Abnormal behavior detection method, system and device, program product and storage medium

Abstract

The invention discloses an abnormal behavior detection method, system and device, a program product and a storage medium. The method comprises the steps of obtaining medical behavior information; taking the medical behavior information as nodes, and constructing edges based on relationships among the nodes to obtain a medical behavior relationship network; utilizing a community detection algorithm to perform community division based on correlation among nodes in the medical behavior relation network, and enabling the divided communities to represent node groups with medical behavior characteristics of which the similarity accords with a preset similarity standard; determining a high-frequency medical behavior event corresponding to each community based on each community obtained by division; and inputting the high-frequency medical behavior event and the cue word into a trained detection model to obtain an abnormal medication behavior in the high-frequency medical behavior event. According to the invention, automatic identification of abnormal medication behaviors in high-frequency medical behavior events is realized, the working efficiency is improved, and the increasingly increasing challenges of medical behavior data can be coped with.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to an abnormal behavior detection method, system, device, program product and storage medium. Background Art

[0002] In recent years, the incidence of chronic diseases has continued to rise, and is showing a trend of younger people. This change has led to the continuous release of protection needs for people with chronic diseases. Chronic disease patients often need to take medication for a long time or even for life. This phenomenon has not only greatly increased the burden on the medical system, but also brought higher claims risks to insurance companies.

[0003] At present, the Medical Insurance Bureau and insurance companies mainly use manual review to identify irrational drug use. Although the manual review method can conduct detailed analysis of individual cases, its limitations are also significant. On the one hand, manual review requires a lot of time and human resources, which is costly; on the other hand, the review process is easily affected by the subjective judgment of the reviewers, resulting in inconsistency and low accuracy in the review results. With the continuous increase in the number of patients with chronic diseases, the speed and efficiency of manual review are even more difficult to meet actual needs. Summary of the invention

[0004] Based on the above problems, the present application provides an abnormal behavior detection method, system, device, program product and storage medium.

[0005] The embodiments of the present application disclose the following technical solutions:

[0006] A first aspect of an embodiment of the present application provides an abnormal behavior detection method, comprising:

[0007] Obtaining information about medical practices;

[0008] Taking the medical behavior information as nodes, building edges based on the relationships between the nodes, and obtaining a medical behavior relationship network;

[0009] Using a community detection algorithm, community division is performed based on the correlation between nodes in the medical behavior relationship network, and the divided communities represent groups of nodes with medical behavior characteristics whose similarity meets a preset similarity standard;

[0010] Based on the divided communities, determine the high-frequency medical behavior events corresponding to each community;

[0011] The high-frequency medical behavior events and prompt words are input into the trained detection model to obtain abnormal medication behaviors in the high-frequency medical behavior events. The detection model is constructed and trained based on a large language model and a disease knowledge graph, and the accuracy of identifying abnormal medication behaviors reaches a preset threshold.

[0012] In a possible implementation, the high-frequency medical behavior events obtained based on the divided communities include:

[0013] For each community, count the frequency of medical behaviors and / or patient characteristics within the community;

[0014] The frequency results obtained based on statistics are compared with the corresponding preset frequency thresholds, and medical behavior events in each community that are higher than the preset frequency thresholds are screened out as high-frequency medical behavior events.

[0015] In a possible implementation, the method for determining the trained detection model includes:

[0016] Obtain information on disease theories;

[0017] The disease knowledge graph is constructed based on disease theory information;

[0018] Graph embedding algorithms are used to convert entities and relationships between entities in the disease knowledge graph into vector representations in high-dimensional space. Large language models are used to convert disease theory information into vector representations in high-dimensional space.

[0019] Aligning vector representations in high-dimensional space from disease knowledge graphs and disease theory information, the alignment is used to achieve a unified representation so that the model can process both structured and unstructured information in the same framework;

[0020] The large language model is trained based on prompt words to obtain a detection model that can achieve a preset threshold accuracy in identifying abnormal medication behavior in high-frequency medical behavior events. The prompt words are used to guide the large language model to generate detection content related to irrational medication behavior.

[0021] In a possible implementation, before the large language model is trained based on the prompt words so as to obtain a detection model with an accuracy rate of identifying abnormal medication behavior in high-frequency medical behavior events reaching a preset threshold, the method further includes:

[0022] In combination with the optimization goal of the disease knowledge graph and the task goal of the large language model, the large language model is trained with the unified goal obtained so that the model can simultaneously achieve the optimization goal of the disease knowledge graph and the task goal of the large language model.

[0023] In a possible implementation, the obtaining of medical behavior information includes:

[0024] Obtaining preliminary data information;

[0025] The prepared data information is cleaned and standardized to obtain medical behavior information, wherein the standardized processing is used to present data information of the same concept in the prepared data information in a consistent manner.

