Method, apparatus and storage medium for generating cancer treatment strategies

Abstract

The application relates to the technical field of auxiliary treatment decision, and discloses a cancer treatment strategy generation method and device and a storage medium. The method comprises the following steps: acquiring clinical pathological feature information of a target patient in multiple dimensions; performing time domain cumulative risk assessment on the clinical pathological feature information to obtain comprehensive risk evaluation information for representing the overall biological malignancy degree of the target patient in a life cycle; performing causal interaction analysis on the comprehensive risk evaluation information and a candidate treatment scheme to identify a treatment recommendation time domain of the target patient and a target treatment scheme corresponding to the treatment recommendation time domain; and generating a cancer treatment strategy based on the treatment recommendation time domain and the target treatment scheme. The embodiment of the application can quantify the benefit of a patient from a specific treatment and provide direct guidance for optimizing the cancer treatment strategy.

Description

Technical Field

[0001] This application relates to the field of auxiliary treatment decision technology, and in particular to a method, apparatus and storage medium for generating cancer treatment strategies. Background Technology

[0002] Cancer treatment strategies can be formulated based on the patient's clinicopathological characteristics and in combination with multiple standard neoadjuvant therapy regimens. For example, for locally advanced esophageal squamous cell carcinoma, the main standard neoadjuvant therapy regimens are neoadjuvant chemoradiotherapy (nCRT) combined with radiotherapy to enhance local tumor control and neoadjuvant chemoimmunotherapy (nCIT) combined with immune checkpoint inhibitors and chemotherapy to enhance systemic micrometastasis control.

[0003] In related technologies, the evaluation or optimization of cancer treatment strategies is usually based on the results of prognostic risk analysis, which predicts how long a patient will live and assists in selecting different treatment strategies based on the potential risks of different patients' prognoses. However, because prognostic risk analysis cannot quantify the benefits a patient receives from a specific treatment, for example, it cannot quantify the additional survival benefits or disadvantages of nCRT compared to nCIT, and therefore cannot provide direct guidance for optimizing treatment strategies for esophageal cancer. Summary of the Invention

[0004] The purpose of this application is to provide a method, apparatus, and storage medium for generating cancer treatment strategies, which can quantify the benefits patients receive from specific treatments and provide direct guidance for optimizing cancer treatment strategies.

[0005] This application provides a method for generating a cancer treatment strategy, including: Obtain clinicopathological features of the target patient from multiple dimensions; A time-domain cumulative risk assessment is performed on the clinicopathological features to obtain comprehensive risk assessment information that characterizes the overall biological malignancy of the target patient over the course of life. A causal interaction analysis is performed on the comprehensive risk assessment information and candidate treatment plans to identify the treatment recommendation time domain for the target patient and the target treatment plan corresponding to the treatment recommendation time domain; A cancer treatment strategy is generated based on the treatment recommendation time domain and the target treatment plan.

[0006] In some embodiments, the step of performing a full-time cumulative risk assessment on the clinicopathological feature information includes: Risk features are extracted from the clinicopathological features to obtain multiple cumulative risk assessment information that characterize the biological malignancy of the target patient across the entire time domain; The cumulative risk assessment information is fitted to obtain the comprehensive risk assessment information.

[0007] In some embodiments, the risk feature extraction of the clinicopathological feature information includes: The random survival forest method is used to select features from the clinical pathological features and obtain the cumulative risk information stored in the leaf nodes of each survival tree to obtain the cumulative risk assessment information.

[0008] In some embodiments, fitting the cumulative risk assessment information includes: Determine the average assessment information of each of the cumulative risk assessment information at the same point in time; The mean evaluation information is discretized into a time series of cumulative risk values. The comprehensive risk assessment information is obtained by summing the values ​​at each discrete time point in the risk accumulation value sequence.

[0009] In some embodiments, the causal interaction analysis of the comprehensive risk assessment information and candidate treatment options includes: Solve the nonlinear interaction relationship between the comprehensive risk assessment information and the candidate treatment plan; Based on the aforementioned nonlinear interaction relationship, a risk ratio function is constructed among the candidate treatment options; Based on the risk ratio function, the comprehensive risk assessment information is defined in the time domain and risk assessed to determine the treatment recommendation time domain and the target treatment plan corresponding to the treatment recommendation time domain.

[0010] In some embodiments, calculating the nonlinear interaction relationship between the comprehensive risk assessment information and the candidate treatment options includes: The comprehensive risk assessment information is subjected to a nonlinear transformation to obtain nonlinear assessment information; Based on the comprehensive risk assessment information, the nonlinear assessment information, and the candidate treatment plans, a weighted proportional risk relationship is constructed. The nonlinear interaction relationship is obtained by weighting each item in the weighted proportional risk relationship using preset inverse processing probability weights; the inverse processing probability weights are obtained by inverse processing probability weighting of the conditional probability of historical patients receiving the candidate treatment plan.

[0011] In some embodiments, the step of defining the comprehensive risk assessment information in the time domain and conducting risk assessment based on the risk ratio function includes: Solve the equation where the risk ratio function equals a unit value, and use the root of the equation as the time-domain boundary threshold; Using the time domain boundary threshold as the boundary, the comprehensive risk assessment information is divided into at least two treatment recommendation time domains, and the target treatment plan corresponding to the treatment recommendation time domain is determined based on the value of the risk ratio function within the treatment recommendation time domain.

[0012] In some embodiments, the method for generating the cancer treatment strategy further includes: Based on the clinicopathological features of the target patient, a preset contribution evaluation strategy is used to determine the contribution of each feature in the clinicopathological features to the comprehensive risk assessment information. Based on the contribution of the feature information to the comprehensive risk assessment information, a contribution waterfall chart is generated to demonstrate the decision-making basis of the cancer treatment strategy.

[0013] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method for generating a cancer treatment strategy.

[0014] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for generating a cancer treatment strategy.

[0015] The beneficial effects of this application are as follows: It introduces a time-domain cumulative risk assessment mechanism and a causal interaction analysis mechanism. Based on the patient's multi-dimensional clinicopathological characteristics, time-domain cumulative risk assessment and causal interaction analysis are performed to generate individualized treatment strategies. Therefore, by introducing time-domain cumulative risk assessment, the overall biological malignancy of the target patient over their lifespan can be dynamically characterized, rather than merely providing a static prognostic judgment. Through causal interaction analysis of this comprehensive risk assessment information and candidate treatment options, the specific impact of different treatment options on the patient's risk curve at different time periods can be directly calculated and quantified. This allows for the quantification of the patient's benefit from a specific treatment, providing direct guidance for optimizing cancer treatment strategies. Attached Figure Description

[0016] Figure 1 This is a diagram illustrating the application environment of the method for generating cancer treatment strategies provided in the embodiments of this application.

[0017] Figure 2 This is a flowchart of a method for generating cancer treatment strategies provided in an embodiment of this application.

[0018] Figure 3 This is a flowchart of a method for performing full-time-domain cumulative risk assessment of clinicopathological feature information, provided in an embodiment of this application.

[0019] Figure 4 This is a flowchart of a method for performing causal interaction analysis on comprehensive risk assessment information and candidate treatment options, provided in an embodiment of this application.

[0020] Figure 5 This is a schematic diagram illustrating the division of the treatment recommendation time domain provided in the embodiments of this application.

[0021] Figure 6 This is a schematic diagram of contribution evaluation provided in the embodiments of this application.

