Causal survival forest-based individualized benefit prediction system and method for postoperative radiotherapy of breast cancer

Abstract

The application discloses a breast cancer postoperative radiotherapy individualized benefit prediction system and method based on a causal survival forest, and belongs to the technical field of artificial intelligence processing of medical diagnosis data. The application relates to a data processing method based on a causal survival forest, and comprises the following steps: acquiring clinical characteristic data of a target object; inputting the clinical characteristic data into a pre-trained causal survival forest model, calculating a survival probability difference of the target object under a first intervention measure and a second intervention measure through the causal survival forest model, and obtaining an individual causal effect value; generating and outputting a classification label or evaluation information of the target object based on the individual causal effect value; wherein the construction process of the causal survival forest model comprises the following steps: based on a preset causal network structure, variables are divided into confounding factors, precision variables, intermediate variables and collision point variables, and the intermediate variables and the collision point variables are removed from a candidate variable set according to the principle of blocking confounding paths without introducing bias, so as to determine an input variable set; based on the input variable set, an integrated learning model containing multiple causal trees is constructed, and in the training process, the principle of honest estimation is adopted, the training data is divided into a first subset for constructing a tree structure and a second subset for estimating leaf node effects.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence processing technology for medical diagnostic data, specifically to a system and method for predicting individualized benefits of postoperative radiotherapy for breast cancer based on causal survival forest. Background Technology

[0002] Post-mastectomy radiotherapy (PMRT) is an important component of comprehensive breast cancer treatment. According to a meta-analysis by the EBCTCG (Early Breast Cancer Trialists' Collaborative Group), PMRT can reduce the risk of local recurrence and improve overall survival. However, existing guidelines (such as NCCN and ESMO) primarily provide treatment recommendations based on population average effects, failing to adequately consider individual patient heterogeneity.

[0003] The limitations of existing technologies include: 1. Traditional predictive models (such as Adjuvant! Online and PREDICT) only predict prognosis and fail to quantify the individualized benefits of treatment; 2. Propensity score matching methods can only estimate the population mean treatment effect (ATE) and cannot distinguish the heterogeneous treatment effects among different patients; 3. Traditional methods such as Cox regression assume proportional hazard (PH assumption), which is often violated in real-world data; 4. There is a lack of clinical decision support tools that combine causal inference with machine learning and can simultaneously handle survival outcomes and heterogeneous treatment effects. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a system and method for predicting individualized benefits of postoperative radiotherapy for breast cancer based on causal survival forest.

[0005] The objective of this invention can be achieved through the following technical solutions: A first aspect of the present invention relates to a data processing method based on causal survival forest, comprising the following steps: Obtain clinical characteristic data of the target subjects; The clinical characteristic data is input into a pre-trained causal survival forest model. The causal survival forest model is used to calculate the difference in survival probability of the target object under the first intervention and the second intervention to obtain the individual causal effect value. Based on the individual causal effect value, generate and output the classification label or evaluation information of the target object; The construction process of the causal survival forest model includes: Based on the pre-defined causal network structure, variables are classified into confounding factors, precision variables, mediating variables, and collision point variables. In accordance with the principle of blocking confounding paths without introducing bias, the mediating variables and collision point variables are removed from the candidate variable set to determine the input variable set. An ensemble learning model containing multiple causal trees is constructed based on the input variable set. During training, the honest estimation principle is adopted, and the training data is divided into a first subset for constructing the tree structure and a second subset for estimating the leaf node effect.

[0006] Optionally, before inputting the clinical feature data into the pre-trained causal survival forest model, the method further includes: A proportional hazards hypothesis test was performed on the dataset to which the clinical feature data belonged. The test included the Schoenfeld residual test and the global χ² test. The causal survival forest model is used for calculation only if the results of the proportional hazards hypothesis test indicate that the data does not conform to the proportional hazards hypothesis.

