Prescription commenting system based on generative artificial intelligence and processing method thereof

Abstract

The invention provides a prescription commenting system based on generative artificial intelligence and a processing method thereof, which are used for specially designing a novel processing commenting system architecture on the basis of the generative artificial intelligence and forming a closed-loop, controllable, explainable and high-precision automatic commenting pipeline in combination with a multi-agent cooperative working mechanism. And further matching optimization is further carried out from each detail aspect, so that high-quality data support can be provided for prescription commenting work. The system adopts a three-layer system architecture design and comprises a knowledge layer, a data layer and an execution layer, the knowledge layer comprises a comment decision path library and a comment rule knowledge library; the data layer is used for acquiring multi-source and multi-mode patient data of a patient corresponding to a prescription; and the execution layer adopts a multi-agent cooperative work design and is used for performing task intention analysis processing on the prescription comment task, selecting a decision path based on a task intention analysis result and scheduling agents.

Description

Technical Field

[0001] This application relates to the field of medical technology, specifically to a prescription review system and processing method based on generative artificial intelligence. Background Technology

[0002] As a type of surgery with a low risk of infection, clean surgery (Class I incision surgery) requires strict adherence to the principle of "prevention first, rational use" in the use of perioperative antibiotics. However, the irrational use of antibiotics for prophylaxis in clean surgery is still quite common in clinical practice, such as using drugs without indication, inappropriate timing of administration, and excessively long courses of treatment. This not only increases the patient's medical expenses but may also lead to the development of drug-resistant bacteria and affect the patient's treatment outcome.

[0003] However, the current review of prescriptions for prophylactic antibiotic use in clean surgeries mainly relies on clinical pharmacists to complete manually. Due to the manual operation method, the scope and number of reviews are limited, and the uniformity and quality of the review results are difficult to guarantee.

[0004] Therefore, how to improve the intelligence level of prescription review for prophylactic antibiotics in clean surgery through artificial intelligence technology has become a key issue in the current management of rational drug use. Summary of the Invention

[0005] This application provides a prescription review system and processing method based on generative artificial intelligence. On the basis of generative artificial intelligence (AI), a novel processing review system architecture is designed and combined with a multi-agent collaborative working mechanism to form a closed-loop, controllable, interpretable and highly accurate automated review pipeline. Furthermore, it has made further supporting optimizations in various details. In this way, it can provide high-quality data support for prescription review work and lay a good foundation for pharmaceutical management and clinical improvement.

[0006] In the first aspect, this application provides a prescription review system based on generative artificial intelligence. The system adopts a three-layer system architecture design, which includes a knowledge layer, a data layer and an execution layer. The knowledge layer includes a review decision path library and a review rule knowledge base. The review decision path library is used to store structured review decision paths for at least two specialties in the form of system workflow. The review rule knowledge base is used to store pharmaceutical knowledge graphs in high-dimensional vector form to provide knowledge sources. The data layer is used to acquire multi-source, multimodal patient data corresponding to the prescription. The execution layer adopts a multi-agent collaborative design to perform task intent parsing for prescription review tasks. Based on the task intent parsing results, it selects decision paths and schedules agents to complete tasks including patient data extraction, structured processing of unstructured data, data structure transformation, knowledge retrieval, prescription suitability review, and writing of review analysis reports.

[0007] Secondly, this application provides a processing method for a prescription review system based on generative artificial intelligence. The method is applied to a prescription review system based on generative artificial intelligence. The system adopts a three-layer system architecture design, including a knowledge layer, a data layer, and an execution layer. The knowledge layer includes a review decision path library and a review rule knowledge base. The review decision path library stores at least two specialized structured review decision paths in the form of a system workflow. The review rule knowledge base stores a high-dimensional vector form of pharmaceutical knowledge graph to provide knowledge sources. The data layer is used to acquire multi-source, multi-modal patient data corresponding to the prescription. The execution layer adopts a multi-agent collaborative work design. The method includes: The execution layer retrieves prescription review tasks; The execution layer performs task intent parsing and processing on the prescription review task; The execution layer selects a decision path and schedules agents based on the task intent parsing results, which respectively complete the processing including patient data extraction, structured processing of unstructured data, data structure transformation, knowledge retrieval, prescription suitability review, and writing of review and analysis reports.

[0008] Thirdly, this application provides a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute the method provided in the second aspect of this application.

