, and the calibration curve was used to evaluate the application value of the model.

[0053] (1) Consistency index (C-index) The C-index is used to measure whether the model can correctly distinguish between high-risk and low-risk patients. For a pair of comparable patients ( i , j If the model predicts and analyzes the risk i > j And the actual observed time of death events i > j This indicates that the model successfully distinguished patients. i and j The risk level is determined by the C-index, which represents the proportion of all comparable patient pairs that the model correctly predicts. The C-index ranges from 0.5 to 1.0, where 0.5 indicates that the model's predictive ability is comparable to random guessing, and 1.0 indicates that the model's prediction is completely correct.

[0054] The calculation formula is:

[0055] in For comparable patient pairs, indicating patients i exist T i An incident occurred, and the patient j At that point in time, the follow-up was still ongoing; This is an indicator function that takes the value 1 when the condition is true and 0 otherwise.

[0056] (2) Area under the time-dependent curve (AUC) AUC is a commonly used metric to measure the ability of a survival prediction model to distinguish between positive and negative samples. It can assess predictive sensitivity and specificity at specific time points. AUC is the area under the receiver operating characteristic (ROC) curve. The ROC curve is plotted with the false positive rate (FPR, 1-specificity) on the x-axis and the true positive rate (TPR, sensitivity) on the y-axis, by varying the threshold of the predicted risk score. The calculation formula is:

[0057] , indicating time t The proportion of patients who actually experienced the event and were identified as high-risk by the model.

[0058] , indicating time t The proportion of patients who were still alive at the time of the incident who were incorrectly classified as high-risk.

[0059] The larger the value of AUC(t), the faster the time... t The stronger the predictive and discriminative ability, the better.

[0060] (3) Calibration Curve: The calibration curve assesses the consistency between the predicted survival probability and the actual observed rate. Patients are divided into 5 groups according to the predicted survival probability, denoted as... G 1 ,..., G 5 For each group g Calculate the average predicted survival rate:

[0061] in Representation group g Number of patients inside the body.

[0062] The Kaplan–Meier method was used to estimate the actual observed survival rate for each group:

[0063] At the point of time u ,Group g The number of patients who are still at risk For at a certain point in time u ,Group g The number of patients with newly occurring events.

[0064] Plot a scatter plot of predicted survival rate vs. actual survival rate, add a y=x reference line, and finally fit a linear relationship for each group to quantify the degree of calibration. , α To calibrate the intercept, the ideal value is 0; β To calibrate the slope, the ideal value is 1; This is the residual term.

[0065] (4) Evaluation results (4.1) Prediction performance Analysis using C-index, ROC curves, and AUC evaluation revealed that the GBM-Cox survival prediction model significantly outperformed the Cox proportional hazards model.

[0066] like Figure 1 As shown, based on ROC curve analysis at the 12-month time point, the GBM-Cox survival prediction model exhibits good predictive discrimination performance, with an area under the curve (AUC) of 0.903, outperforming the Cox proportional hazards model (AUC=0.681). This indicates that when predicting the risk of death 12 months after patients receive immunotherapy, the GBM-Cox survival prediction model can more accurately distinguish between high-risk and low-risk individuals.

[0067] Further C-index validation: The C-index of the GBM-Cox survival prediction model is 0.70, which is much higher than the 0.55 of the Cox model (close to the level of random guessing).

[0068] The consistent results of the two core metrics (AUC and C-index) fully demonstrate that the GBM-Cox survival prediction model significantly outperforms the traditional Cox model in both its discriminative and global ranking abilities in survival prediction. Furthermore, the GBM-Cox survival prediction model's accuracy, sensitivity, positive predictive value, and negative predictive value are 0.786, 0.875, 0.731, 0.667, and 0.905, respectively, which are superior to current Cox survival models.

[0069] Table 1 Performance Comparison of GBM-Cox Model and Cox Survival Model Model AUC Accuracy Sensitivity Specificity Positive predictive value Negative predictive value GBM-Cox model 0.903 0.786 0.875 0.731 0.667 0.905 Cox model 0.681 0.579 0.604 0.564 0.460 0.698 (4.2) Accuracy like Figure 2As shown, the GBM-Cox survival prediction model exhibits acceptable calibration accuracy in predicting 12-month survival rates, with its prediction trend generally matching actual observations (calibration curve closely approximating the reference line). Although there is a slight bias in the high / low risk extremes (slope = 1.182), the prediction in the intermediate risk range is highly reliable and can effectively support clinical stratification decisions.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, alterations, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a personalized survival prediction model for gastric cancer immunotherapy based on the GBM-Cox model, characterized in that, Includes the following steps: (1) Collect baseline data of patients receiving immunotherapy for gastric cancer, including follow-up time. T Event Status δ , and eigenvectors X The feature vector includes ARID1A Gene mutation, TMB, age, metastasis status, and sex; After preprocessing the feature vectors, the sample data of a single patient is represented as follows: The entire patient dataset is represented as ;in T i For patients i Follow-up time, δ i For patients i The status of death events, X i For patients i The preprocessed feature vector data; (2) Construct the GBM-Cox model based on patient characteristics X Using this as input, learn the risk scoring function. and through This is mapped to the relative risk of an individual; in, Reflecting on the patient's time t The instantaneous risk of an event occurring. As the benchmark risk function, It is a risk scoring function; ； in, M This represents the total number of iterations. v Indicates the learning rate; L m For the first m The total number of leaf nodes in the tree; For the first m The trees in the The predicted constant values ​​at each leaf node ; For indicator functions; (3) The negative logarithm of the partial likelihood function of the Cox proportional hazards model is selected as the loss function. The risk scoring function was trained using a gradient boosting iterative algorithm. First, initialize the risk function, then calculate the gradient. and second derivative With pseudo residuals− As a goal, with As weights, a weighted regression tree is fitted, which incorporates the patient feature vectors. X Divided into several leaf node regions In each Internally calculate the optimal output value for this node. ; Based on the weighted regression tree obtained in the current iteration, update the risk scoring function using the following formula: ； Update the model. Let m be the risk scoring function after the m-th iteration. This is the risk scoring function after the (m-1)th iteration; Iterate gradually until completion. M Rounds of iteration to optimize the risk scoring function .

