Intelligent rule reasoning and application system and method based on knowledge graph

Abstract

The invention relates to the technical field of knowledge graph reasoning and application, and discloses an intelligent rule reasoning and application system and method based on a knowledge graph, and the method comprises the steps: executing a query path decision strategy, and calculating the probability that a doctor skips and queries between two knowledge points based on marked knowledge points and a knowledge graph structure; establishing a historical query sequence of doctors; executing a query path probability modeling strategy, and calculating the probability that the doctor skips to query between the two knowledge points according to the historical query sequence of the doctor; obtaining knowledge points which are calculated by executing the query path probability modeling strategy and have the highest probability of jumping from the marked knowledge points to the query, and recording the knowledge points as second-class knowledge points; obtaining the probability between all adjacent knowledge points of the marked knowledge points and the marked knowledge points; and executing a clinical experience enhanced jump strategy, constructing an experience confidence function between the marked knowledge point and the plurality of candidate knowledge points, and taking the value of the experience confidence function as the probability of the candidate knowledge points, thereby improving the medical knowledge query efficiency of the doctor.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of knowledge graph reasoning and application, in particular to an intelligent rule reasoning and application system and method based on a knowledge graph. BACKGROUND

[0002] A medical knowledge graph organizes entities such as diseases, symptoms, and drugs and their relationships in a structured manner, which helps doctors quickly access diagnosis and treatment information. In the face of numerous professional terms and high complexity of knowledge in the medical field, a knowledge graph significantly improves information retrieval efficiency and reduces the misdiagnosis rate, becoming an important tool for modern decision support.

[0003] In the prior art, in the actual use of the knowledge graph by doctors, the knowledge points of the initial query are often not the knowledge content that the doctors actually need to obtain, but are based on fuzzy memory, clinical symptom clues, or temporary associations. Subsequently, doctors often find more relevant knowledge points through multiple jump operations. However, these jump paths are usually not directly reflected in the explicit structure of the knowledge graph, especially in the case where there is a lack of direct connection between nodes or a lack of semantic label support. These jump behaviors cannot be captured by traditional graph structures. The existing technology mainly relies on the structural connection relationship between knowledge points or semantic similarity calculation in the recommendation or navigation mechanism, and it is difficult to identify the "unstructured jump mode" or "potential association path" exhibited by doctors in real operations. This makes the knowledge graph have the use obstacles of inaccurate navigation, ineffective recommendation, and fragmented path in the high complexity and high similarity medical knowledge system, affecting the decision-making efficiency and reference value of doctors in clinical practice. In addition, there are a large number of anti-logic relationships or implicit association knowledge points in the medical knowledge graph, which do not conform to the regular semantic connection rules between these knowledge points. Therefore, in the existing recommendation system based on structure or similarity, these knowledge points are easily ignored or missed. This limits the application depth and intelligent level of the knowledge graph in complex medical scenarios.

[0004] The present application provides an intelligent rule reasoning and application system and method based on a knowledge graph to solve the problems mentioned in the background art. SUMMARY

[0005] The present application provides an intelligent rule reasoning and application system and method based on a knowledge graph to solve the problems mentioned in the background art.

[0006] The present application provides the following technical solution: an intelligent rule reasoning and application method based on a knowledge graph, comprising: establishing a knowledge graph structure, the knowledge graph structure comprising a knowledge point set and an edge set; establishing a semantic structure feature for each knowledge point in the knowledge graph structure; Obtaining a knowledge point that the doctor is currently accessing in the knowledge graph structure, denoted as a marked knowledge point; Executing a query path decision strategy, calculating the probability of the doctor jumping to query between two knowledge points based on the marked knowledge point and the knowledge graph structure; Obtaining the knowledge point with the highest query jump probability from the marked knowledge point calculated by executing the query path decision strategy, denoted as a first type of knowledge point; Establishing a historical query sequence of the doctor, the historical query sequence including knowledge points accessed by the doctor and the time length of staying at the knowledge points; Executing a query path probability modeling strategy, calculating the probability of the doctor jumping to query between two knowledge points according to the historical query sequence of the doctor; Obtaining the knowledge point with the highest query jump probability from the marked knowledge point calculated by executing the query path probability modeling strategy, denoted as a second type of knowledge point; Obtaining all adjacent knowledge points of the marked knowledge point and the probability between the marked knowledge point, comparing the probability with a set probability threshold, obtaining the knowledge points with the probability less than the probability threshold, denoted as candidate knowledge points; Executing a clinical experience reinforced jump strategy, constructing an experience confidence function between the marked knowledge point and the multiple candidate knowledge points, taking the value of the experience confidence function as the probability of the candidate knowledge points; Obtaining the candidate knowledge point with the highest probability, denoted as a third type of knowledge point; Simultaneously displaying the first type of knowledge point, the second type of knowledge point and the third type of knowledge point to the doctor.