[0007] Optionally, generating and outputting the classification label of the target object based on the individual causal effect value specifically includes: A first threshold and a second threshold are preset, wherein the first threshold is 3.5% and the second threshold is 0%; The individual causal effect value is compared with the first threshold and the second threshold: If the individual causal effect value is greater than the first threshold, then a first classification label is generated; If the individual causal effect value is greater than the second threshold and less than or equal to the first threshold, then a second classification label is generated; If the individual causal effect value is less than or equal to the second threshold, a third classification label is generated; The first threshold is a combination of parameters determined based on the population average absolute survival benefit statistics, the number of people requiring treatment, and the clinical benefit scale standard.

[0008] Optionally, outputting the evaluation information of the target object includes generating a multi-dimensional visualization interface, which contains the following information elements displayed collaboratively: The point estimate of the individual causal effect value; The 95% confidence interval of the point estimate; The percentile ranking of the individual causal effect value of the target object in the pre-stored group benefit distribution data; And a population distribution histogram, wherein the histogram is marked with identifiers representing the location of the individual causal effect value of the target object.

[0009] Optionally, the clinical characteristic data includes independent T-stage and N-stage variables, but does not include the composite clinical staging variable resulting from combining T-stage and N-stage.

[0010] Optionally, in the step of performing covariate selection based on a directed acyclic graph, the mediating variable includes a local relapse state variable affected by the first intervention.

[0011] A first aspect of the present invention relates to a data processing system based on causal survival forest, comprising: The data acquisition module is used to acquire clinical characteristic data of the target object; The calculation module is used to store and run a pre-trained causal survival forest model. The clinical feature data is input into the model to calculate the difference in survival probability of the target object under the first intervention and the second intervention, and to obtain the individual causal effect value. The decision support module is used to generate classification labels for the target object based on the comparison results between the individual causal effect value and a preset threshold. The causal survival forest model running in the computation module is trained based on the input variable set after filtering and removing mediator variables and collision point variables through a directed acyclic graph, and the honest estimation principle is adopted during training.

[0012] Optionally, it also includes a visualization output module for outputting a visualization result containing the individual causal effect value and the classification label; A third aspect of the present invention relates to an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0013] A fourth aspect of the present invention relates to a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0014] The beneficial effects of this invention are: This invention overcomes the limitations of traditional survival analysis models (such as Cox regression) that rely on the proportional hazards (PH) assumption. By constructing a DAG causal graph variable screening + CSF causal survival forest architecture, it actively eliminates mediating and collision variables, effectively blocking confounding bias. Combined with the independent input strategy of T / N staging, it achieves a robust, unbiased and causally interpretable accurate estimation of the individualized causal effect (ITE) of breast cancer radiotherapy in complex real-world data.

[0015] This invention innovatively constructs a 0% / 3.5% dual-threshold decision-making system based on clinical evidence-based criteria (EBCTCG / NNT), transforming continuous predicted values ​​into "high / medium / low" benefit stratification. This system can effectively identify patients who are generally recommended by existing guidelines (such as NCCN) but actually benefit only slightly, thereby guiding high-risk patients to receive treatment while safely exempting low-benefit groups and avoiding overtreatment. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a diagram illustrating the overall architecture of the PMRT-ITE prediction system according to an embodiment of this application. Figure 2 This is a DAG diagram representing the causal inference framework of this application embodiment; Figure 3 This is a ranking chart of the importance of variables in this application; Figure 4 This is a histogram of ITE distribution in an embodiment of this application; Figure 5 This is a box plot of ITE groups according to an embodiment of this application; Figure 6 The hierarchical Kaplan-Meier survival curves are shown in the embodiments of this application. Figure 7 Decision curve analysis (DCA) for embodiments of this application; Figure 8 This is the interface of the treatment decision support tool in the embodiments of this application; Figure 9 This is a distribution diagram of PMRT benefit in the treatment decision-making process according to an embodiment of this application; Figure 10 A heatmap comparing the embodiments of this application with the NCCN guidelines; Figure 11 The external verification queue (ITE) distribution is shown in this application embodiment. Figure 12 This is the architecture of the PMRT-ITE personalized benefit prediction system proposed in this application. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention employs the Causal Survival Forest (CSF) algorithm, based on large-scale real-world data (SEER database), to construct a personalized PMRT treatment effect prediction model. Key innovations include: 1. Establishment of a causal inference framework This invention first constructs a directed acyclic graph (DAG) based on domain knowledge, and explicitly defines: Exposure variable (A): Whether or not PMRT is accepted Outcome variable (Y): 5-year overall survival rate Covariate set (Z): 14 clinicopathological variables The selection of these 14 variables follows the core principle of causal inference—blocking confounding paths without introducing bias: 1. True confounding factors (which simultaneously affect PMRT decision-making and survival outcomes): age, marital status, T stage, N stage, pathological N stage, ER / PR / HER2 status, histological grade, and histological type; 2. Precision variable (affects only the outcome, improves estimation accuracy): affected side; 3. Variables related to treatment decision: chemotherapy status, treatment response.

