A method and system for intelligent quality control of electronic medical records based on medical knowledge graphs

Abstract

This invention belongs to the field of medical data processing technology, specifically relating to an intelligent quality control method and system for electronic medical records based on a medical knowledge graph. The method includes: acquiring target medical record data and constructing an initial logical driving vector; calculating the dynamic propagation impedance of each edge in the general medical knowledge graph based on the deviation between the measured values ​​of physiological indicators in the target medical record data and the physiological constraints of edge attributes in the general medical knowledge graph; constructing a transition probability matrix using the dynamic propagation impedance and obtaining the steady-state correlation distribution vector of each node through iterative calculation; calculating a dynamic judgment threshold based on the steady-state correlation distribution vector and performing anomaly verification on the actual treatment instructions in the target medical record data. This invention can dynamically map the real-time physiological state of an individual patient to the constraints of graph reasoning, effectively identifying potential contraindications caused by abnormal physiological indicators, and realizing personalized quality control of electronic medical records.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology. More specifically, this invention relates to an intelligent quality control method and system for electronic medical records based on medical knowledge graphs. Background Technology

[0002] Electronic medical records are the core of digital clinical diagnosis and treatment in medical institutions, and their data quality directly affects medical safety and hospital management. In recent years, medical institutions' requirements for medical record quality have shifted from traditional document integrity checks to a deeper level of logical consistency verification. The core challenge in this process lies in accurately determining whether there is a causal logic between disease diagnosis and treatment measures that conforms to medical standards. This requires not only that the system can identify the indications for medical orders, but also that it comprehensively considers the patient's individual condition to ensure the rationality and safety of the treatment.

[0003] Currently, the industry mainly employs two technical approaches for automated quality control. One is a hard-matching method based on rule bases, which reviews applications by pre-setting mutually exclusive rules such as gender-project and age-drug. However, this method relies on manual maintenance and struggles to handle complex medical logic and cover a vast number of treatment combinations. The other is a path retrieval method based on static medical knowledge graphs. This method constructs medical entities as nodes and determines the rationality of medical orders by querying whether there is a connected path between disease nodes and treatment nodes. For example, if there is a relationship edge between insulin treatment for diabetes in the graph, the system assumes that the medical order is compliant. This method, to some extent, leverages the advantages of knowledge association.

[0004] However, existing quality control methods based on static knowledge graphs have revealed significant limitations in practical applications. The treatment relationships stored in medical knowledge graphs are usually general and idealized medical consensus, ignoring the dynamic impact of individual patient physiological states in specific medical scenarios. In fact, many general treatment pathways are subject to strict physiological indicators. For example, although some drugs correspond to specific diseases, they should be considered contraindicated when patients have abnormal liver and kidney function indicators or coagulation disorders. Existing technologies cannot easily incorporate these real-time, structured vital signs as constraints into the knowledge graph reasoning process, resulting in the inability to identify contraindication risks caused by individual patient differences. This leads to a large number of false positives or false negatives in the quality control results, making it difficult to achieve truly accurate quality control. Summary of the Invention

[0005] To address the technical problem that existing static graph quality control ignores the dynamic constraints of patients' physiological indicators on general pathways, leading to the inability to identify individual contraindications and resulting in a large number of misjudgments, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides an intelligent quality control method for electronic medical records based on medical knowledge graphs, comprising:

[0007] The process involves acquiring target medical record data for quality control and constructing an initial logical driving vector based on the position of disease conclusion entities in the target medical record data within a general medical knowledge graph. It also involves calculating the dynamic propagation impedance of directed edges in the general medical knowledge graph based on the matching degree between measured values ​​of physiological indicators in the target medical record data and physiological constraints recorded in edge attributes. The process then constructs a transition probability matrix for the general medical knowledge graph based on this dynamic propagation impedance and iteratively calculates the matrix using the initial logical driving vector until a steady-state correlation distribution vector is obtained. Finally, it calculates a dynamic judgment threshold based on the medical treatment behavior correlation score of the corresponding actual treatment instruction node in the target medical record data within the steady-state correlation distribution vector, and uses this threshold to determine whether the actual treatment instruction constitutes an abnormal treatment behavior.

[0008] This invention acquires target medical record data and calculates dynamic propagation impedance based on the matching degree between measured values ​​of physiological indicators and physiological constraints recorded in the edge attributes of a general medical knowledge graph. It dynamically maps the real-time physiological state of an individual patient to resistance factors in the graph reasoning process. This suppresses the propagation capability of the corresponding business logic path when the patient's physiological indicators deviate from the normal range and fail to meet the constraints of a specific intervention relationship. Subsequently, iterative calculations are performed using a constructed transition probability matrix combined with the initial logic driving vector, causing the logical association strength to tend to diffuse along a low-impedance path consistent with the current physiological state. This yields a steady-state correlation distribution vector that accurately reflects individual differences. Finally, a dynamic judgment threshold is calculated based on the steady-state correlation distribution vector to perform anomaly verification on actual treatment instructions. This effectively identifies treatment behaviors that, although conforming to general treatment rules, constitute potential contraindications due to abnormal individual patient physiological indicators. This avoids misjudgments common in traditional static graph quality control methods and improves the accuracy and personalized adaptability of electronic medical record quality control.

