Artificial intelligence-based medical monitoring methods and systems for clinical research

Abstract

The present application provides a clinical research medical monitoring method and system based on artificial intelligence, which relates to the field of medical monitoring technology. The method includes: constructing a medical monitoring ontology knowledge base by acquiring relevant medical knowledge texts, and mapping the sample data of the medical monitoring materials to the ontology framework based on the knowledge base to form a medical monitoring knowledge graph. Then, knowledge extraction is performed on the medical monitoring application materials to obtain review elements, and these are mapped to the knowledge graph to obtain relevant ontology concept nodes to form a review element graph. Then, the review element graph is inferred and judged based on the medical monitoring knowledge graph and preset rules to form a preliminary review result. Finally, the preliminary review results are quantitatively scored, and a review decision is made based on the score and the preliminary results to obtain the final review decision result. The above method improves the efficiency of medical monitoring and the quality of review.

Description

Technical Field

[0001] The present application relates to the field of medical monitoring technology, and in particular to a clinical research medical monitoring method and system based on artificial intelligence. Background Art

[0002] Clinical trials are an essential and critical step in the development of new drugs and medical devices. Their purpose is to comprehensively evaluate the safety, efficacy, and quality controllability of candidate drugs or medical devices. To ensure the authenticity, completeness, standardization, and reliability of clinical trial data, rigorous medical monitoring is essential. Medical monitoring aims to review raw data and documents from clinical trial sites, identify any doubts and risks, and promptly address medical concerns to ensure the high-quality completion of clinical trials.

[0003] Traditional medical monitoring relies primarily on manual labor, with professional medical monitors meticulously reviewing and comparing trial protocols, case report forms, medical records, and other documents to identify potential data anomalies, protocol violations, and ethical violations. However, with the rapid growth in the number of new drug developments and the continued expansion of clinical trials, medical monitoring faces numerous challenges, including large amounts of data, a high level of specialized expertise, and a heavy workload.

[0004] Under current technology, medical monitors must invest significant time and effort to systematically review vast amounts of trial site documents and data, relying on their expertise and experience to identify hidden questions and risks. This approach is inefficient, susceptible to human error, and difficult to ensure consistent and stable monitoring quality. This severely impacts both the efficiency and quality of medical monitoring. Summary of the Invention

[0005] This application provides an artificial intelligence-based clinical research medical monitoring method and system to improve medical monitoring efficiency and review quality.

[0006] In a first aspect, the present application provides a method for medical monitoring of clinical research based on artificial intelligence, the method comprising:

[0007] Obtain medical knowledge texts related to medical monitoring and build a medical monitoring ontology knowledge base based on the medical knowledge texts;

[0008] Obtain sample data of medical monitoring materials, and map the sample data of medical monitoring materials to the ontology framework based on the ontology knowledge base to form a semantic network containing concept nodes, attribute edges, and relationship edges to obtain the medical monitoring knowledge graph;

[0009] Obtain medical monitoring application materials, perform knowledge extraction on the medical monitoring application materials, and obtain review elements;

[0010] Map the review elements to the medical monitoring knowledge graph to obtain the review element related ontology concept nodes, and form a review element graph based on the related ontology concept nodes;

[0011] Based on the medical monitoring knowledge graph and preset medical monitoring rules, the review factor graph is inferred and judged to form a preliminary review result;

[0012] Quantitatively score the preliminary review results to obtain a quantitative review score;

[0013] An examination decision is made based on the examination quantitative score and the preliminary examination results to obtain the examination decision result.

[0014] In the above technical solution, by acquiring medical knowledge text related to medical monitoring and constructing a medical monitoring ontology knowledge base based on this medical knowledge text, a solid knowledge foundation is laid for the subsequent construction of a medical monitoring knowledge graph. On this basis, by acquiring sample data of medical monitoring materials and mapping this data into an ontology framework based on the ontology knowledge base, a semantic network consisting of concept nodes, attribute edges, and relationship edges is formed, resulting in a rich and well-structured medical monitoring knowledge graph. This method of transforming unstructured medical knowledge into a structured and semantically represented knowledge greatly improves the efficiency and standardization of processing massive amounts of medical monitoring data.

[0015] Furthermore, after obtaining the medical monitoring application materials, this application uses knowledge extraction technology to intelligently extract the review elements therein. These review elements are then mapped to the medical monitoring knowledge graph, resulting in ontological concept nodes related to the review elements. This process then forms a refined review element graph. This process fully utilizes the semantic association information in the knowledge graph to quickly identify the core review points of the application materials, significantly reducing the burden of manual review.

[0016] On this basis, this application makes inferences and judgments on the review factor graph based on the medical monitoring knowledge graph and preset medical monitoring rules, automatically identifies possible problems in the application materials, and forms preliminary review results. This intelligent reasoning method based on domain knowledge and rules has significantly improved the sensitivity and accuracy of medical monitoring and effectively prevented data quality risks in clinical trials. At the same time, this application also introduces a quantitative scoring mechanism to objectively evaluate the preliminary review results and obtain a quantitative review score. This scoring method establishes a unified monitoring quality assessment standard and improves the consistency and comparability of monitoring results.

[0017] In a second aspect of the present application, a clinical research medical monitoring system based on artificial intelligence is provided, the system comprising:

[0018] An ontology knowledge base construction module is used to obtain medical knowledge texts related to medical monitoring and construct a medical monitoring ontology knowledge base based on the medical knowledge texts;

[0019] The medical monitoring knowledge graph construction module is used to obtain sample data of medical monitoring materials and map the sample data of medical monitoring materials to the ontology framework based on the ontology knowledge base to form a semantic network containing concept nodes, attribute edges, and relationship edges to obtain the medical monitoring knowledge graph;

[0020] The review factor acquisition module is used to obtain medical monitoring application materials and perform knowledge extraction on the medical monitoring application materials to obtain review factors;

[0021] The review factor graph acquisition module is used to map the review factors into the medical monitoring knowledge graph, obtain the review factor-related ontology concept nodes, and form a review factor graph based on the related ontology concept nodes;

[0022] The preliminary review result acquisition module is used to infer and judge the review factor graph based on the medical monitoring knowledge graph and preset medical monitoring rules to form a preliminary review result;

[0023] A quantitative scoring module is used to quantitatively score the preliminary review results and obtain a quantitative review score;

[0024] The review decision module is used to make review decisions based on the review quantitative scores and preliminary review results to obtain review decision results.

[0025] In a third aspect of the present application, a computer storage medium is provided. The computer storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the above method steps.

[0026] In the fourth aspect of the present application, an electronic device is provided, including a processor, a memory, a user interface and a network interface, the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device performs the above method.

[0027] In summary, one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0028] 1. This application establishes a structured and semantic medical monitoring knowledge representation system by acquiring medical knowledge texts related to medical monitoring and constructing a medical monitoring ontology knowledge base based on the medical knowledge texts. The ontology knowledge base organizes the core concepts, relationships and attributes in the field of medical monitoring in a formalized manner, laying a solid knowledge foundation for subsequent intelligent review. By mapping the sample data of medical monitoring materials into the ontology framework, a semantic network containing concept nodes, attribute edges, and relationship edges is formed, and a medical monitoring knowledge graph with rich content and complete structure is further obtained. The knowledge graph not only covers various regulatory requirements for medical monitoring, but also reflects the empirical knowledge contained in the case data, so that the review process can make full use of relevant information of previous cases and improve the comprehensiveness and accuracy of the review.

[0029] 2. This application uses artificial intelligence technologies such as natural language processing and ontology construction to automatically and efficiently mine the intrinsic knowledge of medical knowledge texts, and systematically define concepts and their relationships, thereby constructing a medical monitoring ontology knowledge base, forming a comprehensive, detailed, scientific and standardized regulatory knowledge organization system.

[0030] 3. This application is based on the ontology knowledge base. Through a series of methods such as data preprocessing, concept node mapping, attribute edge mapping, and relationship edge extraction, it can efficiently, accurately, and comprehensively map unstructured medical monitoring data samples into a standardized ontology framework, thereby forming a medical monitoring knowledge graph covering concept connotations, attribute characteristics, and relationship semantics. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A flowchart of an artificial intelligence-based medical monitoring method for clinical research provided in an embodiment of the present application;

[0032] Figure 2 An architectural diagram of an artificial intelligence-based clinical research medical monitoring system provided in an embodiment of the present application;

[0033] Figure 3 This is a schematic diagram of the structure of an electronic device provided in this application. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments.

[0035] In the description of the embodiments of this application, words such as "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "for example" or "for instance" in the embodiments of this application should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "for example" or "for instance" is intended to present the relevant concepts in a concrete manner.

[0036] In the description of the embodiments of the present application, the term "multiple" means two or more. For example, multiple systems refer to two or more systems, and multiple screen terminals refer to two or more screen terminals. In addition, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

[0037] In order to facilitate understanding of the method and system provided by the embodiments of the present application, before introducing the embodiments of the present application, the background of the embodiments of the present application is first introduced.