[0020] Key exclusion: • Mediating variables: Variables that may be affected by PMRT (such as local recurrence status) were not included to avoid over-adjustment. • Collision point variables: Variables simultaneously affected by both A and Y are excluded to avoid introducing collision bias. • Composite variables: Explicitly use T-stage and N-stage, rather than composite clinical stages, to preserve the interpretability of the causal pathway. 2. The overall system architecture of the present invention is as follows: Figure 1 As shown, it includes seven functional modules: ① Data input module: Receives 14 clinical pathology variables; ② Data preprocessing module: integrity check, missing value handling, encoding conversion; The specific implementation of the data preprocessing module includes: (a) Missing value handling: Complete case analysis was used, meaning that during the data quality screening phase, samples with missing values ​​in any modeling variable were excluded without any imputation. This ensured that all samples input into the model were complete across all 14 covariates.

[0021] (b) Encoding Conversion: All encodings use Ordinal Integer Encoding and Binary Encoding; One-Hot Encoding is not used. Specific encoding rules are as follows: Binary variables (0 / 1 coding): marital status (married=1, others=0), histological type (invasive ductal carcinoma IDC=1, others=0), affected side (right side=1, others=0), ER status (positive=1, negative=0), PR status (positive=1, negative=0), HER2 status (positive=1, negative=0), adjuvant chemotherapy (yes=1, no=0).

[0022] Ordinal variables (integer sequence coding): age group (1=<40 years, 2=40-49 years, 3=50-59 years, 4=60-69 years, 5=≥70 years), primary site (0=other / central, 1=medial quadrant, 2=lateral quadrant), histological grade (1=grade I, 2=grade II, 3=grade III, 4=grade IV), T stage (1=T1, 2=T2, 3=T3, 4=T4), clinical N stage (0=N0, 1=N1, 2=N2, 3=N3), pathological N stage (0=pN0, 1=pN1, 2=pN2, 3=pN3), neoadjuvant therapy response (0=no response NR, 1=partial remission PR, 2=pathological complete remission pCR).