[0009] Preferably, the construction of the initial logical driving vector includes: extracting disease conclusion entities, physiological indicators and their corresponding measured values, and actual treatment instructions from the target medical record data; mapping the disease conclusion entities to corresponding nodes in the general medical knowledge graph to determine the disease conclusion node set; constructing an initial logical driving vector with the same dimension as the total number of nodes in the general medical knowledge graph; for each element in the initial logical driving vector, if its corresponding node belongs to the disease conclusion node set, assigning a non-zero initial value, otherwise assigning a zero value.

[0010] Preferably, the elements in the initial logical driving vector satisfy the expression: In the formula, Represents the initial logical driving vector The Middle The value of each element; Represents the set of disease conclusion nodes; Represents the set of disease conclusion nodes The number of elements in the middle; Represents the first in the general medical knowledge graph Each node.

[0011] This invention normalizes the elements in the initial logic driving vector, ensuring that the sum of the initial energies of all disease conclusion nodes remains constant. This avoids fluctuations in the total energy of the logic driving source due to differences in the number of diagnoses in medical records, and ensures that different medical record data have a unified energy benchmark when performing logical reasoning, thereby guaranteeing the numerical stability and convergence of subsequent iterative calculations.

[0012] Preferably, the calculation of the dynamic propagation impedance of directed edges in the general medical knowledge graph includes: in response to a directed edge not being a business logic path connecting a disease conclusion node and a treatment instruction node, setting the dynamic propagation impedance to a default value; otherwise, calculating the dynamic propagation impedance: In the formula, Indicates from node To the node Directed edge Dynamic propagation impedance within the current target medical record data context; Represents a directed edge The set of all physiological indicators contained in the physiological constraint attribute set; Represents the corresponding physiological indicators in the physiological feature vector. The measured values; , These represent physiological indicators within the set of physiological constraint attributes. The ideal mean and allowable standard deviation; Indicators of physiological indicators Standardized tolerance threshold; Indicates the impedance sensitivity coefficient; To prevent extremely small positive numbers with a denominator of 0.

[0013] This invention quantifies the dynamic propagation impedance of directed edges by introducing the deviation between measured values ​​of physiological indicators and preset standards. It can transform the real-time physiological state of an individual patient into resistance on the graph reasoning path. When the patient's indicators show significant abnormalities, the impedance of the corresponding business logic path increases sharply, thereby inhibiting the transmission of association strength to nodes with taboo risks during the logical reasoning process. This effectively solves the problem of quality control misjudgment caused by the inability of traditional static graphs to perceive individual physiological differences.

[0014] Preferably, the transition probability matrix for constructing a general medical knowledge graph includes: the transition probability from any node to another node satisfies the expression: In the formula, Indicates from node To the node The transition probability; Indicates from node To the node Directed edge Dynamic propagation impedance within the current target medical record data context; Represents a node The set of all adjacent nodes; Indicates from node to adjacent nodes Directed edge Dynamic propagation impedance within the current target medical record data context.

[0015] This invention utilizes the reciprocal of dynamic propagation impedance to construct a transition probability matrix, making the propagation probability of logical association strength in the general medical knowledge graph positively correlated with the smoothness of the path. That is, the higher the impedance, the lower the transition probability assigned to the path. This mechanism forces the logical reasoning flow to automatically avoid high-risk paths that do not conform to the patient's current physiological state and tends to diffuse towards low-impedance nodes that conform to clinical norms and are permissible by physiological indicators, thereby achieving implicit screening of the safety of medical treatment behaviors.

[0016] Preferably, the update rule for iterative calculation satisfies the expression: In the formula, Indicates the first Steady-state correlation distribution vector after the next iteration; Indicates the initial logic driving vector; Indicates the first The steady-state correlation distribution vector after the next iteration is defined in the first iteration as follows: equal ; Represents the transition probability matrix. Indicates the transpose symbol; This indicates the probability of restarting.

[0017] Preferably, the calculation of the dynamic judgment threshold includes: extracting the element values ​​of each actual treatment instruction node in the target medical record data in the steady-state correlation distribution vector, as the medical treatment behavior correlation score of the actual treatment instruction node; and calculating the dynamic judgment threshold based on the medical treatment behavior correlation score. In the formula, Indicates a dynamic threshold; The nodes in the set that represent the actual disposal instruction nodes Relevance score of medical treatment behavior; This indicates the number of nodes in the set of nodes that actually handle the instructions; This represents the abnormal deviation coefficient.

