Liver cancer prognosis risk assessment method and system based on data analysis

Abstract

The invention discloses a liver cancer prognosis risk assessment method and system based on data analysis, and relates to the technical field of intelligent medical treatment. The method is realized through the following steps: firstly, acquiring and standardizing clinical, image, genome and other multi-modal data of a patient; then constructing an individualized dynamic causal graph based on the medical knowledge base, and dynamically adjusting a causal edge weight by using a graph attention network; inputting the causal diagram into a prognosis evaluation model to obtain a basic risk trajectory, performing intervention simulation by modifying a treatment node state, and calculating an anti-factual risk trajectory to quantify an individual treatment effect; and finally, outputting a risk assessment result and generating natural language interpretation containing the quantitative contribution degree. The corresponding system comprises a data sensing layer, a causal calculation layer, an intelligent decision-making layer and an interactive presentation layer. According to the method, causal reasoning and deep learning are creatively combined, and dynamic, accurate and interpretable prognosis evaluation and intervention guidance of the liver cancer are realized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent medical treatment, and particularly relates to a liver cancer prognosis risk assessment method and system based on data analysis. BACKGROUND

[0002] Liver cancer is one of the malignant tumors with high morbidity and mortality worldwide. Accurate prognosis assessment is of great significance for developing personalized treatment plans and improving patient survival rates. Currently, the following methods are mainly used for liver cancer prognosis assessment in clinical practice: Traditional clinical staging systems: such as BCLC staging, TNM staging, etc., mainly based on tumor size, number, vascular invasion, and other clinical pathological features. Although these methods are simple and easy to use, they cannot fully utilize multi-modal medical data and lack the ability to predict individualized treatment outcomes.

[0003] Statistical prediction models: such as Cox proportional hazards model, based on historical data to establish statistical regularities. This type of method assumes a linear relationship, making it difficult to handle high-dimensional, non-linear multi-modal data and unable to quantify the intervention effect.

[0004] Machine learning methods: such as random forests, support vector machines, etc., can handle complex data relationships, but are mostly based on data association rather than causal relationships, are easily influenced by confounding factors, and lack interpretability.

[0005] Deep learning models: such as convolutional neural networks, recurrent neural networks, etc., can automatically learn feature representations, but their "black box" nature makes it difficult for doctors to understand and trust the prediction results, and cannot answer the key clinical question of "how would the prognosis change if a different treatment plan was adopted".

[0006] The existing technology mainly has the following defects: unable to effectively integrate multi-modal time-series data (clinical, imaging, genomic), based on association rather than causality, prediction results easily disturbed by confounding factors, lack of quantitative assessment ability of intervention effect, poor model interpretability, difficult to gain the trust of clinicians, and insufficient cross-center generalization ability.

[0007] Therefore, there is an urgent need for a liver cancer prognosis risk assessment scheme that can integrate multi-modal data, be based on causal reasoning, have intervention modeling capability, and have strong interpretability. SUMMARY

[0008] The application aims to provide a liver cancer prognosis risk assessment method and system based on data analysis, which integrates multi-modal data to construct an individualized dynamic causal graph, combines advanced machine learning algorithms and attribution methods, realizes accurate assessment of liver cancer prognosis risk, quantitative simulation of treatment effect, and interpretable output of assessment results, and provides scientific and reliable basis for clinicians to develop personalized treatment plans.

[0009] To solve the above technical problems, the application is realized by the following technical solutions: The application is a liver cancer prognosis risk assessment method based on data analysis, comprising the following steps: S1. Data acquisition and standardization: acquiring time-series multi-modal data of a target patient, including clinical data, imaging data and genomics data; standardizing the data to obtain a standardized feature vector; S2. Individualized dynamic causal graph construction: based on a preset medical causal graph template and the standardized feature vector, an individualized dynamic causal graph is constructed; wherein the standardized feature vector is used as a node feature, a graph attention network is used to calculate the attention coefficient between nodes, and the weight of each causal edge in the causal graph is dynamically adjusted according to the attention coefficient; S3. Risk assessment and intervention simulation: inputting the individualized dynamic causal graph into a prognosis risk assessment model to obtain a basic risk trajectory; based on a treatment effect knowledge base, modifying the state of an intervenable node in the causal graph corresponding to a specific treatment option, while fixing the state of a confounder node, using the prognosis risk assessment model to calculate counterfactual risk trajectories, and generating individualized treatment effect assessment; S4. Result generation and interpretation: outputting the basic risk trajectory, counterfactual risk trajectory and individualized treatment effect assessment, and generating a quantitative causal path explanation based on an attribution algorithm, and converting the quantitative results into natural language descriptions containing contribution information.