[0009] From the above, it can be concluded that this application has the following beneficial effects: To address the goals of prescription review, this application, based on generative artificial intelligence, specifically designs a novel processing review system architecture and combines it with a multi-agent collaborative working mechanism to form a closed-loop, controllable, interpretable, and highly accurate automated review pipeline. Furthermore, it has made further supporting optimizations in various details, thus providing high-quality data support for prescription review and laying a solid foundation for pharmaceutical management and clinical improvement. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of an architecture for the prescription review system based on generative artificial intelligence in this application. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules is not necessarily limited to those explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices. The naming or numbering of steps appearing in this application does not imply that the steps in the method flow must be performed in the chronological / logical order indicated by the naming or numbering. The execution order of named or numbered process steps can be changed according to the desired technical purpose, as long as the same or similar technical effect is achieved.

[0014] The module division described in this application is a logical division. In practical applications, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the coupling or direct coupling or communication connection between modules shown or discussed may be through some interfaces, and the indirect coupling or communication connection between modules may be electrical or other similar forms, none of which are limited in this application. Furthermore, the modules or sub-modules described as separate components may or may not be physically separated, may or may not be physical modules, or may be distributed in multiple circuit modules. Some or all of the modules may be selected to achieve the purpose of the solution in this application according to actual needs.

[0015] First, refer to Figure 1 The diagram shown is an architectural schematic of the prescription review system based on generative artificial intelligence of this application. The prescription review system based on generative artificial intelligence of this application is specially designed for the purpose of prescription review. Specifically, it adopts a three-layer system architecture design, namely, the prescription review system includes three major system layers: knowledge layer, data layer and intelligent agent execution layer.

[0016] It is important to note that the above three system layers are mainly divided and configured from the perspective of software / functional services. In practical applications, these three system layers may be deployed on the same hardware device or on different hardware devices. The hardware devices involved can be different types of devices such as server devices and physical host devices.

[0017] Furthermore, in practical applications, based on the above three system layers, the system can also involve the setting of User Equipment (UE). The UE can be a device that has built the client corresponding to the solution of this application, or it can be a device specifically matched with the application of this solution. In this way, system maintenance personnel, system administrators and system users (usually doctors) can conveniently view the system's working status and the overall / individual review task processing status. It can also involve the management of user permissions to reasonably verify and differentiate the system viewing and control capabilities of different personnel.

[0018] Specifically, the UE can be a terminal device such as a Personal Digital Assistant (PDA), tablet computer, smartphone, or smart bracelet.

[0019] Next, we will introduce the three system levels mentioned above in detail to understand the prescription review system based on generative artificial intelligence provided in this application.

[0020] (I) Knowledge Layer The knowledge layer includes a review decision path library and a review rule knowledge base. The review decision path library is used to store structured review decision paths for at least two specialties in the form of a system workflow. The review rule knowledge base is used to store pharmaceutical knowledge graphs in high-dimensional vector form to provide knowledge sources. Understandably, the knowledge layer is responsible for providing objective data support, and it is clear that this involves two databases: the review decision path database and the review rule knowledge base. In practical applications, these two databases, or data products, can also provide data support for other aspects of data usage.

[0021] The first type of database, namely the review decision path database, stores different pre-planned structured review decision paths or at least two specialized structured review decision paths. It is easy to understand that the intelligent prescription review implemented in this application is based on the selected and adapted structured review decision paths in the specific execution work. Therefore, the pre-configured structured review decision paths can be stored in the review decision path database or a review decision path database can be formed.

[0022] Specifically, the structured review decision-making path is constructed by collecting, extracting, and formalizing data. The original data that forms the basis for this process can be the expert experience and review logic of different clinical pharmacists.

[0023] Correspondingly, in terms of specific implementation, the structured review decision-making path can be constructed by using systematic knowledge engineering methods to collect, extract, and formalize the expert experience and review logic of multiple clinical pharmacists.

[0024] In addition, the structured review decision-making path can be defined by three key aspects: standard review steps, reference basis, and decision-making logic.

[0025] The second type of database, namely the review rule knowledge base, can be understood to store a pharmaceutical knowledge graph in the form of a high-dimensional vector, in order to provide objective data support in a different way than the review decision path database.

[0026] In terms of specific operations, the construction of the review rule knowledge base can first be based on knowledge graph methods to integrate authoritative literature, including drug instructions, clinical practice guidelines and expert consensus, to build a high-quality intermediate pharmaceutical knowledge graph. Then, the intermediate pharmaceutical knowledge graph can be transformed into a high-dimensional vector form of the review rule knowledge base using an embedding model.

