Liver disease diagnosis and treatment strategy recommendation method, medium and device based on knowledge graph

Abstract

This invention discloses a method, medium, and device for recommending liver disease diagnosis and treatment strategies based on a knowledge graph. The method includes: collecting basic user information; constructing a spatiotemporally related individual liver disease knowledge graph based on the user information, including interconnected etiological feature nodes, pathological grading nodes, complication warning nodes, and treatment response nodes; inputting the individual liver disease knowledge graph into a hierarchical reinforcement learning model for joint reasoning, and outputting multiple sets of strategy options including antiviral treatment plans, hepatoprotective drug combinations, and metabolic intervention measures; analyzing the topological propagation path of complication warning nodes through a complication prediction sub-model, and dynamically adjusting the initial set of candidate diagnosis and treatment strategies; and updating the individual liver disease knowledge graph when acquiring new user basic information. This invention achieves multi-dimensional data fusion and dynamic strategy optimization for liver disease diagnosis and treatment, significantly improving the accuracy and timeliness of diagnosis and treatment plans.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, specifically to a method, medium, and device for recommending liver disease diagnosis and treatment strategies based on knowledge graphs. Background Technology

[0002] In the clinical diagnosis and treatment of liver diseases, physicians typically need to comprehensively consider a patient's biochemical indicators, imaging characteristics, and medical history to formulate a treatment plan. The current mainstream treatment model relies primarily on standardized procedures from clinical guidelines and the physician's individual experience, manually comparing various test results to assess disease progression. While the development of precision medicine has enabled electronic medical record systems to integrate multi-source data such as laboratory tests and imaging reports, it remains difficult to achieve dynamic correlation analysis of medical data across different dimensions. This is particularly true when dealing with the complex interactions between etiological characteristics, pathological evolution, and treatment response, where effective decision support tools are lacking. This limits the personalization and timeliness of clinical treatment plans, preventing them from fully adapting to the dynamic characteristics of liver disease progression.

[0003] Although artificial intelligence technology has been applied in the medical field in recent years, most systems lack the ability to provide early warning of complication risks and to optimize multi-strategy collaboration. Especially when dealing with complex liver diseases such as cirrhosis and liver cancer, traditional methods struggle to achieve real-time dynamic adjustments to treatment strategies, thus limiting the accuracy of clinical decision-making and treatment outcomes. Summary of the Invention

[0004] In view of the above problems, the present invention provides a method, medium and device for recommending liver disease diagnosis and treatment strategies based on knowledge graphs, which solves the problems of delayed early warning of complications and lack of dynamic optimization of treatment plans in the existing diagnosis and treatment of liver diseases.

[0005] To achieve the above objectives, in a first aspect, this application provides a knowledge graph-based method for recommending liver disease diagnosis and treatment strategies, including:

[0006] Collect basic user information, including serum test indicators, medical imaging features, gene test information, and lifestyle parameters;

[0007] Based on user basic information, a spatiotemporally related individual liver disease knowledge graph is constructed. The individual liver disease knowledge graph includes interconnected etiology feature nodes, pathological grading nodes, complication warning nodes, and treatment response nodes. Etiology feature nodes are configured to establish connections with viral load and metabolic abnormality markers. Pathological grading nodes are configured to label the grading information of liver fibrosis stage and inflammatory activity. Complication warning nodes are configured to be associated with portal pressure parameters and organ function compensation indicators. Treatment response nodes are configured to record drug sensitivity and adverse reaction events.

[0008] The individual liver disease knowledge graph is input into the hierarchical reinforcement learning model, which performs joint reasoning on the etiology feature node, pathological grade node, and treatment response node, and outputs multiple sets of strategy options including antiviral treatment regimens, hepatoprotective drug combinations, and metabolic intervention measures. Each set of strategy options is accompanied by an expected efficacy assessment curve and a risk probability matrix.

[0009] An initial set of candidate treatment strategies is generated based on multiple sets of strategy options;

[0010] Furthermore, a complication prediction sub-model is established in the hierarchical reinforcement learning model. The complication prediction model analyzes the topological propagation path of the complication warning nodes through a graph convolutional network. When the risk coefficient of esophageal varices exceeds the preset risk threshold or the warning signal of hepatic encephalopathy continues to appear, a preventive treatment plan is dynamically inserted into the initial candidate treatment strategy set to generate and output the final candidate treatment strategy set.

[0011] The next time the user's basic information is retrieved, the individual liver disease knowledge graph corresponding to the current user will be updated.

[0012] Furthermore, a spatiotemporally related individual liver disease knowledge graph is constructed based on user basic information, including:

[0013] Initialize an individual liver disease knowledge graph, which includes nodes for etiological characteristics, pathological grading, complication warning, and treatment response.

[0014] The collected basic user information is mapped to the etiology feature nodes, pathological grading nodes, complication warning nodes, and treatment response nodes of the individual liver disease knowledge graph;

[0015] Align the timestamps of etiology feature nodes, pathological grading nodes, complication warning nodes, and treatment response nodes on the time dimension axis to create a time index;

[0016] Cross-node association weights are calculated using a three-layer attention mechanism, which includes a first operational layer, a second operational layer, and a third operational layer. The first operational layer calculates the pathological influence coefficient of viral load and liver fibrosis stage, the second operational layer calculates the interaction strength of portal pressure and drug response, and the third operational layer integrates the behavioral influence coefficient of lifestyle parameters on the pathological influence coefficient and the interaction strength.

[0017] The next time the user's basic information is retrieved, the individual liver disease knowledge graph corresponding to the current user will be updated, including:

[0018] When the current user inputs their basic information, the update of the viral load time-series curve in the etiology feature node and the recalculation of the inflammation activity grading information in the pathology grading node are triggered, and the updated individual liver disease knowledge graph is output.

[0019] Furthermore, initialize the individual liver disease knowledge graph, including:

[0020] The etiological feature nodes are configured to connect viral load time-series curves and metabolic abnormality marker fluctuation data via a Bayesian network, including:

[0021] A bivariate Gaussian mixture model of viral genotype and liver enzyme profile was constructed. The viral load sequence and metabolic biomarker ratio sequence in serum detection indicators were used as input features. The metabolic regulation condition probability corresponding to different viral load ranges was calculated by the bivariate Gaussian mixture model.

[0022] The variational inference algorithm was used to solve the latent variable posterior distribution of mutation sites in the core protein region and serological conversion events, and the causal weights of core protein mutation and serological conversion were obtained.

[0023] The dynamic propagation equation of metabolic regulation parameters is established based on the conditional probability and causal weight of metabolic regulation. The feedback inhibition effect of steatosis on viral replication rate is quantified. The dynamic propagation equation of metabolic regulation parameters is established based on the conditional probability and causal weight of metabolic regulation. The first output vector containing viral mutation characteristic parameters, metabolic compensation parameters and dynamic causal network parameters is obtained.

[0024] The metabolic compensation parameters in the first output vector are weighted and connected to the inflammation activity calculation module of the pathological grading node.

[0025] The pathological grading nodes were configured to use a semi-supervised clustering algorithm to label the evolution trend of liver fibrosis staging and inflammatory activity, including:

[0026] A joint embedding space of liver tissue elasticity parameters and serum fibrosis indicators is constructed. The elasticity parameter sequence in medical image features and the metabolic compensation parameter in the first output vector are used as input features. The feature dimensionality is reduced by the t-SNE algorithm and the feature representation is optimized by a semi-supervised graph convolutional network to obtain low-dimensional features that integrate metabolic regulation information.

[0027] A dual-channel gated recurrent unit network model was constructed to model inflammation fluctuations and fibrosis progression. Low-dimensional features were input into the dual-channel gated recurrent unit network model to obtain a second output vector containing spatiotemporal joint distribution features.

[0028] The second output vector is input into the multi-organ risk assessment module of the complication early warning node through an attention gating mechanism.

[0029] Furthermore, the complication early warning node is configured to correlate the gradient of portal pressure parameters with the decay rate of organ function compensation indicators via a gated graph neural network, including:

[0030] An organ function association graph of the portal vein-systemic circulation system is constructed. The spatiotemporal joint distribution features in the second output vector are used as topological initialization features. The functional coupling strength of spleen-kidney-brain is dynamically calculated through graph attention mechanism to obtain the organ function association strength matrix.

[0031] Based on hemodynamic principles, the propagation process of pathological changes on organ function correlation diagrams is simulated, the influence coefficient of portal pressure gradient on distal organ function is quantified, and the organ function correlation strength matrix is ​​adjusted.

[0032] When the influence coefficient of the portal pressure gradient on distal organ function exceeds the clinical warning threshold, a third output vector containing multi-organ risk level assessment and priority intervention target identification is generated.

[0033] The third output vector is input into the multi-objective optimization module of the treatment response node through feature cross-fusion with adaptive weights.

[0034] The treatment response node is configured to construct a drug sensitivity prediction model and a causal inference chain for adverse reaction events, including:

[0035] A drug-metabolizing enzyme genotype-plasma drug concentration response surface was constructed. Gene detection information and the third output vector were used as input features. The dose-response relationship between genotype and drug metabolism kinetics was established through piecewise nonlinear regression.