[0010] As a preferred technical solution of the application, the data standardization processing in step S1 comprises: Based on the HL7FHIR standard, the clinical data is standardized in terms of terminology and format; Quantitative imaging features are extracted from DICOM format images and corrected for inter-device differences; Genomics data is converted into pathway activity scores through standard bioinformatics processes.

[0011] As a preferred technical solution of the present application, the medical causal graph template and the treatment effect knowledge base in step S2 are derived from a medical knowledge base; and the pre-defined association between the treatment options and the biomarker changes in the treatment effect knowledge base is a probability distribution or a confidence interval obtained based on meta-analysis or Bayesian statistics on real world data or clinical research literature.

[0012] As a preferred technical solution of the present application, the step S2 uses a graph attention network to calculate the attention coefficient and adjust the edge weight, specifically including: For each pair of adjacent nodes in the causal graph, the feature vectors thereof are spliced or subjected to an inner product operation; The operation result is input into a trainable feedforward neural network, and a LeakyReLU activation function is applied; The obtained result is subjected to softmax normalization to obtain the attention coefficient representing the causal edge weight.

[0013] According to the attention coefficient, the features of the neighborhood nodes are weighted and summed to update the representation of the current node.

[0014] As a preferred technical solution of the present application, the prognosis risk assessment model in step S3 is a graph neural network model pre-trained using multiple medical center data processed by a domain adaptation technology; and the domain adaptation training specifically adopts a domain adversarial neural network or a domain alignment method based on maximum mean discrepancy.

[0015] As a preferred technical solution of the present application, the fixed confounding factor nodes in step S3 include age and liver function Child-Pugh classification nodes, and the fixing is achieved by keeping the states of the nodes unchanged in the counterfactual reasoning process.

[0016] As a preferred technical solution of the present application, the step S4 adopts a Shapley additive explanation method to calculate the marginal contribution of each node and edge in the graph to the final risk assessment result, and quantifies the contribution degree of the key causal path.

[0017] A liver cancer prognosis risk assessment system based on data analysis is used to implement the above method, and includes: A data perception and fusion layer including a multi-source data access module and a standardization and feature engineering module; A causal calculation and model layer including a medical knowledge base, an individualized causal graph construction module, and a prognosis risk assessment model; An intelligent decision and application layer including a dynamic risk assessment module, an intervention simulation and effect quantification module, and an explainability generation module; An interactive presentation layer including a visual decision support interface.

[0018] As a preferred technical solution of the present application, the visual decision support interface comprises: a risk cockpit view for displaying a base risk trajectory and a counterfactual risk trajectory in contrast; a treatment regimen comparison view for visualizing individualized treatment effect values of different treatment options; a causal path explanation view for interactively displaying a key causal path and its quantified contribution, and synchronously displaying a natural language description generated by the system to explain the clinical significance of the path.

[0019] The present application has the following beneficial effects: Improved assessment accuracy: By integrating multi-modal data and performing standardized processing, various types of information in patient data are fully utilized. Combined with individualized dynamic causal graphs and advanced graph neural network models, the causal relationship between individual characteristics and prognosis risk of liver cancer patients can be more accurately captured, significantly improving the accuracy of prognosis risk assessment compared to traditional methods, and providing more reliable basis for clinical diagnosis. Personalized treatment support: intervention simulation based on treatment effect knowledge base can quantify the impact of different treatment options on patient prognosis risk, providing scientific quantitative reference for clinicians to develop personalized treatment plans, which helps to improve the pertinence and effectiveness of treatment plans and improve the treatment effect and quality of life of patients. Enhanced explainability: Shapley additive explanation method is used to generate quantified causal path explanation and convert it into natural language description, making the complex evaluation results have good explainability, which is convenient for doctors to understand the decision basis of the evaluation model, and also helps to explain the patient's condition and treatment plan to the patient, enhancing the communication and trust between doctors and patients.

