An intelligent analysis system driven by a knowledge graph in the medical field

Abstract

The application relates to the technical field of medical information, and discloses a knowledge graph driving intelligent analysis system based on the medical field, which comprises a knowledge graph construction module, which is used for constructing a medical knowledge graph containing subjects, events, object entities and information accessibility relationships; a behavior pattern learning module, which is used for learning a parameterized behavior decision model for the subject entity; a dynamic deduction module, which is used for deducing market changes after injecting events according to the behavior decision model and the information accessibility relationship; and a strategy generation module, which is used for taking the dynamic deduction process as a fitness function to reversely generate a market strategy. The method comprises the following steps: constructing the medical field knowledge graph; learning the behavior decision model of the subject entity based on the graph; dynamically deducing the market changes in a virtual environment; and reversely solving and generating the market strategy. The application can perform forward-looking dynamic deduction on the market and automatically generate an optimized strategy, thereby improving the scientificity and timeliness of decision-making.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and in particular to a knowledge graph-driven intelligent analysis system based on the medical field. Background Technology

[0002] In today's highly competitive and complex pharmaceutical market, pharmaceutical companies must conduct precise market analysis and formulate effective market strategies to maintain and increase their market share. To this end, companies typically utilize a variety of data resources, including drug sales data, clinical application data, market research reports, and industry policy information.

[0003] Existing market analysis techniques primarily rely on statistical analysis of historical data and business intelligence (BI) tools to uncover patterns and correlations in past market performance. These methods can effectively summarize past market trends and monitor key performance indicators. However, such techniques are inherently retrospective and have limitations in predicting the dynamic evolution of the market under the impact of specific events (such as the implementation of new policies or the introduction of new strategies by competitors). Furthermore, while existing analytical methods can reveal static relationships between market entities, they typically struggle to deeply characterize the dynamic behavioral decision-making mechanisms of various market participants (such as doctors and patient groups), and cannot effectively simulate the macro-market chain reactions triggered by the convergence of these micro-behaviors.

[0004] Therefore, when formulating forward-looking market strategies, decision-makers often lack the technical tools to quantitatively extrapolate the potential consequences of different strategy combinations and to automatically optimize them, resulting in the decision-making process still relying heavily on experience and qualitative judgment.

[0005] Therefore, this invention proposes a knowledge graph-driven intelligent analysis system based on the medical field to address the shortcomings of existing technologies. Summary of the Invention

[0006] The purpose of this invention is to provide a knowledge graph-driven intelligent analysis system based on the pharmaceutical field. This system addresses the problem that data analysis techniques in the pharmaceutical field are usually limited to static analysis and relationship mining of historical data, lacking the ability to make forward-looking predictions of future market dynamics and proactively generate optimization strategies, resulting in lag and limitations in decision support.

[0007] To address the aforementioned technical problems, the first aspect of this invention provides a knowledge graph-driven intelligent analysis system based on the pharmaceutical field.

[0008] The system includes: a knowledge graph construction module, a behavior pattern learning module, a dynamic inference module, and a strategy generation module.

[0009] The knowledge graph construction module is used to receive and integrate heterogeneous data from multiple sources, including structured and unstructured data, to construct a knowledge graph for the pharmaceutical field. In one specific implementation, the knowledge graph construction module utilizes natural language processing technology to extract entities and relationships from unstructured data such as medical literature and industry reports. These entities include subject entities that are decision-makers (e.g., doctors, patient groups), event entities that can influence market changes (e.g., policy releases, academic conferences), and objective entities in the market (e.g., drugs, diseases). Simultaneously, the module also constructs information reachability relationships between the subject entities and the event entities. Finally, the extracted information is integrated with structured data to form the pharmaceutical knowledge graph.

[0010] The behavior pattern learning module is used to generate parameterized behavior decision models for each entity based on the medical knowledge graph. This module aggregates information from the medical knowledge graph using a graph neural network, generating an embedding vector for each entity node. The embedding vector integrates the entity's own attributes and its complex relationships within the graph, serving as the basis for determining the specific parameters of its behavior decision model. The graph neural network can be a graph attention network or a graph convolutional network, and its iterative neighborhood information aggregation process can be described by the following formula:

[0011]

[0012] In the formula, Let v be the embedding vector of node v at the k-th layer; f is the embedding vector of node v at the (k-1)th layer; upd For the update function; f agg It is an aggregate function; e is the embedding vector of neighbor node u at layer k-1; uv The feature vector of the edge connecting node u and node v; Let v be the set of neighboring nodes.