[0007] Optionally, the query path probability modeling strategy includes: The knowledge graph structure is represented as , wherein is a knowledge point set, is the i-th knowledge point, is the total number of knowledge points, is an edge set, and an edge represents that the knowledge point and the knowledge point have a semantic relationship r; The historical query sequence of the doctor is established as , , wherein is the time length of the doctor staying at the i-th knowledge point , is the total number of historical query records; The number of times that the doctor jumps from the knowledge point to the knowledge point is denoted as ; Calculate the doctor's knowledge points The average duration of stay, as a result, is used by doctors to assess knowledge points. Average stay time ; Set weight function ,in, Used to indicate that the doctor learns from knowledge points Jump to knowledge point Knowledge points and knowledge points The sum of attention, Used to scale the dwell time, when doctors view knowledge points and knowledge points Increased attention Increase.

[0008] Optionally, the execution of the query path probability modeling strategy, which calculates the probability of a doctor switching between two knowledge points based on the doctor's historical query sequence, further includes: Construct a query path probability model based on doctors' historical query sequences. This indicates that the doctor started from the knowledge points. Jump to knowledge point The probability of; The calculation formula is: ,in, This indicates that the doctor started from the knowledge points Jump to knowledge point The weighted average of the number of times and the weight function, The values ​​used for normalization times and weighted averages of the weighting function, with j in the denominator representing the index of the knowledge points being traversed.

[0009] Optionally, establishing semantic structure features for each knowledge point in the knowledge graph structure includes: The semantic structural features integrate the position of knowledge points in the knowledge graph structure and the semantic tags of adjacent knowledge points; Knowledge points in the knowledge graph structure Set initial features ; Knowledge points Set the self-connection weight matrix of the k-th layer. Used to retain knowledge points Initial characteristics; opposite side Knowledge Points and knowledge points The semantic relationship r between them sets the propagation weight matrix at the k-th layer. This is used to connect adjacent knowledge points of semantic relation r during the propagation process of the k-th layer graph neural network. and Perform a linear transformation; Obtaining the structure and knowledge points of the knowledge graph All adjacent knowledge points with edges form the set of adjacent points. ; Calculation knowledge points Semantic structural features on the (k+1)th layer of the graph neural network ; ,in, This is a normalization factor used to control the weighted balance of multiple adjacent knowledge points. It is the ReLU activation function; Obtain the semantic structure features of all knowledge points and convert the knowledge graph structure into a graph feature matrix.

[0010] Optionally, the execution of the query path decision strategy, based on the labeled knowledge points and the knowledge graph structure, calculates the probability of a doctor jumping between two knowledge points for querying, including: Get the marked knowledge points semantic structural features ; Get the set of adjacent points Any knowledge point in ; Set reward function Among them, when doctors start from knowledge points Jump to knowledge point ,but When doctors start from knowledge points Not redirected to the knowledge point ,but ; Establish strategy function Used to calculate from marked knowledge points Jump to the set of adjacent points Any knowledge point in The probability, For knowledge points semantic structural features; ,in, and These are the weight matrix and bias term of the policy function, respectively. The weight matrix is ​​used to evaluate the semantic structure features. A weighted average is performed, and a bias term is used to adjust the value of the policy function. The function is a normalized function; Obtain the knowledge point with the highest probability calculated by the policy function. ; acquiring the marked knowledge point from the doctor actual knowledge point to jump to wherein, if then if then .

[0011] Optionally, the execution of the query path decision strategy is based on the marked knowledge point and the knowledge graph structure to calculate the probability of the doctor jumping to query between two knowledge points, and further comprises: updating the marked knowledge point to the actual knowledge point, and repeatedly executing the query path decision strategy to acquire all reward functions; setting a discount factor for controlling the influence degree of future rewards on current decisions; setting a target function , for training an optimal strategy function; constructing a loss function , wherein, denotes the probability logarithmic value of the doctor jumping from the knowledge point to the knowledge point ; setting a loss function threshold ; if , updating the weight matrix and the bias term , and the specific method is as follows: , wherein, is a learning rate for controlling the descending rate of the weight matrix and the bias term, is the partial derivative of the loss function with respect to the weight matrix , is the partial derivative of the loss function with respect to the bias term ; substituting the updated weight matrix and the bias term into the strategy function to recalculate the target function until ; using the updated strategy function to calculate the knowledge point with the highest probability.

[0012] Optionally, the execution of the clinical experience reinforcement jump strategy constructs an experience confidence function between the marked knowledge point and the candidate knowledge point, comprising: setting a basic experience confidence between the marked knowledge point and the candidate knowledge point ​ , the basic experience confidence is an expert-labeled abnormal jump query confidence, ; setting a rule item , for indicating whether the marked knowledge point and the candidate knowledge point meet the anti-logic reasoning rule, wherein, when the marked knowledge point and the candidate knowledge point meet the anti-logic reasoning rule, , otherwise, ; setting a case similarity item , specifically: obtaining historical cases in which a doctor jumps from the marked knowledge point to the candidate knowledge point and takes the candidate knowledge point as a target knowledge point; extracting feature parameters of each historical case; obtaining a current case that the doctor is treating, extracting feature parameters of the current case, calculating cosine similarity of the feature parameters of the current case and the feature parameters of each historical case, and calculating a mean value as the case similarity item ; calculating an experience confidence function between the marked knowledge point and the candidate knowledge point ; , wherein, and are respectively a rule item weight coefficient and a case similarity item weight coefficient.