[0023] ③ PH Hypothesis Testing Module: Schoenfeld Residual Test, verifying the necessity of the CSF method; ④ CSF modeling module: 2500 causal trees, based on the principle of honest estimation; The specific implementation parameters of the causal survival forest model are as follows: (a) Forest size: Contains 2500 causal survival trees (num.trees=2500); (b) Subsampling method: Perform independent subsampling without replacement for each tree, and randomly select 30% of the samples in the training set for each tree (sample.fraction=0.3). (c) Honest estimation principle: Enable honest estimation (honesty=TRUE), further divide the subsamples of each tree into two disjoint subsets: 60% is used to build the tree structure (determine the split point), and 40% is used to estimate the causal effect of the leaf nodes (honesty.fraction=0.6). (d) Minimum sample size for leaf nodes: set to 20 (min.node.size=20) to ensure that each leaf node has enough samples for causal effect estimation; (e) Number of candidate variables for each split: Each time a tree node splits, 8 variables are randomly selected from 14 covariates as candidate split variables (mtry=8). (f) Prediction time horizon: set to 60 months (horizon=60), which means predicting the 5-year survival probability; the prediction target is the survival probability (target="survival.probability"). (g) Training strategy: Divide the training data into training set, tuning set and validation set in a ratio of 60% / 20% / 20%; after performing hyperparameter search on the training set, merge the training set and tuning set (total 80%) as the final training set, train the final model with the frozen optimal hyperparameters, and perform internal validation on the remaining 20% ​​independent validation set.

[0024] ⑤ ITE prediction module: Calculate τ(x) and 95% confidence interval; ⑥ Clinical decision support module: benefit grouping and treatment recommendations; ⑦ Web application interface module: Shiny interactive tool.

[0025] 3. Causal Inference Framework (DAG) This invention clarifies causal relationships based on directed acyclic graphs (DAGs), such as... Figure 2 As shown.

[0026] 4. Input variables (14) The system accepts the following 14 clinicopathological variables as input: 5. Importance analysis of variables The model assesses the contribution of each variable to ITE prediction by permutation importance, such as... Figure 3 As shown.

[0027] 6. ITE Calculation and Clinical Grouping The Individualized Therapy Effect (ITE) is defined as: τ(x) = E[Y(1) - Y(0) | X = x] Where Y(1) is the 5-year survival rate of those who received PMRT and Y(0) is the 5-year survival rate of those who did not receive PMRT.

[0028] Based on clinical evidence and statistical considerations, this invention uses a dual threshold of 0% / 3.5% to divide patients into three groups: • High Benefit Group: ITE>3.5%, PMRT is strongly recommended.

[0029] • Moderate Benefit (Moderate Benefit Group): 0% < ITE ≤ 3.5%, comprehensive evaluation is recommended.

[0030] • Low Benefit (Low Benefit Group): ITE ≤ 0%, PMRT exemption can be considered.

[0031] Clinical Basis for Threshold Setting The selection of the 3.5% threshold is based on the following multi-dimensional evidence: 1. Evidence from EBCTCG meta-analysis: The 2014 EBCTCG meta-analysis of 22 RCTs showed that the absolute survival benefit of PMRT was approximately 2 - 4% (PMID: 24656685).

[0032] 2. Clinical decision threshold: Classic clinical decision tools (such as Adjuvant! Online) show that most clinicians tend to recommend adjuvant treatment when the expected benefit is ≥ 4 - 5%.

[0033] 3. Consideration of NNT (Number Needed to Treat): The absolute benefit of 3.5% corresponds to NNT ≈ 29, meaning that treating 29 patients can save 1 additional life.

[0034] 4. ESMO-MCBS framework: According to the ESMO-Magnitude of Clinical Benefit Scale, an absolute survival benefit of 3.5% represents a clinically meaningful improvement.

[0035] 8 Model Validation Results The present invention verified the model performance on three independent data sets: 9 Survival Curve Validation As Figure 6 shown, the stratified Kaplan-Meier survival analysis confirmed the clinical effectiveness of the ITE grouping.

[0036] 10 Decision Curve Analysis Decision Curve Analysis (DCA) shows that the present model is superior to the 'all treat' and 'all not treat' strategies within a wide range of threshold probabilities.

[0037] 11 Clinical Decision Support Tool The present invention developed a web-based interactive clinical decision support tool (Shiny application), the interface is as Figure 8 shown.

[0038] 12 Comparison with Existing Guidelines The analysis results of the above embodiments, compared with the NCCN guidelines, show that the present invention can further refine patient stratification based on the guidelines' recommendations.