[0022] Figure 7 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0024] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and drawings are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application. Furthermore, the information, data, and signals involved in the embodiments of this application are all authorized by relevant parties or have been fully authorized by all parties, and the collection, use, and processing of related data comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0026] In the traditional development of cancer treatment strategies, existing technologies rely on prognostic risk analysis to assess patient survival. However, this method cannot quantify the causal treatment benefits of specific treatment regimens. Specifically, prognostic risk analysis can only provide overall risk stratification for patients, but cannot identify the relative efficacy differences between different treatment regimens (such as neoadjuvant chemoradiotherapy versus neoadjuvant chemoradiotherapy). This problem stems from the fact that prognostic risk analysis models do not establish a causal relationship between treatment regimens and dynamic changes in risk. This further leads to a lack of direct quantitative evidence for the treatment timeframe and regimen selection when optimizing treatment strategies, thus affecting the objectivity and accuracy of treatment decisions. For example, in the clinical practice of locally advanced esophageal squamous cell carcinoma, physicians need to choose between neoadjuvant chemoradiotherapy and neoadjuvant chemoradiotherapy based on the patient's clinicopathological characteristics. Existing technologies predict overall survival risk by analyzing demographic, anatomical, and biological characteristics, but cannot determine the additional survival benefits or disadvantages of neoadjuvant chemoradiotherapy compared to neoadjuvant chemoradiotherapy. Specifically, when patients exhibit high biological malignancy characteristics, existing methods may incorrectly recommend neoadjuvant chemotherapy and immunotherapy, when in fact neoadjuvant chemoradiotherapy may be more effective in a specific time domain. This problem manifests as the inability to resolve the nonlinear interaction between comprehensive risk assessment information and candidate treatment options, resulting in a lack of definition of the treatment recommendation time domain, and ultimately making the treatment strategy unable to adapt to the dynamic risk evolution of the patient.

[0027] If the aforementioned issues are not addressed, the selection of treatment strategies will rely primarily on empirical judgment rather than data-driven quantitative analysis, potentially leading to patients receiving treatment regimens mismatched with the degree of biological malignancy. In particular, the selection of suboptimal treatment regimens can reduce local tumor control and weaken the ability to control systemic micrometastases, further leading to the accumulation of treatment-related risks. This negatively impacts patients' long-term quality of life and hinders the overall effectiveness of cancer treatment systems.

[0028] Based on this, embodiments of this application provide a method, apparatus, and storage medium for generating cancer treatment strategies. Based on the patient's multi-dimensional clinicopathological characteristics, a time-domain cumulative risk assessment and causal interaction analysis are performed to generate individualized treatment strategies. This can accurately quantify patient benefits, provide targeted treatment decisions, and improve the accuracy of clinical decision-making.

[0029] Figure 1 This diagram illustrates the application environment of the cancer treatment strategy generation method provided in this embodiment. (See also...) Figure 1This method is applied to a cancer treatment strategy generation system. The system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be at least one of a mobile phone, tablet, laptop, or vehicle-mounted terminal. The server 120 can be a standalone server or a server cluster consisting of several servers. The terminal 110 sends multi-dimensional clinicopathological feature information of the target patient to the server 120. The server 120 acquires the multi-dimensional clinicopathological feature information of the target patient, performs a time-domain cumulative risk assessment on the clinicopathological feature information to obtain comprehensive risk assessment information characterizing the overall biological malignancy of the target patient over their lifespan, performs causal interaction analysis on the comprehensive risk assessment information and candidate treatment plans to identify the treatment recommendation time domain and the corresponding target treatment plan for the target patient, and generates a cancer treatment strategy based on the treatment recommendation time domain and the target treatment plan.

[0030] It should be understood that Figure 1 The application scenarios shown are merely examples. In practical applications, the cancer treatment strategy generation method provided in this application embodiment can also be applied to other scenarios. For example, the above-described cancer treatment strategy generation method can be directly applied to terminal 110. Terminal 110 is used to acquire multi-dimensional clinicopathological feature information of the target patient, perform time-domain cumulative risk assessment on the clinicopathological feature information to obtain comprehensive risk assessment information that characterizes the overall biological malignancy of the target patient throughout their life cycle, perform causal interaction analysis on the comprehensive risk assessment information and candidate treatment plans to identify the treatment recommendation time domain and the target treatment plan corresponding to the treatment recommendation time domain, and generate a cancer treatment strategy based on the treatment recommendation time domain and the target treatment plan.

[0031] See Figure 2 In one embodiment, a method for generating a cancer treatment strategy is provided. The execution subject of the method is a server or terminal, including but not limited to steps S201 to S204.

[0032] Step S201: Obtain clinicopathological feature information of the target patient from multiple dimensions.

[0033] Clinicopathological features refer to multidimensional data used to describe the disease status and individual characteristics of a target patient. These features can include demographic characteristics (e.g., age, sex), anatomical characteristics (e.g., tumor size, lymph node metastasis), and biological characteristics (e.g., gene mutations, protein expression levels). This information forms the basis for a comprehensive assessment of the patient's condition.

[0034] This clinicopathological information can be obtained by medical institutions through manual input. For example, doctors or nurses manually enter data such as the patient's age, gender, tumor size, lymph node status, and genetic testing results into an electronic medical record system or data management platform. This information can also be obtained by scanning paper medical records or reports and digitizing them using optical character recognition (OCR) technology.

[0035] In some embodiments, if the obtained clinicopathological feature information is missing, the chain equation multiple imputation (MICE) method is used to train a random forest regressor or classifier with other complete variables to predict and impute each missing feature information.

[0036] Step S202: Perform time-domain cumulative risk assessment on clinicopathological feature information to obtain comprehensive risk assessment information that characterizes the overall biological malignancy of the target patient over the course of life.

[0037] Time-domain cumulative risk assessment refers to the analysis of a patient's clinicopathological characteristics to quantify the risk of disease progression or recurrence at different time points or throughout the patient's life. This assessment aims to capture the dynamic changes in the disease and transform them into quantifiable risk indicators.

[0038] Comprehensive risk assessment information is the result of time-domain cumulative risk assessment, designed to characterize the overall biological malignancy of a target patient over their lifespan. This information is typically a comprehensive numerical value or curve that reflects the severity of the patient's disease and prognostic trends.

[0039] Time-domain cumulative risk assessment of clinicopathological features can be performed using various statistical or machine learning models. For example, a Cox proportional hazards model can be used to model patient survival data, predicting the patient's survival probability at different time points and converting it into a cumulative risk value. Alternatively, a decision tree model can be constructed to stratify patients based on their clinical characteristics and calculate the corresponding cumulative risk for each stratum.

[0040] Step S203 involves performing a causal interaction analysis on the comprehensive risk assessment information and candidate treatment plans to identify the treatment recommendation time domain for the target patient and the target treatment plan corresponding to the treatment recommendation time domain.

[0041] Causal interaction analysis is a statistical or machine learning method used to explore whether causal relationships exist and their interactions exist among different factors, such as integrated risk assessment information and candidate treatment options. This analysis can identify and quantify the impact of specific treatment options on patient risk.

[0042] The recommended treatment timeframe refers to a specific treatment plan that can bring the best benefit or the lowest risk during a certain period of time in the patient's entire disease course.

[0043] A target treatment plan refers to the most effective or appropriate treatment method for a target patient within a specific treatment recommendation timeframe. This plan is determined based on the patient's comprehensive risk assessment information and the results of causal interaction analysis.

[0044] Causal interaction analysis of the comprehensive risk assessment information and candidate treatment options can be conducted using traditional statistical methods. For example, survival analysis can be performed on patients in different treatment option groups to compare the differences in their survival curves, thereby making a preliminary judgment on the effectiveness of the treatment options. Alternatively, a regression model can be constructed, using the comprehensive risk assessment information and treatment options as input variables to predict patient treatment outcomes, and the model results can be used to identify potential treatment recommendations and target treatment options in the time domain.

[0045] Step S204: Generate a cancer treatment strategy based on the treatment recommendation time domain and target treatment plan.

[0046] Cancer treatment strategies are a series of treatment plans and interventions developed based on recommended treatment timeframes and target treatment protocols. These strategies aim to provide patients with personalized and precise treatment guidance to achieve the best possible therapeutic outcomes.

[0047] Cancer treatment strategies can be manually developed by clinicians based on identified treatment recommendations and target therapies, combined with their clinical experience and medical guidelines. For example, based on the analysis results, a doctor can recommend chemotherapy for a patient during a specific time period and radiotherapy during another time period, thus forming a complete treatment plan.

[0048] The following example will provide a more detailed explanation of the above technical solution: Suppose there is a cancer patient, labeled User A. User A has been diagnosed with esophageal cancer and requires a personalized treatment strategy. Existing prognostic risk analysis methods cannot quantify the specific benefits or disadvantages User A gains from a particular treatment plan, and therefore cannot provide direct guidance for optimizing User A's cancer treatment strategy.