[0013] The intelligent rule reasoning and application system based on the knowledge graph comprises: a knowledge graph construction module, configured to construct a knowledge graph structure composed of medical knowledge points, and define an edge set composed of a knowledge point set and its semantic relationship; a doctor behavior modeling module, configured to record a historical query path and a stay duration of a doctor in the knowledge graph structure, extract jump trajectory data, and simulate query jump behavior of the doctor; a query path probability modeling module, configured to count a frequency of jump between different knowledge points of the doctor, calculate a jump probability weighted by attention, and reflect a potential query path preference in combination with a stay time; a knowledge point semantic representation module, configured to fuse an adjacent structure and a semantic label of a knowledge point by using a graph neural network, and generate a semantic structure feature used for calculating a strategy function; a query recommendation decision module, configured to predict an optimal jump target based on a marked knowledge point and a semantic feature by using a reinforcement learning method, and form a dynamic query recommendation strategy; The multi-path fusion recommendation module is used for comprehensively fusing the jump probability, the one-class reinforcement learning result, the adjacency filtering and the clinical experience confidence to generate a final recommendation result.

[0014] The present application has the following advantages: 1. The intelligent rule reasoning and application method based on a knowledge graph can help doctors quickly locate key information in a large amount of structured knowledge, thereby improving query efficiency and diagnosis accuracy.

[0015] 2. The intelligent rule reasoning and application method based on a knowledge graph can identify the high-frequency pathways of the doctor's jump by reconstructing the "implicit behavior path" outside the graph structure.

[0016] 3. The intelligent rule reasoning and application method based on a knowledge graph can balance the long time staying and frequent jumping by introducing a log function designed attention weight term, thereby effectively improving the accuracy and generalization ability of the path modeling.

[0017] 4、The intelligent rule reasoning and application method based on knowledge graph can accurately capture the deep correlation between knowledge points in the semantic and structural layers by introducing a neural network to construct a semantic structure vector for each knowledge point in the knowledge graph. Traditional knowledge graphs only rely on explicit edge connections in the graph structure to connect knowledge points, which cannot reflect the semantic similarity between nodes, especially in the medical field. Similar concepts may not be directly connected due to differences in expression and different disciplines, leading to insufficient query coverage or misleading doctors. By using methods such as graph neural networks, each knowledge point is encoded as a vector containing adjacency relationships and semantic labels, enhancing the ability to model the implicit semantic relationships between knowledge points and providing a high-dimensional comparable feature space for similarity calculation, reasoning path expansion, and recommended decision-making.

[0018] 5、The intelligent rule reasoning and application method based on knowledge graph, in actual use, doctors face a huge medical atlas and it is difficult for them to lock in the target information at once, often relying on "trial and error" to gradually narrow the search range. Traditional rule recommendation methods cannot adapt to this highly personalized and dynamic behavior characteristic. By defining the state-action space, reward function, and strategy function, an intelligent recommendation strategy that conforms to the doctor's behavior pattern is trained, which can actively recommend knowledge points that are more likely to be the doctor's final demand. Reinforcement learning has self-adaptive optimization capability and can continuously learn and iterate from new behavior trajectories during operation to improve long-term performance. It significantly enhances the ability to recognize the doctor's intention and has the advantages of dynamic optimization and self-evolution, which is the core technical support for realizing intelligent recommendation paths.

[0019] 6、The intelligent rule reasoning and application method based on knowledge graph, by constructing an experience confidence function to give reasonable weights to candidate knowledge points with low jump probability but high experience support, it can make up for important jump paths that traditional probability models cannot capture. There are a large number of "non-explicit" logical relationships between medical knowledge points, such as the implicit relationship between some rare diseases and specific test results, which may have low frequency in historical jump data but are highly valuable in actual clinical practice. By incorporating historical case similarity and rule knowledge (such as inverse logic relationships) into the decision function, an experience confidence score is generated, allowing the system to recommend not only where doctors often jump, but also where doctors should jump. This reinforcement strategy not only retains the learning ability of statistical models, but also introduces the controllability of knowledge rules and experience judgments, effectively improving the fault tolerance and clinical reasonableness of the system, especially in the case of ambiguous queries, abnormal path jumps, or rare knowledge point recommendations. It significantly enhances the practicality and medical adaptability of the system. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The method flowchart of the present application.