[0026] A first aspect of an embodiment of the present application provides an abnormal behavior detection system, including:

[0027] An acquisition unit, used for acquiring medical behavior information;

[0028] A construction unit, used to use the medical behavior information as nodes, construct edges based on the relationship between each node, and obtain a medical behavior relationship network;

[0029] A division unit, used to perform community division based on the correlation between nodes in the medical behavior relationship network using a community detection algorithm, wherein the divided communities represent groups of nodes having medical behavior characteristics whose similarities meet a preset similarity standard;

[0030] A determination unit, used for determining high-frequency medical behavior events corresponding to each community based on each community obtained by division;

[0031] The detection unit is used to input the high-frequency medical behavior events and prompt words into a trained detection model to obtain abnormal medication behavior in the high-frequency medical behavior events. The detection model is constructed and trained based on a large language model and a disease knowledge graph, and the accuracy of identifying abnormal medication behavior reaches a preset threshold.

[0032] A third aspect of an embodiment of the present application provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the abnormal behavior detection method as described in the first aspect above is implemented.

[0033] A fourth aspect of the embodiments of the present application provides a computer program product. When the computer program product runs on a computer, the computer executes the abnormal behavior detection method as described in the first aspect above.

[0034] A fifth aspect of an embodiment of the present application provides a computer-readable storage medium, in which instructions are stored. When the instructions are executed on a terminal device, the terminal device executes the abnormal behavior detection method as described in the first aspect above.

[0035] Compared with the prior art, this application has the following beneficial effects:

[0036] This application realizes the automated processing of a large amount of medical data by acquiring medical behavior information and constructing a medical behavior relationship network. Compared with the traditional manual case-by-case review method, this method can greatly reduce the time and human resources required. Using community detection algorithms for large-scale data analysis can simultaneously process a large number of medical behavior events and quickly identify high-frequency medical behavior events that may be in abnormal states. The high-frequency medical behavior events and prompt words are input into the trained detection model to obtain abnormal medication behaviors in the high-frequency medical behavior events. The detection model is constructed based on a large language model and a disease knowledge graph. After a lot of training and optimization, it has a high recognition accuracy and stability, and can effectively avoid subjective judgment errors in the manual review process. Therefore, compared with the manual review method, this application can simultaneously process a large amount of medical behavior data, realize the automatic identification of abnormal medication behaviors in high-frequency medical behavior events, which not only improves work efficiency, but also enables the challenge of growing medical behavior data to be met. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0038] Figure 1 A flowchart of an abnormal behavior detection method provided in an embodiment of the present application;

[0039] Figure 2 A schematic diagram of the collaborative fusion process of a large language model and a chronic disease knowledge graph provided in an embodiment of the present application;

[0040] Figure 3 A schematic diagram of a first application scenario of a prompt word provided in an embodiment of the present application;

[0041] Figure 4 A schematic diagram of a second application scenario of a prompt word provided in an embodiment of the present application;

[0042] Figure 5 A structural diagram of an abnormal behavior detection system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0043] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0044] To facilitate understanding of the technical solutions provided by the embodiments of the present application, the background technology involved in the embodiments of the present application will be described below.

[0045] As mentioned above, identifying irrational medication behavior among chronic disease patients is of great significance for effectively controlling and managing chronic disease medical expenses. In order to meet this challenge, the Medical Insurance Bureau and insurance companies currently use three main methods to identify irrational medication behavior: manual review, rule recognition, and supervised learning methods. However, these methods all have obvious limitations in practical applications.

[0046] First, although the manual review method can conduct detailed analysis of individual cases, its limitations are also significant. On the one hand, manual review requires a lot of time and human resources, which is costly; on the other hand, the review process is easily affected by the subjective judgment of the reviewers, resulting in inconsistency and low accuracy of the review results. With the continuous increase in the number of patients with chronic diseases, the speed and efficiency of manual review are even more difficult to meet actual needs.

[0047] Secondly, the rule identification method is highly dependent on the experience and knowledge of medical experts to set the rules. Most of these rules are based on known patterns of irrational medication behavior. However, once the rules are established, they often lack flexibility and are difficult to adjust dynamically. In the face of new irrational medication behaviors, existing rules are often unable to adapt quickly, resulting in delayed updates. In addition, the rule-setting process is often too rigid, making it difficult to personalize considerations for the specific conditions and treatment needs of different patients, which can easily lead to misjudgments or missed judgments during the judgment process.

[0048] Finally, although supervised learning methods have improved the accuracy and efficiency of identifying irrational medication behaviors to a certain extent, their limitations cannot be ignored. This method mainly relies on labeled data to identify potential irrational medication behaviors by analyzing the association between input data and corresponding output labels. However, given the diversity and complexity of chronic disease drug treatments, it is usually difficult to collect a sufficient amount of labeled data for learning. Limited labeled data may lead to model overfitting or bias, thereby affecting the fairness and accuracy of the model. In addition, as chronic disease treatment plans and medication behaviors continue to evolve, supervised learning models need to be regularly updated and retrained, which is also quite challenging in actual operations.

[0049] In summary, existing technologies have obvious limitations in identifying irrational medication behaviors for chronic diseases.