2. The method for constructing a personalized survival prediction model for gastric cancer immunotherapy based on the GBM-Cox model according to claim 1, characterized in that, In step (1), the feature vector X It consists of the patient's genetic and clinical characteristics, among which the genetic characteristics include TMB, ARID1A Mutation status and clinical characteristics include age, metastasis status, and gender; Preprocessing methods include: Z-score standardization of age and TMB, and analysis of migration patterns. ARID1A Mutation state and sex are binary-classified and coded, with the specific coding rules as follows: Transfer status: 0 = not transferred, 1 = transferred; ARID1A Mutation state: 0 = no mutation, 1 = mutation; Gender: 0 = female, 1 = male.

3. The method for constructing a personalized survival prediction model for gastric cancer immunotherapy based on the GBM-Cox model according to claim 1, characterized in that, In step (2), when the patient feature vector X Falling into the corresponding feature space region The value is 1 if it is true, and 0 otherwise.

4. The method for constructing a personalized survival prediction model for gastric cancer immunotherapy based on the GBM-Cox model according to claim 1, characterized in that, In step (3), the loss function for: ； in, Indicates the patient i The state of the death event, time T i The state of death events at time δ i =1 indicates that a death event has occurred, δ i =0 indicates deletion. Indicates the patient i Risk score; Indicates time T i This refers to all patients who are still under follow-up and have not yet experienced any deaths. Indicates the patient j The risk value; express j Take all risk sets Includes patients i All patients, including those in the group; This represents the total risk value for all patients within the risk set.

5. The method for constructing a personalized survival prediction model for gastric cancer immunotherapy based on the GBM-Cox model according to claim 1, characterized in that, In step (3), the gradient boosting iterative algorithm is used to train the risk scoring function. The model parameters are gradually optimized through multiple iterations. The specific iteration process is as follows: (1) Initialize the risk function: Set the initial risk function. ; (2) Calculate the gradient and second derivative: the predicted value for each patient Calculate the loss function right gradient and second derivative The formula is as follows: ； in, Indicates the patient i The deviation between predictions and actual results under the current model; δ represents the curvature of the loss function. i Indicates the patient i The state of death events, δ j Indicates the patient j The status of the death event; Indicates the patient i In patients j The relative risk proportion of risk concentration, among which Indicates in risk set The total risk value for all patients in the group. Indicates the patient i The risk value, Indicates the first k Feature vectors of individual patients.

6. The method for constructing a personalized survival prediction model for gastric cancer immunotherapy based on the GBM-Cox model according to claim 1, characterized in that, In step (3), an iterative update is performed, and the number of iteration trees is set. M =2000, interaction depth is 4, learning rate v It is 0.

01.

7. The GBM-Cox survival prediction model obtained by the method of constructing the individualized survival prediction model for gastric cancer immunotherapy based on the GBM-Cox model as described in any one of claims 1-6.

8. A method for predicting individualized survival probability in gastric cancer immunotherapy based on the GBM-Cox model, characterized in that, The method uses the GBM-Cox survival prediction model obtained by the construction method described in any one of claims 1-6 to predict individualized survival probabilities, and includes the following steps: (1) Input the preprocessed feature vector of each patient into the trained risk scoring function. In the process of calculating individualized risk scores ; (2) Using the Breslow method, estimate the baseline cumulative risk function based on the training set data. The calculation formula is: ； in, u Indicates the time point in the training set where a death event occurs, satisfying the following conditions: u ≤ t ; d ( u ) indicates at a point in time u The number of patients who died; Indicates at a point in time u The risk set consists of all patients who are still under follow-up and have not yet experienced a death. This represents the total risk value for all patients in the risk cluster; (3) At any follow-up time point t Accumulated risk function based on benchmark and individualized risk scoring functions for patients Calculate the individualized survival probability of this patient. The calculation formula is: 。

Images