[0049] The method provided in this embodiment first obtains clinicopathological feature information of user A from multiple dimensions. Specifically, demographic information (e.g., age, gender), anatomical information (e.g., primary tumor site, tumor size, lymph node metastasis), and biological feature information (e.g., gene mutation type, tumor marker levels) can be extracted from user A's electronic medical record system. This information is then integrated into a comprehensive dataset.

[0050] Subsequently, a time-domain cumulative risk assessment is performed on this clinicopathological feature information. For example, a pre-trained machine learning model, built upon a large amount of historical patient data, can be used to predict the risk of disease progression at different future time points based on user A's clinicopathological features. This assessment yields a dynamically changing comprehensive risk assessment, presented as a curve representing the trend of the overall biological malignancy of user A's disease over time throughout their lifespan. For example, the curve might show that user A has a higher risk at the beginning of treatment, gradually decreasing the risk after a certain point in time, or fluctuating the risk within a specific time period.

[0051] Next, a causal interaction analysis is performed on the comprehensive risk assessment information and multiple candidate treatment options. These options may include surgery, chemotherapy, radiotherapy, targeted therapy, or immunotherapy. The analysis aims to quantify the impact of each candidate treatment option on user A's comprehensive risk assessment information. Specifically, the analysis identifies the time periods within user A's comprehensive risk assessment information curve where a particular treatment option significantly reduces risk or prolongs progression-free survival. For example, the analysis might indicate that chemotherapy is most effective in reducing user A's risk within the first six months after diagnosis, while targeted therapy may be more effective between six months and one year. This allows for the identification of recommended treatment timeframes for user A (e.g., 0-6 months, 6-12 months) and the corresponding target treatment options for each timeframe (e.g., chemotherapy recommended for 0-6 months, targeted therapy recommended for 6-12 months).

[0052] Finally, based on the identified treatment recommendation timeframe and target treatment regimen, a corresponding cancer treatment strategy is generated. For example, a phased treatment plan is developed for user A: receiving a specific dose of chemotherapy within the first six months after diagnosis; subsequently, switching to targeted drug therapy within the next six months. This strategy not only clarifies the treatment plan but also specifies the optimal timing for treatment, thus providing user A with personalized and precise treatment guidance.

[0053] Based on the example of User A above, the method provided in this embodiment demonstrates a significant technical contribution in solving the problems of the prior art. Prognostic risk analysis in the prior art can typically only predict a patient's overall survival or recurrence risk, but cannot clearly quantify the specific impact of a particular treatment plan on the individual patient's risk curve, that is, it cannot directly guide the selection and timing of treatment plans.

[0054] This embodiment introduces a time-domain cumulative risk assessment, which dynamically characterizes the overall biological malignancy of the target patient throughout their lifespan, rather than simply providing a static prognostic judgment. Furthermore, by performing causal interaction analysis on this comprehensive risk assessment information and candidate treatment options, this method can directly calculate and quantify the specific impact of different treatment options on the patient's risk curve at different time periods. For example, in the case of user A, this method not only identifies user A's overall risk trend, but more importantly, it clearly indicates that chemotherapy has the most significant risk reduction effect in the 0-6 month period, while targeted therapy is more effective in the 6-12 month period. This direct causal quantitative analysis ensures that the selection of treatment options no longer relies solely on empirical judgment or vague prognostic risks, but is based on data-driven, quantifiable assessments of treatment benefits.

[0055] Therefore, this method can identify specific treatment recommendation time domains and corresponding target treatment plans, thereby generating highly personalized and precise cancer treatment strategies. This contrasts with existing technologies that rely solely on prognostic risk to assist in selecting treatment strategies. This method provides clinicians with clear decision-making criteria for "when to treat" and "how to treat," significantly improving the optimization level of cancer treatment strategies and patient outcomes. This technological concept overcomes the limitation of existing technologies in quantifying the benefits of specific treatments by combining dynamic risk assessment with causal interaction analysis, providing a new approach to precision cancer treatment.

[0056] See Figure 3 In one embodiment, the method for performing full-time cumulative risk assessment of clinicopathological feature information includes, but is not limited to, steps S301 to S302.

[0057] Step S301: Risk features are extracted from clinicopathological information to obtain multiple cumulative risk assessment information for characterizing the biological malignancy of the target patient across the entire time domain.

[0058] Step S302: Fit the cumulative risk assessment information to obtain comprehensive risk assessment information.

[0059] Risk feature extraction refers to the process of identifying and quantifying key risk factors or patterns related to the biological malignancy of a patient from raw clinicopathological information. Its role is to transform high-dimensional, complex clinical data into more representative and interpretable risk indicators. This can be achieved through various machine learning or statistical methods. For example, decision tree-based models (such as survival trees and random forests) can be used to identify feature combinations related to survival risk, or dimensionality reduction techniques such as principal component analysis and factor analysis can be used to extract potential risk factors. Furthermore, deep learning models (such as recurrent neural networks and long short-term memory networks) can be used to capture the complex dependencies of clinicopathological information over time, thereby extracting risk features with time-cumulative effects.

[0060] Cumulative risk assessment information is a set of quantitative indicators used to characterize the biological malignancy of a target patient across the entire time domain, after risk feature extraction. This information is usually in numerical form, reflecting the patient's risk level at different time points or different feature dimensions. For example, it can be the probability of survival, risk of disease progression, risk of recurrence, etc., at a specific time point. This information can be discrete risk scores or continuous risk curves. Its purpose is to provide basic data for subsequent risk fitting, in order to more comprehensively characterize the patient's risk dynamics.

[0061] Fitting refers to the process of integrating, smoothing, or modeling multiple cumulative risk assessment information through mathematical models or algorithms to generate a unified, comprehensive risk assessment that fully reflects the overall biological malignancy of a patient. Its role is to eliminate data noise, fill in missing information, reveal potential trends, and provide a continuous or structured risk assessment result. Fitting methods can include, but are not limited to: using statistical regression models (such as the Cox proportional hazards model and generalized linear models) to model cumulative risk information; using time series analysis methods (such as the ARIMA model and Kalman filtering) to smooth and predict risk data; or using nonparametric methods (such as kernel density estimation and locally weighted regression) to capture the complex patterns of risk distribution.

[0062] Comprehensive risk assessment information is a single or multi-dimensional quantitative result obtained through fitting, used to characterize the overall biological malignancy of a target patient over their lifespan. It is typically an indicator that summarizes the patient's overall risk level, such as a comprehensive risk score, a risk evolution curve, or a risk stratification result. Its purpose is to provide a high-level, easily understood risk overview for subsequent treatment decisions, thereby supporting physicians in developing personalized treatment strategies.

[0063] This application's approach refines the process of conducting full-time-domain cumulative risk assessment of clinicopathological features into two stages: risk feature extraction and cumulative risk assessment information fitting. This allows for more accurate and comprehensive acquisition of integrated risk assessment information for target patients. Specifically, firstly, risk features are extracted from clinicopathological features. This aims to identify and quantify key risk features closely related to the patient's biological malignancy throughout the entire time domain from raw, multi-dimensional clinical data, thereby obtaining multiple cumulative risk assessment information. Subsequently, these extracted cumulative risk assessment information are fitted. By constructing appropriate mathematical models or algorithms, data noise is eliminated, potential risk evolution trends are revealed, and ultimately, a unified integrated risk assessment information that comprehensively reflects the overall biological malignancy of the target patient throughout their lifespan is generated. Through this phased and refined processing flow, this application's approach overcomes the potential inaccuracies and information loss issues associated with directly conducting time-domain cumulative risk assessment. This ensures that the obtained integrated risk assessment information is more accurate and stable, thus providing a more reliable basis for subsequent causal interaction analysis and cancer treatment strategy generation.

[0064] As a specific implementation method, when conducting full-time-domain cumulative risk assessment of clinicopathological feature information, risk feature extraction can be performed first. For example, machine learning-based survival analysis models, such as gradient boosting survival models or deep survival models, can be used to train the patient's demographic, anatomical, and biological characteristics to identify key feature combinations related to the patient's long-term survival risk, disease recurrence risk, or progression risk, and output risk prediction values ​​at different time points. These prediction values ​​constitute multiple cumulative risk assessment information. Subsequently, these cumulative risk assessment information are fitted. For example, a weighted average method can be used, assigning different weights based on the importance or reliability of different cumulative risk assessment information, and then calculating a weighted average to obtain the comprehensive risk value at each time point. Alternatively, curve fitting techniques, such as spline interpolation or polynomial regression, can be used to fit the discrete cumulative risk assessment information into a continuous risk curve. This curve represents the comprehensive risk assessment information, which can intuitively display the patient's risk dynamics throughout their entire lifespan.