[0050] In order to solve the above problems, the embodiment of the present application constructs a medical behavior relationship network and uses a community detection algorithm to divide the community. This process is highly automated and does not require a lot of manual participation. This significantly reduces the investment in time and human resources and improves the processing speed. The detection model constructed by the large language model and the disease knowledge graph can intelligently identify abnormal medication behavior and avoid accuracy problems caused by subjectivity in manual review.

[0051] Compared with the rule-based method, the detection model used in this application has greater flexibility and adaptability. The model can continuously learn and adjust based on new data, so as to quickly adapt to the emerging irrational drug use behavior patterns and avoid the update lag problem in the rule setting method. Through community division, the model can conduct more detailed analysis of node groups with similar medical behavior characteristics, realize personalized considerations, and reduce the possibility of misjudgment or missed judgment.

[0052] This application uses high-frequency medical behavior events and prompt words as input to identify abnormal medication behavior through a trained detection model. This method relies on a large amount of data rather than limited labeled data, thereby reducing the risk of overfitting or bias. As chronic disease treatment plans and medication behaviors continue to evolve, the model can optimize its performance through continuous data input and training to ensure the accuracy and fairness of the model. This solves the problem that the model in the supervised learning method needs to be updated and retrained regularly.

[0053] In summary, this application achieves efficient, accurate, flexible and intelligent analysis of medical behavior by constructing a medical behavior relationship network, using community detection algorithms to divide communities, and using a detection model based on a large language model and disease knowledge graph. This not only solves the problems of large resource investment, strong subjectivity, and low efficiency in manual review methods, but also overcomes the defects of poor flexibility, delayed updates, misjudgments or missed judgments in rule setting methods, and solves the challenges of insufficient data, overfitting or deviation of models, and the need for regular updates in supervised learning methods.

[0054] It should be noted that the abnormal behavior detection method provided in the present application can be applied to the field of computer technology. The above is only an example and does not limit the application field of the abnormal behavior detection method, system, device and medium provided in the present application. In addition, the embodiments of the present application may not limit the execution subject of abnormal behavior detection. For example, the abnormal behavior detection method of the embodiment of the present application can be applied to data processing devices such as terminal devices or servers. Among them, the terminal device can be an electronic device such as a computer, a personal digital assistant (PDA). The server can be a stand-alone server, a cloud server, or a cluster server composed of multiple servers.

[0055] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0056] The following describes an abnormal behavior detection method provided by the present application through an embodiment. Figure 1 ,Should Figure 1 A flowchart of an abnormal behavior detection method provided in an embodiment of the present application, the method comprising:

[0057] S101. Obtain medical behavior information.

[0058] In a possible implementation, the obtaining of medical behavior information includes:

[0059] Acquire the preliminary data information; clean the preliminary data information, and perform standardization on the cleaned preliminary data information to obtain medical behavior information, wherein the standardization is used to present data information of the same concept in the preliminary data information in a consistent manner.

[0060] Collect and integrate data from different sources to obtain complete and structured preliminary data information for subsequent steps. For example, a medical consultation information form: contains the patient's basic information (such as name, gender, age), the time and place of each visit, and the responsible doctor, etc. Disease diagnosis information form: records the patient's disease diagnosis, including diagnosis date, disease code, etc. Cost settlement details table: covers the specific costs of each medical service, drug costs, examination costs, treatment costs, etc., and may include payment methods (self-funded, medical insurance reimbursement, etc.).

[0061] In one possible implementation, the acquired data may be cleaned to ensure data quality, remove erroneous or irrelevant information, and make the data suitable for further analysis.

[0062] Data cleaning steps can include: Removing invalid records, that is, identifying and deleting records that are obviously wrong or insignificant, such as duplicate entries, records with missing key fields, etc. Outlier processing, that is, detecting and processing outliers through statistical methods or other rules, such as extremely high drug doses, illogical dates, etc. Unifying the format. For date formats, all data involving timestamps should be converted to a unified format, such as YYYY-MM-DD, to facilitate calculation of time and sorting. For digital formats, ensure the consistency of numerical data, such as currency amounts, ages, quantities, etc. should use the same units and decimal places.

[0063] In one possible implementation, the acquired data may be subjected to data standardization to normalize the data representation so that data with the same concept are presented in a consistent manner for easy comparison and analysis. This process may include:

[0064] Drug name standardization: Converting a drug’s trade name to a generic name ensures that even different brands of the same drug are correctly identified as belonging to the same class.

[0065] Drug quantity standardization: The quantity of drugs should be converted into numerical values ​​in standard dosage units. For example, if some drugs in the original data are counted in "tablets" and some are counted in "milliliters", they need to be uniformly converted into a standard unit, such as "milligrams", based on the conversion relationship provided in the drug instructions.

[0066] Suppose there is an original record as follows: Drug name: "Enteric-coated Aspirin Tablets", Quantity: "10 tablets", Dosage: "100 mg per tablet". After standardization, the record can be changed to: Drug name: "Aspirin" (generic name), Quantity: "1000 mg" (convert "tablets" to "mg", that is, 10 tablets * 100 mg per tablet).