[0036] A multi-task deep network model incorporating virological response rate and renal function protection was constructed. The input layer of the multi-task deep network model was configured to integrate blood drug concentration prediction results and organ risk characteristics.

[0037] The minimum sufficient causal path between drug combinations and specific adverse reactions is identified by a constrained causal discovery algorithm, resulting in a fourth output vector that includes a combination of time-dependent dosing regimens and risk avoidance strategies.

[0038] The pharmacokinetic parameters in the fourth output vector are updated with the dynamic causal network parameters in the first output vector via a feedback connection.

[0039] Furthermore, the individual liver disease knowledge graph is input into a hierarchical reinforcement learning model to perform joint reasoning on etiological feature nodes, pathological grading nodes, and treatment response nodes. The model outputs multiple strategy options, including antiviral treatment regimens, hepatoprotective drug combinations, and metabolic interventions, including:

[0040] A hierarchical reinforcement learning model with a three-level reasoning framework including a strategy decision layer, a principle optimization layer, and an execution evaluation layer is established. The dynamic causal network parameters in the etiological feature nodes are analyzed through the strategy decision layer, and the long-term evolution trend of viral mutation and treatment pressure is calculated to obtain the strategy objective vector.

[0041] The spatiotemporal joint distribution features of the strategy objective vector and pathological grading nodes are input into the principle optimization layer, and a multi-objective decision space containing efficacy surface and safety boundary is constructed in the principle optimization layer.

[0042] In the evaluation layer, pharmacokinetic parameters are extracted from the treatment response node and mapped to a multi-objective decision space.

[0043] A non-dominated sorting genetic algorithm is used to solve the multi-objective decision space and generate an initial policy set;

[0044] Perform multidimensional evaluation on each initial strategy in the initial strategy set, and output an initial candidate treatment strategy set with priority ranking.

[0045] Furthermore, the multidimensional assessment includes efficacy assessment, safety assessment, and economic assessment dimensions:

[0046] In terms of efficacy evaluation, the inhibitory effect of antiviral treatment regimens on viral load is analyzed, and the trend of virological response is predicted based on viral mutation characteristic parameters of etiological feature nodes.

[0047] In addition, the study evaluated the effect of hepatoprotective drug combinations on improving inflammatory activity and estimated the potential for improving liver fibrosis by combining the spatiotemporal distribution characteristics of pathological grading nodes.

[0048] Generate the first evaluation result;

[0049] In terms of safety assessment, the impact of treatment regimens on renal function is monitored, and the risk of renal function impairment is identified based on pharmacokinetic parameters at treatment response points.

[0050] In addition, it tracks the adverse reaction characteristics of drug combinations and uses the dynamic causal network parameters of etiological feature nodes to predict the probability of occurrence of specific adverse reactions.

[0051] Generate a second evaluation result;

[0052] In terms of economic assessment, the overall cost of treatment options is calculated, integrating drug costs, monitoring costs, and reimbursement policies.

[0053] Assess treatment adherence and predict treatment completion rate by combining medication record characteristics in lifestyle behavior parameters;

[0054] Generate a third evaluation result;

[0055] A strategy priority score is generated based on the results of the first, second, and third assessments.

[0056] Store the policy priority score and the initial policy mapping.

[0057] Furthermore, a complication prediction sub-model is established within the hierarchical reinforcement learning model, including:

[0058] Construct a risk propagation map of complications, using portal pressure parameters and organ function compensation indicators in the early warning nodes of complications as initial node features;

[0059] A functional network connecting the liver, portal system, and distal organs is established using graph convolution operations.

[0060] It receives the inflammation activity grading information and liver fibrosis stage information input from the pathology grading node and calculates the portal hypertension progression rate.

[0061] Obtain drug sensitivity parameters from treatment response points to assess the effectiveness of the current treatment regimen in suppressing complications;

[0062] When the portal pressure gradient exceeds the compensation threshold, the risk factor for esophageal varices is marked.

[0063] When blood ammonia levels remain elevated and are accompanied by neuropsychiatric symptoms, a warning signal for hepatic encephalopathy is triggered.

[0064] Furthermore, a final set of candidate treatment strategies is generated and output, including:

[0065] Based on the risk factor of esophageal varices, a first preventive treatment plan was generated, which included the frequency of endoscopic monitoring, the administration of beta-blockers, and an emergency hemostasis plan.

[0066] In response to early warning signs of hepatic encephalopathy, a second preventative treatment plan was generated, which included adjusting the lactulose dosage, restricting protein intake, and monitoring blood ammonia levels.

[0067] Synergistic optimization of the primary and / or secondary preventative treatment plans with the initial candidate treatment strategies includes:

[0068] The compatibility of the first and / or second preventive measures with the initial candidate treatment strategies is evaluated to obtain the first optimization result;

[0069] The dosage and dosing sequence of the drug combination were adjusted based on the first optimization result to obtain the second optimization result;

[0070] Based on the second optimization results, the risk-benefit ratio of the overall treatment plan is recalculated, and a final candidate treatment strategy with complication prevention and control markers is generated. The final candidate treatment strategy includes efficacy risk assessment information of the basic treatment plan, the execution priority of complication prevention measures, and the time sequence planning diagram of the overall treatment path.

[0071] A final candidate treatment strategy set is generated based on multiple final candidate treatment strategies.

[0072] In a second aspect, the present invention also provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the method described in the first aspect.

[0073] In a third aspect, the present invention also provides an electronic device including a memory and a processor, the memory being used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement the method described in the first aspect.

[0074] Unlike existing technologies, the above technical solution has the following beneficial effects: This invention provides a method, medium, and device for recommending liver disease diagnosis and treatment strategies based on a knowledge graph. The method includes: collecting basic user information including serum test indicators, medical imaging features, gene test information, and lifestyle parameters; constructing a spatiotemporally related individual liver disease knowledge graph, including interconnected etiological feature nodes, pathological grading nodes, complication warning nodes, and treatment response nodes; inputting the individual liver disease knowledge graph into a hierarchical reinforcement learning model for joint reasoning, and outputting multiple sets of strategy options including antiviral treatment plans, hepatoprotective drug combinations, and metabolic intervention measures; analyzing the topological propagation path of complication warning nodes through a complication prediction sub-model, and dynamically adjusting the initial candidate diagnosis and treatment strategy set; and updating the individual liver disease knowledge graph when acquiring new user basic information. This invention achieves multi-dimensional data fusion and dynamic strategy optimization for liver disease diagnosis and treatment, significantly improving the accuracy and timeliness of diagnosis and treatment plans.

[0075] The above description of the invention is merely an overview of the technical solution of this application. In order to enable those skilled in the art to better understand the technical solution of this application and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of this application easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of this application. Attached Figure Description

[0076] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on this application.

[0077] In the accompanying drawings of the instruction manual:

[0078] Figure 1 This is a flowchart illustrating steps S101 to S105 of the method described in the specific implementation.

[0079] Figure 2 This is a flowchart illustrating steps S201 to S205 of the method described in a specific embodiment.

[0080] Figure 3This is a flowchart illustrating steps S301 to S304 of the method described in a specific embodiment.

[0081] Figure 4 This is a flowchart illustrating steps S401 to S406 of the method described in a specific embodiment.

[0082] Figure 5 This is a schematic diagram of the structure of the electronic device described in a specific embodiment.

[0083] The reference numerals used in the above figures are explained as follows:

[0084] 1. Electronic equipment;

[0085] 11. Memory;

[0086] 12. Processor. Detailed Implementation

[0087] To illustrate the possible application scenarios, technical principles, implementable specific solutions, and achievable objectives and effects of this application in detail, the following description, in conjunction with the listed specific embodiments and accompanying drawings, provides a detailed explanation. The embodiments described herein are merely illustrative of the technical solutions of this application and are therefore intended to limit the scope of protection of this application.

[0088] In this document, the term "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The term "embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment, nor does it specifically limit its independence or connection with other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the technical features mentioned in each embodiment can be combined in any way to form corresponding implementable technical solutions.

[0089] Unless otherwise defined, the technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the use of related terms herein is merely for the purpose of describing particular embodiments and is not intended to limit this application.

[0090] In the description of this application, the term "and / or" is used to describe the logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A exists, B exists, and A and B exist simultaneously. Additionally, the character " / " in this document generally indicates that the preceding and following objects have an "or" logical relationship.

[0091] In this application, terms such as “first” and “second” are used only to distinguish one entity or operation from another, and do not necessarily require or imply any actual quantity, hierarchy or order relationship between these entities or operations.

[0092] Without further limitations, the use of terms such as “comprising,” “including,” “having,” or other similar open-ended expressions in this application is intended to cover non-exclusive inclusion, which does not exclude the presence of additional elements in a process, method, or product that includes the stated elements, such that a process, method, or product that includes a list of elements may include not only those defined elements but also other elements not expressly listed, or elements inherent to such a process, method, or product.