[0065] The aforementioned technical solution refines the full-time-domain cumulative risk assessment process into two steps: risk feature extraction and cumulative risk assessment information fitting. This significantly improves the accuracy and comprehensiveness of the integrated risk assessment information. The risk feature extraction step effectively identifies and quantifies key risk factors from complex clinical data, avoiding information redundancy and noise interference, making subsequent risk assessments more targeted. Fitting the cumulative risk assessment information integrates scattered risk indicators into a unified and continuous comprehensive risk assessment, not only smoothing data fluctuations but also revealing the inherent patterns of patient risk changes over time, thus more accurately reflecting the overall biological malignancy of the target patient throughout their lifespan. This provides a more solid and reliable data foundation for subsequent treatment recommendations and the identification of target treatment plans, thereby making the generated cancer treatment strategies more personalized and precise, effectively improving the scientific rigor and effectiveness of treatment decisions.

[0066] In some embodiments, risk feature extraction of clinicopathological feature information includes: using a random survival forest method to select features from the clinicopathological feature information, obtaining the cumulative risk information stored in the leaf nodes of each survival tree, and obtaining cumulative risk assessment information.

[0067] The Random Survival Forest method utilizes an ensemble learning algorithm to process survival data by constructing multiple decision trees, effectively handling censoring. This method can learn nonlinear relationships from complex clinical data and assess the importance of features. Specifically, existing machine learning libraries, such as the `scikit-survival` library in Python, can be used to construct a Random Survival Forest model by configuring parameters such as the number of trees, the minimum number of samples required for each leaf node, and the feature sampling strategy. Alternatively, the algorithm can be implemented custom-, including data bootstrapping, recursive splitting based on survival analysis metrics (such as the log-rank test statistic), and processing of censored data.

[0068] Feature selection for clinicopathological information refers to the process in which, during the construction of a random survival forest, the algorithm selects the optimal feature from a randomly chosen subset of features for splitting at each split node of each decision tree. This inherent mechanism enables the model to automatically identify and focus on features that have the greatest impact on patient survival risk, thereby effectively reducing the dimensionality of the feature space and improving the model's generalization ability. In addition to automatic selection within the model, preliminary feature screening or ranking can be performed before constructing the random survival forest to further optimize the input feature set.

[0069] Obtaining the cumulative risk information stored in the leaf nodes of each survival tree means that after each survival tree in a random survival forest is constructed, each leaf node represents a group of patients with similar survival patterns. For a patient falling into a specific leaf node, their cumulative risk (or cumulative hazard) can be estimated based on the survival data of the training samples within that leaf node. For example, the Nelson-Aalen estimator or the Kaplan-Meier estimator can be used to calculate the cumulative hazard function represented by that leaf node.

[0070] This application's solution significantly improves the accuracy and robustness of cumulative risk assessment information by introducing the random survival forest method to extract risk features from clinicopathological features. Specifically, the random survival forest method can handle high-dimensional, complex clinicopathological features that may contain censored data when extracting risk features. This method constructs a large number of survival trees, each of which is bootstrap sampled from the original data during training, and a subset of features is randomly selected for splitting. This ensemble learning and stochastic mechanism allows the model to effectively avoid overfitting and capture the nonlinear and complex interaction between clinicopathological features and patient survival risk. The leaf nodes of each survival tree store the cumulative risk information of the patient group represented by that node, which is estimated based on actual survival data. By aggregating (e.g., averaging) the predictions of all survival trees, more stable and accurate cumulative risk assessment information can be obtained. This method not only automatically selects features to identify clinical features that have a key impact on patient prognosis but also provides time-dependent risk assessment, i.e., the cumulative risk of patients at different time points. Therefore, compared to simply extracting general risk features, the random survival forest method can generate more refined and predictive cumulative risk assessment information, providing high-quality input for subsequent comprehensive risk assessment information, thereby making the generated cancer treatment strategies more personalized and precise.

[0071] The following is a concrete example to illustrate this. Suppose we need to generate a treatment strategy for a cancer patient. First, we need to obtain the patient's clinicopathological features. To extract risk features from these clinicopathological features, we can use a random survival forest model pre-trained on a large amount of historical cancer patient data. This model may consist of hundreds of decision trees, each of which is constructed using random sampling and random feature selection to enhance its diversity. When the patient's clinicopathological features are input into this random survival forest model, the patient's data will descend along the decision path of each tree, eventually falling into a leaf node. Each leaf node already stores a cumulative risk function calculated based on the training data. For example, a certain leaf node might represent "patients aged over 60 years, with tumors larger than 5cm, and exhibiting a specific gene mutation." This leaf node would have a corresponding cumulative risk function describing the probability of such patients experiencing events (such as recurrence or death) at different time points. By averaging or weighted averaging the cumulative risk information corresponding to the leaf nodes the patient falls into across all survival trees, we can obtain the patient's cumulative risk assessment information. For example, a curve representing the patient's cumulative risk of recurrence over the next 5 years can be obtained.

[0072] The above-described technical solution employs a random survival forest method to extract risk features from clinicopathological information, effectively addressing the common challenges of complexity, high dimensionality, and censoring in clinical data. Through its inherent ensemble learning and stochastic mechanisms, this method automatically selects features, identifying clinical features crucial to patient prognosis and capturing complex nonlinear interactions between features. This results in more accurate and time-dependent cumulative risk assessment information, providing a more nuanced characterization of the biological malignancy of the target patient throughout their lifespan. Consequently, the subsequent generation of comprehensive risk assessment information is based on a more reliable risk evaluation foundation, significantly improving the personalization, precision, and effectiveness of the final cancer treatment strategy, providing patients with more targeted treatment options.

[0073] In some embodiments, fitting the cumulative risk assessment information includes: determining the mean assessment information of each cumulative risk assessment information at the same time point; discretizing the mean assessment information into a risk cumulative value sequence over time; and summing the values ​​at each discrete time point in the risk cumulative value sequence to obtain comprehensive risk assessment information.

[0074] Determining the mean risk assessment information at the same point in time involves averaging these cumulative risk assessment information at the same point in time. This effectively smooths out potential noise or abnormal fluctuations in individual risk assessment information, resulting in a more stable and reliable risk assessment at that point in time. For example, statistical methods such as arithmetic mean, weighted average, or median can be used to calculate the mean risk assessment information.

[0075] Discretizing the mean assessment information into a time series of cumulative risk values ​​transforms continuous or non-uniformly distributed mean assessment information into a series of risk values ​​at preset discrete time points. This discretization process allows risk information to be represented in a standardized, ordered time series format, facilitating subsequent quantitative analysis and cumulative calculations. For example, fixed time intervals (such as monthly, quarterly, or annually) can be set, and the mean assessment information can be interpolated or sampled to generate a series of cumulative risk values ​​corresponding to these discrete time points.

[0076] This application's approach refines multiple cumulative risk assessment data to generate stable and reliable comprehensive risk assessment information. First, for multiple cumulative risk assessment data extracted from clinicopathological features, this approach calculates their mean assessment data at the same time points. This process effectively integrates information from different assessment sources or models, reducing the randomness and uncertainty of individual assessments through averaging, thus obtaining more robust and representative risk estimates at each time point. Subsequently, these mean assessment data obtained at different time points are discretized and organized into a time-series sequence of cumulative risk values. This serialization ensures the continuity and comparability of risk information over time, laying the foundation for subsequent cumulative calculations. Finally, by summing the values ​​at each discrete time point in this cumulative risk value sequence, this approach can aggregate the patient's risk contribution at different time periods into a single, comprehensive, and integrated risk assessment information. This summation method not only provides a quantitative indicator of overall malignancy, but also, through the synergistic effect of the above steps, ensures that the comprehensive risk assessment information can accurately and stably reflect the overall biological malignancy of the target patient throughout their life cycle, providing a solid and reliable data foundation for subsequent causal interaction analysis and the generation of cancer treatment strategies.