[0067] Such a standardization process helps eliminate confusion caused by differences in expression, making subsequent analysis more accurate and reliable. At the same time, it also facilitates data integration across databases or systems.

[0068] S102: Take the medical behavior information as nodes, build edges based on the relationships between the nodes, and obtain a medical behavior relationship network.

[0069] That is, a network structure is constructed to reflect the correlation between patients' medical behaviors. Medical behavior information is used as nodes in the behavior relationship network. The relationship between nodes is used to establish edges in the behavior relationship network. Using the extracted node and edge information, a graph-based medical behavior relationship network is constructed.

[0070] The following are several exemplary nodes. Patient information (patient ID) represents an individual in the network. Chronic disease name indicates the type of chronic disease the patient suffers from. Consultation details include consultation date, medical institution, medical expenses, generic name of medicine, dosage unit, quantity and other information.

[0071] The relationship between nodes is used to establish the edge in the behavior relationship network, for example: patient-hospital, which indicates the medical institution where the patient is treated. Patient-chronic disease, which indicates the name of the chronic disease diagnosed by the patient. Patient-drug, which indicates the generic name of the drug prescribed when the patient is treated. Patient-drug quantity, although the quantity is usually not directly represented as an edge, it can be considered as a weight in subsequent analysis.

[0072] In one possible implementation, the NetworkX algorithm can be used to process medical behavior information, taking the medical behaviors (such as diagnosis, treatment, medication, etc.) indicated by the medical behavior information as nodes, and constructing edges based on the associations between patients to form a detailed medical behavior relationship network. This step provides a solid data foundation for subsequent analysis.

[0073] S103, using a community detection algorithm to divide communities based on the correlation between nodes in the medical behavior relationship network, where the divided communities represent groups of nodes with medical behavior characteristics whose similarities meet a preset similarity standard.

[0074] The correlation can be determined based on the connection strength, common attributes or behavior patterns between nodes. The divided communities represent groups of nodes with similar medical behavior characteristics.

[0075] Before or after community segmentation, a preset similarity standard needs to be set. This standard can be defined based on the similarity of medical behavior characteristics, common characteristics of patient groups (such as age, gender, medical history, etc.), or the correlation between medical events. This standard will be used to evaluate whether the community obtained by segmentation meets expectations, that is, whether it represents a group of nodes with medical behavior characteristics that meet the similarity standards.

[0076] Exemplarily, the community detection algorithm may be the Louvain algorithm. That is, the Louvain algorithm is used to perform community detection on the constructed medical behavior relationship network. The Chinese name of the Louvain algorithm is "Loren algorithm" or more commonly, it is directly transliterated as "Luwan algorithm". The Louvain algorithm is a modular-based community partitioning method that aims to partition the network graph by maximizing modularity. The partitioning function of the Louvain algorithm partitions based on the correlation between nodes to form a community. The divided community information is saved for subsequent analysis.

[0077] Communities are formed based on the maximum correlation between nodes in the network graph. The modularity value (Q value) ranges from [-1 / 2, 1) and is used to evaluate the quality of the division. The higher the Q value, the more reasonable the community division. The divided community information is saved accordingly for subsequent analysis. These communities represent closely connected node groups in the medical behavior relationship network, reflecting a specific medical behavior pattern or event.

[0078] S104. Obtain high-frequency medical behavior events based on the divided communities.

[0079] Further analysis is performed on the divided communities to identify high-frequency medical behavior events. High-frequency medical behavior events include but are not limited to:

[0080] Individual high-frequency events: The number of specific medical behavior events that occur to an individual in a certain period of time is significantly higher than the average. For example, the number of diabetes visits by patient A in a quarter is significantly higher than that of other patients.

[0081] High-frequency group events: The number of medical behavior events of the same type occurring in a certain group of people within a certain period of time is significantly higher than the average level. For example, a group of hypertensive patients frequently visited the doctor and prescribed a large number of sacubitril valsartan sodium tablets in a quarter, which was significantly higher than other patients.

[0082] Hierarchical relationship events: Medical behavior events with hierarchical relationships are summarized through clustering, classification, etc. For example, the number of diabetic patients seen by multiple departments of a hospital surges in the same period of time, and these patients have similar backgrounds or characteristics.

[0083] The divided community information is saved, including the list of nodes contained in each community and its related attributes. The identified high-frequency medical behavior events are recorded in detail, including the time period of the event, the population or individuals involved, the specific description of the event, and the possible cause analysis.

[0084] In a possible implementation, the high-frequency medical behavior events obtained based on the divided communities include:

[0085] For each community, the frequency of occurrence of medical behaviors and / or patient characteristics in the community is counted; based on the statistical frequency results, they are compared with the corresponding preset frequency thresholds, and medical behavior events in each community that are higher than the preset frequency thresholds are screened out as high-frequency medical behavior events.