[0018] This invention calculates a dynamic judgment threshold based on the statistical distribution of the medical treatment behavior relevance scores of all actual treatment instructions in the target medical record data. It abandons the shortcomings of the traditional fixed threshold judgment method, which is easily affected by the complexity of the medical record. It can adaptively adjust the quality control standards according to the overall logical tightness of the current medical record. This allows the system to more accurately identify abnormal treatment behaviors with medical treatment behavior relevance scores that are significantly lower than the overall level when facing medical records with different disease complexities and number of medical orders, thereby improving the robustness and accuracy of defect localization.

[0019] Preferably, the step of using a dynamic judgment threshold to determine whether an actual treatment instruction is an abnormal treatment behavior includes: when the medical treatment behavior relevance score of the actual treatment instruction node is less than the dynamic judgment threshold, the actual treatment instruction is determined to be an abnormal treatment behavior.

[0020] Preferably, the method further includes: if the relevance score of the medical treatment behavior in response to the actual treatment instruction is less than a preset effective connectivity threshold, marking its defect type as an unfounded execution item; if the relevance score of the medical treatment behavior in response to the actual treatment instruction is greater than or equal to the preset effective connectivity threshold but less than the dynamic judgment threshold, marking its defect type as the presence of contraindications or inconsistencies in physiological indicators; generating a structured quality control defect report, wherein when the defect type is the presence of contraindications or inconsistencies in physiological indicators, the structured quality control defect report includes reference physiological indicators that lead to an increase in dynamic propagation impedance; the method for obtaining the reference physiological indicators is: tracing back the business logic path connecting the associated disease conclusion node and the actual treatment instruction node, and filtering out physiological indicator types whose standardized deviation exceeds the standardized tolerance threshold.

[0021] Secondly, the present invention provides an intelligent quality control system for electronic medical records based on medical knowledge graphs, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned intelligent quality control method for electronic medical records based on medical knowledge graphs is implemented.

[0022] By adopting the above technical solution, a computer program is generated from the above-mentioned intelligent quality control method for electronic medical records based on medical knowledge graphs and stored in a memory so that it can be loaded and executed by a processor. In this way, a terminal device can be made based on the memory and the processor for convenient use.

[0023] The beneficial effects of this invention are as follows: This invention acquires target medical record data and calculates dynamic propagation impedance based on the degree of matching between the measured values ​​of physiological indicators and the physiological constraints recorded in the edge attributes of a general medical knowledge graph. This dynamically maps the real-time physiological state of an individual patient into a resistance factor in the graph reasoning process, so that when the patient's physiological indicators deviate from the normal range and do not meet the constraints of a specific intervention relationship, the transmission capability of the corresponding business logic path is suppressed. Furthermore, iterative calculation is performed using a constructed impedance-modulated transition probability matrix combined with the initial logic driving vector, causing the logical association strength to tend to diffuse along a low-impedance path that conforms to the current physiological state, thereby obtaining a steady-state correlation distribution vector that reflects individual differences. Finally, a dynamic judgment threshold is calculated based on the steady-state correlation distribution vector to perform anomaly verification on actual treatment instructions. This can effectively identify treatment behaviors that, although conforming to general treatment rules, constitute potential contraindications due to abnormal individual patient physiological indicators, avoiding misjudgments or omissions caused by ignoring individual differences in traditional static graph quality control methods, and improving the accuracy and personalized adaptability of electronic medical record quality control. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating an intelligent quality control method for electronic medical records based on medical knowledge graphs according to the present invention.

[0025] Figure 2 This is a schematic diagram illustrating the deviation of physiological indicators;

[0026] Figure 3 This is a schematic diagram illustrating the energy propagation process of a random walk;

[0027] Figure 4 This is a schematic diagram illustrating the judgment of abnormal handling behavior. Detailed Implementation

[0028] 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, not all, of the embodiments of the present invention. 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.

[0029] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0030] This invention discloses an intelligent quality control method for electronic medical records based on medical knowledge graphs, referring to... Figure 1 This includes steps S1-S4:

[0031] S1. Obtain the target medical record data to be quality controlled, and construct the initial logical driving vector based on the position of the disease conclusion entity in the target medical record data in the general medical knowledge graph.

[0032] Specifically, the system connects to the electronic medical record system's data interface to obtain the target medical record data to be quality controlled, including chief complaint, present illness history, progress notes, treatment instructions and remarks, laboratory test results, vital signs, and a list of treatment instructions.