[0038] Medical monitoring, an essential and crucial component of scientific research, shoulders the crucial mission of ensuring research compliance, protecting the rights and interests of research subjects, and promoting research integrity. For a long time, medical monitoring has relied primarily on manual processes, requiring reviewers to devote significant time and effort to meticulously reviewing application materials, thoroughly analyzing the compliance of research protocols, comprehensively assessing potential ethical risks, and making prudent review decisions based on these efforts. While this traditional manual review model has played an important role, it has also gradually exposed several issues that require urgent improvement. First, manual review is inefficient and unable to meet the growing demand for review. With the booming development of scientific research, the number of medical monitoring applications has continued to rise, and manual processing is no longer able to adapt to this explosive growth. Reviewers are required to devote significant time and effort to reviewing each application item by item. This high workload not only affects review efficiency but also can lead to a decline in review quality. Second, manual review is inherently subject to subjectivity and inconsistent standards. Reviewers' knowledge, experience, and values ​​can influence review decisions, resulting in a degree of subjectivity in the results. Different reviewers may differ in their judgments on the same case, making it difficult to achieve fully standardized review standards. This subjectivity and inconsistency in standards may affect the fairness and standardization of the review, and cause doubts and dissatisfaction among applicants.

[0039] After the background introduction of the above content, those skilled in the art can understand the problems existing in the prior art. The technical solutions in the embodiments of the present application will be clearly and completely described below in combination 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.

[0040] On the basis of the above background technology, further, please refer to Figure 1 , Figure 1 This is a flow chart of a method for medical monitoring of clinical research based on artificial intelligence provided in an embodiment of the present application. The system can be implemented by a computer program or can be run as an independent tool application. Specifically, in the embodiment of the present application, the method can be applied on a server, but can also be applied to electronic devices such as servers. A method for medical monitoring of clinical research based on artificial intelligence includes the following steps:

[0041] S101, obtaining medical knowledge texts related to medical monitoring, and constructing a medical monitoring ontology knowledge base based on the medical knowledge texts;

[0042] Specifically, by systematically collecting and organizing authoritative textual materials in the field of medical surveillance, including various laws, regulations, departmental rules, industry norms, and guiding principles, a comprehensive and rich corpus of medical knowledge texts will be formed. This medical knowledge text covers the requirements and standards for all aspects of medical surveillance and serves as an important basis for conducting reviews. Based on this acquired medical knowledge text, ontology construction technology is used to conduct in-depth analysis and knowledge extraction of this text, identifying the core concepts, relationships, and attributes in the field of medical surveillance, and formally organizing them into a structured ontology knowledge base.

[0043] Based on the above embodiment, as an optional embodiment, a medical monitoring ontology knowledge base is constructed based on medical knowledge texts, including:

[0044] S201, using natural language processing technology to extract core concepts related to medical monitoring from medical knowledge texts;

[0045] Specifically, medical knowledge text must first be preprocessed, including sentence segmentation, word segmentation, part-of-speech tagging, and named entity recognition. This converts unstructured text into a structured sequence of semantic units, providing a foundation for concept extraction. This preprocessing step is typically implemented using mature natural language processing toolkits or platforms, such as Jieba, THULAC, and Stanford CoreNLP.

[0046] After preprocessing, concept extraction primarily utilizes rule-based, statistical, and deep learning-based methods. Rule-based methods manually define a series of templates and rules based on the grammatical and semantic features of concepts within sentences to match and extract concepts. Common templates include "concept + attributive clause," "concept + de + noun," and "adjective + concept." This method offers high accuracy, but also low recall and a strong reliance on domain expertise. Statistical methods automatically discover concepts within text by measuring statistical metrics such as the frequency of co-occurrence of concepts within documents and lexical composition. Common statistical metrics include TF-IDF, Mutual Information, and the Chi-squared statistic.

[0047] S202, based on the ontology core concepts and the medical knowledge text, determining the semantic relationship between the ontology core concepts, and performing ontology relationship definition based on the semantic relationship between the ontology core concepts to obtain the relationship definition between each ontology core concept;

[0048] Specifically, there are two main strategies for defining concept relationships: one is to use ontology description language to directly formally define the relationship between concepts; the other is to automatically mine the relationship between concepts from the text, then abstract it into ontology relationship and formally define it.

[0049] The first strategy primarily involves collaboration between domain experts and knowledge engineers. By referencing the context of the concepts within medical knowledge texts, they intuitively determine the relationships between concepts. These relationships are then formalized into machine-readable definitions using ontology description languages ​​such as RDF and OWL. For example, a "protection" relationship could be defined between "informed consent" and "subject rights," as in "informed consent protects subject rights." This relationship could then be expressed using the RDF triple (subject, predicate, object) as (informed consent, protection, subject rights). While this manual definition approach is time-consuming and labor-intensive, it provides a high degree of accuracy and consistency in the relationship definitions, better aligning with expert understanding.

[0050] The second strategy first utilizes natural language processing techniques, particularly semantic dependency analysis and open knowledge extraction, to automatically extract semantic dependencies between concepts from medical text. For example, the semantic relation "protect" can be identified from the sentence "When conducting human trials, the privacy and personal information of subjects must be protected." This allows the extraction of the two triples (human trials, protection, subject privacy) and (human trials, protection, personal information) to be extracted. After obtaining preliminary relationship extraction results, experts screen, verify, and abstract them, removing redundant and erroneous relationships. Relationships with similar semantics are grouped into the same type and formalized using an ontology language. For example, relationships such as "protect," "maintain," and "guarantee" can be abstracted into a unified "protect" relationship.

[0051] S203, defining the attributes of each ontology core concept based on the relationship definition between each ontology core concept, to obtain the attribute definition of each ontology core concept;

[0052] Specifically, the definition of concept attributes mainly adopts two approaches: one is to combine the relationship between concepts in the ontology, extract and generalize the intrinsic attributes of the concept from the relationship objects; the other is to refer to the contextual information of the concept in the medical knowledge text, and manually define the external attributes that reflect the characteristics of the concept.

[0053] The first approach is to fully leverage the rich set of already defined conceptual relationships and abstractly extract conceptual attributes from the semantics of these relationships. This approach offers the advantage of highly consistent attribute definitions with relationship definitions, preventing conflicts and errors. Furthermore, by revealing attributes through relationships, the inherent connection between relationships and attributes is demonstrated, enhancing the semantic consistency of the ontology. For example, from the relationship "Ethics Committee consists of medical professionals," the attribute "Members = Medical Professionals" of "Ethics Committee" can be extracted. Similarly, from the relationship "Informed consent must include a statement of research purpose," the attribute "Required elements = Statement of research purpose" of "Informed consent" can be extracted. These attributes extracted from conceptual relationships form an integral part of the concept's connotation.

[0054] The second approach involves domain experts directly referencing medical knowledge text and manually defining extrinsic attributes that reflect the characteristics of concepts. While this background knowledge-based attribute definition approach is more subjective, it can capture conceptual features not covered by relationships and serves as a necessary supplement to relationship-based attribute extraction. By analyzing the semantic context of concepts within medical knowledge text, experts can often identify key attributes that are easy to understand and grasp. For example, by reading relevant clauses, it's easy to identify attributes such as "voluntary participation," "right to be informed," "right to privacy," and "right to withdraw." Similarly, attributes such as "name," "purpose," "start and end dates," "participants," and "funding source" are associated with "research project." These manually defined attributes, based on expert experience and domain common sense, highlight the salient characteristics of concepts and facilitate rapid concept recognition.

[0055] In practice, these two attribute definition methods are usually used interchangeably and in conjunction with each other. Taking the concept of "informed consent" as an example, we can first extract attributes such as "protected object = subject's rights" and "necessary procedures = subject's signature" from relationships such as "informed consent protects the rights and interests of subjects" and "informed consent requires the signature of subjects". Then, experts refer to the original text of the regulations to define external attributes such as "notification content", "notification method", and "signature format". These two types of attributes complement each other and together constitute a comprehensive and detailed "informed consent" attribute spectrum, making the originally abstract concept more three-dimensional, vivid, and perceptible.

[0056] S204: Construct a medical monitoring ontology knowledge base based on the relationship definitions of each ontology core concept and the attribute definitions of each ontology core concept.

[0057] Specifically, the construction of the medical monitoring ontology knowledge base mainly includes two aspects: one is to formalize the definitions of concepts, relationships, and attributes into an ontology knowledge base based on the ontology description language and using ontology development tools; the other is to manually review and evaluate the constructed ontology knowledge base and make necessary revisions, supplements, and extensions to the ontology.

[0058] For the first aspect, ontology description languages ​​such as OWL (Web Ontology Language) are mainly used, and with the help of ontology development tools such as Protégé and TopBraid Composer, the definitions of concepts, relationships, and attributes are converted into formal ontology knowledge base representations. Specifically, the ontology header information is first defined, declaring the ontology's namespace, version, and metadata such as the introduction of external ontologies; then the class (Class) hierarchy is defined, and based on the hierarchical relationship of concepts, the core concepts are organized into a tree-like or graph-like type hierarchy; then the relationship (Property) between types is defined, and based on the conceptual relationship definition, logical properties such as SubProperty, InverseOf, and TransitiveProperty are used to characterize the semantic characteristics of the relationship; finally, the instance (Individual) of the class is defined and the attribute values ​​of the instance are described. Based on the conceptual attribute definition, ObjectProperty, DatatypeProperty, etc. are used to associate type instances with attribute value domains and data types. For example, define a "Subject" class and use SubClassOf to define it as a subclass of the "Research Participant" class. Then, define ObjectProperties such as "Right to Know," "Right to Privacy," and "Benefit Compensation" to associate "Subject" with other related classes. Finally, define instances such as "Zhang San" and "Li Si," describing their properties such as right to know, right to privacy, and benefit compensation. Through this layer-by-layer definition, concepts, relationships, and properties are organically organized into a complete knowledge system, forming an ontology knowledge base with a clear structure and rich semantics.