[0013] In another specific implementation, the behavior pattern learning module is also used to handle cold-start scenarios where the target entity lacks historical data. In this case, the module calculates the similarity between the associated objects of the target entity and existing object entities in the medical knowledge graph, and based on the similarity, selects the parameters of the learned behavior decision model associated with the existing object entity, and generates an initial behavior decision model for the target entity through transfer learning technology.

[0014] The dynamic simulation module is used to instantiate the subject entities in a virtual market environment and dynamically simulate market changes based on the behavioral decision model. When an event entity is injected, the module identifies the set of target subject entities affected by the event entity based on the pre-built information accessibility relationships in the medical field knowledge graph. Subsequently, the module applies an adjustment function to modify the behavioral decision model of the target subject entities in real time to simulate adaptive changes in their behavior. The adjustment process can be described by the following formula:

[0015] π ′ a =g(π) a ,e);

[0016] In the formula, π ′ a The adjusted behavioral decision-making model; π a is the behavioral decision model before adjustment; g is the adjustment function; e is the event entity.

[0017] The strategy generation module is used to inversely solve for and generate market strategies based on the output of the dynamic deduction module. This module defines market strategies (e.g., marketing budgets for different regions, drug pricing) as decision variables in a multi-objective optimization algorithm. It then executes a complete market change deduction process from the initial state to the final state through the dynamic deduction module, using this as a fitness evaluation function to assess the merits of a decision variable. Through iterative optimization of the algorithm, a set of market strategies that ensures the fitness evaluation function results satisfy Pareto optimality is ultimately output.

[0018] In another specific implementation, the system further includes an attribution alignment module. This module is used to record and store the micro-agent behavior chains of all entities during the dynamic inference process. When it is necessary to explain a certain macro-market change, this module can automatically align the key micro-agent behavior chains that contributed to the change with the specific facts and relationships in the medical field knowledge graph, thereby providing a traceable explanatory path from the macro-inference results to the micro-graph facts.

[0019] A second aspect of this invention provides a knowledge graph-driven intelligent analysis method based on the pharmaceutical field. The method is applied to the system described in any of the foregoing embodiments and includes the following steps:

[0020] S1. Integrate multi-source heterogeneous data to construct a medical domain knowledge graph that includes subject entities, event entities, object entities, and the information reachability relationships between the subject entities and the event entities;

[0021] S2. Based on the knowledge graph of the pharmaceutical field, a parameterized behavior decision model is learned and generated for the subject entity through a graph neural network. The parameters of the behavior decision model are determined by the embedding vector of the subject entity in the knowledge graph of the pharmaceutical field.

[0022] S3. Instantiate the subject entity in the virtual market environment, and dynamically deduce market changes based on the behavioral decision model and the target subject entity affected by the injected event entity identified based on the information accessibility relationship; S4. Use the complete market change deduction process of the dynamic deduction module as a fitness evaluation function, solve it in reverse through a multi-objective optimization algorithm, and generate a market strategy.

[0023] In summary, the present invention has at least one of the following beneficial technical effects:

[0024] 1. By constructing a multidimensional medical knowledge graph that includes subject entities, event entities, and object entities, and by learning a data-driven behavioral decision-making model for each subject entity, this invention can deeply depict the complex dynamic interactions of various participants in the market, rather than just mining static relationships, thereby improving the depth and systematicness of analysis when extrapolating market changes.

[0025] 2. The dynamic extrapolation mechanism proposed in this invention can simulate the impact of injected event entities on the behavior of target entities in a virtual market environment, and proactively extrapolate the resulting chain reactions in the market. This method enables the system to extrapolate future market scenarios based on different assumptions, avoiding the limitations of traditional technologies that can only analyze historical data.

[0026] 3. This invention combines a dynamic deduction process with a multi-objective optimization algorithm to construct a reverse strategy generation module, which can automatically solve and generate market strategy combinations based on preset business objectives. This design transforms the system from a passive analysis tool into a proactive decision support system, providing a direct and optimized solution for market strategy formulation.