[0033] For unstructured text such as chief complaints, present medical history, medical records, and treatment instructions and remarks, entity recognition technology based on medical dictionaries is used to extract disease conclusion entities, symptom entities, and surgical entities, and the disease conclusion entities are mapped to the International Classification of Diseases (ICD) code. For structured data such as test results, vital signs, and treatment instruction lists, standardized cleaning is performed, the hospital's private codes are mapped to the national medical insurance standard code, and drug execution entities, test item entities, and their corresponding numerical results are extracted.

[0034] A pre-constructed general medical knowledge graph is introduced, which includes a set of nodes and a set of edges. The set of nodes covers medical concepts such as diseases, drugs, examination items, symptoms, and physiological indicators, while the set of edges... This includes medical relationships such as intervention, contraindications, induction, and monitoring, with each edge in the edge set containing an attribute set describing the physiological constraints that establish the medical relationship. Entities extracted from the target medical record data are used as indexes to perform retrieval and matching within a general medical knowledge graph. Specific nodes in the general medical knowledge graph corresponding to the disease conclusion entity of the current medical record are located, and these matched nodes are identified as the disease conclusion node set. Similarly, nodes corresponding to the treatment instruction entity are located and identified as the treatment instruction node set.

[0035] An initial logical driving vector for the target medical record data is constructed, with its dimension consistent with the total number of nodes in the general medical knowledge graph. Each value in the initial logical driving vector represents the energy value of the corresponding node in the initial stage of inference, satisfying the expression:

[0036]

[0037] In the formula, Represents the initial logical driving vector The Middle The value of each element; Represents the set of disease conclusion nodes The number of elements in the middle; Represents the first in the general medical knowledge graph One node; This represents the set of disease conclusion nodes. When a node... Belongs to the set of disease conclusion nodes When, assign it a non-zero initial value, through Normalization is performed to ensure that the sum of the initial energies of all disease conclusion nodes is 1; when the node Not part of the disease conclusion node set When the value is 0, the initial logic driving vector mathematically represents the business logic driving source of the current medical record, that is, it is assumed that all reasonable medical treatment behaviors in the medical record should be stimulated and driven by the current disease conclusion information.

[0038] S2. Based on the degree of matching between the measured values ​​of physiological indicators in the target medical record data and the physiological constraints recorded in the edge attributes of the general medical knowledge graph, calculate the dynamic propagation impedance of the directed edges in the general medical knowledge graph.

[0039] It should be noted that the directed edges connecting disease conclusion nodes and treatment instruction nodes in the general medical knowledge graph constitute standard business logic paths at the data topology level. These business logic paths represent the general business rule that corresponding treatment instructions should be executed based on a specific disease conclusion, such as drug A being used to intervene in disease B. Although business logic paths are generally considered to be connected by default in the general knowledge base, meaning the rule is universally valid, their effectiveness is limited by the physiological state recorded in the specific target medical record data. If the physiological indicator values ​​in the target medical record data deviate from the normal range, it means that the current state represented by the target medical record data does not meet the preconditions for executing the business rule, thus turning this conventional business logic path into a high-risk forbidden path. Existing technologies often ignore this dynamic constraint caused by individual data differences, leading to misjudgments. Therefore, this invention constructs dynamic propagation impedance by calculating the deviation between the actual indicator values ​​in the target medical record data and the preset standards in the edge attributes of the general medical knowledge graph. This reflects the degree of obstruction of the business logic path in the current medical record context, allowing for the suppression of business logic paths that do not conform to the physiological state in subsequent logical reasoning.

[0040] Specifically, traverse all directed edges in the general medical knowledge graph, and for any directed edge... First, determine the directed edge. Is it a directed edge connecting the disease conclusion node and the treatment instruction node? If so, then the directed edge... For business logic paths, otherwise, there should be directed edges. It does not belong to the business logic path.

[0041] Responding to directed edges For the business logic path, read the directed edge. The attached set of physiological constraint attributes contains several constraint tuples regarding physiological indicator types. Each tuple describes the numerical requirements of the intervention relationship for a specific physiological indicator type, such as an ideal mean serum creatinine of 80 μmol / L, an allowable standard deviation of 15 μmol / L, and a standardized tolerance threshold of 2. Based on the examination and testing items extracted from the target medical record data and their corresponding numerical results, a physiological feature vector is constructed. Each dimension of the physiological feature vector corresponds to a physiological indicator type, and its value is the measured value of that physiological indicator in the target medical record data. The physiological feature vector is then used in conjunction with the directed edge... The degree of matching of the physiological constraint attribute set determines the directed edge. Dynamic propagation impedance for the current target medical record data:

[0042]