[0039] 13 External Validation The distribution of ITE in the independent external validation queue is as follows: Figure 11 As shown, the model's generalization ability is confirmed.

[0040] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0041] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A data processing method based on causal survival forest, characterized in that, Includes the following steps: Obtain clinical characteristic data of the target subjects; The clinical characteristic data is input into a pre-trained causal survival forest model. The causal survival forest model is used to calculate the difference in survival probability of the target object under the first intervention and the second intervention to obtain the individual causal effect value. Based on the individual causal effect value, generate and output the classification label or evaluation information of the target object; The construction process of the causal survival forest model includes: Based on the pre-defined causal network structure, variables are classified into confounding factors, precision variables, mediating variables, and collision point variables. In accordance with the principle of blocking confounding paths without introducing bias, the mediating variables and collision point variables are removed from the candidate variable set to determine the input variable set. An ensemble learning model containing multiple causal trees is constructed based on the input variable set. During training, the honest estimation principle is adopted, and the training data is divided into a first subset for constructing the tree structure and a second subset for estimating the leaf node effect.

2. The data processing method based on causal survival forest according to claim 1, characterized in that, Before inputting the clinical feature data into the pre-trained causal survival forest model, the method further includes: A proportional hazards hypothesis test was performed on the dataset to which the clinical feature data belonged. The test included the Schoenfeld residual test and the global χ² test. The causal survival forest model is used for calculation only if the results of the proportional hazards hypothesis test indicate that the data does not conform to the proportional hazards hypothesis.

3. The data processing method based on causal survival forest according to claim 1, characterized in that, The process of generating and outputting classification labels for the target object based on the individual causal effect value specifically includes: A first threshold and a second threshold are preset, wherein the first threshold is 3.5% and the second threshold is 0%; The individual causal effect value is compared with the first threshold and the second threshold: If the individual causal effect value is greater than the first threshold, then a first classification label is generated; If the individual causal effect value is greater than the second threshold and less than or equal to the first threshold, then a second classification label is generated; If the individual causal effect value is less than or equal to the second threshold, a third classification label is generated; The first threshold is a combination of parameters determined based on the population average absolute survival benefit statistics, the number of people requiring treatment, and the clinical benefit scale standard.

4. The data processing method based on causal survival forest according to claim 1, characterized in that, The output of the evaluation information of the target object includes generating a multi-dimensional visualization interface, which contains the following information elements displayed collaboratively: The point estimate of the individual causal effect value; The 95% confidence interval of the point estimate; The percentile ranking of the individual causal effect value of the target object in the pre-stored group benefit distribution data; And a population distribution histogram, wherein the histogram is marked with identifiers representing the location of the individual causal effect value of the target object.

5. The data processing method based on causal survival forest according to claim 1, characterized in that, The clinical characteristic data includes independent T-stage and N-stage variables, but does not include the composite clinical staging variable formed by combining T-stage and N-stage.

6. The data processing method based on causal survival forest according to claim 1, characterized in that, In the covariate selection step based on the directed acyclic graph, the mediating variable includes the local relapse state variable affected by the first intervention.

7. A data processing system based on causal survival forest, characterized in that, include: The data acquisition module is used to acquire clinical characteristic data of the target object; The calculation module is used to store and run a pre-trained causal survival forest model. The clinical feature data is input into the model to calculate the difference in survival probability of the target object under the first intervention and the second intervention, and to obtain the individual causal effect value. The decision support module is used to generate classification labels for the target object based on the comparison results between the individual causal effect value and a preset threshold. The causal survival forest model running in the computation module is trained based on the input variable set after filtering and removing mediator variables and collision point variables through a directed acyclic graph, and the honest estimation principle is adopted during training.

8. The data processing system based on causal survival forest according to claim 7, characterized in that, It also includes a visualization output module for outputting visualization results containing the individual causal effect value and the classification label.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 6.

Images