[0027] 4. The attribution alignment module introduced in this invention can link macro-level market projection results with underlying micro-level agent behavior chains and knowledge graph facts, opening the black box of complex model analysis. This design provides a traceable explanatory path from phenomenon to cause, enhancing the credibility and interpretability of the analysis results. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the functional module structure of the present invention;

[0029] Figure 2 This is a schematic diagram of the method flow of the present invention;

[0030] Figure 3 This is a schematic diagram of the entity and relation structure of the medical field knowledge graph of the present invention;

[0031] Figure 4 This is a schematic diagram of the interaction process between event injection and behavior adaptation in the dynamic deduction module of the present invention.

[0032] The modules are: 10. Knowledge Graph Construction Module; 20. Behavioral Pattern Learning Module; 30. Dynamic Inference Module; 40. Strategy Generation Module; and 50. Attribution Alignment Module. Detailed Implementation

[0033] The following is in conjunction with the appendix Figure 1 - Appendix Figure 4 The present invention will be further described in detail below.

[0034] See attached document Figure 1 This illustrates the structure of a knowledge graph-driven intelligent analysis system based on a medical field according to an embodiment of the present invention. The system may include: a knowledge graph construction module 10, a behavior pattern learning module 20, a dynamic inference module 30, and a strategy generation module 40. In an optional embodiment, the system may further include an attribution alignment module 50.

[0035] See attached document Figure 2 This illustrates the workflow of a knowledge graph-driven intelligent analysis method based on the pharmaceutical field, according to an embodiment of the present invention. The following will be discussed in conjunction with the appendix. Figure 1 , Figure 3 and Figure 4 The method is explained in detail for each step.

[0036] Step S1: The knowledge graph construction module 10 performs the knowledge graph construction operation. This module receives and integrates multi-source heterogeneous data in the pharmaceutical field, including structured data (e.g., drug registration information, clinical trial databases) and unstructured data (e.g., medical literature, industry reports, news information). (See Appendix) Figure 3 The knowledge graph construction module 10 extracts three types of entities from the data and constructs relationships between them. Specifically, entities include: principal entities that participate in market decision-making (e.g., doctor entities, patient groups entities), event entities that can dynamically influence market changes (e.g., policy release event entities, academic conference event entities), and objective object entities in the market (e.g., drug entities, disease entities). Simultaneously, the knowledge graph construction module 10 also identifies and constructs information reachability relationships between the principal entities and the event entities, defining the scope of entities that a specific event can influence. All extracted entities and relationships are integrated to construct the medical field knowledge graph.

[0037] Step S2: The behavior pattern learning module 20 performs behavior pattern learning operations. This module takes the medical field knowledge graph constructed by the knowledge graph construction module 10 as input, and learns and generates parameterized behavior decision models for each entity in the knowledge graph using a graph neural network. Specifically, the behavior pattern learning module 20 iteratively aggregates neighborhood information in the knowledge graph using a graph neural network to generate embedding vectors for each entity node.

[0038] Step S3: The dynamic simulation module 30 performs a market change dynamic simulation operation. (See attached document.) Figure 4 The module first instantiates the subject entities in the virtual market environment. Each instantiated subject entity is assigned the behavioral decision model generated by the behavioral pattern learning module 20. The inference process proceeds in discrete time steps. When an external event entity is injected into the virtual market environment, the dynamic inference module 30 identifies the target subject entity affected by the event entity based on the information accessibility relationships defined in the knowledge graph. Subsequently, module 30 applies an adjustment function to modify the behavioral decision model of the target subject entity in real time to simulate its adaptive behavioral response to the event entity.

[0039] Step S4: The strategy generation module 40 performs the market strategy generation operation. This module receives the deduction capability of the dynamic deduction module 30 as input, and generates market strategies by inversely solving the problem using a multi-objective optimization algorithm. Specifically, market strategies (such as marketing budgets for different regions, drug pricing, etc.) are defined as decision variables of the multi-objective optimization algorithm. The dynamic deduction module 30 performs a complete market change deduction process from the initial state to the final state, which is used as a fitness evaluation function to evaluate the merits of a decision variable. Through iterative optimization, the strategy generation module 40 finally outputs a set of market strategies that ensure the fitness evaluation function results satisfy Pareto optimality.

[0040] In the above embodiments, the system may further include an attribution alignment module 50. This module is used to record and store the micro-agent behavior chains during the dynamic inference process. When it is necessary to explain a certain macro market change, this module can automatically align the key micro-agent behavior chains that contributed to the change with the specific facts and relationships in the medical field knowledge graph, thereby providing a traceable explanatory path from the macro inference results to the micro-graph facts.