[0020] Of course, implementing any product of the present application does not necessarily require achieving all the advantages described above at the same time. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0022] Figure 1 The flowchart of the method of the present application; Figure 2 The structural schematic diagram of the system in the present application; Figure 3 The flowchart of data acquisition and standardization in the present application; Figure 4 The logic diagram of the construction of individualized dynamic causal graph in the present application; Figure 5 Risk assessment and intervention simulation flowchart in the present application; Figure 6 Visual decision support interface layout diagram in the present application. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0024] In the description of the present application, it should be understood that the terms "opening", "upper", "lower", "thickness", "top", "middle", "length", "inner", "periphery" and the like indicate the orientation or positional relationship, which are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the indicated component or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0025] The present application is a liver cancer prognosis risk assessment method based on data analysis, which specifically comprises the following steps S1. Data acquisition and standardization Multi-source data acquisition: through a multi-source data access module, time-series multi-modal data of a target patient is acquired, covering clinical data, imaging data and genomics data. The clinical data includes the basic information of the patient (such as age, gender), medical history, symptom performance, laboratory examination results (such as alpha-fetal protein AFP level, liver function indicators) and various records during the treatment process, etc.; the imaging data is mainly the DICOM format liver CT or MRI image; the genomics data includes the patient's gene sequencing raw data, such as gene mutation information and gene expression profile.

[0026] Standardization processing: Clinical data standardization: based on the HL7FHIR standard, comprehensive terminology unification and format standardization processing is performed on the clinical data, to ensure that the clinical data of different sources and different formats have consistency and comparability, facilitating subsequent data integration and analysis.

[0027] Imaging data processing: For DICOM format images, first extract quantitative imaging features, including texture features (such as gray level co-occurrence matrix, gray level run length matrix related parameters), shape features (such as tumor volume, perimeter, diameter, etc.) and enhancement mode parameters (such as the degree of enhancement of enhanced scan, enhancement time-density curve features) and so on. Then, using professional inter-device difference correction algorithm, eliminate the data deviation caused by different imaging devices, imaging parameter differences, ensure the accuracy and reliability of the imaging data.

[0028] Genomics data transformation: Through standard bioinformatics process, the original genomics data is deeply analyzed and processed, and is transformed into pathway activity score with biological significance. For example, using gene enrichment analysis and other methods, determine the biological pathways closely related to the occurrence and development of liver cancer (such as Wnt / β-catenin pathway, Hippo pathway, etc.), and calculate the activity level of each pathway, which is used as the quantitative representation of genomics data. Finally, after the above processing, the standardized feature vector is obtained, which provides the basis data for subsequent analysis.

[0029] S2. Individualized dynamic causal graph construction Template and data combination: Based on the preset medical causal graph template and standardized feature vector, an individualized dynamic causal graph is constructed. The medical causal graph template is derived from a medical knowledge base that integrates a large number of clinical guidelines, medical research literature and biological pathway databases, which contains key medical entities related to liver cancer (such as tumor size, tumor stage, gene mutation type, treatment method, etc.) and their causal relationships, providing a basic framework for the construction of individualized causal graph. The standardized feature vector is used as the node feature and is integrated into the causal graph template, so that the causal graph can reflect the individual characteristics of the target patient.

[0030] Graph attention network adjusts weights: Use the graph attention network (GAT) to calculate the attention coefficient between nodes to dynamically adjust the weights of the causal edges in the causal graph. The specific process is as follows: For each pair of adjacent nodes in the causal graph, concatenate or inner product their feature vectors to fuse the information of the two nodes and get a comprehensive feature representation.

[0031] Input the above operation results into a trainable feedforward neural network and apply LeakyReLU activation function to perform nonlinear transformation on the features, enhance the expression ability of the model, and get the preliminary attention score.

[0032] Perform softmax normalization on the obtained results to make the attention score satisfy the probability distribution among adjacent nodes, so as to get the attention coefficient representing the weight of the causal edge, which reflects the importance of the causal relationship between the two nodes.