[0059] S102, obtaining sample data of medical monitoring materials, and mapping the sample data of medical monitoring materials to the ontology framework based on the ontology knowledge base, forming a semantic network including concept nodes, attribute edges, and relationship edges, and obtaining a medical monitoring knowledge graph;

[0060] Specifically, we first collect sample data of medical monitoring documents from historical review cases and typical samples. This includes various types of unstructured text data, such as applications, informed consent forms, study plans, and review opinions. This sample data contains a wealth of review practice experience and case knowledge, and serves as a crucial data source for constructing the knowledge graph. After obtaining the sample data, preprocessing is required, including format conversion, text cleaning, word segmentation, and part-of-speech tagging, to transform the unstructured text into a form that computers can understand and process. Next, based on the established medical monitoring ontology knowledge base, the preprocessed sample data is mapped into the ontology framework. This process is primarily accomplished through natural language processing and knowledge extraction technologies. First, named entity recognition (NER) methods are used to identify entities corresponding to ontology concepts from the sample data, such as "subject," "research purpose," and "informed consent." Then, relationship extraction techniques, such as rule-based or machine learning methods, are used to extract semantic relationships between entities from the sample data, such as "research project - involving - subjects" and "investigator - obtaining - informed consent." Through entity recognition and relationship extraction, the key information in the sample data is converted into a form compatible with the ontology knowledge base.

[0061] After entity and relationship extraction, the extracted entities are mapped to corresponding concept nodes in the ontology knowledge base, and the extracted relationships are mapped to attribute edges or relationship edges between concept nodes, thus forming a semantic network closely connected to the ontology framework. For example, the "subject" entity identified in the sample data is mapped to the "subject" concept node in the ontology knowledge base, and the "research project-involved-subject" relationship extracted from the sample data is mapped to a relationship edge between the "research project" and "subject" concept nodes. Through this mapping process, scattered sample data is integrated into a unified semantic framework, forming a highly connected and semantically rich medical monitoring knowledge graph.

[0062] The construction process of a medical monitoring knowledge graph typically uses a semi-automated approach, combining manual intervention and computer automation. Experts first define mapping rules and strategies to guide the mapping of sample data to the ontology framework. Then, natural language processing and machine learning techniques, such as conditional random fields (CRFs), support vector machines (SVMs), and deep learning, are used to automatically complete tasks such as entity recognition, relationship extraction, and mapping. In addition to automated processing, manual verification and correction by experts are also required to ensure the quality and accuracy of the knowledge graph. If necessary, experts can also expand and optimize the knowledge graph, adding new concept nodes, attribute edges, or relationship edges to make it more complete and comprehensive.

[0063] Based on the above embodiment, as an optional embodiment, the sample data of medical monitoring materials are mapped to the ontology framework based on the ontology knowledge base to form a semantic network containing concept nodes, attribute edges, and relationship edges, thereby obtaining a medical monitoring knowledge graph, including:

[0064] S301, pre-processing the sample data of the medical monitoring information to obtain medical monitoring data;

[0065] Specifically, the preprocessing of medical monitoring data samples mainly includes the following steps: data cleaning, data conversion, data integration, data specification, etc.

[0066] S302, performing data mapping on the medical monitoring data based on the ontology knowledge base, mapping entities in the medical monitoring data to corresponding ontology concept nodes in the ontology knowledge base, and mapping attributes in the medical monitoring data to attribute edges of the ontology concept nodes;

[0067] Specifically, the process of data mapping of medical monitoring data based on the ontology knowledge base mainly includes two aspects: one is to map the entities in the review data to the corresponding concept nodes in the ontology knowledge base; the other is to map the attributes in the medical monitoring data to the attribute edges of the concept nodes.

[0068] S303, performing relationship extraction on the medical monitoring data based on the ontology knowledge base, extracting the semantic relationships between named entities in the medical monitoring data, and mapping the semantic relationships into relationship edges between ontology concept nodes;

[0069] Specifically, relationship extraction of medical monitoring data based on the ontology knowledge base mainly includes two steps: one is to extract the semantic relationships between named entities from the data text; the other is to map the extracted semantic relationships to the relationship edges of concept nodes in the ontology knowledge base.

[0070] The first step, semantic relationship extraction, primarily utilizes natural language processing techniques to identify named entities from the unstructured text of the review data and determine the types of semantic relationships between them. Specifically, named entity recognition models, such as conditional random fields and recurrent neural networks, are first employed to identify multiple named entities, such as persons, institutions, documents, and drugs, from the text. Relation extraction models, such as convolutional neural networks and attention mechanisms, are then employed to determine whether semantic associations exist between identified entity pairs and predict the specific types of associations, such as "author-work," "sponsor-project," and "researcher-informed consent." Currently, most mainstream relationship extraction models employ a distantly supervised learning paradigm, automatically generating large-scale training data from ontology knowledge bases and learning semantic mappings between entity pairs and relationship types at the sentence level. Some work also incorporates pre-trained language models, such as BERT, to enhance the model's understanding of semantic relationships through pre-training. Furthermore, methods such as rule-based template matching and bootstrapping are also used to mine structured relationship patterns between entities from syntactic parse trees. Through these technologies, structured triple knowledge such as "drug A - indication - disease B" and "case C - inclusion criteria - condition D" are identified from the review data text, and the correlation between review elements is preliminarily mined.

[0071] The second step, relationship-edge mapping, mainly involves semantically matching the semantic relationships extracted from the data with the relationship edges defined in the ontology knowledge base to form a unified graphical representation. Specifically, based on methods such as dictionary matching and vector space semantic similarity calculation, the semantic equivalence of the extracted semantic relationships and the ontology relationship edges is determined. In other words, whether the head and tail entities in the triple correspond to the head and tail concepts connected by the relationship edge, and whether the relationship type corresponds to the definition of the relationship edge; for semantically equivalent relationships, they are converted into corresponding relationship edge instances; for relationships for which no direct corresponding relationship can be found, the hierarchical relationship of the existing concept nodes in the ontology is used to infer a relationship path that spans at most one intermediate concept, and the relationship edges in the path are combined to generate a new composite relationship edge. For example, the semantic relationship "Patient A - Enrollment - Trial Protocol B" is identified and mapped to the "Subject - Participation - Clinical Trial" edge defined in the ontology, directly generating an instance of that edge. Another example is the semantic relationship "Researcher C - Record - Adverse Reaction D" is identified, but no corresponding edge can be found in the ontology. However, based on the two existing edges "Researcher - Conducted - Research" and "Research - Involved - Adverse Event", a composite relationship "Researcher - Record - Adverse Reaction" can be inferred, and a new composite edge can be generated accordingly. Through these mapping and reasoning, the semantic relationships extracted from the data are uniformly mapped to the relational framework of the ontology knowledge base, inheriting the structured semantic information of the edge, such as type definition, value range, and constraint rules.

[0072] S304, construct a medical monitoring knowledge graph based on ontology concept nodes, attribute edges and relationship edges.

[0073] Specifically, constructing a medical monitoring knowledge graph based on ontology concept nodes, attribute edges, and relationship edges mainly includes three steps: defining the graph schema, storing graph data, and optimizing graph queries.

[0074] The first step is to define the knowledge graph's schema. This primarily organizes the graph's knowledge organization framework, including metadata such as concept types, relationship types, attribute types, and their semantic constraints, to inherit and extend the semantic system of the ontology knowledge base. Schema definitions typically use ontology description languages ​​such as RDF and OWL. By defining axioms such as domain, range, and cardinality, they standardize the semantic types of concept nodes and relationship edges. For example, core concept classes such as "researcher" and "subject" are defined, mapping essential ontology attributes such as "name" and "institution" to graph data attributes. Core relationship classes such as "participation" and "signature" are defined, clarifying the semantic constraints of their head and tail concept classes. Schema definitions also consider physical storage optimization, employing attribute embedding for frequently co-occurring attributes and implementing sharded storage for concepts with a large number of instances. A well-designed graph schema should strike a balance between semantic completeness and storage efficiency. Secondly, the specific data of the graph, such as entities, attributes, and relationships, is stored. This is primarily achieved through physical storage using mainstream graph databases such as Neo4j and JanusGraph. This uses a node-edge-attribute schema to store semantic network data mapped to the ontology. For storage, database tables are first created based on the schema definition, representing concept nodes, relationship edges, and attributes. ETL tools are then used to convert previously heterogeneous review data into standard graph formats such as RDF and CSV. This data is then imported into the graph database using transactional mechanisms via graph query languages ​​like Cypher and Gremlin. This graphical physical storage allows for seamless connection between nodes through relationship edges, eliminating the iterative foreign key queries required by traditional relational databases. For example, to identify subjects who participated in trial X and experienced adverse reaction Y, the graph database simply uses pattern matching to quickly locate the "trial" and "adverse event" nodes, then traces the relationship edges of their first-degree neighbors to efficiently select the target "subject" node. In fact, the graph database's graph traversal algorithm can transform complex multi-table joins into single-table pattern matching, significantly improving the efficiency of linked data retrieval.