[0086] For all medical behaviors (such as diagnosis, treatment, medication, etc.) in each community, record the number of times these behaviors occur. For example, how many times a certain drug was prescribed in a specific time period; how many surgeries of a certain type were performed, etc. At the same time, statistics can also be collected on relevant characteristics of patients, such as age distribution, gender ratio, types of chronic diseases, etc. This helps to understand which types of patients experience certain medical behaviors more frequently.

[0087] Set a reasonable frequency threshold for each medical behavior and patient characteristic based on historical data or expert opinions. This threshold is used to distinguish the boundary between "normal" and "high frequency". Compare the actual frequency counted with the preset threshold. If the frequency of a medical behavior or patient characteristic exceeds the set threshold, it is considered high frequency.

[0088] Through the above comparison process, it is possible to identify which medical behaviors occur too frequently in a specific community. These behaviors that exceed the threshold are considered high-frequency medical behavior events. For the high-frequency events screened out, a deeper analysis can be performed in combination with contextual information (such as time trends, geographic location, etc.) to explore the reasons behind them and assess whether there may be abnormal situations.

[0089] Therefore, the embodiments of the present application can systematically identify those medical behaviors that occur abnormally frequently in a specific group, thereby helping medical institutions to warn of potential problem areas in advance, such as whether there is overtreatment, irrational use of drugs, etc. At the same time, by comparing and analyzing high-frequency events between different communities, it is also possible to determine the differences and imbalances in medical services.

[0090] In addition, this method is flexible and allows personalized adjustments based on the specific conditions of different regions, hospitals and even departments, ensuring that the threshold setting is more in line with actual conditions and improving the accuracy and practicality of detection.

[0091] S105, inputting the high-frequency medical behavior events and prompt words into the trained large language model to obtain abnormal medication behaviors in the high-frequency medical behavior events.

[0092] By combining the large language model (LLM) and the structured disease knowledge graph, irrational medication behaviors in high-frequency medical behavior events can be identified. This method utilizes the powerful text processing capabilities of the large language model and the accuracy of the knowledge graph in professional fields, improving the semantic understanding and explanatory reasoning capabilities of the model, thereby improving the accuracy of detecting irrational medication behaviors for chronic diseases.

[0093] In one possible implementation, see Figure 2 , Figure 2 A schematic diagram of the collaborative fusion process of a large language model and a chronic disease knowledge graph provided in an embodiment of the present application, wherein the method for determining the trained detection model includes:

[0094] Step A1: Obtain disease theory information.

[0095] Collect chronic disease-related drug instructions, authoritative medical literature, clinical diagnosis and treatment guidelines, and high-quality medical records as disease theory information. Disease theory information provides professional information on drug use, indications, contraindications, side effects, interactions, etc.

[0096] Step A2: construct a disease knowledge graph based on disease theory information.

[0097] Based on the collected disease theory information, a disease knowledge graph is constructed, where entities can include chronic diseases (such as diabetes, hypertension), drugs (such as aspirin), populations (such as elderly patients), etc. Relationships can describe the associations between entities, such as indications, contraindications, side effects, drug interactions, etc. A structured disease knowledge graph is constructed using the above entities and relationships. This graph will serve as one of the key inputs for subsequent operations to provide professional background knowledge support for the model.

[0098] Step A3: Use a graph embedding algorithm to convert entities and relationships between entities in the disease knowledge graph into vector representations in a high-dimensional space; use a large language model to convert disease theory information into vector representations in a high-dimensional space.

[0099] Step A4: align the vector representations from the disease knowledge graph and the disease theory information, wherein the alignment is used to achieve a unified representation so that the model can process both structured and unstructured information within the same framework.

[0100] Using graph embedding methods, entities and relations in the disease knowledge graph are converted into high-dimensional embedding representations. These vectors capture the relationships and attributes between entities, help the model understand the connections between different concepts, and provide a basis for reasoning.

[0101] The pre-trained large language model is used to encode the collected disease theory information and generate a corresponding vector representation in the high-dimensional space. This process not only extracts the semantic features of the text, but also retains the contextual information, enhancing the model's understanding and generalization capabilities.

[0102] Through specific technical means (such as the KEPLER model), the embedding representations from the knowledge graph and disease theory information are aligned to achieve a unified representation. This fusion ensures a seamless connection between the two different types of data, enabling the model to process structured and unstructured information simultaneously within the same framework. Among them, the Chinese name of the KEPLER model is the unified model of knowledge embedding and pre-trained language representation. The model aims to combine knowledge embedding (KnowledgeEmbedding) with pre-trained language representation (Pre-trained Language Representation) to better utilize text information and relational facts in the knowledge graph.

[0103] In one possible implementation, the training process also includes: combining the optimization goal of the disease knowledge graph and the task goal of the large language model, and training the large language model with the unified goal obtained by the combination, so that the model can simultaneously complete the optimization goal of the disease knowledge graph and the task goal of the large language model.