[0033] According to the calculated attention coefficient, the features of the neighborhood nodes are weighted and summed to update the representation of the current node. In this way, individualized adaptation of causal relationships is achieved, enabling the constructed causal graph to more accurately reflect the causal relationships between individual factors of the target patient.

[0034] S3. Risk assessment and intervention simulation Basic risk assessment: input the individualized dynamic causal graph into the prognosis risk assessment model, which is a graph neural network model pre-trained using multiple medical center data processed by domain adaptation technology. Domain adaptation training specifically uses Domain Adversarial Neural Network (DANN) or Maximum Mean Discrepancy (MMD) based domain alignment methods, effectively addressing the problem of large differences in data distribution between different medical centers and improving the model's generalization ability. By analyzing and calculating the individualized dynamic causal graph through this model, the basic risk trajectory is obtained, which is the trend prediction of the patient's prognosis risk over time without receiving specific treatment intervention under the current state.

[0035] Intervention simulation and effect quantification: modify the state of the interventionable node corresponding to the specific treatment option in the causal graph based on the treatment effect knowledge base. The pre-defined association between treatment options and biomarker changes in the treatment effect knowledge base is based on the probability distribution or confidence interval obtained through meta-analysis or Bayesian statistics on a large amount of real-world data or clinical research literature, ensuring the scientificity and reliability of the association. While modifying the state of the interventionable node, the state of the confounder node (such as the age and liver function Child-Pugh classification node) is fixed, and by keeping the state of these nodes unchanged during the counterfactual reasoning process, the interference of confounding factors on treatment effect assessment is excluded. Then, the counterfactual risk trajectory is calculated again using the prognosis risk assessment model after modifying the node state, which is the trend prediction of the patient's prognosis risk over time after receiving specific treatment intervention. Finally, based on the basic risk trajectory and the counterfactual risk trajectory, individualized treatment effect assessment is generated to quantify the impact of specific treatment options on the patient's prognosis risk.

[0036] S4. Result generation and interpretation Result output: through the output module of the system, the basic risk trajectory, counterfactual risk trajectory and individualized treatment effect assessment results are output, providing clinicians with intuitive patient prognosis risk assessment and treatment effect prediction information.

[0037] Causal path explanation: Shapley Additive Explanations (SHAP) is used as the attribution algorithm to quantify the contribution of key causal paths by accurately calculating the marginal contribution of each node and edge in the graph to the final risk assessment result. The quantification results are further converted into natural language descriptions containing contribution information, such as "patient age (contribution 30%) and tumor size (contribution 25%) are the main factors leading to high prognosis risk, while the use of immunotherapy can reduce the risk by about 20% (based on individualized treatment effect evaluation)", to explain the internal causal mechanism of the assessment results to doctors and patients in an easy-to-understand way, and to improve the explainability and clinical practicability of the assessment results.

[0038] The application also provides a liver cancer prognosis risk assessment system based on data analysis, which includes the following four levels: Data perception and fusion layer: including multi-source data access module and standardization and feature engineering module. The multi-source data access module is responsible for obtaining the clinical data, imaging data and genomics data of patients from multiple data sources such as hospital information system (HIS), picture archiving and communication system (PACS) and gene sequencing platform. The standardization and feature engineering module processes the obtained data according to the standardization processing method in S1 step to generate a standardized feature vector.

[0039] Causal calculation and model layer: including medical knowledge base, individualized causal graph construction module and prognosis risk assessment model. The medical knowledge base stores rich medical knowledge to provide basis for individualized causal graph construction and treatment effect evaluation. The individualized causal graph construction module constructs an individualized dynamic causal graph based on the medical causal graph template and the standardized feature vector using the graph attention network according to S2 step. The prognosis risk assessment model uses a graph neural network model pre-trained on multi-center data to perform basic risk assessment and risk calculation after intervention simulation according to S3 step.

[0040] Intelligent decision and application layer: including dynamic risk assessment module, intervention simulation and effect quantification module and explainability generation module. The dynamic risk assessment module inputs the individualized dynamic causal graph into the prognosis risk assessment model to obtain the basic risk trajectory. The intervention simulation and effect quantification module performs intervention simulation and individualized treatment effect evaluation based on the treatment effect knowledge base and the pre-set treatment options. The explainability generation module uses Shapley Additive Explanations (SHAP) to explain and quantify the causal path of the assessment results.