[0043] In the formula, Indicates from node To the node Directed edge Dynamic propagation impedance within the current target medical record data context; Represents a directed edge The set of all physiological indicators contained in the physiological constraint attribute set; Represents the corresponding physiological indicators in the physiological feature vector. The measured values; Physiological indicators representing physiological constraint attribute sets The ideal mean; Physiological indicators representing physiological constraint attribute sets The permissible standard deviation; Indicators of physiological indicators The standardized tolerance threshold, which is a dimensionless value, is used to define the allowable fluctuation range of the Z-Score. The impedance sensitivity coefficient is used to measure the sensitivity of dynamic propagation impedance to deviations from physiological indicators. Based on the system's tolerance for quality control risks, the larger the value of the impedance sensitivity coefficient, the stronger the system's punishment of abnormal indicators, that is, the stricter the interception of potential contraindications. In this embodiment, the impedance sensitivity coefficient is set to 2. In other embodiments, implementers can adjust the impedance sensitivity coefficient according to the actual business scenario's requirements for the strictness of quality control. Represents the maximum value function; Indicates the absolute value symbol; To prevent extremely small positive numbers with a denominator of 0, in this embodiment, it is set to... In other embodiments, implementers may set the parameters according to the actual implementation situation.

[0044] When the measured value of a certain physiological indicator in the target medical record data Standardized deviation Exceeding the standardized tolerance threshold hour, A positive value results in a dynamic propagation impedance. The deviation increases linearly; when the measured value Within tolerance range, dynamic propagation impedance Maintain the base value of 1. Dynamic propagation impedance The larger the value, the more directed the edge. The more severely the business logic path it represents is restricted by the current physiological state, the worse its connectivity.

[0045] It should be noted that the ideal mean The baseline value of this physiological indicator under ideal conditions is set based on the median value of the normal reference range specified in authoritative clinical practice guidelines, or based on the arithmetic mean obtained from historical data of a large-scale healthy population; standard deviation is allowed. It is calculated based on the width of the normal fluctuation range of this physiological indicator as specified in medical standards, or based on the statistical standard deviation of historical data from a large-scale healthy population. This standard deviation is used to normalize deviations of physiological indicators of different dimensions into a unified measurement scale. It should be noted that when the allowable standard deviation is... When the value is 0, this physiological indicator is an absolute indicator and has no allowable fluctuation range. This embodiment uses [the appropriate value] in the dynamic propagation impedance calculation. Using it as the denominator in calculations is to avoid allowing standard deviation. A value of 0 is a special case that leads to calculation anomalies. The standardized tolerance threshold is set based on the principles of medical statistics and is usually set to 2 or 3, corresponding to the 95% or 99.7% confidence interval boundaries in a normal distribution, respectively. It is used to define whether the physiological indicator has a significant abnormality. In this embodiment, it is set to 3. In other embodiments, implementers can set the standardized tolerance threshold according to the actual implementation situation. For example, Figure 2 This diagram illustrates the deviation of physiological indicators, showing the specific causes of high dynamic propagation impedance. The horizontal axis indicates the specific correlation constraint of serum creatinine for metformin, and the vertical axis represents the standardized deviation. Figure 2 The results showed that the standardized deviation of the patient's serum creatinine level was as high as 6.7, which was significantly higher than the standardized tolerance threshold of 2.

[0046] Responding to directed edges If it does not belong to the business logic path, the dynamic propagation impedance of the directed edge is set to the default value of 1.

[0047] Through the above traversal calculations, a definite dynamic propagation impedance value is assigned to each directed edge in the general medical knowledge graph, thereby constructing a complete dynamic impedance topology.

[0048] S3. Construct the transition probability matrix of the general medical knowledge graph based on dynamic propagation impedance, and perform iterative calculations in conjunction with the initial logical driving vector until convergence is obtained to obtain the steady-state correlation distribution vector.

[0049] It should be noted that, in order to determine whether the treatment instructions in the target medical record data are reasonably driven by the disease conclusion, it is necessary to calculate the transmission distribution of logical association strength on the general medical knowledge graph. The logical association strength characterizes the degree of causal support between the disease conclusion as the logical starting point and the treatment instruction as the logical ending point. Traditional random walk algorithms assume that all business logic paths have the same transmission capability and cannot perceive the physiological constraints on the business logic paths. This can easily lead to the judgment that the relevant treatment instructions are reasonable even when the current state represented by the target medical record data does not meet the execution premise (i.e., there is a contraindication). Therefore, this invention uses the calculated dynamic propagation impedance to correct the transfer probability and constructs an impedance-modulated transfer mechanism. Under this transfer mechanism, admittance is used to characterize the transmission capability of the business logic path to the logical association strength. That is, the higher the impedance, the lower the admittance, and the less likely the logical association strength is to be transmitted through the business logic path. Through multiple rounds of mathematical iteration calculations, the logical association strength tends to diffuse along the low-impedance business logic path that conforms to the physiological state of the target medical record data, thereby obtaining an accurate medical treatment behavior relevance score.