[0021] Figure 2This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1, refer to Figure 1 Knowledge graph-based intelligent rule reasoning and application methods include: Establish a knowledge graph structure, which includes a set of knowledge points and a set of edges; In this example, a doctor sees a 40-year-old male patient in the internal medicine outpatient clinic, complaining of "sudden onset of chest pain lasting 30 minutes," accompanied by mild shortness of breath. The doctor logs into the hospital's knowledge graph platform, hoping to assist in determining the possible causes of the chest pain and providing treatment suggestions. The system supports intelligent recommendations based on historical behavior analysis, knowledge graph structure, reinforcement learning reasoning, and clinical experience.

[0024] The doctor began researching information related to "chest pain," and the knowledge graph contained the following information: v1: Definition and causes of chest pain; v2: Diagnostic criteria for myocardial infarction; v3: Diagnostic process for pulmonary embolism; v4: Symptoms and treatment of gastroesophageal reflux disease; Establish a doctor's historical query sequence, which includes the knowledge points visited by the doctor and the duration of time spent on those knowledge points; The query path probability modeling strategy is implemented to calculate the probability that a doctor will switch between two knowledge points based on the doctor's historical query sequence. knowledge graph structure Represented as ,in, A collection of knowledge points For the i-th knowledge point, The total number of knowledge points. Let be the set of edges, and the edges Representing knowledge points and knowledge points There is a semantic relationship r between them; Establish a historical query sequence for doctors , ,in, For the doctor at the i-th knowledge point Duration of stay on Total number of historical query records; The number of times the doctor jumps from a knowledge point to a knowledge point is recorded as ; The average of the length of time the doctor stays at a knowledge point is calculated, and the result is the average length of time the doctor stays at a knowledge point ; ; A weight function is set, where , is used to represent the sum of the attention of the doctor to a knowledge point and a knowledge point , is used to scale the length of stay, and when the doctor's attention to a knowledge point and a knowledge point increases, increases.

[0025] The doctor's historical query records are as follows: Query "Definition and Causes of Chest Pain" (v1), stay for 3 minutes.

[0026] Query "Diagnosis Criteria for Myocardial Infarction" (v2), stay for 5 minutes.

[0027] Query "Symptoms and Treatment of Gastroesophageal Reflux Disease" (v4), stay for 1 minute.

[0028] Query "Diagnosis Process of Pulmonary Embolism" (v3), stay for 4 minutes.

[0029] Statistics on the number of times the doctor jumps from v1: From "Definition and Causes of Chest Pain" (v1) to "Diagnosis Criteria for Myocardial Infarction" (v2), there are 4 jumps.

[0030] From v1 to "Symptoms and Treatment of Gastroesophageal Reflux Disease" (v4), there are 2 jumps.

[0031] From v2 to v3, there are 5 jumps.

[0032] Average length of stay: 3 minutes (definition of chest pain); 5 minutes (myocardial infarction); 4 minutes (pulmonary embolism); 1 minute (gastroesophageal reflux); Calculate the weight function: ; ; ​​The execution query path probability modeling strategy comprises the following steps: According to the historical query sequence of the doctor, the probability of the doctor jumping to query between two knowledge points is calculated. According to the historical query sequence of the doctor, a query path probability model is constructed , which represents the probability of the doctor jumping from knowledge point to knowledge point The calculation formula is as follows: , wherein represents the weighted average of the number of times and the weight function of the doctor jumping from knowledge point to knowledge point , is used to normalize the value of the number of times and the weighted average of the weight function, and j in the denominator represents the index of traversing the knowledge points.

[0033] The jump probability is calculated as follows: ; ; Explanation: After querying the definition of chest pain, the doctor is more inclined to jump to the knowledge point of myocardial infarction diagnosis.

[0034] Existing knowledge graph structure connection: v1 (chest pain) is usually strongly connected with v2 (myocardial infarction) in structure; v1 is usually not connected with v4 (gastroesophageal reflux); v2 is weakly connected with v3 (pulmonary embolism) in structure; Semantic similarity: v1 and v2 are highly semantically related (such as "heart", "pain", etc.); v1 and v4 have weak semantic relevance (few "reflux" or "digestive system" keywords); v2 and v3 have general semantic relevance (both are "acute chest pain" diseases); Failure of existing recommendation methods: If only relying on structure recommendation, only v2 will be recommended, and it cannot be known that users often jump to v4.

[0035] If only relying on semantics, v4 has weak semantic relevance with v1 and will be excluded or ranked last by the system.

[0036] If the jump frequency and dwell time are not combined, v2 cannot be identified as an important knowledge point with high jump frequency and long reading time.

[0037] Although v3 (pulmonary embolism) has no connection with v1, it has a high frequency of jumping through v2 and a dwell time of 4 minutes. If the path chain is not tracked, the system cannot perceive its importance.

[0038] Therefore, the real access path of the doctor to the knowledge points has cross-layer jumps, out-of-structure jumps, and weakly related semantic jumps, which reflect the "non-explicit behavior path" that the traditional knowledge graph cannot capture.