[0075] S103, obtaining medical monitoring application materials, and performing knowledge extraction on the medical monitoring application materials to obtain review elements;

[0076] Specifically, the first step is to obtain the complete medical monitoring application materials, including the application form, study plan, informed consent form, and investigator qualification certificate. These materials, typically submitted in paper or electronic format, cover all aspects of the research project and serve as the foundation for conducting medical monitoring. After obtaining the application materials, they undergo necessary preprocessing, such as format conversion, text extraction, and image recognition, to transform the unstructured application materials into a computer-readable and processable form. Next, knowledge extraction is performed on the preprocessed application materials, aiming to identify review elements closely related to medical monitoring. This process primarily utilizes natural language processing and text mining techniques. First, named entity recognition (NER) is used to identify key entities from the application materials, such as "study purpose," "subject selection criteria," "informed consent process," and "data confidentiality measures." Because review elements often involve specialized terminology and concepts within specific fields, a pre-built domain dictionary and rule base are required to improve the accuracy and recall of entity recognition.

[0077] Building on entity recognition, we further utilize techniques such as keyword extraction and phrase mining to extract keywords and phrases related to review elements from the application materials, such as "risk," "subject rights," and "privacy protection." These keywords and phrases often reflect the application's key points and risks in medical monitoring, providing important clues for subsequent review and judgment.

[0078] Based on the above embodiment, as an optional embodiment, knowledge extraction is performed on the medical monitoring application materials to obtain review elements, including:

[0079] S401: Structuring the medical monitoring application materials to obtain structured application materials, and performing text parsing on the structured application materials to obtain key information elements;

[0080] Specifically, knowledge extraction of medical monitoring application materials to obtain review elements mainly includes two steps: structuring the application materials and text parsing the structured application materials to extract key information elements.

[0081] The first step, structuring, involves converting the original application documents in formats like PDF and Word into structured examination information, providing standardized input for subsequent text analysis. Specifically, document parsing tools such as Apache PDFBox and Apache POI are used to convert the original PDF and Word documents into structured HTML or XML formats, preserving important document structural information such as text content, heading hierarchies, tables, and lists. The converted structured documents are then parsed to extract structural elements such as title trees, paragraph text, table cells, and list items. A unique location expression is generated for each structural element using mechanisms such as XPath or CSS selectors. Finally, the extracted structured information is stored in a document database such as MongoDB, where a structured index based on "document ID-location path-text content" is established, providing fast random access for subsequent analysis and extraction. The key to this step is to remove noise from the original documents while preserving as much semi-structured information as possible that is crucial for semantic understanding, such as the field names on the examination form and the chapter headings in the research proposal. This allows subsequent analysis to focus on areas with higher information density. Furthermore, it is important to establish a unified structural index for multiple application documents for the same review project (such as research plans, informed consent forms, and medical monitoring application forms) to support the extraction of associations from multiple sources of information. Practice has shown that using document structured preprocessing can increase the speed of subsequent text parsing by 1-2 orders of magnitude.

[0082] The second step, extracting key information elements, primarily utilizes natural language processing and machine learning technologies to semantically parse and extract information from structured review materials, identifying key information elements necessary for medical monitoring decisions. Specifically, the structured text undergoes NLP preprocessing, including sentence segmentation, word segmentation, part-of-speech tagging, named entity recognition, and syntactic dependency analysis, to build foundational features at the lexical and syntactic levels for semantic analysis. A text parsing engine is then developed for review elements, primarily employing rule-based, template-based, and knowledge-based approaches to extract review elements at the lexical, syntactic, and semantic levels. At the lexical level, methods such as dictionary matching and regular expressions are primarily used to quickly identify keywords and phrases related to review elements, such as "informed consent," "research purpose," and "subject rights," thereby preliminarily locating sentences containing these elements. At the syntactic level, techniques such as syntactic analysis and dependency analysis are primarily employed to extract descriptive modifiers of review elements, such as "written consent" and "clear research purpose," from relevant sentences, based on syntactic rules such as subject-verb structure and attributive structure. At the semantic level, predefined review element ontologies and extraction rules are used to map the extracted vocabulary and syntactic units into structured representations of review elements. For example, the sentence "This project ensures that subjects are fully aware of the research objectives, processes, and risks by signing a written informed consent form" is identified, and based on the extraction template of "research project - through - written informed consent - ensure - subjects - are aware - research objectives / processes / risks", a structured representation of "informed consent method: written; informed content: research objectives, research processes, research risks" is formed. The above three levels complement each other. The lexical level realizes the coarse-grained identification and positioning of key information, the syntactic level realizes the structured organization of modifying and limiting phrases, and the semantic level realizes the ontology mapping of structured semantic components to review elements, together forming a progressive extraction from unstructured to structured.

[0083] S402: Name entities based on key information elements and structured application materials to obtain key entities corresponding to the review elements;

[0084] Specifically, entity naming is performed based on key information elements and structured application materials to identify key entities corresponding to review elements. This approach primarily employs a dictionary- and rule-based naming approach, supplemented by a machine learning model trained on a small amount of manually annotated data. First, a named entity dictionary covering synonyms, variants, and aliases for core entity types in the medical monitoring knowledge graph, such as "research project," "research institution," "research subject," "informed consent," "research risk," and "conflict of interest," is constructed. Key entities belonging to these types are then quickly identified through string matching. For example, keywords such as "clinical trial," "randomized controlled trial," and "research project" are mapped to the "research project" entity; "patient," "subject," and "volunteer" are mapped to the "research subject" entity; and "informed consent form" and "informed consent document" are mapped to the "informed consent" entity. While dictionary matching is simple and efficient, it is susceptible to cross-contextual ambiguity and requires supplementary contextual rules to filter results. Therefore, based on dictionary matching, a series of entity naming rules are manually defined to target specific combination patterns of review elements, leveraging contextual features to improve naming accuracy. For example, it stipulates that only "subjects" in contexts such as "The subjects of this study are healthy volunteers over the age of 18" can be named "research subjects," excluding other situations such as "subject rights" and "subject informed consent." Another example is that sentences in which the entities "researcher" and "conflict of interest" co-occur are labeled as "conflict of interest" elements, while general declarative sentences in which "researcher" appears alone are excluded. Building on the dictionary and rules, machine learning sequence labeling models such as conditional random fields (CRFs) and recurrent neural networks (RNNs) are further introduced. By manually annotating a small amount of training data, the distribution patterns of review elements in context are learned, further improving the generalization and robustness of name recognition.

[0085] S403, performing named entity recognition on the structured application materials to extract key entities corresponding to the review elements;

[0086] Specifically, named entity recognition (NER) is performed on structured application materials to extract key entities corresponding to review elements. This approach primarily employs a multi-strategy approach based on dictionary matching, rule filtering, and sequence annotation. First, domain dictionaries covering terminology, aliases, and abbreviations for common entity types in the medical monitoring field, such as "study design," "statistical analysis," and "credentials," are constructed. These dictionaries are then used to quickly match structured text to initially identify candidate entities. Next, a series of manually defined filtering rules are used to leverage contextual features such as part of speech and syntax to eliminate ambiguous names and irrelevant entities, improving recognition accuracy. For example, the phrase "subjects" in the phrase "subjects of this study signed the informed consent form" specifies that only "subjects" refers to the research subject, excluding distracting terms such as "subjects' rights and interests." Another example is that sentences co-occurring with "researcher" and "conflict of interest" entities are annotated as conflict of interest elements, while general statements containing "researcher" alone are excluded. Building on this NER and rule filtering, machine learning sequence annotation models such as CRF and BiLSTM are further introduced. By manually annotating a small amount of training data, these models learn the distribution patterns of review entities in context, overcoming the limitations of insufficient dictionary coverage and weak rule generalization. Specifically, syntactic features such as part of speech and dependency analysis are extracted and automatically combined into high-dimensional feature vectors that depict the entity context using feature templates. Combined with manually annotated corpus, discriminant models such as CRF are used to learn feature weights, resulting in a more domain-adaptive entity recognizer. Furthermore, active learning and human-computer interaction are employed to optimize the manual annotation process, obtaining the most representative training samples with minimal annotation cost, further improving model performance.

[0087] S404, performing type definition on the key entities to obtain defined key entities, and performing relationship extraction based on the defined key entities to determine the semantic associations between the defined key entities;

[0088] Specifically, key entities are defined by type, and relationships are extracted based on these defined key entities to determine the semantic connections between them. This approach primarily employs a knowledge base and rule-based approach, supplemented by a machine learning model trained on a small amount of annotated data. First, an entity type system is constructed, referencing the knowledge graph in the medical monitoring field. For example, entities are divided into broad categories such as research projects, research institutions, research subjects, and risks, and further subdivided into subcategories such as cross-sectional studies and drug clinical trials. Next, a series of type mapping rules are manually defined, leveraging the hierarchical relationships and attribute characteristics of similar entities in the knowledge base. For example, the research design entity in "This study adopts a randomized controlled trial design" is mapped to the "randomized controlled trial" type, and the institution entity in "The research team is from Peking University First Hospital" is defined as the "research hospital" type. Furthermore, methods such as dictionary matching are used to quickly align the type information of existing entities in the knowledge base. In addition, for the tail and new entities that are insufficiently covered by the knowledge base, active learning is used to guide experts to label the types of a small number of seed entities. Then, through remote supervision technologies such as bootstrpping, type reference patterns in large-scale unstructured texts are automatically mined, the annotation scale is expanded, and machine learning classification models such as CRF and CNN are trained to automatically classify new entities.