[0027] Therefore, providing the system with authoritative knowledge sources that can be retrieved quickly and accurately is the key to achieving enhanced retrieval and ensuring the accuracy and traceability of the output results.

[0028] Furthermore, it can be noted that although the original intention of this application was to solve the problems of limited scope, limited quantity, unstable uniformity of review results, and unstable review quality associated with the existing prescription review methods for perioperative antibiotics in clean surgery (Class I incision surgery) which rely on manual operation, it can also be applied to other aspects of prescription review in hospital clinical work. Therefore, the review decision-making paths in the review decision-making path library can cover multiple specialties and specialties, and each specialty and specialty can be further subdivided into different types of review decision-making paths.

[0029] As an example, there are at least two specific structured review decision paths here, which may include review decision paths involved in the prescription review of antimicrobial drugs, antitumor drugs and cardiovascular drugs, among which antimicrobial drugs specifically include perioperative antimicrobial drugs for clean surgery.

[0030] (II) Data Layer The data layer is used to acquire multi-source, multimodal patient data corresponding to the prescription.

[0031] As can be seen, the data layer is responsible for providing data support for patient-side data, and in detail, it involves multimodal / multidimensional / multiscale patient data, as well as different data sources. This provides rich data support for the specific prescription review and processing of the subsequent execution layer, and provides the agent with comprehensive and well-organized "working materials".

[0032] Specifically, the multi-source, multimodal patient data involved here can be divided into two types: structured data and unstructured data. Among them are: 1) Structured data specifically includes examination and testing data, examination result data, vital signs and prescription data, etc. The prescription data here can be either the prescription that needs to be intelligently reviewed now, or the patient's previous prescriptions.

[0033] Understandably, structured data refers to standardized data that hospitals record uniformly in relevant online work systems or in daily clinical work, and it has a specific / fixed data format.

[0034] 2) Unstructured data specifically includes long text-type data such as medical records, surgical records, ward round records, and nursing records.

[0035] It is understandable that data such as medical records, surgical records, ward round records, and nursing records are relatively flexible. Some of these records are written down or handwritten due to urgent circumstances, thus exhibiting a flexible and varied data format. Furthermore, depending on the individual's handwriting style, speaking style, or work style, even more complex characteristics may emerge.

[0036] Specifically, unstructured data can be stored in specific data formats such as voice, video, and scanned documents.

[0037] Regarding specific data acquisition methods, the data layer can be used to aggregate multi-source, multi-modal patient data from multiple independent information systems within the hospital and store it in an intermediate database in the form of a relational database (MySQL). The multiple independent information systems specifically include the Hospital Information System (HIS), the Hospital Laboratory Information System (LIS), the Electronic Medical Record (EMR), the surgical anesthesia system, the nursing system, the rational drug use system, and the pharmacy management system.

[0038] This may involve the application of standardized data interfaces such as the Model Context Protocol (MCP).

[0039] It is understandable that obtaining the multi-source, multi-modal patient data required for this application through the aforementioned specific hospital's independent information system has the advantages of facilitating the integration of this application's solution into practical application scenarios and making data extraction convenient.

[0040] At the same time, it should be noted that obtaining multimodal patient data through hospital information systems or other independent information systems does not mean that the multimodal patient data required by this application is directly stored in the existing system. For some patient data, the system can perform some patient data storage work as needed for the use of this application solution. In other words, the data processing work of the system side is included in the scope of obtaining and processing some patient data specifically introduced in this application.

[0041] Furthermore, the raw data obtained by the data layer can be temporarily stored in a unified intermediate database to prepare for subsequent processing by the intelligent agent.

[0042] (III) Execution Layer The execution layer adopts a multi-agent collaborative design to perform task intent parsing for prescription review tasks. Based on the task intent parsing results, it selects decision paths and schedules agents to complete tasks including patient data extraction, structured processing of unstructured data, data structure transformation, knowledge retrieval, prescription suitability review, and writing of review analysis reports.

[0043] Understandably, this application involves the application of multiple intelligent agents in its specific implementation. Specifically, the complex cognitive task of prescription review is decomposed into a series of refined sub-tasks, and under the scheduling and organization of the central workflow engine, multiple related large language model intelligent agents cooperate in sequence to complete the required prescription review processing.