[0039] A semantic structure feature is established for each knowledge point in the knowledge graph structure, including: The semantic structure feature integrates the position of the knowledge point in the knowledge graph structure and the semantic labels of adjacent knowledge points; The semantic structure feature of the knowledge point in the knowledge graph structure is calculated as follows: An initial feature is set ; The kth layer self-connection weight matrix is set for the knowledge point , which is used to retain the initial feature of the knowledge point ; The kth layer propagation weight matrix is set for the semantic relationship r between the knowledge point and the knowledge point , which is used to perform linear transformation on the adjacent knowledge points and connected by the semantic relationship r in the kth layer graph neural network propagation process; The knowledge graph structure and all adjacent knowledge points of the knowledge point are obtained, and an adjacent point set is formed ; The semantic structure feature of the knowledge point on the k+1th layer graph neural network is calculated ; The semantic structure feature of the knowledge point on the k+1th layer graph neural network is calculated ; wherein, is a normalization factor used to control the weighted balance of multiple adjacent knowledge points, is a ReLU activation function, and the prior art is The semantic structure features of all knowledge points are obtained, and the knowledge graph structure is converted into a graph feature matrix.

[0040] The prior art "semantic label" refers to the semantic information carried by the knowledge point, which can be represented as: A set of keywords (extracted by humans or models); Word embedding vectors (such as BERT, Word2Vec encoding); Label classification of symptoms or causes (such as ICD labels, SNOMED classification); Text summary (semantic fragments extracted from the text); Medical entity description word vector combination.

[0041] Here are some examples of "semantic tag" for several knowledge points: v1 = chest pain, semantic tags are "chest", "pain", "compression", "heart area", "radiating pain"; v2 = myocardial infarction, semantic tags are "heart", "ischemic injury", "ECG changes", "chest pain", "troponin elevation"; v3 = gastroesophageal reflux, semantic tags are "poststernal burning", "worsened after eating", "gastric acid reflux", "nighttime cough"; These tags can be extracted from medical dictionaries, mapped from medical knowledge bases (such as UMLS), extracted from text encoding models (such as BERT), or formed from a feature set annotated by experts.

[0042] Fusing the semantic tags of adjacent nodes means: Not only use the semantic content of the current knowledge point itself (such as the "chest pain" keyword), but also encode the semantic tags of its neighbor nodes (such as "myocardial infarction", "pneumonia", "gastroesophageal reflux", etc.) into the vector of the current knowledge point. The semantic structure features formed in this way can better reflect the context semantics and structural environment in the entire knowledge graph.

[0043] Establish semantic structure features for each knowledge point in the knowledge graph structure, which fuse the position of the knowledge point in the knowledge graph structure and the semantic tags of adjacent knowledge points; When a doctor queries a knowledge point in the knowledge graph structure: Get the knowledge point the doctor is currently accessing, denoted as the marked knowledge point; Execute the query path decision strategy, based on the marked knowledge point and the knowledge graph structure to calculate the probability of the doctor jumping between two knowledge points during query, including: Get the semantic structure features of the marked knowledge point ; ; Get any one of the adjacent point set ; ; Set the reward function , where , when the doctor jumps from knowledge point to knowledge point , then , when the doctor does not jump from knowledge point to knowledge point ; Establish a strategy function to calculate the jump from the marked knowledge point to the adjacent point set Any knowledge point in The probability, For knowledge points semantic structural features; ,in, and These are the weight matrix and bias term of the policy function, respectively. The weight matrix is ​​used to evaluate the semantic structure features. A weighted average is performed, and a bias term is used to adjust the value of the policy function. The function is a normalized function; Obtain the knowledge point with the highest probability calculated by the policy function. ; Obtain doctors from marked knowledge points The actual knowledge point to jump to from the starting point , among which, if ,but ,like ,but .

[0044] And using loss functions to optimize a class of knowledge points also includes: Update the marked knowledge points to the actual knowledge points, repeatedly execute the query path decision strategy, and obtain all reward functions; Set discount factor This is used to control the degree to which future rewards affect current decisions; Set the objective function , This is used to train the optimal policy function; Constructing the loss function , ,in, This indicates that the doctor started from the knowledge points Jump to knowledge point The logarithmic value of the probability; Set the threshold for the loss function ; like Then update the weight matrix. and bias terms The details are as follows: , ,in, The learning rate controls the rate at which the weight matrix and bias terms decrease. The loss function is applied to the weight matrix. Find the partial derivative. For the loss function with respect to the bias term Find the partial derivative; The updated weight matrix and bias terms Substitute the policy function and recalculate the objective function until... ; The updated policy function is used to calculate the knowledge point with the highest probability.

[0045] Reward function settings: If the doctor actually jumps to v2 (myocardial infarction), the reward is 1; otherwise, it is 0.