[0077] In one specific embodiment, the clinicopathological features of the target patient Each tree is assigned to a specific leaf node, and each leaf node corresponds to a Nelson-Aalen cumulative risk estimation curve. The formula for determining the mean value of each cumulative risk assessment information at the same point in time is: , in, To accumulate the average risk assessment information at the same point in time, Let be the total number of decision trees in the random survival forest. For the first The cumulative risk estimate output by the survival tree (based on the corresponding leaf node to which the patient falls).

[0078] In a specific embodiment, to transform complex time-related risks into a single scalar for causal analysis, the mean assessment information is discretized into a sequence of cumulative risk values ​​over time. This can be achieved by integrating the mean assessment information over the entire follow-up time domain, and then summing the values ​​at each discrete time point in the cumulative risk value sequence to obtain comprehensive risk assessment information. , in, Comprehensive risk assessment information for target patients, For target patients at specific time points in the defined risk cumulative value sequence The cumulative risk estimate above, Let j be the j-th discrete time point within the follow-up time domain.

[0079] Through the above technical solution, this application can effectively integrate cumulative risk assessment information generated from multiple sources or models. By averaging, the randomness and uncertainty of individual assessments are reduced, improving the stability and accuracy of risk assessment. Discretizing the mean assessment information into a time series and summing the results allows the final comprehensive risk assessment information to fully and quantitatively reflect the overall biological malignancy of the target patient throughout their lifespan. This not only overcomes the limitations of single or fragmented risk assessments but also provides a more reliable and robust input for subsequent causal interaction analysis, thereby significantly improving the scientific rigor and precision of the cancer treatment strategy generation process and contributing to providing patients with more personalized and effective treatment options.

[0080] See Figure 4 In one embodiment, the method for performing causal interaction analysis on comprehensive risk assessment information and candidate treatment options includes, but is not limited to, steps S401 to S403.

[0081] Step S401: Solve the nonlinear interaction relationship between the comprehensive risk assessment information and the candidate treatment plan.

[0082] Step S402: Based on the nonlinear interaction relationship, construct the risk ratio function between candidate treatment options.

[0083] Step S403: Based on the risk ratio function, the comprehensive risk assessment information is defined in the time domain and the risk is assessed to determine the treatment recommendation time domain and the target treatment plan corresponding to the treatment recommendation time domain.

[0084] Determining the nonlinear interaction between comprehensive risk assessment information and candidate treatment options refers to using mathematical models to reveal the complex, disproportionate, or nonlinear relationship between the overall biological malignancy of a patient and the effectiveness of different treatment options. This can be achieved in various ways. For example, deep learning models, such as recurrent neural networks or long short-term memory networks, can be used to capture nonlinear dependencies over time by learning from large amounts of historical patient data. Alternatively, nonlinear regression models, such as generalized additive models or kernel regression methods, can be used to model comprehensive risk assessment information and candidate treatment options to identify their complex nonlinear modes of action.

[0085] Constructing a hazard ratio function based on nonlinear interaction relationships refers to quantifying the relative risks or benefits of different treatment options under a specific risk context after clarifying the nonlinear relationship between comprehensive risk assessment information and candidate treatment options. This hazard ratio function can be expressed as the ratio of the incidence rates of specific events (such as disease progression, relapse, or death) under different treatment options. For example, the hazard ratio of any two candidate treatment options at a specific time point can be calculated based on the survival curves or risk curves predicted by the nonlinear interaction model for each treatment option; alternatively, a function can be constructed that takes the patient's comprehensive risk assessment information as input and outputs the hazard ratio of each candidate treatment option relative to a benchmark treatment option.

[0086] Based on the hazard ratio function, the comprehensive risk assessment information is time-domain defined and risk-assessed to determine the recommended treatment time domain and the corresponding target treatment plan. This involves dynamically dividing the patient's life cycle into time periods where different treatment plans may have the best effect or lowest risk using the hazard ratio function, and recommending the most suitable treatment plan for each time period. This can be achieved by analyzing the trend of the hazard ratio function over time. For example, when the hazard ratio function value of a certain treatment plan changes significantly at a certain point in time, that point in time can be used as the time domain boundary; alternatively, a predetermined hazard ratio threshold can be set, triggering a redefinition of the time domain and adjustment of the treatment plan when the hazard ratio function value crosses this threshold.

[0087] This application's approach, by introducing the solution to the nonlinear interaction between comprehensive risk assessment information and candidate treatment options, can more precisely characterize the dynamic risk changes of an individual patient under different treatment plans. This nonlinear modeling method transcends the limitations of traditional linear assumptions, leading to a deeper understanding of complex biological processes. Based on this, by constructing a risk ratio function between candidate treatment plans, a quantitative basis for comparing the merits of different treatment plans is provided, enabling treatment decisions to move beyond empirical judgment and rely on data-driven, precise calculations. Furthermore, by using this risk ratio function to perform time-domain definition and risk assessment of comprehensive risk assessment information, dynamic optimization of treatment plans is achieved. This means that the system can intelligently identify the most suitable treatment timing and specific plan for a patient at different life stages based on real-time changes in the patient's comprehensive risk assessment information and the risk ratio functions of each treatment plan. This method enables the aforementioned cancer treatment strategy generation method to provide more personalized and precise treatment recommendations, effectively avoiding the problem of poor treatment effects or increased side effects caused by a mismatch between the treatment plan and the patient's risk status, thereby significantly improving the effectiveness and safety of treatment.

[0088] The following is a concrete example to illustrate this. When calculating the nonlinear interaction between comprehensive risk assessment information and candidate treatment options, a deep learning-based causal inference model can be used. For example, a neural network with multiple hidden layers can be constructed. This network takes the comprehensive risk assessment information of the target patient (such as a time-varying risk score sequence) and the encoding of candidate treatment options as input, and learns to predict the patient's long-term survival probability or disease progression time under different treatment options through training. The nonlinear activation function and multi-layer structure of this neural network enable it to capture the complex nonlinear interaction patterns between comprehensive risk assessment information and treatment options. Based on this, a risk ratio function between the candidate treatment options is constructed based on the nonlinear interaction relationship. For example, if the above deep learning model can predict that the patient's mortality risk in the next five years is RA under treatment option A and RB under treatment option B, then a risk ratio function RR(A,B)=RA / RB can be constructed. This function quantifies the relative risk of treatment option A compared to treatment option B under the current comprehensive risk assessment information. Furthermore, based on the risk ratio function, the comprehensive risk assessment information is defined in a time domain and risk assessed to determine the recommended treatment time domain and the corresponding target treatment regimen. For example, the patient's comprehensive risk assessment information can be continuously monitored, and risk ratio functions between different treatment regimens (such as chemotherapy, targeted therapy, and immunotherapy) can be calculated. Suppose that in the early stages of treatment, the risk ratio function value of chemotherapy relative to immunotherapy is less than 1 (i.e., chemotherapy is superior), but as the patient's condition progresses and the comprehensive risk assessment information changes, at a certain point in time (e.g., 6 months after treatment), the risk ratio function value of immunotherapy relative to chemotherapy becomes less than 1 (i.e., immunotherapy is superior). At this point, the system can define two recommended treatment time domains: the first time domain is from the initial treatment to 6 months, with chemotherapy as the recommended target treatment regimen; the second time domain is after 6 months, with immunotherapy as the recommended target treatment regimen. In this way, dynamic adjustment and optimization of the treatment regimen are achieved.

[0089] Through the above technical solution, this application overcomes the limitations of traditional causal interaction analysis in handling complex nonlinear relationships. By solving the nonlinear interaction between comprehensive risk assessment information and candidate treatment options, the impact of different treatment options on individual patient risk can be understood more accurately and comprehensively, thus avoiding misjudgments caused by simplified models. Based on this, the constructed risk ratio function provides clinicians with intuitive and quantitative decision-making basis, making the comparison of the advantages and disadvantages of different treatment options more objective. Finally, by using the risk ratio function to perform time-domain definition and risk assessment of comprehensive risk assessment information, dynamic optimization and precise recommendation of treatment options are achieved, ensuring that the most appropriate treatment can be obtained at different stages of the patient's life cycle, significantly improving the personalization level and clinical effectiveness of cancer treatment strategies, thereby effectively improving patient prognosis.