[0089] S405, defining attributes based on the defined key entities to obtain attributes of each defined key entity;

[0090] Specifically, attribute definition is performed based on the defined key entities to obtain the attributes of each defined key entity. This approach primarily utilizes a knowledge base and context-based approach, supplemented by active learning and manual annotation. First, referring to the medical monitoring knowledge ontology, common attributes of key entities are enumerated and an attribute dictionary is constructed. For example, the attribute dictionary for the "researcher" entity includes name, affiliation, and title. Next, the dictionary is automatically matched to identify attribute mentions of each entity. Based on contextual features such as syntax and semantics, irrelevant attributes are filtered out to eliminate cross-context attribute ambiguity. For example, the "Professor" attribute value for "Wang Ming" in "Professor Wang Ming of Peking University School of Pharmacy" is specified as his affiliation, the "School of Pharmacy," eliminating redundant attributes related to other part-time affiliations. Furthermore, attribute values ​​are automatically extracted from the corresponding paragraphs in the application materials. For example, from the sentence "This trial has passed the review of the ethics committee, and the approval number is 2022-010," "2022-010" is extracted as the "Approval Number" attribute value for the "Medical Monitoring" entity. For long-tail attribute values ​​in unstructured expressions, we use methods such as regular expressions and dependency analysis to extract the grammatical form of attribute values. For new and sparse attribute types, active learning methods guide experts in annotation. Through distributed representation, we learn the semantic similarity between attribute mentions and attribute types, and automatically mine training corpus for annotating new attributes.

[0091] S406: Organize the defined key entities, the attributes of each defined key entity, and the semantic associations between the defined key entities into a structured knowledge representation to obtain review elements.

[0092] Specifically, during the knowledge extraction process for medical monitoring application documents, organizing the defined key entities, their attributes, and the semantic associations between them into a structured knowledge representation to obtain review elements is a crucial step. This is because review elements are essentially structured, standardized knowledge representations extracted from the medical monitoring application documents, encompassing the core information elements of the application documents. Only by systematically organizing and representing key entities, attributes, and semantic associations can the content of the application documents be fully and accurately described, providing a clear and complete information foundation for subsequent review decisions. Specifically, in the previous steps, the medical monitoring application documents were structured to obtain key information elements. Entities were then named based on these key information elements, resulting in the key entities corresponding to the review elements. Next, types were defined for the key entities, determining the type of each key entity. Based on these type definitions, relationships were further extracted from the defined key entities, clarifying the semantic associations—that is, the relationships—between them. Furthermore, attributes were defined for the defined key entities, determining the attributes of each key entity. On this basis, the defined key entities, the attributes of each defined key entity and the semantic associations between the defined key entities are organized into a structured knowledge representation, which can be done specifically in the following ways: first, each defined key entity is used as the basic unit of knowledge representation, and each key entity is represented by a standardized term or identifier; second, for each key entity, its attributes are associated with it in the form of attribute name and attribute value, where the attribute name represents the type or name of the attribute, and the attribute value gives the specific value of the entity on the attribute; third, for the semantic associations between key entities, the two key entities are connected by a relationship name to represent the relationship between them, and the relationship consists of three elements: the starting entity, the relationship name and the ending entity; finally, the key entities, attributes and semantic associations are organically organized in the form of triples (entity-attribute-attribute value, or entity-relationship-entity) to form a complete structured knowledge representation, namely the review element.

[0093] S104, mapping the review elements to the medical monitoring knowledge graph, obtaining the review element related ontology concept nodes, and forming a review element graph based on the related ontology concept nodes;

[0094] Specifically, the extracted review elements are first mapped one by one with the ontology concepts in the medical monitoring knowledge graph. This process is mainly achieved by using semantic matching and similarity calculation technology. For each review element, the knowledge graph is searched for semantically similar or related ontology concept nodes, and the similarity or correlation between them is calculated. The similarity calculation can consider multiple factors such as the literal similarity, semantic similarity, and hierarchical relationship between the review element and the ontology concept to comprehensively evaluate their matching degree. Usually, ontology concept nodes whose similarity or correlation exceeds a certain threshold are considered to be related to the review element and are selected as the mapping result.

[0095] After mapping review elements to ontology concept nodes, a subgraph closely related to the review elements is formed within the medical monitoring knowledge graph based on these relevant ontology concept nodes. This is known as the review element graph. Specifically, with the mapped relevant ontology concept nodes as the center, their adjacent nodes, attribute edges, and relationship edges are extracted from the knowledge graph to form a connected and compact semantic subnetwork. This subnetwork encompasses concepts, attributes, and relationships directly or indirectly related to the review elements, reflecting the semantic context and associative environment of the review elements within the knowledge graph.

[0096] Based on the above embodiment, as an optional embodiment, the review elements are mapped to the medical monitoring knowledge graph to obtain the review element related ontology concept nodes, and a review element graph is formed based on the related ontology concept nodes, including:

[0097] S501, mapping the review elements to the ontology concept nodes in the medical monitoring knowledge graph to obtain the review element-related ontology concept nodes;

[0098] Specifically, the review elements are mapped onto ontology concept nodes in the medical monitoring knowledge graph to obtain the relevant ontology concept nodes. This approach primarily employs semantic similarity and graph random walks. First, the ontology concept nodes in the medical monitoring knowledge graph are traversed, and the semantic similarity between the review elements and the concept nodes is calculated using pre-trained word embeddings such as BERT and Word2Vec. Considering the diversity of conceptual representations in the medical ethics field, character-level features such as glyph similarity are incorporated into the semantic similarity calculation to mitigate the out-of-place (OOV) problem. Next, high-level semantic representations of ontology concepts are learned using methods such as topic models and knowledge distillation. A semantic association matrix between concepts is constructed, enabling the mapping of review elements to capture the underlying thematic meaning of the terms. For example, for the review element "The project team obtained a list of research subjects from the village committee," even though its explicit vocabulary differs significantly from the concept of "informed consent," analyzing the corresponding themes reveals a deeper connection between the two concepts in terms of "privacy protection." After calculating the semantic similarity matrix, strategies such as breakpoint filtering and top-K are used to select the top K most relevant ontology concept nodes to the review element. Furthermore, given that relying solely on semantic similarity struggles to reveal indirect connections, we further utilize graph random walks to explore the multi-hop propagation paths of related concepts, capturing the higher-order connections between review elements within the ontology network. For example, multi-hop association analysis reveals an implicit violation relationship between the element "conflict of interest between the research team and participating enterprises" and the concept "researcher's duty of integrity." Finally, we aggregate the ontology concept nodes directly or indirectly related to the review elements to obtain a set of ontology concept nodes related to the review elements.

[0099] S502, expanding the relevant nodes in the medical monitoring knowledge graph according to the review factor-related ontology concept nodes to obtain the review factor expansion nodes;

[0100] Specifically, the method expands relevant ontology concept nodes within the medical monitoring knowledge graph to obtain expanded nodes based on review factor concepts. This method primarily employs graph neural networks and cross-graph mapping. First, with the review factor-related ontology concept node as the central node in the medical monitoring knowledge graph, a graph neural network method is used to learn a low-dimensional representation of the node within the graph, ensuring that semantically related nodes in the embedding space are closer in the vector space. Considering the heterogeneity of entity-relationship-entity triples, heterogeneous graph neural networks such as R-GCN are used to distinguish the semantic contributions of different relationship types to node embeddings. Furthermore, a graph attention network such as GAT is used to model the mutual influence between concept nodes in the graph. Attention weights are used to adjust aggregate node neighbor information and highlight the semantic contributions of strongly related nodes. After learning the node embeddings, a nearest neighbor search method based on Euclidean distance, cosine similarity, and other methods is used to obtain the top N most relevant expanded nodes to the central node. Furthermore, to expand knowledge coverage, a cross-knowledge graph node expansion method is introduced. By manually mapping seed nodes and learning joint representations of cross-graph nodes, we construct a semantic mapping between the medical monitoring knowledge graph and knowledge graphs in fields like medicine and law. This allows us to extend the scope of related concept nodes to multiple external knowledge graphs. For example, by linking to the legal domain graph, we can expand the concept of "informed consent" to include legal concepts such as "contracting capacity" and "contract validity." Ultimately, by integrating and deduplicating the extended nodes from the internal and external graphs of the ontology, we obtain the expanded node set of review elements.

[0101] S503, extracting a connected review factor subgraph from the medical monitoring knowledge graph based on the review factor extension node and the review factor-related ontology concept nodes;

[0102] Specifically, a connected subgraph of review elements is extracted from the medical monitoring knowledge graph based on the review element extension nodes and review element-related ontology concept nodes. This approach primarily employs graph community discovery and a minimum spanning tree approach. First, the review element-related ontology concept nodes and the extended node set are merged to form a candidate node set. Next, a graph community discovery algorithm, such as the Louvain algorithm based on modularity optimization, is run on the medical monitoring knowledge graph to partition the graph into several node communities with dense interiors and loose exteriors. Next, each graph community is traversed, and the top-K communities containing the largest number of candidate nodes are selected. Considering that relying solely on local communities may miss critical paths, a minimum spanning tree algorithm is further employed to extract the shortest connected subgraph of candidate nodes in the knowledge graph, aiming to connect as many candidate nodes as possible with as few edges as possible. This approach requires balancing the information content and noise of the subgraph. Thresholds for the number of nodes and edges can be set to appropriately constrain the subgraph size. Finally, the graph community is combined with the minimum spanning tree to prune redundant nodes and edges. Local search is then used to expand one-hop neighbors to improve information completeness, ultimately resulting in a connected subgraph of review elements. It is worth mentioning that for complex scenarios involving multiple review elements, connected subgraphs of each element can be constructed in parallel, and alignment and fusion between subgraphs can be achieved through iterative optimization of graph matching and other technologies, thereby constructing an association network covering multiple review elements and internal semantic connectivity.