[0046] The policy function predicts the jump probability based on the semantic features of v1 and v2.

[0047] Through repeated training, the system learns how doctors typically jump from the definition of chest pain to the knowledge point of myocardial infarction.

[0048] Loss function optimization reinforces this behavior, making recommendations more aligned with doctors' habits. The knowledge point with the highest probability of jumping from the marked knowledge point to the query path is denoted as the second-class knowledge point, calculated by the probability modeling strategy for obtaining the execution query path. Set a probability threshold; Obtain the probability between all adjacent knowledge points and the marked knowledge point, compare the probability with a probability threshold, and select knowledge points with a probability less than the probability threshold as candidate knowledge points; The implementation of the clinical experience-enhanced jump strategy involves constructing an experience confidence function between labeled knowledge points and candidate knowledge points, including: Set up and mark knowledge points and candidate knowledge points Basic empirical confidence level The basic experience confidence level is the confidence level of anomalous redirect queries annotated by experts. ; Set rule items Used to indicate marked knowledge points and candidate knowledge points Does it conform to the rules of anti-logical reasoning, where, when the knowledge point is marked and candidate knowledge points If it conforms to the rules of anti-logical reasoning, then ,otherwise, ; Set case similarity item Specifically: Retrieve historical doctors from marked knowledge points Jump to candidate knowledge points and candidate knowledge points Historical cases serving as target knowledge points; Extract the feature parameters for each historical case; Obtain the current case being treated by the doctor, extract the feature parameters of the current case, calculate the cosine similarity between the feature parameters of the current case and the feature parameters of each historical case, calculate the mean, and use the result as the case similarity item. Among them, cosine similarity is a current technology; Calculate the empirical confidence function between labeled knowledge points and candidate knowledge points. ; ,in, and These are the weight coefficients for the rule item and the case similarity item, respectively.

[0049] Empirical confidence level: Expert feedback indicates that although there is no direct structural connection between "chest pain" and "gastroesophageal reflux disease", they are often considered together by doctors in clinical practice. The empirical confidence level is set at 0.4.

[0050] Rule item: If it meets the "exclusion method" anti-logic reasoning rule (because gastroesophageal reflux can cause chest pain and the symptoms are similar, heart disease needs to be ruled out), it is assigned a value of 1.

[0051] Case similarity: The proportion of doctors who jumped from chest pain to gastroesophageal reflux in similar historical cases was calculated to be 0.6.

[0052] The total confidence score, calculated using weights, is approximately 0.88.

[0053] This high confidence level means that despite the low probability of redirection, the system still recommends information about "gastroesophageal reflux disease" to help doctors with diagnosis.

[0054] Although "gastroesophageal reflux disease (v4)" and "chest pain (v1)" are relatively distant semantically and not directly connected in the knowledge graph structure, it is often referred to as an exclusion item for "non-cardiac chest pain" in clinical diagnosis, and has the following three characteristics: Clinical necessity: V4 is an important "differential diagnosis" for chest pain, used to rule out the possibility of non-cardiac diseases, such as chest pain caused by acid reflux. Although this non-cardiac chest pain is atypical, it should not be overlooked in the initial diagnosis of chest pain.

[0055] Behavioral concealment: Doctors usually only quickly scan or briefly look at v4-related content after screening for heart diseases (such as v2 myocardial infarction), resulting in short dwell time and low semantic connectivity, making it easy for semantic vector models or structural path mining to ignore.

[0056] Explicitizing experience: For this reason, the system introduces an "experience confidence function" to quantify the doctor's actual path (i.e., the jump frequency of v1→v4 + dwell traces), and incorporates these "knowledge points that are not the main path of cognition but must be considered in clinical practice" into the recommendation system.

[0057] v4 represents a type of knowledge point that is not particularly prominent in terms of structure and semantics, but is "repeatedly consulted and easily overlooked by traditional recommendations" during actual clinical practice. Strengthening the recommendation of this type of knowledge point helps the system to be closer to the actual clinical cognitive path and improves the clinical credibility and practicality of the recommendations.

[0058] The candidate knowledge points with the highest probability of being obtained are categorized into three types of knowledge points. The first, second, and third categories of knowledge points are presented to the doctor simultaneously.

[0059] After consulting the doctor about the definition and common causes of chest pain, the following recommendations were made: The diagnostic criteria for myocardial infarction (v2) are recommended based on: historical query sequence + jump probability; The presentation and examination of pulmonary embolism (v3) are recommended because of the highest path statistics, the highest semantic relevance, and the reinforcement strategy. Symptoms and treatment of gastroesophageal reflux disease (v4), recommended because of its clinical screening relationship and high empirical confidence. Compared to traditional algorithms, it significantly improves the ability to identify and recommend knowledge points that are structurally weakly connected, semantically distant but clinically necessary, making the knowledge graph system more in line with medical diagnosis and treatment processes and closer to doctors' actual thinking paths, greatly enhancing its application value in highly complex scenarios.