[0044] In this process, based on the data support of the knowledge layer and the data layer, the prescription review task is processed by analyzing the task intent. Based on the knowledge layer, the decision path and intelligent agent adapted to the task intent analysis results are obtained to perform detailed processing of prescription review work such as patient data extraction, unstructured data structuring, data structure transformation, knowledge retrieval, prescription suitability review, and review analysis report writing, so as to complete an intelligent prescription review operation.

[0045] Each agent is responsible for different specialized tasks and performs data processing for the corresponding task processing nodes through a specific Large Language Model (LLM) that specializes in a particular function.

[0046] Intelligent agents typically form a seamless automated pipeline in a sequential manner. Of course, the possibility of intelligent agent deployment schemes involving nesting, parallelism, etc., cannot be ruled out.

[0047] As an example, the large language model used in this application may specifically involve large language models such as qwen3.

[0048] Understandably, in practical applications, the large language models involved in different agents can be either the same type of large language model or different types of large language models. Similar to the case of agent granularity, the same or different types of large language models can also be used for different processing nodes involved in the same agent.

[0049] In practical applications, the deployment of large language models can be conveniently carried out using specific large language model deployment tools such as DIFY (Do It For You). The DIFY tool, which is exemplified here, integrates functions such as workflow orchestration, model management, and agent development. It supports seamless integration with a large number of types of large language models, and provides private deployment and enterprise-level management functions, which facilitates the application of multi-agent / large language models in this application solution.

[0050] As an example of prescription review for perioperative antibiotics in clean surgery, the local treatment nodes may involve determining whether there is a preoperative infection. Specifically, this may involve reviewing information such as daily medical history, medical records, laboratory test results, admission records, preoperative medical history descriptions, and preoperative imaging examination descriptions. This information can be provided by the data layer.

[0051] Furthermore, it is understandable that large language models, while providing comments / conclusions, can also provide corresponding reasoning processes / evidence, making each step of the conclusion deduction traceable. This breaks the "black box" limitations of traditional AI systems, constructs a collaborative model of "system as the main body and pharmacist as the auxiliary body," enhances clinical trust, and thus has better interpretability and visualization, providing better data support.

[0052] Specifically, as a preferred implementation scheme, the different agents configured in the agent execution layer here, according to their functional types, may include a central decision-making agent, a data processing agent, a prescription suitability review agent, and a report analysis agent. Different types of large language models are selected to perform specific tasks based on the task characteristics of each agent. Specifically, these include: 1. The central decision-making agent is used to perform task intent parsing and processing for prescription review tasks, call decision paths and knowledge bases, coordinate other agents to complete the review tasks, and select a large language model with strong reasoning ability; Specifically, the large language model used here can be deepseek-r1:671b.

[0053] Specifically, the central decision-making agent can receive review task instructions initiated by pharmacists, such as the task of "reviewing the rationality of emergency antibacterial drug prescriptions".

[0054] Of course, the specific way to initiate a task is quite flexible. It can be initiated manually by relevant personnel, received from other devices, or initiated automatically by the system according to relevant task initiation strategies. It can be configured according to actual needs.

[0055] After receiving the prescription review task, the central decision-making agent can attempt to understand the task intent and match and activate the most suitable review decision path from the review decision path library. This involves corresponding path matching strategies, such as combining relevant prompts with Euclidean distance for path matching. Simultaneously, the central decision-making agent can also generate a specific data requirement list based on the selected review decision path to plan the scope and focus of subsequent data extraction.

[0056] 2. The data processing agent is used for data extraction, data structuring, and data format conversion. Specifically, based on the data processing requirements proposed by the central decision-making agent, it generates structured query language query statements, extracts multi-source and multimodal patient data from the intermediate database of the data layer, calls a large language model with optimized Chinese understanding to read long text-type unstructured data, extracts key information from it and generates structured data, and finally, through code execution, transforms the complete structured data from the structured query language form into a JavaScript object representation data format that is more user-friendly for large language model recognition, and finally outputs structured patient medication data in JavaScript object representation format. Specifically, the large language model here can be qwen3, Structured Query Language (SQL), a relational database, JavaScript Object Notation (JSON), or JSON object notation.

[0057] Specifically, the data processing agent undertakes the tasks of data preprocessing and deep understanding. Based on the data processing requirements given by the central decision-making agent, it accurately queries and reads the required multi-source and multimodal patient information from the intermediate database of the data layer through interfaces / protocols such as MCP.