[0041] Interactive presentation layer: contains a visual decision support interface, which includes risk cockpit view, treatment plan comparison view and causal path explanation view. The risk cockpit view is used to display the base risk trajectory and counterfactual risk trajectory in an intuitive chart form, showing the patient's prognosis risk trend under different conditions; the treatment plan comparison view is used to visualize the individualized treatment effect value of different treatment options, helping doctors quickly compare the effect difference of different treatment plans; the causal path explanation view is used to interactively display the key causal path and its quantitative contribution, and simultaneously display the natural language description generated by the system to explain the clinical significance of the path, making it easy for doctors to understand the internal reasons of the evaluation results.

[0042] The specific implementation steps of the method are as follows: Data acquisition and preprocessing: The multi-source data access module obtains the time series multi-modal data of the patient from various information systems and data platforms in the hospital, and the standardization and feature engineering module immediately performs standardization processing on the data to generate standardized feature vectors, which are transmitted to the causal calculation and model layer.

[0043] Individualized dynamic causal graph construction: The individualized causal graph construction module receives the standardized feature vectors, combines the medical causal graph templates in the medical knowledge base, calculates the attention coefficients between nodes using graph attention network, dynamically adjusts the weights of each edge in the causal graph, and completes the construction of the individualized dynamic causal graph.

[0044] Risk assessment and intervention simulation: The dynamic risk assessment module inputs the individualized dynamic causal graph into the prognosis risk assessment model to calculate the base risk trajectory. The doctor selects a specific treatment option according to the patient's condition and clinical needs in the intervention simulation and effect quantification module, which modifies the state of the corresponding intervention node in the causal graph based on the treatment effect knowledge base, fixes the state of the confounder node, and again uses the prognosis risk assessment model to calculate the counterfactual risk trajectory and generate individualized treatment effect assessment.

[0045] Result presentation and explanation: The visual decision support interface of the interactive presentation layer receives the base risk trajectory, counterfactual risk trajectory, individualized treatment effect assessment and causal path explanation results, and displays them to the clinician in an intuitive and easy-to-understand manner through different views, assisting the doctor in making clinical decisions.

[0046] One specific application of this embodiment is: S1, data acquisition and standardization (I) Multi-source data acquisition In a clinical application scenario in a certain third-grade hospital, the multi-source data access module obtains the clinical data of patient A through the hospital information system (HIS), including age 65, male, hepatitis B history 20 years, alpha-fetoprotein (AFP) baseline value 500 ng / mL, liver function Child-Pugh classification B level, etc.; obtains the liver enhanced CT image (DICOM format, layer thickness 1 mm, matrix 512x512) of patient A from the picture archiving and communication system (PACS); obtains the genomics data of patient A through the gene sequencing platform, and the original data is the FASTQ file obtained by Illumina HiSeq sequencing, which contains whole exome sequencing information.

[0047] (II) Standardization processing Clinical data standardization: using the fhir.resources library of Python, according to the HL7 FHIR R5 standard, the clinical data of patient A is converted into a unified resource format. For example, "hepatitis B history 20 years" is converted into a history resource conforming to the FHIR standard, and the AFP value, liver function classification and other data are standardized into standard terms and formats to ensure the consistency and comparability of the data.

[0048] Imaging data processing: Feature extraction: use the PyRadiomics library to extract radiomics features from the CT image. Set the parameters as follows: when extracting texture features, set the offset distance of gray level co-occurrence matrix (GLCM) to [1,2], and the directions include 0°, 45°, 90°, 135°; the gray level run length matrix (GLRLM) adopts the fixed bin width discretization method, which discretizes the original CT image (12-bit depth, 4096 gray levels) into 64 gray levels, and the run length range is [1,10]. A total of 120 radiomics features including entropy, contrast, energy, etc. are extracted.

[0049] Difference correction: use the device difference correction algorithm based on CycleGAN. First, collect liver cancer images from 3 different hospitals using different CT devices as training data, and train the CycleGAN model. Input the CT image of patient A into the trained model, and output the corrected image to eliminate the data bias caused by device difference.