[0041] See attached document Figure 3 This diagram illustrates the entity and relation structure of a medical field knowledge graph according to an embodiment of the present invention. This section will describe in detail the specific implementation process of the knowledge graph construction module 10.

[0042] This knowledge graph construction module 10 is responsible for constructing a medical field knowledge graph that serves as the information foundation and rule base for the entire system. Its detailed implementation process includes the following steps.

[0043] First, the knowledge graph construction module 10 receives and processes multi-source heterogeneous data. The data sources can be divided into two categories: one is structured data, such as drug registration information databases, clinical trial results databases, and market sales record data; the other is unstructured data, such as publicly published medical journal articles, pharmaceutical industry news and information, and policy documents issued by government departments.

[0044] Next, the knowledge graph construction module 10 processes the received unstructured data. Using natural language processing techniques, it identifies and extracts predefined entities from the text. This process can employ a Named Entity Recognition (NER) model, which, after training, can identify three core entity types:

[0045] 1. The main entity, i.e., the decision-making participants in the market, such as a specific doctor or a patient group in a specific region;

[0046] 2. Event entities, namely, various events that can have a dynamic impact on the market, such as the convening of a specific academic conference or the release of a policy to adjust the coverage of medical insurance;

[0047] 3. Object entities, that is, objective things in the market environment, such as specific drug names or disease names.

[0048] Following entity extraction, the knowledge graph construction module 10 further utilizes relation extraction (RE) technology to identify and extract relationships between entities. This includes not only static relationships between object entities (e.g., the therapeutic relationship between a drug and its indication), but also a key relationship type in this invention: information reachability relationships. These relationships define the connections between subject entities and event entities, specifically indicating which subject entities can perceive or directly influence a particular event. For example, by analyzing the list of participants or topics at an academic conference, an information reachability relationship can be constructed between the conference event entity and physician subject entities in a specific specialty field.

[0049] Meanwhile, the knowledge graph construction module 10 processes the structured data. Using preset mapping rules, it converts each record in the structured data table into nodes and attributes in the graph. For example, it converts a row of data in a drug database into a drug entity node, and uses the columns of that row (such as generic name, approval number, and manufacturer) as attributes of that node.

[0050] Finally, the knowledge graph construction module 10 performs entity alignment and data fusion operations. Entities extracted from unstructured data are linked and aligned with entities transformed from structured data to ensure that entities pointing to the same real-world object (e.g., a drug mentioned in a news article and the same drug in a database) are represented as the same unique node in the graph. After these steps, all processed entities, attributes, and relationships are loaded and stored in a graph database, forming the final medical domain knowledge graph that can be used by other modules of the system.

[0051] The core task of this behavioral pattern learning module 20 is to generate a personalized, data-driven behavioral decision-making model for each entity in the medical field knowledge graph generated by the knowledge graph construction module 10. The detailed implementation process of this module includes the following steps.

[0052] First, to convert the symbolic information in the knowledge graph into a form that can be numerically computed, the behavior pattern learning module 20 employs graph neural network technology to generate a low-dimensional, dense real-valued vector, or embedding vector, for each entity node in the graph. This embedding vector aims to encode the entity's own attributes and its contextual information within the graph's topology.

[0053] To obtain the embedding vector, the behavior pattern learning module 20 iteratively aggregates neighborhood information from the knowledge graph. In each iteration, each node aggregates information from its neighbors to update its own embedding vector. After multiple iterations, the embedding vector of each node can fuse structural and feature information from its multi-hop neighbors. This iterative neighborhood information aggregation process can be described by the following equation:

[0054]

[0055] In the formula, Let v be the embedding vector of node v at the k-th layer; f is the embedding vector of node v at the (k-1)th layer; upd For the update function; f agg It is an aggregate function; e is the embedding vector of neighbor node u at layer k-1; uv The feature vector of the edge connecting node u and node v; Let v be the set of neighboring nodes.

[0056] After completing the above iterative process, the behavior pattern learning module 20 obtains the final embedding vector for each subject entity. This embedding vector is then used to parameterize the subject's behavior decision model. This parameterization mechanism ensures that the specific form of the behavior decision model directly depends on the subject's comprehensive information in the knowledge graph, thus guaranteeing the model's personalization.