[0050] Specifically, a transition probability matrix based on dynamic propagation impedance is constructed. This transition probability matrix describes the probability distribution of logical association strength from one node to another in a general medical knowledge graph. For each node... For each starting edge, its admittance is defined as the reciprocal of the dynamic propagation impedance of that edge, for nodes. The admittances of all outgoing edges are row normalized to obtain the transition probabilities from node u to another node:

[0051]

[0052] In the formula, Indicates from node To the node The transition probability; Indicates from node To the node Directed edge Dynamic propagation impedance within the current target medical record data context; Represents a node The set of all adjacent nodes; Indicates from node to adjacent nodes Directed edge Dynamic propagation impedance within the current target medical record data context. When dynamic propagation impedance... When the admittance is larger, The smaller the value, the lower the transition probability. The smaller the value, the less likely it is to prevent business logic from flowing to high-impedance nodes.

[0053] An iterative algorithm is used to calculate the steady-state relevance distribution vector of each node in a general medical knowledge graph. Each element in the steady-state relevance distribution vector represents the logical correlation between the corresponding node and the initial disease conclusion. The iterative process continuously updates the value of the steady-state relevance distribution vector until convergence. The update rule satisfies the expression:

[0054]

[0055] In the formula, Indicates the first Steady-state correlation distribution vector after the next iteration; Indicates the first The steady-state correlation distribution vector after the next iteration is defined in the first iteration as follows: equal ; Represents the transition probability matrix. Indicates the transpose symbol. Represents the transition probability matrix transpose; Indicates the initial logic driving vector; This represents the restart probability, used to adjust the calculation results for the initial logical driving vector. The degree of dependence.

[0056] It should be noted that this invention simulates the logical association strength starting from the disease conclusion node and affected by the transition probability matrix through iterative updates. The diffusion process within the general medical knowledge graph is controlled. Convergence is considered achieved when the changes in the values ​​of all elements in the steady-state correlation distribution vector before and after iteration are both below a preset iteration convergence threshold. At this point, the steady-state correlation distribution vector... The values ​​of each element in the table represent the relevance scores of the medical treatment behavior at the corresponding node. Among these, the restart probability... The key lies in balancing the global exploration of the topological structure of a general medical knowledge graph with the local preservation of initial disease conclusions. When the weight is set to a larger value, the iteration results tend to retain the original weights of the disease conclusion nodes, resulting in a smaller logical diffusion range; when When the setting is smaller, logical associations tend to spread more widely within the general medical knowledge graph, and the restart probability... The empirical value is 0.15. In other embodiments, implementers can adjust the restart probability based on the sparsity of the general medical knowledge graph and the business scenario's requirements for the depth of association mining. The iteration convergence threshold is used to determine whether the iterative calculation has reached a convergence state. Since general medical knowledge graphs typically contain millions of nodes, their average node probability weights are... To ensure the algorithm can capture subtle logical changes with real physical meaning, the iteration convergence threshold is set to [a specific value] in this implementation. ,at the same time, The machine minimum value, which is higher than that of a single-precision floating-point number, effectively avoids the iteration process from getting caught in meaningless floating-point rounding errors and computational noise, ensuring the numerical stability of the convergence determination. In other embodiments, implementers can set the iteration convergence threshold according to their specific needs for computational accuracy and efficiency.

[0057] For example, Figure 3 This diagram illustrates the energy propagation process of a random walk, showing the changing trends of the relevance scores of medical treatment behaviors at each node during the iterative calculation. As the iteration progresses, the energy from type 2 diabetes as an energy source gradually diminishes and diffuses; the score of the insulin node steadily increases and tends to converge; while the score of the metformin node is suppressed by high impedance and remains at a low level close to 0; and the score of aspirin remains 0 due to the lack of pathway connectivity.

[0058] S4. Based on the medical treatment behavior correlation score of the actual treatment instruction node in the target medical record data corresponding to the steady-state correlation distribution vector, calculate the dynamic judgment threshold, and use the dynamic judgment threshold to determine whether the actual treatment instruction is an abnormal treatment behavior.

[0059] It should be noted that after impedance-modulated random walk calculation, each node in the general medical knowledge graph will obtain a value representing the degree of logical closeness between it and the disease conclusion node, namely, the medical treatment behavior relevance score. For reasonable treatment instructions that actually exist in the target medical record data, since they are located on the low-impedance business logic path, they should obtain a high logical correlation strength; while for treatment instructions that are unrelated to the current disease conclusion (no path connection) or have physiological contraindications (path blocked by high impedance), their scores will be extremely low. Since there are significant differences in the complexity of the disease and the number of treatment instruction nodes in different target medical record data, it is easy to make mistakes by using a fixed threshold. Therefore, this invention calculates a dynamic judgment threshold based on the statistical distribution of the current medical record score, adaptively identifying abnormal treatment behaviors with scores significantly lower than the overall level, thereby achieving accurate defect localization.