[0104] In order to further enhance the performance of the model, the optimization objectives of the knowledge graph (such as link prediction and triple classification) can be combined with the tasks of the large language model (such as question-answering system and text generation) as a unified goal of synergistic fusion. Doing so can make full use of the advantages of both and make up for their respective shortcomings, especially in the reasoning ability and semantic understanding of professional fields.

[0105] Step A5: training the large language model based on prompt words to obtain a detection model that can identify abnormal medication behavior in high-frequency medical behavior events with an accuracy rate reaching a preset threshold, wherein the prompt words are used to guide the large language model to generate detection content related to irrational medication behavior.

[0106] Design prompt words for specific areas, which can include drug information, patient medical history, medication records and other dimensions to ensure that the large model can fully identify potential risks of irrational medication. For example, design corresponding prompt words for scenarios such as "overdose", "drug interactions", "contraindications", and "drug duplication". Figure 3 , Figure 3 A schematic diagram of the first prompt word application scenario provided in an embodiment of the present application.

[0107] Input prompt: I currently have the following chronic disease medication case: Patient A visited XX Hospital in September 24, diagnosed with hypertension, and prescribed Sacubitril Valsartan Sodium Tablets. Please help me extract the main method of this medication case, and refer to the knowledge graph to detect whether there is unreasonable medication behavior? If so, please mark it in the output method and provide detailed description.

[0108] The training process outputs the following response: Based on the medication cases you provided, we have extracted the main methods and detected whether there are unreasonable medication behaviors by referring to the knowledge graph. The following are the detailed analysis results:

[0109] ###Main method refinement

[0110] 1. Name: Patient A;

[0111] 2. Date of consultation: September 24;

[0112] 3. Hospital visited: Shanghai XX Hospital;

[0113] 4. Diagnostic diseases: hypertension;

[0114] 5. Medication regimen: Sacubitril / valsartan sodium tablets.

[0115] ###Detection of irrational drug use

[0116] 1. Sacubitril / valsartan sodium tablets: a drug used to treat hypertension that lowers blood pressure by inhibiting the action of angiotensin II.

[0117] 2. Indications: Hypertension;

[0118] 3. Dosage: 1 to 2 tablets each time, twice a day;

[0119] 4. Precautions: Monitor blood pressure and renal function during use and avoid taking with potassium supplements;

[0120] Conclusion: According to the detection results of the knowledge graph, no obvious irrational drug use behavior was found.

[0121] These prompt words are used to guide the large language model to generate detection content about irrational drug use behavior. By repeatedly optimizing and iterating the answers output by the model, its accuracy, sensitivity and specificity in detecting abnormal drug use behavior are continuously improved.

[0122] In this case, in order to detect irrational medication use, the following domain-specific prompts are included:

[0123] Drug information prompt words: including drug name (sacubitril valsartan sodium tablets), indications, usage and dosage, precautions, etc. Medication record prompt words: used to guide the large model to analyze whether the patient's current medication regimen conflicts or repeats with previous medication records.

[0124] The large model first extracts the main information of the medication case, including the patient's name, visit time, hospital, diagnosed disease and medication regimen. Sacubitril valsartan sodium tablets are confirmed as a drug for the treatment of hypertension, and the indications are checked. General usage and dosage information is provided (1 to 2 tablets each time, twice a day), but no personalized analysis is performed based on the patient's specific situation (such as weight, renal function, etc.). In actual applications, the patient information can be further combined to determine whether the usage and dosage are reasonable. Since the patient's detailed medical history and medication history are not provided in the case, the large model cannot directly determine whether there are contraindications or risks of drug interactions. In actual applications, this information should be supplemented for a more comprehensive analysis.

[0125] In actual application scenarios, in order to continuously improve the accuracy, sensitivity and specificity of the large model in detecting abnormal medication behavior, the answers output by the model can be repeatedly optimized and iterated. Specific methods include: adding more diverse medication cases for training to improve the generalization ability of the model. Figure 4 , Figure 4 A schematic diagram of a second prompt word application scenario provided in an embodiment of the present application.

[0126] Further feedback prompt: Thank you for your analysis. Please further optimize the model to detect the following potential irrational medication behaviors: 1. Overdose: Assuming that patient A was prescribed 16,800 mg of sacubitril valsartan sodium tablets in September, there may be a risk of overdose. 2. Drug interactions: Assuming that patient A was also prescribed other drugs in September, such as non-steroidal anti-inflammatory drugs (such as celecoxib, loxoprofen sodium), there may be a risk of drug interactions.

[0127] Based on your feedback, we have optimized the model. The following are the updated analysis and test results:### Detection of unreasonable medication behavior Based on the test results of the knowledge graph, the following potential unreasonable medication behaviors were found: 1. Overdose:

[0128] -Test results: Assuming that Patient A was prescribed 16,800 mg of sacubitril valsartan sodium tablets in September, and the standard dosage is a maximum daily dose of 400 mg, then the total maximum dose of Patient A in September (30 days) is 12,000 mg. 16,800 mg obviously exceeds the upper limit, so there is a risk of overdose.