[0057] The behavioral decision-making model is used to calculate the probability distribution of various possible actions taken by the subject in a given market scenario. Based on the final embedding vector of the subject entity obtained above, the behavioral decision-making model can be formally defined by the following equation:

[0058] P(a|s,h v ) = softmax(W(h v )·s+b(h v ));

[0059] In the formula, P(a|s,h) v ) represents the embedding vector h under a given market scenario s. v The parameterized entity v represents the probability vector of various behaviors it may take; a is the set of all possible behaviors of the entity; s is the state vector describing the current market situation, which may include market price, competitor information, etc.; h v W(h) is the final embedding vector of the main entity v; v ) is a weight matrix, the specific values ​​of which are determined by a weight matrix with h as the weight. v Generate the input function; b(h) v ) is a bias vector, the specific value of which is also determined by a bias vector with h. v The function is generated for the input; softmax(·) is a standard function that transforms a real vector into a probability distribution.

[0060] Furthermore, in one embodiment, for newly added target entities lacking historical data, the behavior pattern learning module 20 also provides a cold start processing mechanism. The behavior pattern learning module 20 calculates the similarity between the target associated object and existing object entities in the knowledge graph, and based on the similarity, uses transfer learning techniques and parameters of the learned behavior decision model associated with the existing object entities to generate an initial behavior decision model for the target entity.

[0061] See attached document Figure 4 This figure illustrates the interaction process of event injection and behavior adaptation in a dynamic inference module according to an embodiment of the present invention. This section will describe in detail the specific implementation process of the dynamic inference module 30.

[0062] This dynamic simulation module 30, based on the output of the behavior pattern learning module 20, constructs a multi-agent simulation environment for simulating market dynamics. Its detailed implementation includes the following steps.

[0063] First, the dynamic inference module 30 initializes the simulation environment. This process instantiates the main entities in the knowledge graph within the virtual market environment. Each instantiated entity is assigned a corresponding parameterized behavioral decision model generated by the behavior pattern learning module 20. Simultaneously, based on the needs of the analysis task, the initial state of the market is set, such as the initial market share and price of each drug.

[0064] Next, the dynamic inference module 30 drives the simulation in discrete time steps. Within each time step, each instantiated entity first perceives the current market state, then makes a decision based on its own behavioral decision-making model, outputting the action it will take in that time step. For example, a doctor entity might, based on current market information and its decision-making model, output the action of prescribing a specific drug. After all entities have completed their decisions, the dynamic inference module 30 aggregates the results of all actions and updates the macro-state of the entire virtual market environment to proceed to the next time step.

[0065] At any time step during the simulation, this dynamic simulation module 30 supports the injection of external event entities. (See attached...) Figure 4 When an event entity is injected into the virtual market environment, the dynamic inference module 30 first performs a target identification operation. Specifically, the dynamic inference module 30 queries and determines the set of target entities that the event entity can influence, based on the predefined information accessibility relationships in the knowledge graph.

[0066] After identifying the target entities, the dynamic inference module 30 adjusts the behavioral decision models of these entities in real time to simulate their adaptive behavioral response to the event. This adjustment process is implemented through a preset adjustment function, which takes the behavioral decision model of the entity before adjustment and the event entity as input, and outputs the adjusted behavioral decision model. The dynamic adjustment process of this behavioral decision model can be described by the following formula:

[0067] π ′ a =g(π) a ,e);

[0068] In the formula, π ′ a The adjusted behavioral decision-making model; π a is the behavioral decision model before adjustment; g is the adjustment function, the specific form of which can be pre-configured according to the type of event entity. For example, for policy events, g can be a set of hard-coded rules, while for academic conference events, g can be a small neural network; e is the injected event entity containing specific information.

[0069] The behavioral decision-making model of the target entity was adjusted to π. ′ a Subsequently, the subject will use this new model to make decisions in subsequent simulation time steps. Through the iterative execution of the above steps, the dynamic simulation module 30 can completely output the trajectory of the future evolution of the market under specific initial conditions, market strategies, and external event shocks.

[0070] This strategy generation module 40 transforms the system from a forward deduction tool into a decision engine capable of inversely solving for the optimal strategy. Its detailed implementation includes the following steps.