[0060] Specifically, from the steady-state correlation distribution vector Extract the set of all actual treatment instruction nodes from the target medical record data. The corresponding medical treatment behavior relevance score is used to calculate the dynamic judgment threshold:

[0061]

[0062] In the formula, Indicates a dynamic threshold; Represents the set of actual disposal instruction nodes. Middle node Relevance score of medical treatment behavior; Represents the set of actual disposal instruction nodes. The number of nodes in the middle; This represents the abnormal deviation coefficient. This represents the average relevance score of all actually executed treatment instructions in the current target medical record data. It reflects the overall logical coherence of the current medical record. The higher the average relevance score, the higher the dynamic threshold. The corresponding increase indicates that quality control standards are dynamically adjusted according to the quality of medical records. Abnormal deviation coefficient. This is used to adjust the system's tolerance threshold for low-relevance behaviors, set based on the business scenario's need to balance recall and accuracy in quality control. When the threshold is set too high, dynamically determine the threshold. As the number of variables increases, the system's tolerance for weak logical relationships decreases, and quality control standards become more stringent; when... When the setting is small, the system tends to only intercept extreme anomalies. In this embodiment, the anomaly deviation coefficient... The empirical value is 0.1. In other embodiments, implementers can adjust the abnormal deviation coefficient based on the statistical distribution characteristics of historical quality control data.

[0063] The relevance score of medical treatment behavior in response to actual treatment instructions is less than the dynamic judgment threshold. If so, the actual handling instruction is determined to be an abnormal handling behavior.

[0064] Furthermore, an effective connectivity threshold is introduced to classify abnormal handling behaviors by attribution: For the actual handling instruction corresponding to the abnormal handling behavior, if the relevance score of the medical handling behavior in response to the actual handling instruction is less than the preset effective connectivity threshold, it indicates that there is no effective connectivity path from the disease conclusion node to the actual handling instruction in the general medical knowledge graph, that is, there is no logical connection between the two in the topological structure, and the system marks its defect type as an unfounded execution item; if the relevance score of the medical handling behavior in response to the actual handling instruction is greater than or equal to the preset effective connectivity threshold but less than the dynamic judgment threshold, it indicates that there is a business logic path between the two, but the dynamic propagation impedance on the business logic path is high, which leads to the obstruction of the transmission of logical association strength, and the system marks its defect type as the presence of contraindications or inconsistencies in physiological indicators.

[0065] The effective connectivity threshold is used to define the effective signal boundaries of the system and eliminate computational noise. Due to the inherent precision limitations of computer floating-point arithmetic, extremely small numerical fluctuations often represent computational residuals without physical meaning rather than actual logical transmissions. For example, values ​​less than [a certain value]... To avoid misjudging such computational noise as weak business logic connections, a clear noise filtering threshold must be set. Therefore, this invention sets the effective connectivity threshold to... This ensures that only actual treatment instructions with a medical treatment behavior relevance score exceeding this threshold are considered to have a topological connection.

[0066] For example, Figure 4 This diagram illustrates the assessment of abnormal treatment behaviors. The horizontal axis represents medical order items, and the vertical axis represents the relevance score of the medical treatment behavior. Figure 4 The insulin score was above the dynamic threshold, indicating a pass; the metformin score was significantly below the dynamic threshold but above the effective connectivity threshold. The behavior was deemed abnormal, with the defect type being the presence of contraindications or inconsistent physiological indicators; the aspirin score was 0, which is below the effective connectivity threshold. This was determined to be an abnormal handling behavior, and the defect type was "execution without basis".

[0067] Furthermore, a structured quality control defect report is generated, which includes defect handling items, related disease conclusions, inferred causes, and reference physiological indicators. This report is then fed back to the electronic medical record system via a data interface. The inferred cause is the defect type of the abnormal handling behavior. The reference physiological indicators are specific physiological indicator types that cause a significant increase in dynamic propagation impedance on the business logic path, along with their measured values ​​in the target medical record data. Output is only provided when the inferred cause is the presence of contraindications or a mismatch in physiological indicators. The specific acquisition method is as follows: by tracing back the business logic path connecting the related disease conclusion node and the abnormal handling behavior node, the physiological constraint attribute set on the business logic path is traversed to filter out specific physiological indicator types whose standardized deviation exceeds the standardized tolerance threshold, and the measured values ​​corresponding to these physiological indicators in the target medical record data are extracted.

[0068] This invention also discloses an intelligent quality control system for electronic medical records based on medical knowledge graphs, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement an intelligent quality control method for electronic medical records based on medical knowledge graphs according to this invention.