[0050] Genomics data transformation: Standard bioinformatics pipelines were executed by Snakemake workflow management system. FASTQ files were aligned to human reference genome GRCh38 using BWA software, and variant calling was performed using GATK to obtain genetic mutation information. From KEGG, Hallmark, and other authoritative databases, we screened a number of biological pathways that were significantly related to the occurrence and development of liver cancer (screening criteria: confirmed to be closely related to liver cancer in at least 3 high-impact literature, and the sample size of related research in public databases was greater than 500 cases). Using the MsigDB database and GSVA algorithm, we calculated the activity scores of these key signaling pathways. Finally, we obtained a number of key signaling pathway activity scores related to liver cancer, such as the Wnt / β-catenin pathway activity score of 0.75 (value range 0-1, the higher the score, the stronger the pathway activity).

[0051] After the above processing, the clinical data, the corrected imaging features, and the genomics pathway activity scores were combined to generate a standardized feature vector as input for subsequent steps.

[0052] S2, Individualized dynamic causal graph construction (I) Template and data combination The pre-stored medical causal graph template in the medical knowledge base contains 100 key medical entity nodes (such as tumor size, tumor stage, gene mutation type, treatment method, etc.) and 200 causal edges (such as "tumor size → prognosis risk" and "gene mutation type → treatment response"). The standardized feature vector of patient A is integrated into the causal graph template as node features. For example, the tumor size of patient A (tumor volume is 50 cm 3 , AFP value, and Wnt / β-catenin pathway activity score are mapped to the corresponding nodes in the causal graph.

[0053] (II) Graph attention network weight adjustment The graph attention network (GAT) is implemented using the PyTorch-Geometric library. For each pair of adjacent nodes in the causal graph, such as the "tumor size" node and the "prognosis risk" node: Feature fusion: The feature vectors of the two nodes are concatenated. Assuming that the "tumor size" node feature vector has a dimension of 10 and the "prognosis risk" node feature vector has a dimension of 10, the combined feature representation has a dimension of 20.

[0054] Nonlinear transformation: The combined feature representation is input into a trainable feedforward neural network (containing 1 hidden layer with 16 hidden neurons) and a LeakyReLU activation function (negative slope of 0.2) is applied to obtain the preliminary attention score.

[0055] Normalization: The preliminary attention scores are normalized by softmax to obtain the attention coefficients representing the causal edge weights. For example, the attention coefficient of the "tumor size" node pointing to the "prognosis risk" node is calculated to be 0.8, indicating that the tumor size has a higher impact on the prognosis risk.

[0056] Node update: According to the calculated attention coefficients, the features of the neighboring nodes are weighted and summed to update the representation of the current node. Through the stacking of multiple GAT layers (3 layers in this embodiment), individualized adaptation of causal relationships is achieved, and the construction of individualized dynamic causal graphs is completed.

[0057] S3, risk assessment and intervention simulation (I) Basic risk assessment The prognosis risk assessment model uses a graph neural network model built based on the PyTorch framework. The model is pre-trained using data from 5 medical centers of liver cancer patients processed by domain adversarial neural network (DANN). In the data preparation stage, a total of 3000 cases of liver cancer patients from 5 medical centers provided multi-modal data (data volume of each medical center is between 500-700 cases). In the training process, 5-fold cross-validation is used to optimize the model parameters. After DANN processing, the cross-center prediction AUC of the model on the independent test set fluctuates from ±0.1 to ±0.03, significantly improving the generalization ability of the model.

[0058] The individualized dynamic causal graph of patient A is input into the model, and the model calculates the prognosis risk of patient A in the next 12 months through forward propagation, obtaining the recurrence risk probability of patient A from 0.6 at baseline to gradually fluctuate, rising to 0.92 at the 12th month. During this period, due to the influence of physical condition and potential factors, the risk value shows reasonable fluctuations, such as decreasing to 0.62 at the 5th month and then gradually increasing.

[0059] (II) Intervention simulation and effect quantification The treatment effect knowledge base predefines the association between treatment options such as surgical resection, targeted therapy (sorafenib), immunotherapy (pembrolizumab), and biomarker changes. Taking targeted therapy as an example, through Bayesian statistical analysis of real-world data, the probability distribution of sorafenib treatment and changes in biomarkers such as AFP value reduction and tumor angiogenesis inhibition is obtained.