[0058] Next, the data processing agent transforms the complex structured query language into a clear, large language model-friendly structured data format, namely JavaScript object representation. On the other hand, for unstructured text, it uses deep natural language understanding technology to extract crucial clinical indicators for medication decisions (such as "presence of infection focus", "description of liver and kidney function status", "history of allergies", etc.) from long texts such as medical records, surgical records, ward round records, and nursing records, and transforms them into structured Boolean values ​​or enumeration values. It can also extract structured text from scanned documents, videos, and audio data through text recognition, image recognition, and speech recognition technologies.

[0059] Ultimately, the data processing agent outputs a comprehensive and structured clinical profile of the patient, which integrates all the raw data and derived indicators—that is, structured patient medication data.

[0060] 3. The prescription suitability review agent is used to review prescription information one by one based on structured patient medication data, call the review rule knowledge base, start retrieval enhancement generation, and review prescription suitability one by one. The prescription suitability review specifically includes indication review, contraindication review, usage and dosage review, and interaction review. In detailed operations, the structured patient medication data includes information such as diagnosis, symptoms, and test results. At this point, the prescription appropriateness review agent can call the review rule knowledge base, automatically start the search enhancement generation process, and independently judge the indications and necessity of each prescription drug used by the current patient based on the prescription information, and give a preliminary conclusion of "indication" or "no indication". In addition, when giving whether there is an indication or not, the prescription appropriateness review agent can also give specific prescription items / contents that are indicated.

[0061] Next, for prescriptions deemed "indicative," the system can use keywords such as the prescription drug name and key information (e.g., age, weight, liver and kidney function indicators) as search criteria to retrieve the most relevant knowledge (e.g., medication guidelines and instruction leaflets) from the review rules knowledge base in real time. Then, based on this retrieved knowledge / authoritative evidence, a detailed prescription suitability review process can be performed, including review of indications, contraindications, dosage and administration, and interactions.

[0062] 4. The report analysis agent is used to summarize and analyze the review results of the prescription suitability review agent, call the corresponding large language model to statistically analyze the review results according to the established template, generate special review and analysis reports according to the template, call the code to execute and draw appropriate statistical analysis charts, and generate an interactive business intelligence interface.

[0063] Among them, the Business Intelligence (BI) interface provides review and analysis reports that can be tailored to individual levels or to the overall / batch level.

[0064] The review report clearly adopts a strict structured format, explicitly listing the review conclusions, specific problem categories, detailed causal analysis, and authoritative sources cited, ensuring the professionalism and verifiability of the results.

[0065] For review and analysis reports targeting the overall / batch level, in practice, the structured review results of all individual prescriptions involved can be aggregated and analyzed.

[0066] In this way, the report analysis agent can perform statistical analysis, identify commonalities and trends in irrational drug use, and automatically generate a comprehensive thematic analysis report, providing specific data support and decision-making suggestions for pharmaceutical management and clinical improvement.

[0067] In addition, the report analysis agent can also generate a user-friendly and interactive business intelligence interface by drawing corresponding statistical analysis charts based on the review analysis report, so as to achieve an interactive display effect of review results on the user side.

[0068] As can be seen from the above preferred implementation scheme, the multiple agents, driven by the workflow engine, form a seamless automated production line. From the moment the task is triggered, the data is processed and value-added step by step on the production line, ultimately producing multi-level, high-value intelligent review results from micro (single prescription) to macro (special report).

[0069] As an example of a final prescription review conclusion, the following could be included: "The conclusion is that this patient does not belong to a high-risk group, for the following reasons:" Age over 60 years: The patient is 52 years old, which is under 60 years old and does not meet this high-risk factor.

[0070] Nutritional status: According to the medical record description, "Nutrition: Good", BMI is calculated as follows: height 165cm (1.65m), weight 75kg; Body Mass Index (BMI) = 75 / (1.65×1.65)≈27.5; Although the BMI is greater than 24, combined with the clinical description of "good nutrition" and no obvious signs of malnutrition, it is not considered a high-risk factor for malnutrition.

[0071] Whether the operation time exceeded 3 hours: The operation started at 10:43 on December 23, 2024 and ended at 12:02, with an operation time of approximately 1 hour and 19 minutes, which did not exceed 3 hours and did not meet the high-risk factor.

[0072] Did the blood loss during the operation exceed 1500 ml? The description of the operation did not mention a large amount of blood loss, nor were there any records of blood transfusion. The blood loss should be much less than 1500 ml, which does not meet the criteria for this high-risk factor.

[0073] Is there any history of use of glucocorticoids or immunosuppressants in the medical history? If there is no mention of use of glucocorticoids or immunosuppressants in the past medical history, it does not meet this high-risk factor.