[0103] S504: Perform redundant pruning on the review factor subgraph to obtain a review factor graph.

[0104] Specifically, the censorship factor subgraph is pruned of redundancies to generate the censorship factor graph, primarily using methods based on betweenness centrality and structural holes. First, the betweenness centrality of each node in the subgraph is calculated—that is, the frequency with which a node serves as an intermediary for the shortest path between other nodes—to characterize the node's importance in information dissemination within the graph structure. Furthermore, structural hole theory is used to measure the sparsity of a node's neighborhood—that is, the ease with which different communities connected by the node can be connected via other paths—to characterize the node's ability to control information within the graph structure. Next, subgraph nodes are jointly scored using a combination of betweenness centrality and structural hole metrics to balance their information dissemination and control capabilities. Given the reliance of association analysis on core nodes, a threshold is set to retain the top-K% of high-scoring nodes as backbone nodes. Furthermore, an optimization method similar to PCSF is employed to minimize the total edge weight of the subgraph while maintaining the backbone nodes unchanged, aiming to extract the densest and concise subset of node connections. Drawing on the concept of the minimum spanning tree, the local topology is regenerated from the backbone node set, with parameters set to balance edge size and path length, striving to connect the most key nodes with the fewest edges. Through iterative optimization, we ultimately obtain a concise, compact, and semantically rich review factor graph. In the case of multiple review factors, we further apply frequent subgraph mining to each factor graph to extract co-occurring concept combination patterns and reveal general rules of multi-factor association.

[0105] S105: Based on the medical monitoring knowledge graph and the preset medical monitoring rules, the review factor graph is inferred and judged to form a preliminary review result;

[0106] Specifically, the pre-defined medical monitoring rules must first be represented in a form compatible with the knowledge graph, such as Horn clauses and predicate logic. These review rules are derived from laws, regulations, ethical guidelines, and expert consensus related to medical monitoring, and cover all aspects of medical monitoring requirements and standards. For example, "All human research must obtain informed consent from subjects" can be represented as a Horn clause: "Human research (X) ∧ involving (X, subjects) → informed consent (X, subjects)." By formalizing the review rules, they can be directly matched and interoperated with ontological concepts and relationship edges in the knowledge graph, providing a basis for reasoning and judgment based on the knowledge graph. Next, the review factor graph is fused with the medical monitoring knowledge graph to form a more complete and interconnected knowledge network. Specifically, the concept nodes, attribute edges, and relationship edges in the review factor graph are mapped to the corresponding ontological elements in the knowledge graph, achieving semantic alignment and connectivity between the two graphs. This fusion process can be automated using techniques such as semantic similarity matching and ontology mapping, but it also requires necessary manual review and confirmation by experts. The fused knowledge graph contains a seamless combination of specific information of the application project and knowledge in the field of medical monitoring, providing a comprehensive and accurate knowledge basis for reasoning and judgment.

[0107] Based on the fusion of knowledge graphs, tools such as ontology reasoning engines or rule engines are used to perform reasoning and judgment on the fused knowledge graph. The reasoning process primarily consists of two steps: rule matching and conflict resolution. First, the reasoning engine traverses the knowledge graph, searching for matching graph segments based on pre-set medical monitoring rules. For example, for the rule "Human subjects research must obtain informed consent," the reasoning engine searches the graph for nodes labeled "Human subjects research" and checks whether they have an "involved" edge with the "subjects" node, as well as a corresponding "informed consent" node and edge. Successful rule matching indicates that the proposed project meets ethical requirements in the corresponding aspects, while unmatched rules indicate potential ethical risks or deficiencies.

[0108] After completing rule matching, the inference engine further performs conflict detection and resolution on the matching results. Because medical monitoring rules may overlap, conflict, or have exceptions, simple rule matching can lead to contradictory or erroneous judgments. Therefore, it's necessary to utilize mechanisms such as priority setting and exception handling to comprehensively analyze and adjust the rule matching results to eliminate conflicts and inconsistencies. For example, if the "Informed Consent" rule is successfully matched but the "Subject's Insufficient Comprehension Ability" exception rule is also triggered, further determination is needed to determine whether the subject meets the exception criteria and adjust the review results accordingly.

[0109] After matching rules and resolving conflicts, the inference engine ultimately produces a preliminary review result, including a compliance assessment of the application's various medical monitoring aspects, risk warnings, and improvement suggestions. This preliminary review result, presented in a structured and organized format, clearly identifies the application's medical monitoring priorities, compliance evidence, and areas requiring supplementation, providing a comprehensive and reliable reference for expert review.

[0110] S106, quantitatively scoring the preliminary review results to obtain a quantitative review score;

[0111] Specifically, it is first necessary to develop a set of scientific and standardized scoring standards and rules for medical monitoring. These scoring standards address key dimensions of medical monitoring, such as informed consent, risk control, and subject protection, and define several evaluation indicators and weighting coefficients for each dimension. For example, the "informed consent" dimension could include indicators such as "completeness of informed consent" and "understandability of informed consent," with each indicator assigned a different weight based on its importance. The scoring rules define the scoring level, score range, and evaluation criteria for each indicator. For example, the "completeness of informed consent" indicator could be assigned three levels: "complete (5 points)," "basically complete (3 points)," and "incomplete (0 points)," with detailed criteria for determining each level. The development of these scoring standards and rules requires full input from stakeholders, including medical monitoring experts, legal professionals, and the public, to ensure their scientific validity, rationality, and widespread acceptance.

[0112] After clarifying the scoring criteria and rules, the preliminary review results will be matched and evaluated item by item to obtain the original score of each indicator. Specifically, for each medical monitoring dimension and indicator involved in the review results, the corresponding scoring level will be determined according to the degree to which it complies with the scoring rules, and the corresponding score will be assigned. This matching and evaluation process can be completed manually by experts or automated through technologies such as natural language processing and pattern matching. Regardless of the method adopted, it is necessary to base it on the preliminary review results, strictly follow the scoring criteria and rules, and try to avoid subjectivity or arbitrariness. After evaluating all indicators, a set of original scores will be obtained, which reflects the basic situation of the application project in various medical monitoring aspects.

[0113] In order to obtain a quantitative review score for the overall application project, it is also necessary to perform weighted synthesis based on the original score. Specifically, according to the weight coefficient of each indicator, its original score is weighted and calculated to obtain the weighted score of the indicator. Then, the weighted scores of all indicators are added together to obtain the comprehensive score of the application project in the entire medical monitoring dimension. This weighted synthesis process usually adopts a linear weighted method, that is, the weight of each indicator is multiplied by the original score and then summed. However, in practice, other weighting strategies can also be adopted as needed, such as geometric mean, logistic regression, etc., to better reflect the interaction and combination effects between different indicators.

[0114] For example, suppose the "Informed Consent" dimension includes three indicators with weights of 0.5, 0.3, and 0.2, respectively. After evaluation, the original scores are 4, 3, and 5, respectively. The weighted scores of these three indicators are 2, 0.9, and 1, respectively, for a total of 3.9, which is the application's score for the "Informed Consent" dimension. If the total weight of the "Informed Consent" dimension is 0.4, and the weights of the other four medical monitoring dimensions are 0.3, 0.1, 0.1, and 0.1, respectively, and the weighted composite scores are 4.2, 3.5, 4.0, and 3.8, respectively, the application's overall review score is 3.99 (i.e., 0.43.9 + 0.34.2 + 0.13.5 + 0.14.0 + 0.1 * 3.8). This comprehensive score intuitively reflects the application's ethical compliance level and provides an important quantitative reference for the final review outcome.

[0115] Based on the above embodiment, as an optional embodiment, a review decision is made based on the review quantitative score and the preliminary review result, and the review decision result is obtained, including:

[0116] S601: Analyze the review quantitative score and the preliminary review results to obtain the review quantitative score;

[0117] Specifically, the quantified review score and preliminary review results are fused and analyzed to produce a fused quantitative review score. This is primarily based on the DS evidence theory and the Analytic Hierarchy Process (AHP). First, the qualitative descriptions in the preliminary review results are mapped to corresponding quantitative sub-scores. By constructing a rule base that maps descriptive statements to sub-scores, natural language processing techniques are used to automatically extract key statements from the review results and map them to corresponding sub-scores, thus performing structured quantitative processing on the review results. Simultaneously, the credibility of each sub-score dimension is modeled using DS evidence theory. The uncertainty of each score is characterized by setting a basic probabilistic confidence level for each sub-score. Based on this, the DS evidence synthesis rule is applied to combine the quantitative scores and confidence levels of each dimension to form the initial value of the fused score. Considering the varying importance of each scoring dimension, the AHP is further employed to determine the weights of each dimension. A pairwise comparison judgment matrix is ​​used to relatively measure the importance of each dimension, and the eigenvalue method is used to calculate weights to balance the influence of each dimension. Finally, the initial fused score value and dimension weights are combined to form a weighted sum, resulting in a fused score that comprehensively considers both the quantitative score and qualitative analysis. This score is then fed back into the quantitative review score. In addition, to enhance the interpretability of the fusion analysis, methods such as fuzzy comprehensive evaluation are used to compare the score differences before and after fusion, trace the key factors behind the differences, and automatically generate text explanations to intuitively present the impact of qualitative analysis on the score.