[0090] In some embodiments, solving the nonlinear interaction relationship between comprehensive risk assessment information and candidate treatment options includes: performing a nonlinear transformation on the comprehensive risk assessment information to obtain nonlinear assessment information; constructing a weighted proportional risk relationship based on the comprehensive risk assessment information, the nonlinear assessment information, and the candidate treatment options; weighting each item in the weighted proportional risk relationship using a preset inverse processing probability weight to obtain a nonlinear interaction relationship; the inverse processing probability weight is obtained by inverse processing probability weighting of the conditional probability of historical patients receiving candidate treatment options.

[0091] Nonlinear transformation of comprehensive risk assessment information aims to map the original information to a new feature space, enabling it to better reveal potential nonlinear relationships and thus improve the model's expressive power and prediction accuracy. For example, polynomial transformations can be used to square, cube, or combine cross-terms on the original information; alternatively, kernel function transformations, such as radial basis function (RBF) kernels or polynomial kernels, can be used to map the data to a higher-dimensional space; or, activation functions in neural networks (such as ReLU, Sigmoid, and Tanh) can be used to perform nonlinear processing on the information.

[0092] Constructing a weighted proportional hazards relationship based on comprehensive risk assessment information, nonlinear assessment information, and candidate treatment options aims to establish a statistical model between a patient's overall risk and different treatment options. This model can quantify the impact of different treatment options on patient risk. Introducing nonlinear assessment information is intended to more comprehensively capture the complex interaction between risk and treatment options. For example, the Cox proportional hazards model can be extended by using comprehensive risk assessment information, nonlinear assessment information, and candidate treatment options as covariates and introducing interaction terms to construct the model. Alternatively, a generalized linear model (GLM) or a generalized additive model (GAM) can be used, linking the linear predictor to the risk rate through a link function and considering nonlinear terms.

[0093] By using pre-defined inverse processing probability weights to weight each item in the weighted proportional hazards relationship, the aim is to correct for selection bias in clinical data, so that the data from observational studies can more closely approximate the effects of randomized controlled trials, thereby more accurately estimating the true causal effect of treatment regimens.

[0094] This application's solution generates nonlinear evaluation information by performing a nonlinear transformation on comprehensive risk assessment information, thereby capturing more complex nonlinear patterns between patient risk and treatment plans. Based on this, a weighted proportional risk relationship is constructed by combining the original comprehensive risk assessment information, nonlinear evaluation information, and candidate treatment plans. This allows the model to more comprehensively characterize the patient's risk status and its interaction with treatment plans. To address the selection bias and confounding factors commonly found in clinical data, this application further utilizes pre-defined inverse processing probability weights to weight the terms in the weighted proportional risk relationship. These inverse processing probability weights are obtained by inversely processing the conditional probabilities of historical patients receiving candidate treatment plans, effectively correcting for biases caused by non-random treatment allocation. Through these steps, this application can accurately and robustly calculate the nonlinear interaction relationship between comprehensive risk assessment information and candidate treatment plans, laying a solid foundation for subsequent construction of the risk ratio function and determination of the treatment recommendation time domain and target treatment plan, thus ensuring higher accuracy and reliability of the generated cancer treatment strategy.

[0095] The following is a specific example to illustrate this. As a concrete implementation method, when calculating the nonlinear interaction between comprehensive risk assessment information and candidate treatment options, a polynomial kernel function can first be used to nonlinearly transform the comprehensive risk assessment information. For example, the original comprehensive risk assessment information can be squared, cubed, etc., and its interaction terms can be generated to obtain nonlinear assessment information. Next, based on the comprehensive risk assessment information, the nonlinear assessment information, and the candidate treatment options, an extended Cox proportional hazards model can be constructed as a weighted proportional hazards relationship. This model can include comprehensive risk assessment information, nonlinear assessment information, and linear and interaction terms of candidate treatment options. To correct for selection bias, the likelihood function of this Cox model can be weighted using preset inverse processing probability weights. For example, based on the patient's historical clinical characteristics such as age, tumor stage, and comorbidities, a logistic regression model can be constructed to predict the conditional probability of a patient receiving a specific candidate treatment option. For patients who actually receive treatment A, the inverse treatment probability weight can be calculated as `1 / P(A|clinical characteristics)`; for patients who actually receive treatment B, the inverse treatment probability weight can be calculated as `1 / (1-P(A|clinical characteristics))`. Applying these weights to the parameter estimation process of the Cox model above yields a corrected and more accurate nonlinear interaction relationship.

[0096] In one specific embodiment, the comprehensive risk assessment information is nonlinearly transformed using a restricted cubic spline function, with the position set at the 5th, 35th, 65th, and 95th percentiles of the distribution of the comprehensive risk assessment information, in order to flexibly fit the complex nonlinear relationship between risk and benefit. The expression for the constructed nonlinear interaction relationship is as follows: , in, Let i be the risk function of individual i at time t. For the baseline risk function, As a candidate treatment option, To obtain the nonlinear evaluation information, a restricted cubic spline nonlinear transformation function is used to perform a nonlinear transformation on the comprehensive risk assessment information. The main effect regression coefficients for the candidate treatment options are: The regression coefficient represents the interaction between candidate treatment options and evaluation information. Other confounding covariate factors used for correction and their corresponding coefficient vectors , , and All are calculated based on the inverse processing probability weights.

[0097] The inverse processing probability weights are obtained by inversely processing the conditional probabilities of historical patients receiving candidate treatment options. This can be achieved by first calculating a propensity score for each candidate treatment option, and then using logistic regression to calculate the nCIT for each patient. ) or nCRT( The probability of 0) Then, inverse processing probability weighting (IPTW) is performed to calculate the standardized inverse processing probability weights. : , in, This is the inverse processing of probability weights. Propensity score for accepting the target treatment plan.

[0098] Through the aforementioned technical solution, this application can more comprehensively and precisely capture the complex nonlinear patterns between patient risk and treatment plans, avoiding the simplification and information loss that may result from traditional linear models. By introducing nonlinear evaluation information and constructing a weighted proportional hazards relationship, this application can establish a more expressive model, thereby more accurately characterizing the patient's overall risk status and its interaction with different treatment plans. Crucially, by using pre-defined inverse processing probability weights to weight the weighted proportional hazards relationship, selection bias and confounding factors commonly found in clinical observation data are effectively corrected, making the calculated nonlinear interaction relationship closer to the true causal effect. This significantly improves the accuracy and reliability of the causal effect assessment of treatment plans, enabling the identification of more precise and personalized treatment recommendations and target treatment plans for target patients, ultimately generating more effective and safer cancer treatment strategies.

[0099] In some embodiments, the comprehensive risk assessment information is defined in the time domain and risk assessed based on the risk ratio function, including: solving the equation root of the risk ratio function equal to a unit value as the time domain boundary threshold; dividing the comprehensive risk assessment information into at least two treatment recommendation time domains using the time domain boundary threshold as the boundary, and determining the target treatment plan corresponding to the treatment recommendation time domain based on the value of the risk ratio function within the treatment recommendation time domain.

[0100] Solving the equation for the hazard ratio function to a unit value means mathematically finding the solution where the hazard ratio function equals 1 at a specific time point. The hazard ratio function is typically used to quantify the relative risks or benefits of different treatment options at different time points. A hazard ratio function of 1 usually indicates that the risks or benefits of two or more treatment options have reached a balance point, or that the risk / benefit curve of a particular treatment option has crossed a critical clinical threshold. The solution can be obtained using various numerical analysis methods, such as Newton's iteration method, bisection method, or secant method, to find a solution that meets the conditions within a predetermined accuracy range. Alternatively, if the hazard ratio function has an analytical expression, it can also be solved directly using algebraic methods.

[0101] Temporal boundary thresholds are one or more specific time points used to divide a patient's entire lifespan or disease progression into different stages. These thresholds are objective and data-driven, marking critical points where a patient's relative response to different treatment options or risk characteristics significantly change. In addition to directly using equation roots as thresholds, these roots can be fine-tuned using clinical experience or statistical methods, for example, adjusting them to time points consistent with common treatment cycles or disease stages, to enhance their clinical interpretability and applicability.