[0129] -Conclusion: Unreasonable

[0130] 2. Drug interactions:

[0131] -Test results: Assume that patient A was also prescribed celecoxib or loxoprofen sodium in September. These drugs may increase renal impairment when used in combination with sacubitril / valsartan sodium tablets, so there is a risk of drug interaction.

[0132] -Conclusion: Unreasonable

[0133] Irrational medication behavior may cause damage to patients' health or even endanger their lives. In order to improve medical safety, we use the cue word engineering to train a large model to identify irrational medication behaviors from high-frequency medical behavior events. These behaviors include but are not limited to overdose, drug interactions, contraindications, and repeated medication.

[0134] By repeatedly optimizing and iterating the answers output by the large model, the accuracy, sensitivity and specificity of the model in detecting abnormal medication behavior are continuously improved. This optimization process includes training on more diverse medication cases and fine-tuning the model in combination with expert knowledge and clinical data.

[0135] In a possible implementation, the method further includes:

[0136] Step B1: Obtain user feedback information.

[0137] Collect feedback from end users (such as doctors, nurses, pharmacists, or patients). This feedback is crucial for evaluating and improving the detection model, helping to identify possible blind spots or misjudgments of the model and providing a basis for subsequent optimization.

[0138] The acquisition method may include at least one of the following:

[0139] Online Questionnaire: Design a short and targeted list of questions to invite users to share the problems and suggestions they encounter during use.

[0140] Built-in feedback mechanism: Embed convenient feedback portals in applications or system interfaces to facilitate users to submit opinions instantly.

[0141] Manual review team: Establish a dedicated manual review team to review the results of the model output, especially for complex or edge cases, and provide professional judgment and suggestions.

[0142] Step B2: Adjusting the model parameters of the detection model based on the user feedback information

[0143] By analyzing the feedback data provided by users, especially those that point out model prediction errors or inaccuracies, model parameters can be adjusted in a targeted manner to improve its performance. This includes but is not limited to adjusting hyperparameters such as weights, learning rates, regularization coefficients, or modifying the specific implementation details of certain algorithms.

[0144] Step B3: training the adjusted detection model based on the case information in the user feedback information.

[0145] The specific cases in user feedback not only provide valuable negative samples (i.e., situations that the model fails to correctly identify), but also reveal some special cases or rules that have not been considered. This information can enrich the training set and improve the generalization ability of the model.

[0146] Representative cases are extracted from user feedback and annotated to clearly indicate which are abnormal medication behaviors and which are not. The number and diversity of training samples can be increased through data augmentation techniques such as text generation, simulation of different variants, etc. These newly collected data are added to the existing training set to further fine-tune the adjusted detection model. After training is completed, an independent validation set is used to evaluate the new performance of the model to ensure that its improvements meet the expected goals and do not introduce new problems.

[0147] Continuously adjust and optimize model parameters based on actual feedback to improve long-term model performance. Verify model output results through offline supervision and manual review to ensure their authenticity and reliability.

[0148] Therefore, the embodiments of the present application have the following beneficial effects:

[0149] The entire process uses unsupervised learning methods and does not rely on manually labeled data. This reduces the manpower input in the data labeling process, saves time waiting for manual labeling, and improves the system's response speed and flexibility.

[0150] Compared with the Girvan-Newman algorithm and the label propagation algorithm, the Louvain algorithm can provide more reasonable and moderate community division results, and its accuracy is better than the K-means clustering algorithm and the hierarchical clustering algorithm. Therefore, it shows higher performance and accuracy when mining high-frequency medical behavior events.

[0151] This application combines the collaborative fusion model of the large language model and the chronic disease knowledge graph to improve the expression and reasoning ability of chronic disease drug treatment knowledge. This method achieves more accurate and comprehensive understanding and reasoning, and can effectively identify irrational drug use behaviors for chronic diseases, thereby assisting medical institutions in formulating more effective intervention measures, promoting rational drug use, and ensuring patient safety.

[0152] The model output is verified through offline supervision and manual review, and user feedback is collected. This feedback is not only used to directly improve the model parameters, but also input into the system as new training data to continuously optimize the model performance and ensure its long-term stability and adaptability.

[0153] The above are some specific implementations of the abnormal behavior detection method provided in the embodiment of the present application. Based on this, the present application also provides a corresponding abnormal behavior detection system. The system provided in the embodiment of the present application will be introduced from the perspective of functional modularization. Figure 5 A structural diagram of an abnormal behavior detection system provided in an embodiment of the present application.