[0093] As understood in the Examination Guidelines, in this application, expressions such as "greater than," "less than," and "exceeding" are understood to exclude the stated number; expressions such as "above," "below," and "within" are understood to include the stated number. Furthermore, in the description of the embodiments in this application, "multiple" means two or more (including two), and similar expressions related to "multiple" are also understood in this way, such as "multiple groups" and "multiple times," unless otherwise explicitly specified.

[0094] In the description of the embodiments of this application, the space-related expressions used, such as "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "vertical," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," indicate the orientation or positional relationship based on the orientation or positional relationship shown in the specific embodiments or drawings. They are only for the purpose of describing the specific embodiments of this application or for the reader's understanding, and do not indicate or imply that the device or component referred to must have a specific position, a specific orientation, or be constructed or operated in a specific orientation. Therefore, they should not be construed as limitations on the embodiments of this application.

[0095] Please see Figure 1 In a first aspect, this embodiment provides a knowledge graph-based method for recommending liver disease diagnosis and treatment strategies, including:

[0096] S101. Collect basic user information, which includes serum test indicators, medical imaging features, gene test information, and lifestyle parameters.

[0097] S102. Construct a spatiotemporally related individual liver disease knowledge graph based on user basic information. The individual liver disease knowledge graph includes interconnected etiology feature nodes, pathological grading nodes, complication warning nodes, and treatment response nodes. Etiology feature nodes are configured to establish connections with viral load and metabolic abnormality markers. Pathological grading nodes are configured to label the grading information of liver fibrosis stage and inflammatory activity. Complication warning nodes are configured to be associated with portal pressure parameters and organ function compensation indicators. Treatment response nodes are configured to record events of drug sensitivity and adverse reactions.

[0098] S103. Input the individual liver disease knowledge graph into the hierarchical reinforcement learning model, perform joint reasoning on the etiology feature node, pathological grade node and treatment response node, and output multiple sets of strategy options including antiviral treatment plan, liver protection drug combination and metabolic intervention measures. Each set of strategy options is accompanied by the expected efficacy assessment curve and risk probability matrix.

[0099] S104. Generate an initial set of candidate treatment strategies based on multiple sets of strategy options;

[0100] Furthermore, a complication prediction sub-model is established in the hierarchical reinforcement learning model. The complication prediction model analyzes the topological propagation path of the complication warning nodes through a graph convolutional network. When the risk coefficient of esophageal varices exceeds the preset risk threshold or the warning signal of hepatic encephalopathy continues to appear, a preventive treatment plan is dynamically inserted into the initial candidate treatment strategy set to generate and output the final candidate treatment strategy set.

[0101] S105. When obtaining the current user's basic information again, update the individual liver disease knowledge graph corresponding to the current user.

[0102] In step S101, the user's basic information forms the data foundation for constructing a personalized treatment strategy. This includes: serum testing indicators, which can be obtained through clinical laboratory testing and contain biomarkers reflecting liver function and viral activity; medical imaging features, acquired through medical imaging equipment, characterizing changes in liver tissue structure; gene testing information, obtained through molecular biology techniques, recording genetic characteristics related to drug metabolism; and lifestyle behavioral parameters, collected through medical records, reflecting behavioral factors that may affect treatment outcomes. Furthermore, this data is integrated and managed through a medical information system.

[0103] In step S102, the spatiotemporally correlated individual liver disease knowledge graph is constructed using a graph structure data model. In this model, etiology feature nodes connect virological parameters and metabolic indicators, pathological grading nodes label the degree of liver damage, complication warning nodes are associated with portal venous system parameters and organ function indicators, and treatment response nodes record drug treatment-related events. Preferably, the relationships between nodes are established through time index alignment and feature weight calculation.

[0104] In step S103, the hierarchical reinforcement learning model jointly analyzes the etiology feature nodes, pathological grading nodes, and treatment response nodes of the individual liver disease knowledge graph. Preferably, the hierarchical reinforcement learning model generates multiple sets of strategy options and their risk-benefit assessments for antiviral treatment regimens, hepatoprotective drug combinations, and metabolic interventions through a three-layer cascaded reasoning process: a strategy layer analyzes etiology, an optimization layer assesses pathological grading, and an execution layer integrates treatment responses. Further, each set of strategy options is accompanied by an expected efficacy assessment curve and a risk probability matrix. The expected efficacy assessment curve reflects the time-effect characteristics of the treatment regimen, and the risk probability matrix quantifies the potential adverse consequences during the treatment process.

[0105] In step S104, the complication prediction sub-model analyzes the topological relationships of complication warning nodes. When the risk coefficient of esophageal varices exceeds a preset risk threshold or a warning signal for hepatic encephalopathy persists, corresponding preventive measures are added to the basic treatment plan. Preferably, the risk coefficient of esophageal varices is calculated using portal pressure parameters to trigger endoscopic monitoring; the preset risk threshold is set based on clinical guidelines; and the warning signal for hepatic encephalopathy can be determined by a combination of blood ammonia and neurological symptoms. Esophageal variceal rupture and hepatic encephalopathy are major fatal complications of cirrhosis, and dynamically inserting preventive measures can reduce the risk of sudden onset. Furthermore, the insertion process of preventive measures considers the synergistic effect with existing treatments.

[0106] In step S105, the knowledge graph update is triggered by the newly acquired user basic information, focusing on updating dynamically changing feature parameters such as viral load and inflammatory activity, thereby tracking disease progression in real time and ensuring that treatment strategies respond promptly to changes in the patient's condition.

[0107] This embodiment integrates multidimensional clinical data through a knowledge graph and utilizes a hierarchical reasoning mechanism to dynamically optimize treatment strategies. For example, when a patient's medical imaging shows progression of fibrosis, the system can reassess the risk-benefit ratio of a combination of hepatoprotective drugs by incorporating the latest serum test indicators. This embodiment integrates multi-source heterogeneous data such as serum test indicators, medical imaging features, gene testing information, and lifestyle parameters through a spatiotemporally correlated individual liver disease knowledge graph, achieving multi-dimensional and accurate recommendations for liver disease treatment strategies. A hierarchical reinforcement learning model is used to jointly reason about etiological feature nodes, pathological grading nodes, and treatment response nodes, ensuring that the generated antiviral treatment plans, hepatoprotective drug combinations, and metabolic interventions consider the synergistic effects of each treatment dimension while avoiding excessive intervention with a single treatment method. A complication prediction sub-model dynamically monitors the risk coefficient of esophageal varices and early warning signals of hepatic encephalopathy, intelligently inserting preventative treatment plans into the initial candidate treatment strategy set. This allows the system to continuously adapt to changes in the patient's condition, preserving the normative requirements of clinical guidelines while achieving the proactive defense advantage based on individualized risk warnings, effectively improving the accuracy of liver disease treatment strategies and the timeliness of complication prevention and control.

[0108] Please see Figure 2 In some embodiments, a spatiotemporally related individual liver disease knowledge graph is constructed based on basic user information, including:

[0109] S201. Initialize the individual liver disease knowledge graph, which includes etiology characteristic nodes, pathological grading nodes, complication warning nodes, and treatment response nodes.

[0110] S202. Map the collected user basic information to the etiology feature node, pathological grading node, complication warning node and treatment response node of the individual liver disease knowledge graph;

[0111] S203. Align the timestamps of etiology feature nodes, pathological grading nodes, complication warning nodes, and treatment response nodes on the time dimension axis to establish a time index;

[0112] S204. Cross-node association weights are calculated through a three-layer attention mechanism, which includes a first operational layer, a second operational layer, and a third operational layer. The first operational layer calculates the pathological influence coefficient of viral load and liver fibrosis stage, the second operational layer calculates the interaction strength of portal pressure and drug response, and the third operational layer integrates the behavioral influence coefficient of lifestyle parameters on the pathological influence coefficient and the interaction strength.

[0113] S205. When retrieving the current user's basic information again, update the individual liver disease knowledge graph corresponding to the current user, including:

[0114] When the current user inputs their basic information, it triggers an update of the viral load time-series curve in the etiology feature node and a recalculation of the inflammation activity grading information in the pathology grading node, outputting an updated individual liver disease knowledge graph. When the time difference between the newly input serum test indicators and existing data is less than 24 hours, it triggers an incremental update of the nodes.

[0115] Etiological characteristic nodes update viral load decay rate;

[0116] Recalculate the slope of inflammatory activity at pathological grading nodes;

[0117] Output an individual liver disease knowledge graph with version identifiers.

[0118] In step S201, preferably, the initialization of the individual liver disease knowledge graph adopts a standardized graph structure template, wherein the etiology feature node presets a data interface for virology and metabolic abnormalities, the pathological grading node presets a grading standard interface for liver fibrosis and inflammatory activity, the complication warning node presets a monitoring parameter interface for the portal system and organ function, and the treatment response node presets a data recording interface for drug response events. Preferably, each node retains a null placeholder during initialization for subsequent data filling.