[0060] The doctor selects an intervention simulation of targeted therapy (sorafenib) for patient A. The intervention simulation and the effect quantification module modify the state of the "treatment method" node in the causal graph to "sorafenib treatment" based on the treatment effect knowledge base, while fixing the states of the "age" and "liver function Child-Pugh classification" and other confounder nodes. The prognosis risk assessment model is calculated again to obtain the counterfactual risk trajectory. Taking the 6th month as an example, the individual treatment effect value = the 6th month risk value of the baseline risk trajectory - the 6th month risk value of the counterfactual risk trajectory, that is, sorafenib treatment is expected to reduce the prognosis risk of patient A by a certain value at the 6th month, and the specific value is calculated according to the calculation.

[0061] S4, result generation and interpretation The explainability generation module uses the SHAP library of Python to realize the Shapley additive explanations method. The marginal contribution of each node and edge in the individualized dynamic causal graph to the final risk assessment result is calculated, and the contribution degree of the key causal path is quantified. For example, it is determined that "tumor size (contribution degree 35%) ", "Wnt / β-catenin pathway activity (contribution degree 25%) " and "AFP value (contribution degree 20%) " are the key causal paths. The quantification result is converted into a natural language description: "The high prognosis risk of patient A is mainly caused by large tumor (contribution degree 35%), strong Wnt / β-catenin pathway activity (contribution degree 25%) and high AFP value (contribution degree 20%); the use of sorafenib treatment is expected to reduce the risk, which is mainly achieved by inhibiting tumor angiogenesis and regulating related biological pathways (based on individualized treatment effect evaluation)".

[0062] The visualization decision support interface of the interactive presentation layer is realized by using the Echarts library. In the risk cockpit view, the baseline risk trajectory and the counterfactual risk trajectory are compared and displayed in the form of a line chart; the treatment scheme comparison view visualizes the individualized treatment effect value of different treatment options (surgical resection, targeted therapy, immunotherapy) through a column chart; the causal path explanation view uses a sankey chart to display the key causal path and its quantified contribution degree, and synchronously displays the above natural language description, which helps doctors understand the evaluation results.

[0063] Five, specific implementation of system architecture (1) Data perception and fusion layer The multi-source data access module uses the requests library of Python to communicate with the APIs of various information systems in the hospital to realize data acquisition. The standardization and feature engineering module is deployed on a high-performance server with parallel computing capability, and uses Docker container technology to realize environment isolation, ensuring the stability and portability of data processing.

[0064] (2) Causal calculation and model layer The medical knowledge base is stored in a Neo4j graph database, which is convenient for managing complex relationship data in the causal graph template and treatment effect knowledge base. The individualized causal graph construction module and prognosis risk assessment model are deployed on servers equipped with computing acceleration cards, which use GPU acceleration to improve processing efficiency.

[0065] (Three) intelligent decision-making and application layer The dynamic risk assessment module, intervention simulation and effect quantification module, and explainability generation module build API interfaces through the Flask framework to realize data interaction and function calls between modules.

[0066] (Four) interactive presentation layer The visual decision support interface is developed using the Vue.js front-end framework and deployed on a web server. Doctors can access the system through a browser to achieve intuitive display and interactive operation of the assessment results.

[0067] In the description of the present specification, the description of the terms "one embodiment", "example", "specific example" and the like means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are contained in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0068] The preferred embodiments of the application disclosed above are only used to help explain the application. The preferred embodiments do not describe all the details and do not limit the application to the specific embodiments described. Obviously, many modifications and changes can be made according to the content of the present specification. The present specification selects and describes these embodiments in order to better explain the principles and practical applications of the application, so that those skilled in the art can well understand and utilize the application. The application is limited only by the claims and their full scope and equivalents.