[0102] Using time-domain boundary thresholds as boundaries, the comprehensive risk assessment information is divided into at least two treatment recommendation time domains, dividing the patient's entire treatment timeline into multiple consecutive time periods. For example, if there is one time-domain boundary threshold, the timeline can be divided into two time domains: "before the threshold" and "after the threshold"; if there are multiple thresholds, it can be divided into more time domains. This division method ensures that each treatment recommendation time domain has relatively homogeneous risk characteristics and treatment response patterns. This division process can be implemented using simple interval judgment logic; for example, all time points less than the first threshold are grouped into one time domain, time points between two thresholds are grouped into another time domain, and so on.

[0103] Within each recommended treatment timeframe, the target treatment plan is determined based on the value of the hazard ratio function. This means that for each defined time period, the hazard ratio function values ​​of all candidate treatment plans are re-evaluated. By comparing these values, the treatment plan with the lowest risk, greatest benefit, or best overall effect for the patient within that specific timeframe can be identified. For example, the plan with the lowest hazard ratio function value can be selected as the target treatment plan, or the plan that performs best within a specific hazard ratio interval can be selected according to pre-defined clinical decision rules. This dynamic selection mechanism ensures that the treatment plan can be adjusted as the patient's condition evolves and progresses over time, thereby achieving truly personalized and phased treatment.

[0104] This application's approach, by solving the equation where the hazard ratio function equals a unit value, objectively identifies key treatment turning points in a patient's life cycle and uses them as time-domain boundary thresholds. These thresholds allow for the precise division of comprehensive risk assessment information into at least two treatment recommendation time domains, overcoming the lack of fine-grained segmentation in the time dimension of traditional treatment plans. Within each defined treatment recommendation time domain, the optimal target treatment plan for that time domain can be dynamically determined by re-evaluating the value of the hazard ratio function. This method enables cancer treatment strategies to be adjusted based on the patient's biological malignancy and relative response to treatment at different time stages, thereby providing more targeted and timely treatment recommendations.

[0105] The following is a concrete example to illustrate this. Suppose that for a specific target patient, the aforementioned steps have already constructed a risk ratio function between two candidate treatment options (Option A and Option B). Where t represents time. To determine when to switch from option A to option B, or vice versa, the solution is... The equation roots can be determined. For example, if the solution yields t = 12 months, then 12 months is the time-domain boundary threshold. Based on this threshold, the patient's treatment cycle can be divided into two recommended treatment time domains: the first domain is 0-12 months, and the second domain is after 12 months. Within the 0-12 month domain, if the value of the hazard ratio function R(t) is generally less than 1 (e.g., the risk of option A is lower than option B), then option A is determined as the target treatment option for this domain. Conversely, within the domain after 12 months, if the value of R(t) is generally greater than 1 (e.g., the risk of option A is higher than option B), then option B is determined as the target treatment option for this domain. This allows for the generation of a phased cancer treatment strategy, for example, using option A for the initial 12 months and then switching to option B. This process can be executed by a dedicated decision module that receives the hazard ratio function and threshold information and outputs the phased treatment plan.

[0106] In one specific embodiment, the risk ratio function is expressed as follows: , like Figure 5 As shown, solve The roots of the equation 1 yield two key cross thresholds: (116.5) and (145.0), thus deriving the recommended treatment time zone 1 (Low Risk, <116.5), <1, Superior efficacy (nCIT recommended), treatment recommendation is Zone 2 (Intermediate Risk, 116.5≤ ≤145.0), >1, High-risk (recommended nCRT, i.e., treatment window) and recommended treatment time zone Zone 3 (High Risk) >145.0), <1, Highly effective (nCIT recommended).

[0107] The aforementioned technical solutions enable more refined and dynamic cancer treatment strategies for target patients. This method objectively identifies key time points for treatment regimen switching and recommends the most suitable treatment plan based on the patient's risk characteristics at different time periods. This avoids the limitations that may arise from using a single treatment plan throughout the entire treatment process, thus improving the precision and effectiveness of treatment.

[0108] In some embodiments, the method for generating the cancer treatment strategy further includes: determining the contribution of each feature in the clinical pathological feature information to the comprehensive risk assessment information based on the clinicopathological feature information of the target patient using a preset contribution evaluation strategy; and generating a contribution waterfall chart to display the decision basis of the cancer treatment strategy based on the contribution of the feature information to the comprehensive risk assessment information.

[0109] Contribution assessment strategies are methods used to quantify the influence of various features in the clinicopathological information of a target patient on comprehensive risk assessment information. Their core function is to reveal the internal logic of model decision-making, thereby improving the interpretability of the generated cancer treatment strategy. One implementation approach is model-based. For example, for tree-based models (such as random forests and gradient boosting trees), contribution can be assessed by calculating the importance of features in tree splits (e.g., Gini importance, gain); for linear models, it can be measured by the magnitude of feature coefficients. Another approach is model-agnostic methods, such as interpretable AI techniques like LIME (Local Interpretable Model-agnostic Explanations) or SHAP (SHapley Additive Explanations), which calculate feature contribution by perturbing input features and observing changes in model output.

[0110] A contribution waterfall plot is a visualization tool used to visually represent the positive or negative impact of various clinicopathological features on the final comprehensive risk assessment, and the magnitude of such impacts. It clearly shows which factors increase risk and which decrease risk, thus providing transparent information for cancer treatment strategy decisions. One implementation method is to use data visualization libraries (such as Matplotlib or Seaborn in Python, or D3.js in JavaScript) to create a waterfall plot based on the calculated contribution values ​​of each feature. The plot typically starts with a baseline value and then uses a series of rising or falling bars to represent the contribution of each feature, ultimately reaching the predicted value. Another implementation method is to integrate it into existing clinical decision support systems, using a user interface module to transform the calculated contribution data into an interactive waterfall plot display, allowing doctors or patients to view detailed contribution values ​​for different features.

[0111] This application's solution, building upon the generation of cancer treatment strategies, further enhances the transparency and interpretability of the decision-making process. Specifically, after acquiring clinicopathological feature information of the target patient and obtaining comprehensive risk assessment information, a pre-defined contribution evaluation strategy is used to conduct in-depth analysis of this clinicopathological feature information. This strategy aims to quantify the impact of each specific clinicopathological feature (e.g., patient age, tumor size, gene mutation type, etc.) on the final comprehensive risk assessment information. In this way, it is possible to identify which features are key factors leading to increased or decreased patient risk. Subsequently, based on the contribution of these quantified feature information to the comprehensive risk assessment information, the system generates a contribution waterfall chart. This waterfall chart graphically and intuitively shows the specific role played by each clinicopathological feature (whether it increases or decreases risk, and the magnitude of its impact) during the change from baseline risk to the final comprehensive risk assessment information. This mechanism ensures that the generated cancer treatment strategy is no longer a "black box" decision, but rather clearly presents the underlying decision-making logic and basis. In this way, the proposed solution not only provides a treatment strategy, but also an explanation of the strategy, thereby effectively solving the problem of providing only a strategy without a basis for decision-making, and greatly improving the scientific nature of clinical decision-making and patient compliance.

[0112] The following is a concrete example. Suppose that during the generation of a cancer treatment strategy, the clinicopathological characteristics of the target patient have been obtained, and their comprehensive risk assessment information has been calculated. To provide a basis for decision-making, such as... Figure 6As shown, the SHAP (SHapley Additive ex Planations) algorithm can be used as a contribution evaluation strategy. This algorithm calculates the Shapley value for each clinicopathological feature (e.g., age, tumor stage, lymph node metastasis, specific gene mutation status, etc.) to assess the overall risk. These Shapley values ​​quantify the average marginal contribution of each feature to risk prediction across all possible combinations of features. For example, if a patient has a high tumor stage, their Shapley value may be positive and large, indicating that the feature significantly increases the patient's overall risk. Conversely, if a patient has a favorable biomarker, their Shapley value may be negative, indicating that the feature reduces risk. After calculating the Shapley values ​​for all features, a contribution waterfall plot can be generated. This waterfall plot starts with a baseline risk value (e.g., the average risk of all patients) and then displays the Shapley value for each clinicopathological feature sequentially through a series of upward or downward bars. The height of each bar represents the magnitude of the feature's contribution to risk, and the direction represents the positive or negative of the contribution. Ultimately, the sum of all feature contributions will yield comprehensive risk assessment information for the target patient. This visualization allows physicians to clearly see which features are the primary drivers of the current risk assessment results, providing strong explanation and support for subsequent cancer treatment strategies.