[0074] Whether or not an organ transplant patient: There is no information related to organ transplantation in the medical history, so this does not meet the high-risk factor.

[0075] Whether you have been diagnosed with diabetes or have a history of diabetes: If your medical history does not mention diabetes or related conditions, you do not meet this high-risk factor. In addition, it is understandable that after obtaining the template production special review and analysis report, in addition to optimizing the display / presentation through the business intelligence interface to help pharmacists review it as soon as possible, other forms of output processing can also be involved.

[0076] For example, it can perform local storage, off-site storage, output completion feedback prompts, result push notifications, or further data processing and analysis. Obviously, the specific output processing can be flexibly configured according to actual needs.

[0077] In the output stage, the relevant hardware structure can be either the system's own or external. Taking result display as an example, it can use the system's own display screen (including touch screen) or external display devices or other devices with display screens to achieve the display goal of the relevant required content (such as prescription review results, prescription review process, etc.).

[0078] Similarly, the data acquisition aspects mentioned above can involve both real-time collection and extraction of existing data, and the relevant hardware structure required for data acquisition operations can be either internal to the system or external to the system.

[0079] At the same time, it is understood that, as previously demonstrated, for the different intelligent agents involved in this application, each processing node of the different intelligent agents is handled by the corresponding large language model.

[0080] The large language model involved can be specifically trained on multi-source, multimodal patient data that has not been labeled with prescription review results.

[0081] In other words, in the conventional approach to using deep learning models, when introducing a relevant large language model into the application scenario of this application, relevant multi-source and multimodal patient data will be configured and labeled (the labeling results are the true values ​​predicted by the model), and then the training samples will be used to train the model.

[0082] In contrast, this application argues that training based on training samples limits the large language model to the training samples, leading to bias and a loss of objectivity. Therefore, this application adopts a training-free mechanism for the large language model. This approach avoids the training costs associated with model training in practical applications. By using multiple large language models in a training-free manner, the multi-agent collaborative work involved in this application can be achieved, resulting in high-performance prescription review.

[0083] In conclusion, regarding the above solutions, this application, based on generative artificial intelligence, has designed a novel processing and review system architecture specifically for prescription review, and combined it with a multi-agent collaborative working mechanism to form a closed-loop, controllable, interpretable, and highly accurate automated review pipeline. Furthermore, it has made further supporting optimizations in various details, thus providing high-quality data support for prescription review and laying a solid foundation for pharmaceutical management and clinical improvement.

[0084] The above is an introduction to the prescription review system based on generative artificial intelligence provided in this application. Correspondingly, this application also provides a processing method for the prescription review system based on generative artificial intelligence from the perspective of system workflow. This method is obviously applicable to the prescription review system based on generative artificial intelligence.

[0085] In summary, the prescription review system based on generative artificial intelligence adopts a three-layer system architecture design, comprising a knowledge layer, a data layer, and an execution layer. The knowledge layer includes a review decision path library and a review rule knowledge base. The review decision path library stores structured review decision paths for at least two specific areas in the form of a system workflow, while the review rule knowledge base stores a high-dimensional vector-based pharmaceutical knowledge graph to provide knowledge sources. The data layer acquires multi-source, multi-modal patient data corresponding to the prescription. The execution layer employs a multi-agent collaborative design.

[0086] Based on the above brief system architecture, the processing method of the prescription review system based on generative artificial intelligence provided in this application may specifically include the following steps: The execution layer retrieves prescription review tasks; The execution layer performs task intent parsing and processing on the prescription review task; The execution layer selects a decision path and schedules agents based on the task intent parsing results, which respectively complete the processing including patient data extraction, structured processing of unstructured data, data structure transformation, knowledge retrieval, prescription suitability review, and writing of review and analysis reports.

[0087] As an exemplary implementation, the structured review decision-making path is constructed by collecting, extracting and formalizing the expert experience and review logic of multiple clinical pharmacists through systematic knowledge engineering methods. The structured review decision-making path is specifically defined by three key aspects: standard review steps, reference basis and decision logic.

[0088] As another exemplary embodiment, the review rule knowledge base is constructed by first integrating authoritative literature, including drug instructions, clinical practice guidelines and expert consensus, based on the knowledge graph method to build an intermediate pharmaceutical knowledge graph. Then, the intermediate pharmaceutical knowledge graph is transformed into a high-dimensional vector form of the review rule knowledge base by using an embedding model to convert the entities and relationships in the graph.