Claims

1. A data analysis-based method for assessing the prognostic risk of liver cancer, characterized in that, Includes the following steps: S1. Data Acquisition and Standardization: Acquire time-series multimodal data of the target patient, including clinical data, imaging data, and genomic data; perform standardization processing on the data to obtain standardized feature vectors; S2. Construction of Personalized Dynamic Causal Graph: Based on the preset medical causal graph template and the standardized feature vector, a personalized dynamic causal graph is constructed; wherein, the standardized feature vector is used as node features, a graph attention network is used to calculate the attention coefficient between nodes, and the weight of each causal edge in the causal graph is dynamically adjusted according to the attention coefficient; S3. Risk Assessment and Intervention Simulation: Input the individualized dynamic causal graph into the prognostic risk assessment model to obtain the basic risk trajectory; modify the state of the interventionable nodes corresponding to specific treatment options in the causal graph based on the treatment effect knowledge base, while fixing the state of the confounding factor nodes; calculate the counterfactual risk trajectory using the prognostic risk assessment model and generate an individualized treatment effect assessment. S4. Result Generation and Interpretation: Output the basic risk trajectory, counterfactual risk trajectory, and individualized treatment effect assessment, and generate a quantitative causal path explanation based on the attribution algorithm, converting the quantitative results into a natural language description containing contribution information.

2. The method for assessing the prognostic risk of liver cancer based on data analysis according to claim 1, characterized in that, The data standardization process in step S1 includes: Standardize the terminology and format of clinical data based on the HL7FHIR standard; Quantitative radiomics features were extracted from DICOM format images and inter-device differences were corrected. Genomic data is converted into pathway activity scores using standard bioinformatics processes.

3. The method for assessing the prognostic risk of liver cancer based on data analysis according to claim 1, characterized in that, In step S2, the medical causal graph template and the treatment effect knowledge base are derived from a medical knowledge base. The predefined association between treatment options and changes in biomarkers in the treatment effect knowledge base is based on probability distributions or confidence intervals obtained after meta-analysis or Bayesian statistics on real-world data or clinical research literature.

4. The method for assessing the prognostic risk of liver cancer based on data analysis according to claim 3, characterized in that, Step S2 uses a graph attention network to calculate attention coefficients and adjust edge weights, specifically including: For each pair of adjacent nodes in the causal graph, their feature vectors are concatenated or inner producted. The operation result is input into a trainable feedforward neural network, and the LeakyReLU activation function is applied. The results are normalized using softmax to obtain attention coefficients representing the weights of causal edges.

5. Based on the attention coefficient, perform a weighted summation of the features of the neighboring nodes to update the representation of the current node.

6. The method for assessing the prognostic risk of liver cancer based on data analysis according to claim 1, characterized in that, The prognostic risk assessment model in step S3 is a graph neural network model pre-trained using data from multiple medical centers processed by domain adaptation technology; the domain adaptation training specifically adopts a domain adversarial neural network or a domain alignment method based on the maximum mean difference.

7. The method for assessing the prognostic risk of liver cancer based on data analysis according to claim 5, characterized in that, The confounding factor nodes in step S3 include age and Child-Pugh classification nodes for liver function. This fixation is achieved by keeping the state of these nodes unchanged during counterfactual reasoning.

8. The method for assessing the prognostic risk of liver cancer based on data analysis according to claim 1, characterized in that, In step S4, the attribution algorithm uses the Shapley sum interpretation method, which quantifies the contribution of key causal paths by calculating the marginal contribution of each node and edge in the graph to the final risk assessment result.

9. A data analysis-based liver cancer prognostic risk assessment system, used to implement the method of any one of claims 1-7, characterized in that, include: The data perception and fusion layer includes a multi-source data access module and a standardization and feature engineering module; The causal calculation and model layer includes a medical knowledge base, a personalized causal graph construction module, and a prognostic risk assessment model; The intelligent decision-making and application layer includes a dynamic risk assessment module, an intervention simulation and effect quantification module, and an interpretability generation module; The interactive presentation layer includes a visual decision support interface.

10. A liver cancer prognostic risk assessment system based on data analysis according to claim 8, characterized in that, The visual decision support interface includes: Risk cockpit view, used to compare and display basic risk trajectories with counterfactual risk trajectories; Treatment plan comparison view, used to visualize the individualized treatment effect values ​​of different treatment options; The causal pathway explanation view is used to interactively display key causal pathways and their quantitative contributions, and simultaneously display natural language descriptions generated by the system that explain the clinical significance of the pathways.

Images