[0071] First, the strategy generation module 40 formally defines the optimization problem. This process transforms the specific market strategy problem into a standard multi-objective optimization problem. Specifically, the market strategies to be formulated, such as marketing resource allocation in various regions and drug pricing, constitute a decision variable vector. Simultaneously, the desired business objectives, such as maximizing future market share and maximizing sales profit, are defined as one or more objective functions.

[0072] The core of this strategy generation module 40 lies in the fact that it does not directly define the analytical expression of the objective function. Instead, it uses the dynamic simulation module 30 to execute a complete market change simulation process as a fitness evaluation function connecting the decision variables and the objective function. That is, for any given set of decision variables (a specific market strategy), the system will run a complete simulation through the dynamic simulation module 30 and calculate the function values ​​of each objective function based on the final market state at the end of the simulation.

[0073] Therefore, the market strategy optimization problem to be solved can be formally described as:

[0074] maximize / minimize F(X)=(f1(X),f2(X),...,f m (X));

[0075] subject to g j (X)≤0,j=1,...,p;

[0076] X min ≤X≤X max ;

[0077] In the formula, X is a vector of decision variables, representing a specific market strategy; F(X) is a multi-objective vector function; f i (X) is the i-th objective function, whose value is determined by the market result after performing a complete dynamic simulation with strategy X as the input parameter; g j(X) represents the j-th inequality constraint function, such as the total budget constraint; p represents the total number of inequality constraints; X min and X max These are the lower and upper bounds of the decision variable vector, respectively.

[0078] Secondly, the strategy generation module 40 uses a multi-objective optimization algorithm to iteratively solve the formalized problem described above. In a specific embodiment, the Non-Dominated Sort Genetic Algorithm II (NSGA-II) can be used. This algorithm first randomly generates an initial population consisting of multiple different decision variable vectors X. In each iteration, the algorithm performs a fitness evaluation on each decision variable vector X in the population (i.e., calls the dynamic inference module 30 to perform a complete simulation and calculate the objective function value), and then selects superior individuals based on selection operations such as non-dominated sorting and crowding calculation, and generates the next generation population through genetic operators such as crossover and mutation.

[0079] Finally, after satisfying the preset termination conditions (e.g., reaching the maximum number of iterations or solution set convergence), the strategy generation module 40 outputs the final solution result. Since this problem is a multi-objective optimization problem, its solution is usually not unique. Therefore, the final output of the strategy generation module 40 is a set of Pareto optimal solutions, also known as the Pareto optimal front. Each solution in this set corresponds to a specific, executable market strategy and represents a specific trade-off between different business objectives.

[0080] To enhance the interpretability of the system's deduction results, this embodiment may further include an attribution alignment module 50. Its detailed implementation includes the following steps.

[0081] First, the attribution alignment module 50 records the behavior of micro-entities synchronously throughout the market simulation process executed by the dynamic inference module 30. For each time step in the simulation, the attribution alignment module 50 records and stores the behavior performed by all entities, forming a chain of micro-entity behaviors that includes timestamps, entity identifiers, behavior types, and market state information on which the behavior was based.

[0082] When an explanation is needed for a macro-market change (e.g., a significant increase in a drug's market share over a specific period), the attribution alignment module 50 initiates an attribution analysis process. This process first calculates the contribution of each specific action to the macro-market change based on the recorded chain of micro-agent behaviors.

[0083] To quantify this contribution, the attribution alignment module 50 calculates a contribution score for each action. For example, for macroeconomic market changes ΔM... k That is, the change in indicator k, for each behavior a i Its contribution score S c (ai ,ΔM k This score can be calculated. This score can be obtained by analyzing behavior a. i The attribution alignment module 50 ranks all relevant behaviors based on the calculated contribution scores and selects the behaviors with the highest contribution scores to form a chain of micro-entity behaviors that contribute to key market changes.

[0084] Subsequently, the attribution alignment module 50 performs the core alignment operation. For each behavior in the key behavior chain selected in the previous step, the attribution alignment module 50 extracts the entity identifier of the subject performing the behavior, as well as the object entity identifier associated with the behavior (e.g., a doctor entity prescribes a certain drug entity).

[0085] The attribution alignment module 50 uses these identifiers to query the medical knowledge graph generated by the knowledge graph construction module 10. This query aims to retrieve statically stored facts and relationships related to these key principal entities and object entities. For example, the module might query the attributes of a doctor entity (such as their department and title) and its connections within the graph (such as whether they participated in an academic conference or have a collaboration with a key opinion leader).