[0089] As another exemplary embodiment, multi-source multimodal patient data is divided into two types: structured data and unstructured data; Structured data specifically includes examination and testing data, examination result data, vital signs, and prescription data; Unstructured data specifically includes long text-type data such as medical records, surgical records, ward round records, and nursing records.

[0090] As another exemplary embodiment, the data layer is used to aggregate multi-source, multimodal patient data from multiple independent information systems in the hospital and store it in an intermediate database in the form of a relational database; The multiple independent information systems specifically include the hospital information system, the hospital laboratory information system, electronic medical records, surgical anesthesia system, nursing system, rational drug use system, and pharmacy management system.

[0091] As another exemplary embodiment, the different intelligent agents configured in the intelligent agent execution layer are classified according to their functional types, including central decision-making intelligent agents, data processing intelligent agents, prescription suitability review intelligent agents, and report analysis intelligent agents. Different types of large language models are selected to perform specific tasks based on the task characteristics of different intelligent agents. The central decision-making agent is used to perform task intent parsing and processing for prescription review tasks, call decision paths and knowledge bases, coordinate other agents to complete the review tasks, and select a large language model with strong reasoning ability. The data processing agent is used for data extraction, data structuring, and data format conversion. Specifically, based on the data processing requirements proposed by the central decision-making agent, it generates structured query language query statements, reads long text-type unstructured data from the intermediate database of the data layer (multi-source, multimodal patient data), calls a large language model with optimized Chinese understanding capabilities to read the data, extracts key information and generates structured data, and finally, through code execution, transforms the complete structured data from the structured query language form into a JavaScript object representation data format that is more user-friendly for large language model recognition, and finally outputs structured patient medication data in JavaScript object representation format. The prescription suitability review agent is used to review prescription information one by one based on structured patient medication data, call the review rule knowledge base, start retrieval enhancement generation, and review prescription suitability one by one. The prescription suitability review specifically includes indication review, contraindication review, usage and dosage review, and interaction review. The report analysis agent is used to summarize and analyze the review results of the prescription suitability review agent, call the corresponding large language model to statistically analyze the review results according to the established template, generate special review and analysis reports according to the template, call code to execute and draw appropriate statistical analysis charts, and generate an interactive business intelligence interface.

[0092] As another exemplary embodiment, each processing node of different intelligent agents performs business processing by the corresponding large language model; The large language model is specifically a model trained on multi-source, multimodal patient data without labeled prescription review results.

[0093] As another exemplary embodiment, at least two specialized structured review decision paths are included, specifically review decision paths involved in the prescription review of antimicrobial drugs, antitumor drugs, and cardiovascular drugs, wherein antimicrobial drugs specifically include perioperative antimicrobial drugs for clean surgery.

[0094] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the generative artificial intelligence-based prescription review system described above can be found in, for example... Figure 1 The description of the prescription review system based on generative artificial intelligence in the corresponding embodiment will not be repeated here.

[0095] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0096] To this end, this application provides a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute the steps of the processing method of the prescription review system based on generative artificial intelligence in the above embodiments of this application. For specific operations, please refer to the description of the processing method of the prescription review system based on generative artificial intelligence in the above embodiments, which will not be repeated here.

[0097] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0098] Since the instructions stored in the computer-readable storage medium can execute the steps of the processing method of the prescription review system based on generative artificial intelligence in the above embodiments of this application, the beneficial effects that the processing method of the prescription review system based on generative artificial intelligence in the above embodiments of this application can achieve can be realized, as detailed in the foregoing description, and will not be repeated here.

[0099] The foregoing has provided a detailed description of the generative artificial intelligence-based prescription review system, the processing method of the generative artificial intelligence-based prescription review system, and the computer-readable storage medium provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A prescription review system based on generative artificial intelligence, characterized in that, The system adopts a three-layer system architecture design, which includes a knowledge layer, a data layer, and an execution layer. The knowledge layer includes a review decision path library and a review rule knowledge base. The review decision path library is used to store at least two specialized structured review decision paths in the form of a system workflow. The review rule knowledge base is used to store a pharmaceutical knowledge graph in high-dimensional vector form to provide knowledge sources. The data layer is used to acquire multi-source, multimodal patient data corresponding to the prescription. The execution layer adopts a multi-agent collaborative design to perform task intent parsing for prescription review tasks, select decision paths and schedule agents based on the task intent parsing results, and complete processes including patient data extraction, structured processing of unstructured data, data structure transformation, knowledge retrieval, prescription suitability review and review analysis report writing.