[0154] The system comprises:

[0155] An acquisition unit 110, used for acquiring medical behavior information;

[0156] A construction unit 111 is used to use the medical behavior information as nodes, construct edges based on the relationship between each node, and obtain a medical behavior relationship network;

[0157] A division unit 112 is used to perform community division based on the correlation between nodes in the medical behavior relationship network using a community detection algorithm, wherein the divided communities represent groups of nodes having medical behavior characteristics whose similarity meets a preset similarity standard;

[0158] A determination unit 113, configured to determine, based on the divided communities, high-frequency medical behavior events corresponding to each community;

[0159] The detection unit 114 is used to input the high-frequency medical behavior events and prompt words into the trained detection model to obtain the abnormal medication behavior in the high-frequency medical behavior events. The detection model is constructed and trained based on a large language model and a disease knowledge graph, and the accuracy of identifying abnormal medication behavior reaches a preset threshold.

[0160] The embodiments of the present application also provide corresponding devices and computer storage media for implementing the test method solutions provided in the embodiments of the present application.

[0161] The device includes a memory and a processor, the memory is used to store instructions or codes, and the processor is used to execute the instructions or codes so that the device executes the test method described in any embodiment of the present application.

[0162] The computer storage medium stores codes, and when the codes are executed, a device executing the codes implements the test method described in any embodiment of the present application.

[0163] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other. For the system or device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0164] It should be understood that in the present application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0165] It should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0166] It should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0167] It should also be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0168] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0169] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting abnormal behavior, characterized in that: include: Obtaining information about medical practices; Taking the medical behavior information as nodes, building edges based on the relationships between the nodes, and obtaining a medical behavior relationship network; Using a community detection algorithm, community division is performed based on the correlation between nodes in the medical behavior relationship network, and the divided communities represent groups of nodes with medical behavior characteristics whose similarity meets a preset similarity standard; Based on the divided communities, determine the high-frequency medical behavior events corresponding to each community; The high-frequency medical behavior events and prompt words are input into the trained detection model to obtain abnormal medication behaviors in the high-frequency medical behavior events. The detection model is constructed and trained based on a large language model and a disease knowledge graph, and the accuracy of identifying abnormal medication behaviors reaches a preset threshold.

2. The method according to claim 1, characterized in that Based on the divided communities, high-frequency medical behavior events are obtained, including: For each community, count the frequency of medical behaviors and / or patient characteristics within the community; The frequency results obtained based on statistics are compared with the corresponding preset frequency thresholds, and medical behavior events in each community that are higher than the preset frequency thresholds are screened out as high-frequency medical behavior events.

3. The method according to claim 1, characterized in that The method for determining the trained detection model includes: Obtain information on disease theories; The disease knowledge graph is constructed based on disease theory information; Graph embedding algorithms are used to convert entities and relationships between entities in the disease knowledge graph into vector representations in high-dimensional space. Large language models are used to convert disease theory information into vector representations in high-dimensional space. Aligning vector representations in high-dimensional space from disease knowledge graphs and disease theory information, the alignment is used to achieve a unified representation so that the model can process both structured and unstructured information in the same framework; The large language model is trained based on prompt words to obtain a detection model that can achieve a preset threshold accuracy in identifying abnormal medication behavior in high-frequency medical behavior events. The prompt words are used to guide the large language model to generate detection content related to irrational medication behavior.

4. The method according to claim 3, characterized in that Before the large language model is trained based on the prompt words to obtain a detection model whose accuracy rate of identifying abnormal medication behavior in high-frequency medical behavior events reaches a preset threshold, the method further includes: In combination with the optimization goal of the disease knowledge graph and the task goal of the large language model, the large language model is trained with the unified goal obtained so that the model can simultaneously achieve the optimization goal of the disease knowledge graph and the task goal of the large language model.

5. The method according to claim 1, characterized in that The obtaining of medical behavior information includes: Obtaining preliminary data information; The prepared data information is cleaned and standardized to obtain medical behavior information, wherein the standardized processing is used to present data information of the same concept in the prepared data information in a consistent manner.

6. The method according to claim 3, characterized in that The method further comprises: Obtain user feedback information; Adjusting model parameters of the detection model based on the user feedback information; The adjusted detection model is trained based on the case information in the user feedback information.

7. An abnormal behavior detection system, characterized in that: include: An acquisition unit, used for acquiring medical behavior information; A construction unit, used to use the medical behavior information as nodes, construct edges based on the relationship between each node, and obtain a medical behavior relationship network; A division unit, used to perform community division based on the correlation between nodes in the medical behavior relationship network using a community detection algorithm, wherein the divided communities represent groups of nodes having medical behavior characteristics whose similarities meet a preset similarity standard; A determination unit, used for determining high-frequency medical behavior events corresponding to each community based on each community obtained by division; The detection unit is used to input the high-frequency medical behavior events and prompt words into a trained detection model to obtain abnormal medication behavior in the high-frequency medical behavior events. The detection model is constructed and trained based on a large language model and a disease knowledge graph, and the accuracy of identifying abnormal medication behavior reaches a preset threshold.

8. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the abnormal behavior detection method according to any one of claims 1 to 6 is implemented.

9. A computer program product, characterized in that When the computer program product runs on a computer, the computer executes the abnormal behavior detection method according to any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed on the terminal device, the terminal device executes the abnormal behavior detection method according to any one of claims 1 to 6.

Images