[0060] Example 2, refer to Figure 2 A knowledge graph-based intelligent rule reasoning and application system includes: The knowledge graph construction module is used to build a knowledge graph structure composed of medical knowledge points and define the edge set composed of the knowledge point set and its semantic relationship. The doctor behavior modeling module is used to record the doctor's historical query paths and dwell time in the knowledge graph structure, extract jump trajectory data, and simulate the doctor's query jump behavior; The query path probability modeling module is used to count the frequency of doctors jumping between different knowledge points, and calculate the attention-weighted jump probability by combining the dwell time, reflecting the potential query path preferences. The knowledge point semantic representation module is used to fuse the adjacency structure and semantic labels of knowledge points using graph neural networks to generate semantic structure features for calculating policy functions; The query recommendation decision module is used to predict the optimal jump target based on labeled knowledge points and semantic features, and to form a dynamic query recommendation strategy. The multi-path fusion recommendation module integrates jump probability, reinforcement learning results, neighbor selection, and clinical experience confidence to generate the final recommendation result.

[0061] It should be noted that, in this document, relational 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 such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0062] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based intelligent rule reasoning and application method, characterized in that, include: Establish a knowledge graph structure, which includes a set of knowledge points and a set of edges; Establish semantic structure features for each knowledge point in the knowledge graph structure; Retrieve the knowledge points that the doctor is currently accessing in the knowledge graph structure and label them as knowledge points. The system executes a query path decision strategy, calculating the probability that a doctor will jump between two knowledge points based on the labeled knowledge points and the knowledge graph structure. The knowledge points with the highest probability of jumping from the marked knowledge points to the query path are identified as a class of knowledge points based on the query path decision strategy calculation. Establish a doctor's historical query sequence, which includes the knowledge points visited by the doctor and the duration of time spent on those knowledge points; The query path probability modeling strategy is implemented to calculate the probability that a doctor will switch between two knowledge points based on the doctor's historical query sequence. The knowledge point with the highest probability of jumping from the marked knowledge point to the query path is denoted as the second-class knowledge point, calculated by the probability modeling strategy for obtaining the execution query path. Obtain the probability between all adjacent knowledge points and the marked knowledge point, compare the probability with a set probability threshold, and select knowledge points with a probability less than the probability threshold as candidate knowledge points; A clinical experience-enhanced jump strategy is implemented, and an experience confidence function is constructed between the labeled knowledge point and multiple candidate knowledge points. The value of the experience confidence function is used as the probability of the candidate knowledge point. The candidate knowledge points with the highest probability of being obtained are categorized into three types of knowledge points. The first, second, and third categories of knowledge points are presented to the doctor simultaneously.

2. The intelligent rule reasoning and application method based on knowledge graphs according to claim 1, characterized in that, The execution query path probability modeling strategy calculates the probability that a doctor will switch between two knowledge points based on the doctor's historical query sequence, including: knowledge graph structure Represented as ,in, A collection of knowledge points For the i-th knowledge point, The total number of knowledge points. Let be the set of edges, and the edges Representing knowledge points and knowledge points There is a semantic relationship r between them; Establish a historical query sequence for doctors , ,in, For the doctor at the i-th knowledge point Duration of stay on This represents the total number of historical query records. Doctors from knowledge points Jump to knowledge point The number of times is recorded as ; Calculate the doctor's knowledge points The average duration of stay, as a result, is used by doctors to assess knowledge points. Average stay time ; Set weight function ,in, Used to indicate that the doctor learns from knowledge points Jump to knowledge point Knowledge points and knowledge points The sum of attention, Used to scale the dwell time, when doctors view knowledge points and knowledge points Increased attention Increase.

3. The intelligent rule reasoning and application method based on knowledge graphs according to claim 2, characterized in that, The execution query path probability modeling strategy, which calculates the probability of a doctor switching between two knowledge points based on the doctor's historical query sequence, also includes: Construct a query path probability model based on doctors' historical query sequences. This indicates that the doctor started from the knowledge points. Jump to knowledge point The probability of; The calculation formula is: ,in, This indicates that the doctor started from the knowledge points Jump to knowledge point The weighted average of the number of times and the weight function, The values ​​used for normalization times and weighted averages of the weighting function, with j in the denominator representing the index of the knowledge points being traversed.

4. The intelligent rule reasoning and application method based on knowledge graphs according to claim 1, characterized in that, The step of establishing semantic structure features for each knowledge point in the knowledge graph structure includes: The semantic structural features integrate the position of knowledge points in the knowledge graph structure and the semantic tags of adjacent knowledge points; Knowledge points in the knowledge graph structure Set initial features ; Knowledge points Set the self-connection weight matrix of the k-th layer. Used to retain knowledge points Initial characteristics; opposite side Knowledge Points and knowledge points The semantic relationship r between them sets the propagation weight matrix at the k-th layer. This is used to connect adjacent knowledge points of semantic relation r during the propagation process of the k-th layer graph neural network. and Perform a linear transformation; Obtaining the structure and knowledge points of the knowledge graph All adjacent points with edges form the set of adjacent points. ; Calculation knowledge points Semantic structural features on the (k+1)th layer of the graph neural network ; ,in, This is a normalization factor used to control the weighted balance of multiple adjacent knowledge points. It is the ReLU activation function; Obtain the semantic structure features of all knowledge points and convert the knowledge graph structure into a graph feature matrix.