2. The system according to claim 1, characterized in that, In the construction process of the structured review decision-making path, a systematic knowledge engineering approach is used to collect, extract, and formalize the expert experience and review logic of multiple clinical pharmacists to construct the decision-making path. The structured review decision-making path is specifically defined by three key aspects: standard review steps, reference basis, and decision logic.

3. The system according to claim 1, characterized in that, In the construction process of the review rule knowledge base, the authoritative literature, including drug instructions, clinical practice guidelines and expert consensus, is first integrated based on the knowledge graph method to construct an intermediate pharmaceutical knowledge graph. Then, the intermediate pharmaceutical knowledge graph is transformed into the review rule knowledge base in high-dimensional vector form using an embedding model.

4. The system according to claim 1, characterized in that, The multi-source, multimodal patient data is divided into two types: structured data and unstructured data. The structured data specifically includes examination and testing data, examination result data, vital signs, and prescription data; The unstructured data specifically includes long text-type data such as medical records, surgical records, ward round records, and nursing records.

5. The system according to claim 1, characterized in that, The data layer is used to aggregate the multi-source, multimodal patient data from multiple independent information systems in the hospital and store it in an intermediate database in the form of a relational database. The aforementioned multiple independent information systems specifically include a hospital information system, a hospital laboratory information system, electronic medical records, a surgical anesthesia system, a nursing system, a rational drug use system, and a pharmacy management system.

6. The system according to claim 1, characterized in that, The different agents configured in the agent execution layer are classified according to their function type, including central decision-making agent, data processing agent, prescription suitability review agent and report analysis agent. Different types of large language models are selected to perform specific tasks according to the task characteristics of the different agents. The central decision-making agent is used to perform task intent parsing processing on the prescription review task, call decision paths and knowledge bases, coordinate other agents to complete the review task, and select the large language model with strong reasoning ability. The data processing agent is used for data extraction, data structuring, and data format conversion. Specifically, based on the data processing requirements proposed by the central decision-making agent, it generates structured query language query statements, reads long text-type unstructured data from the intermediate database of the data layer (multi-source, multimodal patient data), calls the large language model optimized for Chinese understanding, extracts key information from the data, generates structured data, and finally, through code execution, converts the complete structured data from the structured query language form into a JavaScript object representation data format that is more user-friendly for the large language model, and finally outputs structured patient medication data in JavaScript object representation format. The prescription suitability review agent is used to call the review rule knowledge base based on the structured patient medication data, initiate retrieval enhancement generation, and review the prescription suitability of each prescription information. The prescription suitability review specifically includes indication review, contraindication review, usage and dosage review, and interaction review. The report analysis agent is used to summarize and analyze the review results of the prescription suitability review agent, call the corresponding large language model to statistically analyze the review results according to the predetermined template, generate a special review and analysis report according to the template, call the code to execute and draw appropriate statistical analysis charts, and generate an interactive business intelligence interface.

7. The system according to claim 6, characterized in that, Each processing node of the different intelligent agents is processed by the corresponding large language model; The large language model is specifically a model trained on multi-source, multimodal patient data that has not been labeled with prescription review results.

8. The system according to claim 1, characterized in that, The structured review decision paths for at least two specific areas include review decision paths for prescription reviews of antimicrobial drugs, antitumor drugs, and cardiovascular drugs, wherein the antimicrobial drugs specifically include perioperative antimicrobial drugs for clean surgery.

9. A processing method for a prescription review system based on generative artificial intelligence, characterized in that, The method is applied to a prescription review system based on generative artificial intelligence. The system adopts a three-layer system architecture design, which includes a knowledge layer, a data layer, and an execution layer. The knowledge layer includes a review decision path library and a review rule knowledge base. The review decision path library is used to store at least two specialized structured review decision paths in the form of system workflow. The review rule knowledge base is used to store a pharmaceutical knowledge graph in high-dimensional vector form to provide knowledge sources. The data layer is used to acquire multi-source, multimodal patient data corresponding to the prescription. The execution layer employs a multi-agent collaborative work design; the method includes: The execution layer obtains prescription review tasks; The execution layer performs task intent parsing and processing on the prescription review task; The execution layer selects a decision path and schedules intelligent agents based on the task intent parsing results, respectively completing processes including patient data extraction, structured processing of unstructured data, data structure transformation, knowledge retrieval, prescription suitability review, and writing of review and analysis reports.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to execute the method of claim 9.

Images