[0118] S602: Perform review decision based on the preset scoring threshold and review quantitative score to obtain a review decision result.

[0119] Specifically, review decisions are made based on preset scoring thresholds and quantitative review scores to arrive at the final review decision. This approach primarily employs a decision tree and fuzzy comprehensive evaluation. First, based on historical review data and expert experience, thresholds are set for the integrated quantitative review scores, dividing the score range into discrete intervals. Furthermore, clear decision labels are assigned to each score interval, forming a rule base mapping quantitative scores to review decisions. Next, a decision tree algorithm is used to train and optimize this mapping rule base. By constructing a decision tree with score intervals as non-leaf nodes and decision labels as leaf nodes, and performing node splitting and pruning based on criteria such as information entropy and the Gini index, concise and efficient decision rules are formed. Based on this, the quantitative scores of the items to be reviewed are input into the decision tree, automatically inferring the corresponding review decision results. Considering the potential for decision uncertainty when the score is at the critical point of the interval, a fuzzy comprehensive evaluation method is further employed to assist in the decision-making process. A membership function is used to measure the proximity of the score to the critical point, and a membership threshold is set. Samples at the critical point are then labeled as requiring review, prompting manual intervention. In addition, for projects that require rectification and re-examination, a method similar to functional magnetic resonance imaging is used to compare the state matrices before and after, quantitatively evaluate the degree of project optimization, and provide a basis for adjusting the review strategy.

[0120] S107, making a review decision based on the review quantitative score and the preliminary review result to obtain a review decision result.

[0121] Specifically, it is first necessary to set scientific and reasonable decision thresholds and judgment rules based on the quantitative review scores. Usually, corresponding score ranges can be set for different scoring levels (such as excellent, good, qualified, unqualified, etc.), and the type of review result corresponding to each range can be clearly defined. For example, a score of 90 or above can be defined as "excellent", corresponding to direct pass; 80-90 points can be defined as "good", corresponding to pass after modification; 60-80 points can be defined as "qualified", corresponding to conditional pass; and below 60 points can be defined as "unqualified", corresponding to failure. When determining thresholds and judgment rules, it is necessary to comprehensively consider factors such as the legal and regulatory requirements of medical supervision, the characteristics of the subject area, and the social ethical environment, and to widely absorb expert opinions and practical experience to ensure its scientific nature and operability.

[0122] Based on the above embodiment, as an optional embodiment, the preliminary review results are quantitatively scored to obtain a review quantitative score, including:

[0123] S701: Determine the quantitative dimensions and quantitative indicators of the quantitative scoring and set scoring rules for each quantitative indicator based on the requirements and standards of medical monitoring;

[0124] Specifically, based on the requirements and standards of medical monitoring, the quantitative dimensions and indicators for the quantitative scoring were determined, and scoring rules were established for each indicator. This was primarily achieved using a hierarchical analysis and Delphi method. First, a literature review systematically reviewed the general requirements for medical monitoring, and several core elements were summarized from the perspectives of scientificity, ethics, and compliance. Field experts were invited to supplement and revise the criteria, ultimately finalizing the top-level quantitative dimensions. Within each top-level dimension, specific assessment points were further refined. Multiple rounds of expert consultation were conducted through the Delphi method, resulting in a convergence of actionable and assessable quantitative indicators. Based on expert judgment, machine learning methods were used to optimize the indicator set using techniques such as cluster analysis and association analysis. The analytic hierarchy process was used to determine indicator weights to balance the importance of each factor. Next, scoring rules were developed for each quantitative indicator. Using a fuzzy comprehensive evaluation approach, the corresponding membership level was set based on the degree of compliance of the indicator with the key review points, and then the score range was mapped to a percentage scale. Furthermore, additional bonus and deduction points were added based on the completeness and authenticity of the review materials to address special circumstances. On this basis, we further developed a scorecard template, clarifying the scoring method for each indicator, the list of required materials, and the key review points, and ultimately established a structured quantitative scoring system. For dynamically adjusted scoring rules, we promptly updated the scorecard and related guidelines to ensure synchronized iteration of the system.

[0125] S702, assigning weights to the quantitative dimensions and quantitative indicators to obtain a weight ratio table;

[0126] Specifically, weights were assigned to quantitative dimensions and indicators to create a weighting table, primarily using methods based on the Analytic Hierarchy Process (AHP) and machine learning. First, based on the "dimension-indicator-rule" framework, the AHP approach was used to construct a three-level hierarchical weighting model. Through pairwise comparisons, the relative importance of each top-level dimension to the review decision was determined at a macro level. A judgment matrix was constructed and consistency tested. Based on the dimension-level weights, the relative importance of each indicator within each dimension was further compared, and indicator-level weights were determined using a similarity approach. The top-level dimension weights were comprehensively analyzed with the corresponding indicator-level weights and normalized, ultimately forming a top-down hierarchical weighting scheme. Given that the AHP relies primarily on expert experience and is subject to certain subjectivity, machine learning methods were further introduced to optimize it. A representative sample of historical review projects was selected, and their quantitative scoring data was extracted. Based on the final review results, the projects were labeled into different categories, such as approved, rectified, and rejected. Furthermore, using common metrics such as information gain and the Gini index, algorithms such as classification decision trees and logistic regression were used to analyze the correlation between the quantitative scores of different indicators and the review results. The expert-assigned weightings were then calibrated accordingly. The weights derived from qualitative analysis and quantitative calculations are combined to create a dynamically integrated weighted ratio table through weighted averaging and other methods. Considering the dynamic nature of review requirements, new review data is regularly collected to continuously optimize the weight ratios and continuously improve the scientific nature of weight setting. Furthermore, in response to dynamic adjustments to weights, the scoring system is updated simultaneously to achieve automatic changes in score calculations, ensuring synchronization between scoring rules and weight ratios.

[0127] S703, setting scoring rules and weight ratio table based on quantitative indicators, quantitatively scoring the preliminary review results, and obtaining the review quantitative score.

[0128] Specifically, based on the scoring rules and weighting table for the quantitative indicators, the preliminary review results are quantitatively scored to obtain a quantitative review score. This is primarily achieved through a tiered scoring method based on weighted summation. First, the preliminary review results of the subject to be reviewed are matched against the scoring rules table. The degree of compliance with the corresponding requirements for each indicator is assessed item by item, and a score is assigned based on the rules table. For elements that are difficult to assess subjectively, a combination of qualitative and quantitative methods, such as fuzzy comprehensive evaluation, can be used to assign scores as appropriate, in conjunction with expert opinion. With each indicator having a clear score, a weighted summation method is used, referring to the weighting table, to calculate the weighted average of the scores for each top-level dimension to serve as the composite score for that dimension. Furthermore, using the same method, the scores for each top-level dimension are summed according to their weights to obtain a preliminary quantitative review score for the subject. Based on the common weightings, appropriate adjustments can be made to the specific circumstances, taking into account the characteristics of different review items. For example, for projects involving human subjects, the weighting of informed consent can be increased; for big data application research, greater emphasis should be placed on data security and privacy protection. The same tiered weighting calculation method is used for the adjusted personalized weights to obtain a final quantitative score tailored to the specific review subject. During the calculation process, care should be taken to uniformly convert percentage scores to decimals between 0 and 1, retaining two to three decimal places in the result to ensure accuracy. Furthermore, appropriate constants can be set to avoid extreme cases where a single indicator receiving a zero score results in a low overall score.

[0129] On the other hand, this application also provides a clinical research medical monitoring system based on artificial intelligence, such as Figure 2 , the system comprises:

[0130] Ontology knowledge base construction module 1 is used to obtain medical knowledge texts related to medical monitoring and construct a medical monitoring ontology knowledge base based on the medical knowledge texts;

[0131] Medical monitoring knowledge graph construction module 2 is used to obtain sample data of medical monitoring materials and map the sample data of medical monitoring materials into the ontology framework based on the ontology knowledge base to form a semantic network including concept nodes, attribute edges, and relationship edges to obtain a medical monitoring knowledge graph;

[0132] Review factor acquisition module 3, used to obtain medical monitoring application materials and perform knowledge extraction on the medical monitoring application materials to obtain review factors;

[0133] An audit factor graph acquisition module 4 is configured to map the audit factors into the medical monitoring knowledge graph, obtain the audit factor-related ontology concept nodes, and form an audit factor graph based on the related ontology concept nodes;

[0134] A preliminary review result acquisition module 5 is used to infer and judge the review factor graph based on the medical monitoring knowledge graph and preset medical monitoring rules to form a preliminary review result;

[0135] Quantitative scoring module 6, used for quantitatively scoring the preliminary review results to obtain a review quantitative score;

[0136] The review decision module 7 is used to make a review decision based on the review quantitative score and the preliminary review result to obtain a review decision result.

[0137] Please refer to Figure 3 The present application also discloses an electronic device. Figure 3 The electronic device 300 may include: at least one processor 301 , at least one network interface 304 , a user interface 303 , a memory 305 , and at least one communication bus 302 .

[0138] The communication bus 302 is used to implement the connection and communication between these components.