5. The knowledge graph-based intelligent rule reasoning and application method according to claim 4, characterized in that, The execution query path decision strategy, based on the labeled knowledge points and the knowledge graph structure, calculates the probability that a doctor will jump between two knowledge points for a query, including: Get the marked knowledge points semantic structural features ; Get the set of adjacent points Any knowledge point in ; Set reward function Among them, when doctors start from knowledge points Jump to knowledge point ,but When doctors start from knowledge points Not redirected to the knowledge point ,but ; Establish strategy function Used to calculate from marked knowledge points Jump to the set of adjacent points Any knowledge point in The probability, For knowledge points semantic structural features; ,in, and These are the weight matrix and bias term of the policy function, respectively. The weight matrix is ​​used to evaluate the semantic structure features. A weighted average is performed, and a bias term is used to adjust the value of the policy function. The function is a normalized function; Obtain the knowledge point with the highest probability calculated by the policy function. ; Obtain doctors from marked knowledge points The actual knowledge point to jump to from the starting point , among which, if ,but ,like ,but .

6. The intelligent rule reasoning and application method based on knowledge graphs according to claim 5, characterized in that, The execution query path decision strategy, which calculates the probability of a doctor jumping between two knowledge points based on labeled knowledge points and the knowledge graph structure, also includes: Update the marked knowledge points to the actual knowledge points, repeatedly execute the query path decision strategy, and obtain all reward functions; Set discount factor This is used to control the degree to which future rewards affect current decisions; Set the objective function , This is used to train the optimal policy function; Constructing the loss function , ,in, This indicates that the doctor started from the knowledge points Jump to knowledge point The logarithmic value of the probability; Set the threshold for the loss function ; like Then update the weight matrix. and bias terms The details are as follows: , ,in, The learning rate controls the rate at which the weight matrix and bias terms decrease. For the loss function on the weight matrix Find the partial derivative. For the loss function with respect to the bias term Find the partial derivative; The updated weight matrix and bias terms Substitute the policy function and recalculate the objective function until... ; The updated policy function is used to calculate the knowledge point with the highest probability.

7. The intelligent rule reasoning and application method based on knowledge graphs according to claim 1, characterized in that, The implementation of the clinical experience-enhanced jump strategy involves constructing an experience confidence function between labeled knowledge points and candidate knowledge points, including: Set up and mark knowledge points and candidate knowledge points Basic empirical confidence level The basic experience confidence level is the confidence level of anomalous redirect queries annotated by experts. ; Set rule items Used to indicate marked knowledge points and candidate knowledge points Does it conform to the rules of anti-logical reasoning, where, when the knowledge point is marked and candidate knowledge points If it conforms to the rules of anti-logical reasoning, then ,otherwise, ; Set case similarity item Specifically: Retrieve historical doctors from marked knowledge points Jump to candidate knowledge points and candidate knowledge points Historical cases serving as target knowledge points; Extract the feature parameters for each historical case; Obtain the current case being treated by the doctor, extract the feature parameters of the current case, calculate the cosine similarity between the feature parameters of the current case and the feature parameters of each historical case, calculate the mean, and use the result as the case similarity term. ; Calculate the empirical confidence function between labeled knowledge points and candidate knowledge points. ; ,in, and These are the weight coefficients for the rule item and the case similarity item, respectively.

8. A knowledge graph-based intelligent rule reasoning and application system, applied to the knowledge graph-based intelligent rule reasoning and application method described in claims 1-7, characterized in that, include: The knowledge graph construction module is used to build a knowledge graph structure composed of medical knowledge points and define the edge set composed of the knowledge point set and its semantic relationship. The doctor behavior modeling module is used to record the doctor's historical query paths and dwell time in the knowledge graph structure, extract jump trajectory data, and simulate the doctor's query jump behavior; The query path probability modeling module is used to count the frequency of doctors jumping between different knowledge points and calculate the attention-weighted jump probability by combining the dwell time. The knowledge point semantic representation module is used to fuse the adjacency structure and semantic labels of knowledge points using graph neural networks to generate semantic structure features for calculating policy functions; The query recommendation decision module is used to predict the optimal jump target based on labeled knowledge points and semantic features, and to form a dynamic query recommendation strategy. The multi-path fusion recommendation module integrates jump probability, reinforcement learning results, neighbor selection, and clinical experience confidence to generate the final recommendation result.

Images