[0086] Finally, the attribution alignment module 50 integrates and presents the key behaviors captured during the dynamic deduction process with the static facts and relationships retrieved from the knowledge graph. Thus, the system can output a complete and traceable explanatory path from the macro-level deduction result (increased market share) to the micro-level facts (a certain group of doctors increased their prescribing behavior), and finally to the static knowledge in the knowledge graph (this group of doctors recently participated in an academic promotion conference for a certain drug).

[0087] To further clarify how the various modules of the system work together to solve practical problems, this section will use a specific application scenario as an example to describe the complete workflow of the system from receiving tasks to outputting the final strategy.

[0088] For example, suppose a pharmaceutical company faces the market problem of: within budget constraints, how to develop the optimal marketing resource allocation strategy for one of its core drugs to cope with an upcoming adjustment policy (event) to health insurance coverage, with the ultimate goal of maximizing market share in the coming year. The system will execute the following steps.

[0089] First, the system initializes and defines the problem. Users input the optimization task through the strategy generation module 40. Decision variables are defined as marketing resource allocation vectors for different regions and channels. The objective function is defined as the final market share of the core drug after a one-year dynamic simulation, with the optimization direction being maximization. Constraints are defined as total marketing resource investment not exceeding the preset total budget.

[0090] Upon receiving the task, the system confirms that the medical knowledge graph generated by the knowledge graph construction module 10 contains all relevant information, including the attributes of the core drug, entity information of doctors and patient groups in various regions, and detailed information on the upcoming medical insurance coverage adjustment policy as an event entity. Based on this graph, the behavior pattern learning module 20 has generated baseline behavioral decision models for all relevant doctor and patient group entities.

[0091] Subsequently, the strategy generation module 40 initiates its internal multi-objective optimization algorithm to begin iterative optimization. In each iteration, the algorithm generates one or more candidate marketing resource allocation strategies. For each candidate strategy, the system performs a complete dynamic inference and evaluation.

[0092] During the evaluation process, the dynamic simulation module 30 is invoked. The dynamic simulation module 30 sets the initial parameters of the simulation environment based on the candidate strategy and begins time-step simulation. When the simulation reaches the preset policy implementation time point, the medical insurance coverage adjustment policy event entity is injected. Based on the information accessibility relationships in the knowledge graph, the dynamic simulation module 30 identifies all entities affected by the policy and dynamically adjusts the formulas using behavioral decision-making models, modifying the behavioral decision-making models of these entities in real time. The simulation continues until the total duration of one year is reached.

[0093] After the simulation, the dynamic inference module 30 outputs the final market share result. This result is returned to the strategy generation module 40 as a fitness evaluation value for this candidate strategy. Based on this evaluation value and the evaluation values ​​of other candidate strategies, the optimization algorithm generates the next generation of candidate strategies through operations such as selection, crossover, and mutation, and repeats the above evaluation process.

[0094] Once the optimization algorithm meets the termination condition, the strategy generation module 40 outputs a set of Pareto optimal marketing resource allocation strategies. Each strategy is a specific resource allocation scheme that can be executed within budget constraints.

[0095] If a user needs to understand why a particular optimal strategy is effective, they can invoke the attribution alignment module 50. This module retrieves the micro-agent behavior chains recorded during the dynamic deduction of the strategy, identifying the key behaviors that contribute the most to market share. Subsequently, the attribution alignment module 50 aligns these key behaviors with static facts in the knowledge graph. For example, it might reveal a traceable explanation: although a region is negatively affected by policies, the physicians in that region are marked as highly focused on academic promotion in the knowledge graph, and the strategy happens to allocate higher academic promotion resources in this region, thus offsetting the negative impact of the policies and ultimately increasing prescription volume.