[0119] In step S202, the mapping process of user basic information can be implemented through a feature matching engine. Specifically, virological parameters from serum testing indicators are mapped to etiology feature nodes, fibrosis scores from medical imaging features are mapped to pathological grading nodes, portal venous system parameters are mapped to complication warning nodes, and drug metabolism-related SNP sites obtained from gene testing are mapped to treatment response nodes. Furthermore, lifestyle behavior parameters are converted into standardized feature vectors by a behavioral feature encoder and then simultaneously associated with etiology feature nodes and treatment response nodes.

[0120] In step S203, timestamp alignment is achieved through a distributed time series database. Preferably, event time is used as the baseline time axis, and the viral load detection time in the etiology feature node, the imaging examination time in the pathological grading node, the portal vein pressure measurement time in the complication warning node, and the medication record time in the treatment response node are uniformly aligned to the patient's visit time coordinate system. Furthermore, the time index is established using a sliding time window mechanism, and the window size is dynamically adjusted according to the clinical follow-up cycle.

[0121] In step S204, the specific implementation of the three-layer attention mechanism includes: the first operational layer uses attention calculation guided by prior pathological knowledge, determining the pathological influence coefficient through the covariance matrix of the logarithmic transformation value of viral load and the stage of liver fibrosis; the second operational layer uses cross-correlation analysis of portal pressure gradient and drug blood concentration to calculate the interaction strength, wherein the portal pressure parameter is standardized by Z-score; the third operational layer uses tensor product operation of behavioral feature vector and pathological influence coefficient, combined with partial correlation analysis of interaction strength, to finally output a comprehensive weight matrix containing behavioral influence coefficient. Preferably, each operational layer uses residual connections to maintain gradient flow.

[0122] In step S205, the knowledge graph update mechanism includes two modes: full update and incremental update. When the time difference between new input data and actual data is ≥24 hours, a full update process is triggered, recalculating the decay trend of the viral load time-series curve and the dynamic changes in the inflammation activity grade. When the time difference is <24 hours, an incremental update strategy is adopted, adjusting only the instantaneous change value of the viral load decay rate and the recent gradient of the inflammation activity slope. A data time difference exceeding 24 hours may lead to significant changes in disease status; a full update ensures that the knowledge graph accurately reflects the latest disease condition. Incremental updates are suitable for short-term, small fluctuations, maintaining computational efficiency. Preferably, the version identifier uses a timestamp hash value to ensure the traceability of the graph version. Furthermore, the original topological relationships between nodes remain unchanged during the update process; only the feature weight values ​​are modified.

[0123] This embodiment employs a three-layer attention mechanism to accurately calculate the pathological influence coefficients of viral load and liver fibrosis stage, the interaction strength between portal pressure and drug response, and the behavioral influence coefficients of lifestyle parameters, achieving dynamic quantification of cross-node relationships in an individual liver disease knowledge graph. A timestamp alignment mechanism on the time dimension axis ensures temporal consistency among etiological feature nodes, pathological grading nodes, complication warning nodes, and treatment response nodes, enabling the knowledge graph to accurately reflect the spatiotemporal characteristics of disease development. By distinguishing between full and incremental update modes, intelligent allocation of computational resources is achieved during the updating of viral load time-series curves and the recalculation of inflammation activity grading information, ensuring timely integration of new data while avoiding redundant calculations. Combined with a version-identified knowledge graph output mechanism, the system can accurately track the data version upon which each diagnostic and treatment decision is based, improving the timeliness of liver disease treatment strategies while ensuring the traceability of the decision-making process, effectively optimizing the update efficiency of the knowledge graph and the reliability of clinical applications.

[0124] Please see Figure 3 In some embodiments, initializing an individual liver disease knowledge graph includes:

[0125] S301, the etiology feature node is configured to connect viral load time-series curves and metabolic abnormality marker fluctuation data via a Bayesian network, including:

[0126] A bivariate Gaussian mixture model of viral genotype and liver enzyme profile was constructed. The viral load sequence and metabolic biomarker ratio sequence in serum detection indicators were used as input features. The metabolic regulation condition probability corresponding to different viral load ranges was calculated by the bivariate Gaussian mixture model.

[0127] The variational inference algorithm was used to solve the latent variable posterior distribution of mutation sites in the core protein region and serological conversion events, and the causal weights of core protein mutation and serological conversion were obtained.

[0128] The dynamic propagation equation of metabolic regulation parameters is established based on the conditional probability and causal weight of metabolic regulation. The feedback inhibition effect of steatosis on viral replication rate is quantified. The dynamic propagation equation of metabolic regulation parameters is established based on the conditional probability and causal weight of metabolic regulation. The first output vector containing viral mutation characteristic parameters, metabolic compensation parameters and dynamic causal network parameters is obtained.

[0129] S302. The metabolic compensation parameters in the first output vector are connected to the inflammation activity calculation module of the input pathological grading node through weighted connection.

[0130] S303, the pathological grading node, is configured to use a semi-supervised clustering algorithm to label the evolution trend of liver fibrosis staging and inflammatory activity, including:

[0131] A joint embedding space of liver tissue elasticity parameters and serum fibrosis indicators is constructed. The elasticity parameter sequence in medical image features and the metabolic compensation parameter in the first output vector are used as input features. The feature dimensionality is reduced by the t-SNE algorithm and the feature representation is optimized by a semi-supervised graph convolutional network to obtain low-dimensional features that integrate metabolic regulation information.

[0132] A dual-channel gated recurrent unit network model was constructed to model inflammatory fluctuations and fibrosis progression. Low-dimensional features were input into the dual-channel gated recurrent unit network model to obtain a second output vector containing spatiotemporal joint distribution features, including:

[0133] Short-term inflammatory fluctuation channel: Processing ALT / AST fluctuation sequences from the three most recent tests;

[0134] Long-term fibrosis progression pathway: Analysis of liver stiffness measurement trends over 12 weeks;

[0135] A second output vector containing spatiotemporal joint distribution features is obtained by fusing dual-channel features;

[0136] S304. Input the second output vector into the multi-organ risk assessment module of the complication early warning node through the attention gating mechanism.

[0137] In step S301, preferably, the construction of the bivariate Gaussian mixture model uses the EM algorithm to optimize the covariance matrix, wherein the viral load sequence is subjected to Box-Cox transformation to eliminate the influence of dimensions, and the metabolic marker ratio sequence is standardized by interquartile range.

[0138] Preferably, the variational inference algorithm uses reparameterization techniques to process discrete variables of mutation sites in the core protein region, approximates the true posterior distribution of serological conversion events by minimizing KL divergence, quantifies the driving effect of core protein mutations on serological conversion, and provides causal basis for subsequent targeted therapy and prognosis prediction.

[0139] Preferably, the establishment of the dynamic propagation equation introduces a time-delay differential operator, in which the feedback inhibition effect of fatty degeneration can be modeled by the Hill equation; the viral mutation characteristic parameters include the quasi-species complexity index to quantify the diversity of viral mutations; the metabolic compensation parameters integrate the branched-chain amino acid / aromatic amino acid ratio to quantify the degree of metabolic imbalance.

[0140] In step S302, preferably, the weighted connection employs an adaptive feature selection mechanism. The metabolic compensation parameters are first normalized using a sigmoid function, and then gated and weighted with the baseline features of the inflammation activity calculation module. Preferably, the weight coefficients are obtained through supervised learning using gold standard data from liver biopsy.

[0141] In step S303, the construction of the joint embedding space preferably adopts a contrastive learning strategy, wherein the elasticity parameter sequence is subjected to wavelet transform to extract multi-scale features, which form positive sample pairs with serum fibrosis indicators. The semi-supervised graph convolutional network uses a label propagation algorithm to enhance feature representation in sparse labeling scenarios, and the node features include spatial location encoding and diffusion representation of metabolic compensation parameters.

[0142] In the dual-channel gated recurrent unit network, the short-term inflammation fluctuation channel uses one-dimensional causal convolution to capture local temporal patterns, while the long-term fibrosis progression channel introduces a self-attention mechanism to model long-distance dependencies. The feature fusion layer achieves cross-scale feature alignment through a learnable weight matrix.

[0143] In step S304, the attention gating mechanism is implemented using a query-key-value memory network, where the second output vector serves as the query vector, and the historical warning records of the multi-organ risk assessment module constitute a dynamic key-value library. Preferably, the attention score calculation incorporates the anatomical distance of the organ system as a spatial prior constraint, and the gating weight is dynamically adjusted through portal vein hemodynamic parameters.

[0144] This embodiment establishes a multi-level analytical framework from viral metabolic interactions to clinical phenotypes by initializing an individual liver disease knowledge graph. Through joint probabilistic modeling of viral genotype and liver enzyme profiles, the regulatory effect of metabolic conditional probability on viral replication is precisely quantified. By synergistic analysis of dynamic causal network parameters and spatiotemporal joint distribution characteristics, the evolution trend of inflammatory activity and fibrosis staging characteristics are simultaneously captured. An attention gating mechanism based on anatomical distance constraints effectively improves the clinical relevance of multi-organ risk assessment. This embodiment achieves a systematic correlation analysis from molecular mechanisms to histopathology while maintaining the integrity of viral mutation characteristic parameters and metabolic compensation parameters.