[0113] In a specific embodiment, the expression for determining the contribution of each feature in the clinicopathological feature information to the comprehensive risk assessment information is as follows: , in, Basic contribution The contribution of feature information to the comprehensive risk assessment information. For feature information, This refers to the number of feature information entries in the clinicopathological feature information. .

[0114] Through the aforementioned technical solution, this application can determine the contribution of each feature to the comprehensive risk assessment information based on the clinicopathological characteristics of the target patient using a pre-defined contribution evaluation strategy, and generate a contribution waterfall chart to illustrate the decision-making basis for cancer treatment strategies. This makes the generated cancer treatment strategy no longer a simple recommendation, but rather includes clear and intuitive decision-making rationale. Physicians can then explain to patients in detail why a specific treatment plan is recommended, for example, which specific clinical indicators (such as tumor size, gene mutation type, patient age, etc.) are key factors leading to the current risk assessment and treatment recommendations. This transparent decision-making process greatly enhances physicians' understanding and trust in the treatment plan, while also helping patients better understand their condition and the necessity of treatment, thereby improving patient compliance and treatment outcomes. Therefore, the solution presented in this application significantly improves the clinical interpretability and practical value of cancer treatment strategies.

[0115] Figure 7 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application.

[0116] The following reference Figure 7 To describe an electronic device 700 according to such an embodiment of the present disclosure. Figure 7 The electronic device 700 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0117] like Figure 7 As shown, the electronic device 700 is presented in the form of a general-purpose computing device. The components of the electronic device 700 may include, but are not limited to: at least one processing unit 710, at least one storage unit 720, a bus 730 connecting different system components (including storage unit 720 and processing unit 710), a display unit 740, etc.

[0118] The storage unit stores program code, which can be executed by the processing unit 710, causing the processing unit 710 to perform the steps described in the above-described method for generating cancer treatment strategies according to various exemplary embodiments of this disclosure.

[0119] Storage unit 720 may include a readable medium in the form of a volatile storage unit, such as random access memory (RAM) 7201 and / or cache memory 7202, and may further include a read-only memory (ROM) 7203.

[0120] The storage unit 720 may also include a program / utility 7204 having a set (at least one) program module 7205, such program module 7205 including but not limited to: an operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0121] Bus 730 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0122] Electronic device 700 can also communicate with one or more external devices 700' (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 700, and / or with any device that enables electronic device 700 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 750. Furthermore, electronic device 700 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 760. Network adapter 760 can communicate with other modules of electronic device 700 via bus 730. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 700, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0123] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0124] The cancer treatment strategy generation method, apparatus, and storage medium provided in this application introduce a time-domain cumulative risk assessment mechanism and a causal interaction analysis mechanism. Based on the patient's multi-dimensional clinicopathological characteristics, time-domain cumulative risk assessment and causal interaction analysis are performed to generate individualized treatment strategies. Therefore, by introducing time-domain cumulative risk assessment, the overall biological malignancy of the target patient throughout their lifespan can be dynamically characterized, rather than merely providing a static prognostic judgment. Through causal interaction analysis of this comprehensive risk assessment information and candidate treatment options, the specific impact of different treatment options on the patient's risk curve at different time periods can be directly calculated and quantified. This allows for the quantification of the patient's benefit from a specific treatment, providing direct guidance for optimizing cancer treatment strategies.

[0125] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, or network device, etc.) to execute the methods described above according to the embodiments of this disclosure.

[0126] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0127] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0128] Those skilled in the art will understand that the above modules can be distributed in the device as described in the embodiments, or they can be modified accordingly and placed in one or more devices that are unique to this embodiment. The modules in the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0129] Exemplary embodiments of this disclosure have been specifically shown and described above. It should be understood that this disclosure is not limited to the detailed structures, arrangements, or implementations described herein; rather, this disclosure is intended to cover various modifications and equivalent arrangements contained within the spirit and scope of the appended claims.

Claims

1. A method for generating a cancer treatment strategy, characterized in that, include: Obtain clinicopathological features of the target patient from multiple dimensions; A time-domain cumulative risk assessment is performed on the clinicopathological features to obtain comprehensive risk assessment information that characterizes the overall biological malignancy of the target patient over the course of life. A causal interaction analysis is performed on the comprehensive risk assessment information and candidate treatment plans to identify the treatment recommendation time domain for the target patient and the target treatment plan corresponding to the treatment recommendation time domain; A cancer treatment strategy is generated based on the treatment recommendation time domain and the target treatment plan.

2. The method for generating a cancer treatment strategy according to claim 1, characterized in that, The full-time-domain cumulative risk assessment of the clinicopathological feature information includes: Risk features are extracted from the clinicopathological features to obtain multiple cumulative risk assessment information that characterize the biological malignancy of the target patient across the entire time domain; The cumulative risk assessment information is fitted to obtain the comprehensive risk assessment information.

3. The method for generating a cancer treatment strategy according to claim 2, characterized in that, The step of extracting risk features from the clinicopathological information includes: The random survival forest method is used to select features from the clinical pathological features and obtain the cumulative risk information stored in the leaf nodes of each survival tree to obtain the cumulative risk assessment information.

4. The method for generating a cancer treatment strategy according to claim 2, characterized in that, The fitting of the cumulative risk assessment information includes: Determine the average value of each of the cumulative risk assessment information at the same point in time; The mean evaluation information is discretized into a time series of cumulative risk values. The comprehensive risk assessment information is obtained by summing the values ​​at each discrete time point in the risk accumulation value sequence.

5. The method for generating a cancer treatment strategy according to claim 1, characterized in that, The causal interaction analysis of the comprehensive risk assessment information and candidate treatment options includes: Solve the nonlinear interaction relationship between the comprehensive risk assessment information and the candidate treatment plan; Based on the aforementioned nonlinear interaction relationship, a risk ratio function is constructed among the candidate treatment options; Based on the risk ratio function, the comprehensive risk assessment information is defined in the time domain and risk assessed to determine the treatment recommendation time domain and the target treatment plan corresponding to the treatment recommendation time domain.

6. The method for generating a cancer treatment strategy according to claim 5, characterized in that, The calculation of the nonlinear interaction between the comprehensive risk assessment information and the candidate treatment plan includes: The comprehensive risk assessment information is subjected to a nonlinear transformation to obtain nonlinear assessment information; Based on the comprehensive risk assessment information, the nonlinear assessment information, and the candidate treatment plans, a weighted proportional risk relationship is constructed. The nonlinear interaction relationship is obtained by weighting each item in the weighted proportional risk relationship using preset inverse processing probability weights; the inverse processing probability weights are obtained by inverse processing probability weighting of the conditional probability of historical patients receiving the candidate treatment plan.

7. The method for generating a cancer treatment strategy according to claim 5, characterized in that, The step of defining the time domain and assessing the risk based on the risk ratio function includes: Solve the equation where the risk ratio function equals a unit value, and use the root of the equation as the time-domain boundary threshold; Using the time domain boundary threshold as the boundary, the comprehensive risk assessment information is divided into at least two treatment recommendation time domains, and the target treatment plan corresponding to the treatment recommendation time domain is determined based on the value of the risk ratio function within the treatment recommendation time domain.

8. The method for generating a cancer treatment strategy according to claim 1, characterized in that, Also includes: Based on the clinicopathological features of the target patient, a preset contribution evaluation strategy is used to determine the contribution of each feature in the clinicopathological features to the comprehensive risk assessment information. Based on the contribution of the feature information to the comprehensive risk assessment information, a contribution waterfall chart is generated to demonstrate the decision-making basis for the cancer treatment strategy.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method for generating a cancer treatment strategy according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method for generating a cancer treatment strategy according to any one of claims 1 to 8.

Images