[0096] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A knowledge graph-driven intelligent analysis system based on the pharmaceutical field, characterized in that, The system includes: The knowledge graph construction module is used to integrate multi-source heterogeneous data to construct a medical field knowledge graph that includes subject entities, event entities, object entities, and the information reachability relationships between the subject entities and the event entities; The behavior pattern learning module is used to learn and generate parameterized behavior decision models for the subject entity based on the knowledge graph of the pharmaceutical field through a graph neural network. The parameters of the behavior decision models are determined by the embedding vector of the subject entity in the knowledge graph of the pharmaceutical field. A dynamic deduction module is used to instantiate the subject entity in a virtual market environment and dynamically deduce market changes based on the behavioral decision model and the target subject entity affected by the injected event entity identified based on the information accessibility relationship. The dynamic deduction module includes: in the virtual market environment, identifying the target subject entity affected by the event entity according to the information accessibility relationship defined in the medical field knowledge graph; and applying an adjustment function to instantly modify the behavioral decision model of the target subject entity to simulate the adaptive behavioral response of the target subject entity to the event entity. The calculation formula of the adjustment function includes: In the formula, The adjusted behavioral decision-making model; The behavioral decision-making model before adjustment; The adjustment function; The event entity; The strategy generation module is used to take the complete market change simulation process of the dynamic simulation module as the fitness evaluation function, solve it in reverse through a multi-objective optimization algorithm, and generate market strategies. The attribution alignment module records the micro-level subject behavior chains in the process of the dynamic inference module inferring market changes; and automatically aligns the micro-level subject behavior chains that contribute to key market changes with the facts and relationships in the pharmaceutical knowledge graph to provide a traceable explanation from macro-level inference results to micro-level facts.

2. The knowledge graph-driven intelligent analysis system based on the pharmaceutical field according to claim 1, characterized in that, The knowledge graph construction module includes: Natural language processing techniques are used to extract the subject entities, event entities, object entities, and information reachability relationships from unstructured data, and then integrate them with structured data to construct the medical knowledge graph.

3. The knowledge graph-driven intelligent analysis system based on the pharmaceutical field according to claim 2, characterized in that, The main entities include physician entities and patient groups entities; the event entities include policy release event entities and academic conference event entities; the object entities include drug entities and disease entities.

4. The knowledge graph-driven intelligent analysis system based on the pharmaceutical field according to claim 1, characterized in that, The behavior pattern learning module includes: The graph neural network is used to iteratively aggregate neighborhood information in the medical knowledge graph, generating the embedding vector for each entity. The calculation formula for the iterative neighborhood information aggregation includes: ； In the formula, For nodes In the Layer embedding vector; For nodes In the Layer embedding vector; For update functions; It is an aggregate function; For neighboring nodes In the Layer embedding vector; For connecting nodes With nodes eigenvectors of the edges; For nodes The set of neighboring nodes.

5. A knowledge graph-driven intelligent analysis system for the pharmaceutical field according to claim 4, characterized in that, The graph neural network is either a graph attention network or a graph convolutional network.

6. A knowledge graph-driven intelligent analysis system for the pharmaceutical field according to claim 5, characterized in that, The behavior pattern learning module includes: When the target entity lacks historical data, the similarity between the target associated object and the existing object entity is calculated in the medical knowledge graph. Based on the similarity, an initial behavior decision model is generated for the target entity using transfer learning techniques and the parameters of the learned behavior decision model associated with the existing object entity.

7. The knowledge graph-driven intelligent analysis system based on the pharmaceutical field according to claim 1, characterized in that, The strategy generation module includes: The market strategy is defined as the decision variable of the multi-objective optimization algorithm; The complete market change simulation process of the dynamic simulation module is used as the fitness evaluation function of the multi-objective optimization algorithm; Through iterative optimization, a set of market strategies that satisfy Pareto optimality is output.

8. A knowledge graph-driven intelligent analysis method based on the pharmaceutical field, applied to the system described in any one of claims 1-7, characterized in that, The method includes the following steps: S1. Integrate multi-source heterogeneous data to construct a medical domain knowledge graph that includes subject entities, event entities, object entities, and the information reachability relationships between the subject entities and the event entities; S2. Based on the knowledge graph of the pharmaceutical field, a parameterized behavior decision model is learned and generated for the subject entity through a graph neural network. The parameters of the behavior decision model are determined by the embedding vector of the subject entity in the knowledge graph of the pharmaceutical field. S3. Instantiate the subject entity in the virtual market environment, and dynamically deduce market changes based on the behavioral decision model and the target subject entity affected by the injected event entity identified based on the information accessibility relationship; S4. The complete market change simulation process of the dynamic simulation module is used as the fitness evaluation function, and the market strategy is generated by solving it in reverse through a multi-objective optimization algorithm.

Images