[0145] Please see Figure 4 In some embodiments, the complication warning node is configured to correlate the gradient of portal pressure parameters with the decay rate of organ function compensation indicators via a gated graph neural network, including:

[0146] S401. Construct an organ function association graph of the portal vein-systemic circulation system, take the spatiotemporal joint distribution features in the second output vector as the topological structure initialization features, and dynamically calculate the functional coupling strength of spleen-kidney-brain through graph attention mechanism to obtain the organ function association strength matrix.

[0147] S402. Based on the principle of hemodynamics, simulate the propagation process of pathological changes on the organ function correlation map, quantify the influence coefficient of portal pressure gradient on distal organ function, and adjust the organ function correlation strength matrix.

[0148] S403. When the influence coefficient of the portal pressure gradient on the function of distal organs exceeds the clinical warning threshold, a third output vector containing multi-organ risk level assessment and priority intervention target identification is generated.

[0149] S404. Input the third output vector into the multi-objective optimization module of the treatment response node through feature cross-fusion with adaptive weights.

[0150] S405, The treatment response node is configured to construct a drug sensitivity prediction model and a causal inference chain for adverse reaction events, including:

[0151] A drug-metabolizing enzyme genotype-plasma drug concentration response surface was constructed. Gene detection information and the third output vector were used as input features. The dose-response relationship between genotype and drug metabolism kinetics was established through piecewise nonlinear regression.

[0152] A multi-task deep network model incorporating virological response rate and renal function protection was constructed. The input layer of the multi-task deep network model was configured to integrate blood drug concentration prediction results and organ risk characteristics.

[0153] The minimum sufficient causal path between drug combinations and specific adverse reactions is identified by a constrained causal discovery algorithm, resulting in a fourth output vector that includes a combination of time-dependent dosing regimens and risk avoidance strategies.

[0154] S406. Update the dynamic causal network parameters in the first output vector using the pharmacokinetic parameters in the fourth output vector via a feedback connection. The dynamic causal network parameters include:

[0155] The correlation coefficient between viral mutation rate and drug selection pressure;

[0156] Metabolic compensation effect: a corrective factor for drug metabolism;

[0157] Weighting of the impact of treatment regimen adjustments on the probability of core protein mutations.

[0158] In step S401, preferably, the construction of the organ function association map employs anatomical prior constraints, wherein the topology of the portal vein-systemic circulation system is initialized using angiographic data. The graph attention mechanism adopts a multi-head attention architecture, with each attention head corresponding to the blood perfusion characteristics of a specific organ system, and the calculation of functional coupling strength incorporates a distance attenuation factor based on the number of vascular branch generations.

[0159] In step S402, preferably, the hemodynamic simulation employs a simplified model based on the Navier-Stokes equations, where the portal pressure gradient is discretized using the finite element method. The influence coefficient quantification process integrates the organ function reserve index, and the clinical warning threshold is dynamically adjusted according to the Child-Pugh classification.

[0160] In step S403, preferably, the multi-organ risk level assessment is implemented using a fuzzy logic classifier, and the intervention target identification is determined through critical path analysis. The third output vector contains a visual encoding of the portal pressure gradient propagation path and a time-series prediction curve of organ function decline rate.

[0161] In step S404, preferably, the feature cross-fusion is achieved using tensor decomposition technology, and the weight adaptive mechanism is dynamically calibrated through the portal vein pressure-organ function response curve. The fusion process preserves the interaction channels between viral mutation feature parameters and drug metabolism parameters.

[0162] In step S405, preferably, the dose-response relationship modeling uses spline interpolation to handle the discrete variables of genotype, and the blood drug concentration response surface is smoothed using Lagrange interpolation. The loss function of the multi-task deep network model integrates the hinge loss of virological inhibition and the Wasserstein distance metric for renal function protection. The constrained causal discovery algorithm is optimized using the PC-stable algorithm, and the minimum sufficient causal path includes the time delay parameter of drug interaction.

[0163] In step S406, preferably, the feedback connection is implemented using a residual learning architecture, and the momentum gradient descent algorithm is introduced into the dynamic causal network parameter update process. The correlation coefficient calculation uses cross-wavelet transform to analyze the phase coupling relationship between viral quasi-species dynamics and blood drug concentration fluctuations. The correction factor is calibrated through metabolic flux balance analysis, and the influence weight update follows a Markov decision process.

[0164] This embodiment achieves precise prevention and control of liver disease complications by constructing a closed-loop system of organ function correlation diagrams and multi-objective optimization modules. Through topological modeling of the portal vein-systemic circulation system, the cascade damage effect of portal pressure gradient on the spleen-kidney-brain is accurately quantified; pathological propagation simulation based on hemodynamic principles dynamically predicts the trajectory of multi-organ function decline; and by combining a genotype-drug metabolism kinetic model and a constrained causal discovery algorithm, the risk of treatment-related adverse reactions is significantly reduced while ensuring antiviral efficacy. This embodiment establishes a decision-making closed loop from complication early warning to treatment plan optimization while maintaining dynamic updates of viral mutation characteristic parameters and metabolic compensation parameters.

[0165] In some embodiments, an individual liver disease knowledge graph is input into a hierarchical reinforcement learning model to perform joint reasoning on etiological feature nodes, pathological grading nodes, and treatment response nodes, outputting multiple sets of strategy options including antiviral treatment regimens, hepatoprotective drug combinations, and metabolic interventions, including:

[0166] A hierarchical reinforcement learning model with a three-level reasoning framework including a strategy decision layer, a principle optimization layer, and an execution evaluation layer is established. The dynamic causal network parameters in the etiological feature nodes are analyzed through the strategy decision layer, and the long-term evolution trend of viral mutation and treatment pressure is calculated to obtain the strategy objective vector.

[0167] The spatiotemporal joint distribution features of the strategy objective vector and pathological grading nodes are input into the principle optimization layer, and a multi-objective decision space containing efficacy surface and safety boundary is constructed in the principle optimization layer.

[0168] In the evaluation layer, pharmacokinetic parameters are extracted from the treatment response node and mapped to a multi-objective decision space.

[0169] A non-dominated sorting genetic algorithm is used to solve the multi-objective decision space and generate an initial policy set;

[0170] Perform multidimensional evaluation on each initial strategy in the initial strategy set, and output an initial candidate treatment strategy set with priority ranking.

[0171] In this embodiment, the hierarchical reinforcement learning model comprises a three-level inference framework: a policy decision layer, a principle optimization layer, and an execution evaluation layer. In the policy decision layer, preferably, the analysis of dynamic causal network parameters employs a temporal convolutional network architecture, where the evolutionary trends of viral mutation and treatment pressure are modeled using LSTM units. The policy objective vector includes dimensions such as the viral quasi-species complexity decay rate and the gradient of core protein mutation entropy changes; the weights of each dimension can be optimized using the Shapley value allocation algorithm.

[0172] In the principle optimization layer, the preferred method for constructing the multi-objective decision space is the Pareto frontier search method. The efficacy surface is fitted to clinical trial data using support vector regression, and the safety boundary is determined using fuzzy C-means clustering. The decision space dimensions include three orthogonal coordinate axes: virological response rate, liver function improvement index, and metabolic disorder correction degree.

[0173] In the evaluation layer, preferably, the mapping process for pharmacokinetic parameters employs an equidistant logarithmic ratio transformation to maintain the component correlation between parameters. The resolution of the multi-objective decision space is dynamically adjusted based on the feedback error of the treatment response node, and the grid size is proportional to the drug concentration fluctuation range.

[0174] In the initial strategy generation, the preferred fitness function of the non-dominated sorting genetic algorithm includes:

[0175] Treatment efficacy indicators: integrating the rate of decrease in viral load and the time to normalization of liver enzymes;

[0176] Safety tolerance index: risk probability based on the causal inference chain of adverse reaction events;

[0177] Metabolic synergistic indicators: quantifying the improvement in the ratio of branched-chain amino acids to aromatic amino acids;

[0178] The crossover operator uses simulated binary crossover, and the mutation operator uses polynomial mutation.

[0179] In multidimensional evaluation, the preferred evaluation dimensions include:

[0180] Clinical feasibility dimension: assessed through treatment guideline compliance scoring;

[0181] Cost-effectiveness dimension: Calculate the quality-adjusted lifetime year (QALY) gain;

[0182] Personalized adaptation dimension: response prediction probability based on patient genomic characteristics;

[0183] Priority ranking is performed using the TOPSIS algorithm, and the weights of each dimension are determined using the analytic hierarchy process.

[0184] This embodiment achieves intelligent generation of liver disease diagnosis and treatment strategies through a three-level inference framework of a hierarchical reinforcement learning model. The strategy decision-making layer accurately predicts the evolutionary path of viral mutations by analyzing dynamic causal network parameters; the principle optimization layer constructs a multi-objective decision space to effectively balance efficacy requirements and safety constraints; and the execution evaluation layer's multi-dimensional ranking mechanism ensures that the output strategy possesses both clinical feasibility and individualized adaptability. This embodiment provides comprehensive strategy options covering antiviral therapy, liver protection, and metabolic regulation while maintaining the coordinated updating of etiological characteristics, pathological grading, and treatment response parameters.