[0139] The user interface 303 may include a display screen (Display) and a camera (Camera). Optionally, the user interface 303 may also include a standard wired interface and a wireless interface.

[0140] The network interface 304 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0141] The processor 301 may include one or more processing cores. Using various interfaces and circuits, the processor 301 connects to various components within the server. It executes instructions, programs, code sets, or instruction sets stored in the memory 305, as well as accesses data stored in the memory 305, to perform various server functions and process data. Optionally, the processor 301 may be implemented using at least one of the following hardware forms: a digital signal processing (DSP), a field-programmable gate array (FPGA), or a programmable logic array (PLA). The processor 301 may integrate one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. The CPU primarily processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing content displayed on the display screen; and the modem handles wireless communications. It is understood that the modem may not be integrated into the processor 301 but implemented as a separate chip.

[0142] Among them, the memory 305 may include a random access memory (Random Access Memory, RAM) and may also include a read-only memory (Read~Only Memory). Optionally, the memory 305 includes a non-transitory computer-readable medium. The memory 305 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 305 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various method embodiments, etc. The memory 305 may also optionally be at least one storage system located away from the aforementioned processor 301. Reference Figure 3 , as a computer storage medium, the memory 305 may include an operating system, a network communication module, a user interface module, and an application program for a clinical research medical monitoring method based on artificial intelligence.

[0143] exist Figure 3In the electronic device 300 shown, the user interface 303 is mainly used to provide an input interface for the user and obtain the data input by the user; and the processor 301 can be used to call the application program storing the road assessment method in the memory 305. When executed by one or more processors 301, the electronic device 300 executes one or more methods in the above-mentioned embodiments. It should be noted that for the aforementioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should know that this application is not limited to the described order of actions, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required for this application. In the above-mentioned embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0144] In the several embodiments provided in this application, it should be understood that the disclosed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some service interface, and the indirect coupling or communication connection of the system or unit can be electrical or other forms.

[0145] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0146] The present application also provides a computer storage medium that can store multiple instructions, which are suitable for being loaded and executed by a processor as described above. Figure 1 The road assessment method of the embodiment shown, the specific execution process can be found in Figure 1 The detailed description of the illustrated embodiment will not be repeated here.

[0147] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0148] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of this application, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the method of this application. The aforementioned memory includes various media that can store program code, such as USB flash drives, mobile hard drives, magnetic disks, or optical disks.

[0149] The above are merely exemplary embodiments of the present disclosure and are not intended to limit the scope of the present disclosure. In other words, any equivalent variations and modifications made in accordance with the teachings of the present disclosure are still within the scope of the present disclosure. Those skilled in the art will readily conceive of other embodiments of the present disclosure after considering the disclosure and the practical implications thereof.

[0150] This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art not described herein. The description and examples are to be considered as exemplary only, and the scope and spirit of the present disclosure are to be defined by the claims.

Claims

1. A clinical research medical monitoring method based on artificial intelligence, characterized in that the method include: Acquire medical knowledge texts related to medical monitoring, and construct a medical monitoring ontology knowledge base based on the medical knowledge texts; Acquire sample data of medical monitoring materials, and map the sample data of medical monitoring materials into an ontology framework based on the ontology knowledge base to form a semantic network including concept nodes, attribute edges, and relationship edges, thereby obtaining a medical monitoring knowledge graph; Obtaining medical monitoring application materials, and performing knowledge extraction on the medical monitoring application materials to obtain review elements; Mapping the review elements to the medical monitoring knowledge graph to obtain ontology concept nodes related to the review elements, and forming a review element graph based on the related ontology concept nodes, including: mapping the review elements to ontology concept nodes in the medical monitoring knowledge graph to obtain ontology concept nodes related to the review elements; Expanding relevant nodes in the medical monitoring knowledge graph according to the ontology concept nodes related to the review factors to obtain review factor expansion nodes; Extracting a connected review factor subgraph from the medical monitoring knowledge graph based on the review factor extension node and the review factor-related ontology concept node; The review factor graph is obtained by performing redundant trimming on the review factor subgraph. For complex scenarios involving multiple review factors, connected subgraphs of each review factor are constructed in parallel, and alignment and fusion between subgraphs are achieved through iterative optimization of graph matching technology. The extension scope of the review factor-related ontology concept nodes is expanded to multiple external knowledge graphs, and the extension nodes of the internal and external graphs of the ontology are integrated and deduplicated to obtain the review factor extension node set. Performing reasoning and judgment on the review factor graph based on the medical monitoring knowledge graph and preset medical monitoring rules to form a preliminary review result, wherein the review factor graph is integrated with the medical monitoring knowledge graph, and reasoning and judgment are performed based on the integrated knowledge graph to obtain the preliminary review result; Quantitatively scoring the preliminary review results to obtain a quantitative review score; An examination decision is made based on the examination quantitative score and the preliminary examination result to obtain an examination decision result.

2. The method according to claim 1, characterized in that The step of constructing a medical monitoring ontology knowledge base based on the medical knowledge text includes: Using natural language processing technology to extract core concepts related to medical monitoring from the medical knowledge text; Based on the ontology core concepts and the medical knowledge text, determining the semantic relationship between the ontology core concepts, and performing ontology relationship definition according to the semantic relationship between the ontology core concepts to obtain the relationship definition between each ontology core concept; Defining attributes of each ontology core concept based on the relationship definition between each ontology core concept to obtain attribute definitions of each ontology core concept; The medical monitoring ontology knowledge base is constructed based on the relationship definitions of each ontology core concept and the attribute definitions of each ontology core concept.

3. The method according to claim 1, characterized in that The medical monitoring material sample data is mapped to the ontology framework based on the ontology knowledge base to form a semantic network including concept nodes, attribute edges, and relationship edges, thereby obtaining a medical monitoring knowledge graph, including: Preprocessing the sample data of the medical monitoring information to obtain medical monitoring data; Performing data mapping on the medical monitoring data based on the ontology knowledge base, mapping entities in the medical monitoring data to corresponding ontology concept nodes in the ontology knowledge base, and mapping attributes in the medical monitoring data to attribute edges of the ontology concept nodes; Performing relationship extraction on the medical monitoring data based on the ontology knowledge base, extracting semantic relationships between named entities in the medical monitoring data, and mapping the semantic relationships into relationship edges between the ontology concept nodes; A medical monitoring knowledge graph is constructed based on the ontology concept nodes, the attribute edges and the relationship edges.

4. The method according to claim 1, wherein The knowledge extraction of the medical monitoring application materials is performed to obtain review elements, including: Structuring the medical monitoring application materials to obtain structured application materials, and performing text parsing on the structured application materials to obtain key information elements; Perform entity naming based on the key information elements and the structured application materials to obtain key entities corresponding to the review elements; Perform named entity recognition on structured application materials to extract key entities corresponding to review elements; Performing type definition on the key entities to obtain defined key entities, and performing relationship extraction based on the defined key entities to determine semantic associations between the defined key entities; Defining attributes based on the defined key entities to obtain attributes of each of the defined key entities; The defined key entities, the attributes of each of the defined key entities and the semantic associations between the defined key entities are organized into a structured knowledge representation to obtain the review elements.

5. The method according to claim 1, wherein The review decision is made based on the review quantitative score and the preliminary review result, and the review decision result is obtained, including: Performing a fusion analysis on the review quantitative score and the preliminary review results to obtain a fused review quantitative score; An examination decision is made based on a preset scoring threshold and the fused examination quantitative score to obtain the examination decision result.

6. The method according to claim 1, characterized in that The quantitative scoring of the preliminary review results to obtain the review quantitative score includes: Determine the quantitative dimensions and quantitative indicators of the quantitative scoring and set scoring rules for each of the quantitative indicators according to the requirements and standards of medical monitoring; Assigning weights to the quantitative dimensions and the quantitative indicators to obtain a weight ratio table; The scoring rules and the weight ratio table are set based on the quantitative indicators, and the preliminary review results are quantitatively scored to obtain the review quantitative score.

7. An artificial intelligence-based clinical research medical monitoring system for implementing the method according to any one of claims 1 to 6, characterized in that: include: An ontology knowledge base construction module is used to obtain medical knowledge texts related to medical monitoring and construct a medical monitoring ontology knowledge base based on the medical knowledge texts; A medical monitoring knowledge graph construction module is used to obtain sample data of medical monitoring materials and map the sample data of medical monitoring materials into an ontology framework based on the ontology knowledge base to form a semantic network including concept nodes, attribute edges, and relationship edges to obtain a medical monitoring knowledge graph; An examination element acquisition module is used to obtain medical monitoring application materials and perform knowledge extraction on the medical monitoring application materials to obtain examination elements; An audit factor graph acquisition module is used to map the audit factors into the medical monitoring knowledge graph, obtain the ontology concept nodes related to the audit factors, and form an audit factor graph based on the related ontology concept nodes; A preliminary review result acquisition module is used to infer and judge the review factor graph based on the medical monitoring knowledge graph and preset medical monitoring rules to form a preliminary review result; A quantitative scoring module, configured to quantitatively score the preliminary review results to obtain a review quantitative score; The review decision module is used to make a review decision based on the review quantitative score and the preliminary review result to obtain a review decision result.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executed by a method according to any one of claims 1 to 6.

9. An electronic device, characterized in that: The electronic device comprises a processor, a memory and a transceiver, wherein the memory is used to store instructions, the transceiver is used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes the method according to any one of claims 1 to 6.

Images