[0069] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

Claims

1. An electronic medical record intelligent quality control method based on a medical knowledge graph, characterized in that, include: The process involves acquiring the target medical record data to be quality controlled, and constructing an initial logical driving vector based on the position of the disease conclusion entities in the target medical record data within the general medical knowledge graph. This includes: extracting disease conclusion entities, physiological indicators and their corresponding measured values, and actual treatment instructions from the target medical record data; mapping the disease conclusion entities to corresponding nodes in the general medical knowledge graph to determine the set of disease conclusion nodes; and constructing an initial logical driving vector with dimensions consistent with the total number of nodes in the general medical knowledge graph. For each element in the initial logical driving vector, if its corresponding node belongs to the set of disease conclusion nodes, a non-zero initial value is assigned; otherwise, a zero value is assigned. Based on the degree of matching between the measured values ​​of physiological indicators in the target medical record data and the physiological constraints recorded in the edge attributes of the general medical knowledge graph, the dynamic propagation impedance of directed edges in the general medical knowledge graph is calculated, including: in response to the directed edge not being a business logic path connecting the disease conclusion node and the treatment instruction node, the dynamic propagation impedance is set to a default value; otherwise, the dynamic propagation impedance is calculated. In the formula, Indicates from node To the node directed edges Dynamic propagation impedance within the current target medical record data context; Represents a directed edge The set of all physiological indicators contained in the physiological constraint attribute set; Represents the corresponding physiological indicators in the physiological feature vector. The measured values; , These represent physiological indicators within the set of physiological constraint attributes. The ideal mean and allowable standard deviation; Indicators of physiological indicators Standardized tolerance threshold; Indicates the impedance sensitivity coefficient; To prevent extremely small positive numbers with a denominator of 0; The transition probability matrix for constructing a general medical knowledge graph based on dynamic propagation impedance includes: the transition probability from any node to another node satisfies the expression: In the formula, Indicates from node To the node The transition probability; Represents a node The set of all adjacent nodes; Indicates from node to adjacent nodes directed edges The dynamic propagation impedance in the current target medical record data context is calculated iteratively in combination with the initial logical driving vector until the steady-state correlation distribution vector is obtained. Based on the medical treatment behavior relevance score of the actual treatment instruction node in the target medical record data corresponding to the steady-state relevance distribution vector, a dynamic judgment threshold is calculated, including: extracting the element value of each actual treatment instruction node in the target medical record data in the steady-state relevance distribution vector as the medical treatment behavior relevance score of the actual treatment instruction node; and calculating the dynamic judgment threshold based on the medical treatment behavior relevance score. In the formula, Indicates a dynamic threshold; The nodes in the set that represent the actual disposal instruction nodes Relevance score of medical treatment behavior; This indicates the number of nodes in the set of nodes that actually handle the instructions; The abnormal deviation coefficient is represented, and a dynamic judgment threshold is used to determine whether the actual treatment instruction is an abnormal treatment behavior, including: when the medical treatment behavior relevance score in response to the actual treatment instruction node is less than the dynamic judgment threshold, the actual treatment instruction is determined to be an abnormal treatment behavior.

2. The intelligent quality control method for electronic medical records based on medical knowledge graphs according to claim 1, characterized in that, The elements in the initial logical driving vector satisfy the expression: ； In the formula, Represents the initial logical driving vector The Middle The value of each element; Represents the set of disease conclusion nodes; Represents the set of disease conclusion nodes The number of elements in the middle; Represents the first in the general medical knowledge graph Each node.

3. The intelligent quality control method for electronic medical records based on medical knowledge graphs according to claim 1, characterized in that, The update rule for iterative computation satisfies the expression: ； In the formula, Indicates the first Steady-state correlation distribution vector after the next iteration; Indicates the initial logic driving vector; Indicates the first The steady-state correlation distribution vector after the next iteration is defined in the first iteration as follows: equal ; Represents the transition probability matrix. Indicates the transpose symbol; This indicates the probability of restarting.

4. The intelligent quality control method for electronic medical records based on medical knowledge graphs according to claim 1, characterized in that, Also includes: If the relevance score of a medical treatment action in response to an actual treatment instruction is less than the preset effective connectivity threshold, its defect type is marked as an unfounded execution item. If the relevance score of a medical treatment behavior in response to an actual treatment instruction is greater than or equal to a preset effective connectivity threshold but less than a dynamic judgment threshold, its defect type is marked as having contraindications or inconsistent physiological indicators; a structured quality control defect report is generated. When the defect type is having contraindications or inconsistent physiological indicators, the structured quality control defect report includes reference physiological indicators that lead to an increase in dynamic propagation impedance; the method for obtaining the reference physiological indicators is: tracing back the business logic path connecting the associated disease conclusion node and the actual treatment instruction node, and filtering out physiological indicator types whose standardized deviation exceeds the standardized tolerance threshold.

5. An intelligent quality control system for electronic medical records based on medical knowledge graphs, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement an intelligent quality control method for electronic medical records based on a medical knowledge graph, as described in any one of claims 1-4.

Images