[0185] In some embodiments, the multidimensional assessment includes efficacy assessment, safety assessment, and economic assessment dimensions:

[0186] In terms of efficacy evaluation, the inhibitory effect of antiviral treatment regimens on viral load is analyzed, and the trend of virological response is predicted based on viral mutation characteristic parameters of etiological feature nodes.

[0187] In addition, the study evaluated the effect of hepatoprotective drug combinations on improving inflammatory activity and estimated the potential for improving liver fibrosis by combining the spatiotemporal distribution characteristics of pathological grading nodes.

[0188] Generate the first evaluation result;

[0189] In terms of safety assessment, the impact of treatment regimens on renal function is monitored, and the risk of renal function impairment is identified based on pharmacokinetic parameters at treatment response points.

[0190] In addition, it tracks the adverse reaction characteristics of drug combinations and uses the dynamic causal network parameters of etiological feature nodes to predict the probability of occurrence of specific adverse reactions.

[0191] Generate a second evaluation result;

[0192] In terms of economic assessment, the overall cost of treatment options is calculated, integrating drug costs, monitoring costs, and reimbursement policies.

[0193] Assess treatment adherence and predict treatment completion rate by combining medication record characteristics in lifestyle behavior parameters;

[0194] Generate a third evaluation result;

[0195] A strategy priority score is generated based on the results of the first, second, and third assessments.

[0196] Store the policy priority score and the initial policy mapping.

[0197] In this embodiment, the multidimensional assessment includes efficacy assessment, safety assessment, and economic assessment. In the efficacy assessment dimension, the prediction of virological response trends is based on viral mutation characteristic parameters in the etiological feature nodes, achieved by establishing a mapping relationship between the quasi-species diversity index and drug target affinity. Preferably, the viral load inhibition effect is assessed using a time-dependent covariate Cox model, incorporating core protein mutation sites as time-varying covariates into the analysis. The assessment of the potential for liver fibrosis improvement is achieved through the spatiotemporal joint distribution characteristics of pathological grading nodes, specifically by multimodal fusion of enhanced MRI texture features and serum fibrosis markers, and calculating the median change in liver stiffness values ​​using elastic network regression.

[0198] In the safety assessment dimension, the identification of renal function impairment risk relies on pharmacokinetic parameters recorded at treatment response points, with a focus on monitoring the correlation between the dynamic change curve of glomerular filtration rate and the area under the drug plasma concentration-time curve. Preferably, adverse reaction feature tracking employs a Bayesian structural equation model in a dynamic causal network, analyzing the interaction terms between drug exposure dose and specific single nucleotide polymorphism (SNP) sites. The warning signal trigger threshold can be set based on clinical trial data in the drug's package insert.

[0199] In the economic assessment dimension, comprehensive cost calculation integrates three elements: drug costs, monitoring costs, and reimbursement policies. Preferably, drug costs are obtained through real-time price data from the hospital's procurement system, monitoring costs are calculated based on the examination frequency stipulated in the clinical pathway, and reimbursement policies are automatically calculated through a medical insurance catalog matching engine.

[0200] Treatment adherence assessment is based on lifestyle behavioral parameters such as electronic pillbox opening records and outpatient follow-up intervals. Survival analysis is used to predict the risk of treatment discontinuation. Survival analysis involves using statistical models to analyze time data "from treatment initiation to discontinuation" to assess the impact of various factors (such as medication records and side effects) on treatment continuity. Here, it refers to using historical patient behavior data to predict the probability of premature treatment termination.

[0201] Preferably, the strategy priority scoring generation adopts a multi-criteria decision analysis framework, where the weight allocation of each dimension follows the efficacy-first principle recommended by clinical guidelines. The scoring mapping process is implemented through vector space projection, transforming each strategy option in the initial strategy set into a feature vector containing three dimensions: efficacy, safety, and cost, which is ultimately stored in the graph structure nodes of the strategy knowledge base.

[0202] This embodiment achieves precise quantitative evaluation of treatment strategies by constructing a three-dimensional evaluation system encompassing efficacy, safety, and cost-effectiveness. A response prediction model based on viral mutation characteristics significantly improves the prognostic prediction capability of antiviral treatment regimens; a monitoring system integrating pharmacokinetics and genetic characteristics provides molecular-level decision-making basis for safety assessment; and the combination of an automated medical insurance policy matching engine and behavioral data analysis makes cost-effectiveness assessment more clinically practical. While maintaining a balance among the evaluation dimensions, this embodiment provides objective quantitative evidence for strategy optimization through a standardized scoring system.

[0203] In some embodiments, a complication prediction sub-model is established in the hierarchical reinforcement learning model, including:

[0204] Construct a risk propagation map of complications, using portal pressure parameters and organ function compensation indicators in the early warning nodes of complications as initial node features;

[0205] A functional network connecting the liver, portal system, and distal organs is established using graph convolution operations.

[0206] It receives the inflammation activity grading information and liver fibrosis stage information input from the pathology grading node and calculates the portal hypertension progression rate.

[0207] Obtain drug sensitivity parameters from treatment response points to assess the effectiveness of the current treatment regimen in suppressing complications;

[0208] When the portal pressure gradient exceeds the compensation threshold, the risk factor for esophageal varices is marked.

[0209] When blood ammonia levels remain elevated and are accompanied by neuropsychiatric symptoms, a warning signal for hepatic encephalopathy is triggered.

[0210] In this embodiment, preferably, the portal vein pressure parameter in the complication early warning node specifically refers to the hepatic vein pressure gradient value measured by hepatic vein catheterization, and the organ function compensation index uses serum albumin, bilirubin, and prothrombin time parameters from the Child-Pugh scoring system. The construction process of the complication risk transmission map strictly follows the anatomical characteristics of the portal system, wherein the connection relationship between the liver node and the spleen node is weighted using portal vein hemodynamic parameters.

[0211] The graph convolution operation is implemented based on the predefined vascular topology in the individual liver disease knowledge graph. The adjacency matrix used in the operation is consistent with the node connection relationship of the individual liver disease knowledge graph in the previous embodiment. The functional association network of liver-portal system-distal organs is processed through a three-layer graph structure: the first layer processes the local connections of intrahepatic vascular branches, the second layer integrates the hydrodynamic parameters of the main portal vein and hepatic artery, and the third layer outputs the coupling state representation of the portal system and systemic circulation.

[0212] Preferably, the inflammation activity grading information input to the pathological grading node comes from the METAVIR score of liver biopsy, and the liver fibrosis staging information is obtained from the liver stiffness value measured by transient elastography. The portal hypertension progression rate calculation module outputs clinically validated results by comparing the changes in hepatic vein pressure gradient values ​​from two consecutive follow-up tests.

[0213] Preferably, the drug sensitivity parameter fed back by the treatment response node specifically refers to the percentage change in hepatic vein pressure gradient before and after the use of non-selective β-blockers; the inhibition effect assessment module has a built-in response standard (a decrease in hepatic vein pressure gradient ≥10% is considered effective), and when the standard is not met in two consecutive assessments, the hierarchical reinforcement learning model strategy adjustment mechanism is triggered.

[0214] This embodiment constructs an anatomically constrained complication risk propagation map, transforming portal pressure parameters and organ function compensation indicators into graph-structured node features. A functional association network is established using a graph convolutional network to connect the liver, portal system, and distal organs, enabling dynamic quantitative assessment of complication risk. By integrating inflammatory activity and liver fibrosis staging information from pathological grading nodes, combined with drug sensitivity parameters from treatment response nodes, the system can accurately calculate the rate of portal hypertension progression and assess the inhibitory effect of treatment regimens on complications in real time. When the portal pressure gradient exceeds the compensation threshold or abnormal blood ammonia levels are accompanied by neuropsychiatric symptoms, the hierarchical reinforcement learning model automatically triggers esophageal variceal risk markers and hepatic encephalopathy warning mechanisms. This embodiment ensures that the complication prediction sub-model retains the anatomical features of portal system hemodynamics while fully integrating clinical pathological grading and treatment response data. It can identify high-risk complications in advance during the initial candidate treatment strategy generation stage, providing precise decision-making basis for the dynamic insertion of preventative treatment plans and effectively improving the proactive defense capability and clinical safety of liver disease treatment strategies.

[0215] In some embodiments, generating and outputting a final set of candidate treatment strategies includes:

[0216] Based on the risk factor of esophageal varices, a first preventive treatment plan was generated, which included the frequency of endoscopic monitoring, the administration of beta-blockers, and an emergency hemostasis plan.

[0217] In response to early warning signs of hepatic encephalopathy, a second preventative treatment plan was generated, which included adjusting the lactulose dosage, restricting protein intake, and monitoring blood ammonia levels.

[0218] Synergistic optimization of the primary and / or secondary preventative treatment plans with the initial candidate treatment strategies includes:

[0219] The compatibility of the first and / or second preventive measures with the initial candidate treatment strategies is evaluated to obtain the first optimization result;

[0220] The dosage and dosing sequence of the drug combination were adjusted based on the first optimization result to obtain the second optimization result;

[0221] Based on the second optimization results, the risk-benefit ratio of the overall treatment plan is recalculated, and a final candidate treatment strategy with complication prevention and control markers is generated. The final candidate treatment strategy includes efficacy risk assessment information of the basic treatment plan, the execution priority of complication prevention measures, and the time sequence planning diagram of the overall treatment path.

[0222] A final candidate treatment strategy set is generated based on multiple final candidate treatment strategies.

[0223] In this embodiment, when generating the first preventive treatment plan, the frequency of endoscopic monitoring is determined based on the classification results of the risk coefficient of esophageal varices, with higher-risk patients receiving more intensive monitoring cycles; the β-blocker administration plan uses portal pressure change data recorded at treatment response points for personalized dose titration; and the emergency hemostasis plan integrates the combined treatment process of endoscopic banding and drug hemostasis.

[0224] When generating the second preventative treatment plan, the lactulose dosage was dynamically adjusted based on blood ammonia levels, the protein restriction diet was implemented according to the liver function compensation capacity classification, and the blood ammonia monitoring plan was calibrated in conjunction with the ammonia metabolism parameters in serum test indicators.

[0225] The collaborative optimization process is achieved through the strategy layer of the hierarchical reinforcement learning model: the compatibility assessment stage first detects the interaction risk between the preventive treatment plan and antiviral drugs and hepatoprotective drugs; the dosage adjustment stage optimizes the dosing sequence based on the liver metabolic capacity data provided by the pathological grading node; the risk-benefit ratio calculation module simultaneously integrates the time series features of the efficacy evaluation curve and the complication warning signal to ensure that the execution priority of the label matches the basic treatment stage.

[0226] This embodiment deeply integrates complication prevention and control measures into basic treatment strategies through a hierarchical optimization mechanism. For example, when a patient has both high viral load and esophageal varices, the system automatically avoids metabolic conflicts between non-selective beta-blockers and nucleoside analogs. The final output time-series planning graph achieves three-dimensional synergy between antiviral therapy, liver function maintenance, and complication prevention. While retaining the core advantages of personalized knowledge graphs, it further enhances the clinical feasibility and safety of multimodal treatment strategies.

[0227] In a second aspect, this embodiment also provides a computer-readable storage medium storing computer program instructions thereon, which, when executed by a processor, implement the method described in the first aspect.

[0228] The computer program involved in this embodiment can be stored in a computer device readable storage medium, which includes, but is not limited to, disks, magnetic tapes, magnetic cards, floppy disks, flash memory, optical disks, optical cards, read-only memory (ROM), random access memory (RAM), erasable programmable ROM (EPROM), and electrically erasable programmable ROM (EEPROM), etc. It also includes other biological, physical, or chemical structures capable of performing similar or equivalent functions to the storage media listed above, such as DNA, RNA, proteins, and other units with information storage capabilities. In specific embodiments, the storage medium involved can be one of the above-mentioned media types or a combination of the above media types. In different embodiments, the computer program involved in the embodiment can be centrally stored in a single medium or distributed across multiple media. The memory containing the computer device readable storage medium can be non-volatile memory or random access memory. These computer device readable storage media can be built into the device or connected to the device involved in the embodiment as an external device or part of an external device. In some embodiments, the memory having a computer device readable storage medium is deployed locally; in other embodiments, the memory may be deployed remotely from the processor, for example, as a network-attached memory accessed via RF circuitry or an external port and a communication network, wherein the communication network may be the Internet, one or more intranets, a local area network (LAN), a wide area network (WLAN), a storage area network (SAN), or a suitable combination thereof, as long as computer device access to the memory is enabled. Furthermore, the computer program involved in the embodiments may be stored in plaintext / ciphertext form, or it may be designed as training data, integrated and recombined through model training and implicitly stored in the parameter states of a deep neural network or other machine learning model.

[0229] Please see Figure 5 In a third aspect, this embodiment also provides an electronic device 1, including a memory 11 and a processor 12, wherein the memory 11 is used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor 12 to implement the method described in the first aspect.

[0230] The processor described in this embodiment can be implemented by hardware, firmware, software, or a combination thereof. It can be a circuit, one or more of an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, or a microprocessor. It also includes other physical, biological, or chemical structures that can implement the same or equivalent functions as the processors listed above, such as biological neurons, quantum computing units, DNA computing units, etc., so that the processor can execute some or all of the steps in the computer program or method involved in the various embodiments of this application, or any combination of the steps mentioned therein.

[0231] Unlike existing technologies, the above technical solution has the following beneficial effects:

[0232] This invention constructs a spatiotemporally correlated individual liver disease knowledge graph, integrating multi-source data such as serum test indicators and medical imaging features to establish a dynamic correlation system including etiological feature nodes, pathological grading nodes, complication early warning nodes, and treatment response nodes. This knowledge graph retains molecular features such as viral load and metabolic abnormality markers while incorporating clinical indicators such as liver fibrosis staging and portal pressure parameters, providing a structured decision-making basis for hierarchical reinforcement learning models. The hierarchical reinforcement learning model employs a three-level reasoning framework consisting of a strategy decision layer, a principle optimization layer, and an execution evaluation layer to achieve synergistic optimization of antiviral treatment regimens, hepatoprotective drug combinations, and metabolic interventions. By analyzing viral mutation trends, constructing a multi-objective decision space, and combining pharmacokinetic parameters, an initial set of priority-ranked candidate treatment strategies is generated. The complication prediction sub-model establishes the functional correlation between the liver and portal system through a graph convolutional network, dynamically monitors the progression of portal hypertension, and intelligently inserts preventative treatment plans to ensure seamless integration of complication prevention and basic treatment.

[0233] This invention achieves continuous evolution of diagnosis and treatment strategies through the dynamic updating mechanism of knowledge graphs and the closed-loop optimization of hierarchical reinforcement learning models. It can respond to changes in patients' conditions in real time, providing clinicians with decision support that combines standardized guidelines with individualized risk warnings, and effectively improving the prognosis management and complication prevention of liver disease treatment.

[0234] Finally, it should be noted that although the above embodiments have been described in the text and drawings of this application, this should not limit the scope of patent protection of this application. Any technical solutions that are based on the essential concept of this application and utilize the content described in the text and drawings of this application, resulting in equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this application.

Claims

1. A knowledge graph-based method for recommending liver disease diagnosis and treatment strategies, characterized in that, include: Collect basic user information, including serum test indicators, medical imaging features, gene test information, and lifestyle parameters; Based on user basic information, a spatiotemporally related individual liver disease knowledge graph is constructed. The individual liver disease knowledge graph includes interconnected etiology feature nodes, pathological grading nodes, complication warning nodes, and treatment response nodes. Etiology feature nodes are configured to establish connections with viral load and metabolic abnormality markers. Pathological grading nodes are configured to label the grading information of liver fibrosis stage and inflammatory activity. Complication warning nodes are configured to be associated with portal pressure parameters and organ function compensation indicators. Treatment response nodes are configured to record drug sensitivity and adverse reaction events. The individual liver disease knowledge graph is input into the hierarchical reinforcement learning model, which performs joint reasoning on the etiology feature node, pathological grade node, and treatment response node, and outputs multiple sets of strategy options including antiviral treatment regimens, hepatoprotective drug combinations, and metabolic intervention measures. Each set of strategy options is accompanied by an expected efficacy assessment curve and a risk probability matrix. An initial set of candidate treatment strategies is generated based on multiple sets of strategy options; Furthermore, a complication prediction sub-model is established in the hierarchical reinforcement learning model. The complication prediction model analyzes the topological propagation path of the complication warning nodes through a graph convolutional network. When the risk coefficient of esophageal varices exceeds the preset risk threshold or the warning signal of hepatic encephalopathy continues to appear, a preventive treatment plan is dynamically inserted into the initial candidate treatment strategy set to generate and output the final candidate treatment strategy set. The individual liver disease knowledge graph corresponding to the current user will be updated the next time the user's basic information is retrieved. A spatiotemporally related individual liver disease knowledge graph is constructed based on basic user information, including: Initialize an individual liver disease knowledge graph, which includes etiological feature nodes, pathological grading nodes, complication warning nodes, and treatment response nodes; The collected basic user information is mapped to the etiology feature nodes, pathological grading nodes, complication warning nodes, and treatment response nodes of the individual liver disease knowledge graph; Align the timestamps of the etiology feature nodes, pathology grading nodes, complication warning nodes, and treatment response nodes on the time dimension axis to create a time index; Cross-node association weights are calculated using a three-layer attention mechanism, which includes a first operational layer, a second operational layer, and a third operational layer. The first operational layer calculates the pathological influence coefficient of viral load and liver fibrosis stage, the second operational layer calculates the interaction strength between portal pressure and drug response, and the third operational layer integrates the behavioral influence coefficient of lifestyle parameters on the pathological influence coefficient and the interaction strength. The next time the user's basic information is retrieved, the individual liver disease knowledge graph corresponding to the current user will be updated, including: When the current user inputs their basic information, the viral load time-series curve in the etiology feature node is updated and the inflammation activity grading information in the pathology grading node is recalculated, resulting in the output of an updated individual liver disease knowledge graph.

2. The method for recommending liver disease diagnosis and treatment strategies based on knowledge graphs according to claim 1, characterized in that, Initialize the individual liver disease knowledge graph, including: The etiological feature nodes are configured to connect viral load time-series curves and metabolic abnormality marker fluctuation data via a Bayesian network, including: A bivariate Gaussian mixture model of viral genotype and liver enzyme profile was constructed. The viral load sequence and metabolic biomarker ratio sequence in serum detection indicators were used as input features. The metabolic regulation condition probability corresponding to different viral load ranges was calculated through the bivariate Gaussian mixture model. The variational inference algorithm was used to solve the latent variable posterior distribution of mutation sites in the core protein region and serological conversion events, and the causal weights of core protein mutation and serological conversion were obtained. The dynamic propagation equation of metabolic regulation parameters is established based on the conditional probability and causal weight of metabolic regulation to quantify the feedback inhibition effect of steatosis on viral replication rate. The dynamic propagation equation of metabolic regulation parameters is established based on the conditional probability and causal weight of metabolic regulation to obtain a first output vector containing viral mutation characteristic parameters, metabolic compensation parameters and dynamic causal network parameters. The metabolic compensation parameters in the first output vector are weighted and connected to the inflammation activity calculation module of the pathological grading node. The pathological grading nodes were configured to use a semi-supervised clustering algorithm to label the evolution trend of liver fibrosis staging and inflammatory activity, including: A joint embedding space of liver tissue elasticity parameters and serum fibrosis indicators is constructed. The elasticity parameter sequence in medical image features and the metabolic compensation parameter in the first output vector are used as input features. The feature dimensionality is reduced by the t-SNE algorithm and the feature representation is optimized by a semi-supervised graph convolutional network to obtain low-dimensional features that integrate metabolic regulation information. A dual-channel gated recurrent unit network model is constructed to model inflammation fluctuations and fibrosis progression. The low-dimensional features are then input into the dual-channel gated recurrent unit network model to obtain a second output vector containing spatiotemporal joint distribution features. The second output vector is input into the multi-organ risk assessment module of the complication early warning node through an attention gating mechanism.

3. The method for recommending liver disease diagnosis and treatment strategies based on knowledge graphs according to claim 2, characterized in that, The complication early warning node is configured to correlate the gradient of portal pressure parameters with the decay rate of organ function compensation indicators via a gated graph neural network, including: An organ function association graph of the portal vein-systemic circulation system is constructed. The spatiotemporal joint distribution features in the second output vector are used as topological initialization features. The functional coupling strength of spleen-kidney-brain is dynamically calculated through graph attention mechanism to obtain the organ function association strength matrix. Based on hemodynamic principles, the propagation process of pathological changes on the organ function correlation map is simulated, the influence coefficient of portal pressure gradient on distal organ function is quantified, and the organ function correlation strength matrix is ​​adjusted. When the influence coefficient of the portal pressure gradient on distal organ function exceeds the clinical warning threshold, a third output vector containing multi-organ risk level assessment and priority intervention target identification is generated. The third output vector is input into the multi-objective optimization module of the treatment response node through feature cross-fusion with adaptive weights. The treatment response node is configured to construct a drug sensitivity prediction model and a causal inference chain for adverse reaction events, including: A drug-metabolizing enzyme genotype-plasma drug concentration response surface was constructed. Gene detection information and the third output vector were used as input features. The dose-response relationship between genotype and drug metabolism kinetics was established through piecewise nonlinear regression. A multi-task deep network model incorporating virological response rate and renal function protection was constructed. The input layer of the multi-task deep network model was configured to integrate blood drug concentration prediction results and organ risk characteristics. The minimum sufficient causal path between drug combinations and specific adverse reactions is identified by a constrained causal discovery algorithm, resulting in a fourth output vector that includes a combination of time-dependent dosing regimens and risk avoidance strategies. The pharmacokinetic parameters in the fourth output vector are updated with the dynamic causal network parameters in the first output vector via feedback connections.

4. The method for recommending liver disease diagnosis and treatment strategies based on knowledge graphs according to claim 1, characterized in that, The individual liver disease knowledge graph is input into a hierarchical reinforcement learning model, which performs joint reasoning on etiology feature nodes, pathological grading nodes, and treatment response nodes. The model outputs multiple strategy options, including antiviral treatment regimens, hepatoprotective drug combinations, and metabolic interventions. A hierarchical reinforcement learning model is established, which includes a three-level reasoning framework of strategy decision-making layer, principle optimization layer and execution evaluation layer. The dynamic causal network parameters in the etiological feature nodes are analyzed through the strategy decision-making layer, the long-term evolution trend of virus mutation and treatment pressure is calculated, and the strategy target vector is obtained. The spatiotemporal joint distribution features of the strategy objective vector and pathological grading nodes are input into the principle optimization layer, and a multi-objective decision space containing efficacy surface and safety boundary is constructed in the principle optimization layer. In the execution evaluation layer, pharmacokinetic parameters are extracted from the treatment response node and mapped to the multi-objective decision space; A non-dominated sorting genetic algorithm is used to solve the multi-objective decision space and generate an initial strategy set. Each initial strategy in the initial strategy set is evaluated in multiple dimensions, and an initial candidate treatment strategy set with priority ranking is output.

5. The method for recommending liver disease diagnosis and treatment strategies based on knowledge graphs according to claim 4, characterized in that, The multidimensional assessment includes efficacy assessment, safety assessment, and economic assessment dimensions: In terms of efficacy evaluation, the inhibitory effect of antiviral treatment regimens on viral load is analyzed, and the trend of virological response is predicted based on viral mutation characteristic parameters of etiological feature nodes. In addition, the study evaluated the effect of hepatoprotective drug combinations on improving inflammatory activity and estimated the potential for improving liver fibrosis by combining the spatiotemporal distribution characteristics of pathological grading nodes. Generate the first evaluation result; In terms of safety assessment, the impact of treatment regimens on renal function is monitored, and the risk of renal function impairment is identified based on pharmacokinetic parameters at treatment response points. In addition, it tracks the adverse reaction characteristics of drug combinations and uses the dynamic causal network parameters of etiological feature nodes to predict the probability of occurrence of specific adverse reactions. Generate a second evaluation result; In terms of economic assessment, the overall cost of treatment options is calculated, integrating drug costs, monitoring costs, and reimbursement policies. Assess treatment adherence and predict treatment completion rate by combining medication record characteristics in lifestyle behavior parameters; Generate a third evaluation result; A strategy priority score is generated based on the first evaluation result, the second evaluation result, and the third evaluation result; The policy priority score is mapped and stored with the initial policy.

6. The method for recommending liver disease diagnosis and treatment strategies based on knowledge graphs according to claim 1, characterized in that, A sub-model for predicting complications is established within the hierarchical reinforcement learning model, including: Construct a risk propagation map of complications, using portal pressure parameters and organ function compensation indicators in the early warning nodes of complications as initial node features; A functional network connecting the liver, portal system, and distal organs is established using graph convolution operations. It receives the inflammation activity grading information and liver fibrosis stage information input from the pathology grading node and calculates the portal hypertension progression rate. Obtain drug sensitivity parameters from treatment response points to assess the effectiveness of the current treatment regimen in suppressing complications; When the portal pressure gradient exceeds the compensation threshold, the risk factor for esophageal varices is marked. When blood ammonia levels remain elevated and are accompanied by neuropsychiatric symptoms, a warning signal for hepatic encephalopathy is triggered.

7. The method for recommending liver disease diagnosis and treatment strategies based on knowledge graphs according to claim 6, characterized in that, Generate and output the final set of candidate treatment strategies, including: Based on the risk factor of esophageal varices, a first preventive treatment plan was generated, which included the frequency of endoscopic monitoring, the administration of beta-blockers, and an emergency hemostasis plan. In response to early warning signs of hepatic encephalopathy, a second preventative treatment plan was generated, which included adjusting the lactulose dosage, restricting protein intake, and monitoring blood ammonia levels. Co-optimizing the first and / or second preventative treatment plans with the initial candidate treatment strategies includes: The compatibility of the first and / or second preventive measures with the initial candidate treatment strategies is evaluated to obtain the first optimization result; The dosage and dosing sequence of the drug combination are adjusted based on the first optimization result to obtain the second optimization result; Based on the second optimization result, the risk-benefit ratio of the overall treatment plan is recalculated, and a final candidate treatment strategy with complication prevention and control markers is generated. The final candidate treatment strategy includes efficacy risk assessment information of the basic treatment plan, execution priority of complication prevention measures, and time-series planning diagram of the overall treatment path. The final candidate treatment strategy set is generated based on multiple final candidate treatment strategies.

8. A computer-readable storage medium storing computer program instructions thereon, characterized in that, The computer program instructions, when executed by a processor, implement the method as described in any one of claims 1 to 7.

9. An electronic device comprising a memory and a processor, characterized in that, The memory is used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement the method as described in any one of claims 1 to 7.

Images