Main diagnosis rationality judgment method and system based on artificial intelligence

Abstract

The invention relates to the technical field of medical information processing, in particular to a main diagnosis rationality judgment method and system based on artificial intelligence. The method comprises the following steps: acquiring diagnosis request data, and performing candidate main diagnosis generation on the diagnosis request data by using preset medical knowledge graph data and preset intelligent detection rule data to obtain candidate main diagnosis data; performing diagnosis execution parameter generation according to the candidate main diagnosis data to obtain candidate main diagnosis parameter data; acquiring historical diagnosis request data and historical candidate main diagnosis data, and performing diagnosis linkage processing according to the historical diagnosis request data and the historical candidate main diagnosis data to obtain diagnosis linkage network data; and performing node interaction processing on the candidate main diagnosis data according to the diagnosis linkage network data to obtain candidate main diagnosis weight data. According to the method, the intelligent level of diagnosis rationality judgment, the credibility of a scoring result and the adaptability of a clinical system are improved.

Description

Technical Field

[0001] The present invention relates to the field of medical information processing technology, and in particular to a method and system for judging the rationality of a primary diagnosis based on artificial intelligence. Background Art

[0002] With the continuous improvement of the level of medical informationization, electronic medical record systems, clinical pathway management systems and intelligent auxiliary diagnosis systems have gradually become popular in various medical institutions. Traditional auxiliary diagnosis methods mostly rely on rule bases, disease knowledge tables or statistical models to recommend or verify the main diagnosis. Some systems also introduce artificial intelligence algorithms to automatically identify diagnostic information. However, in actual application, the existing technologies generally lack the interpretability of the diagnostic result generation process. Some models rely on black box algorithms and cannot provide a clear logical basis for diagnostic recommendations, which is not conducive to medical staff's acceptance. Summary of the invention

[0003] In order to solve the above-mentioned technical problems, the present invention proposes a method and system for judging the rationality of a main diagnosis based on artificial intelligence to solve at least one of the above-mentioned technical problems.

[0004] The present application provides a method for judging the rationality of a primary diagnosis based on artificial intelligence, the method comprising: S1. Obtain diagnosis request data, and generate candidate primary diagnosis for the diagnosis request data using preset medical knowledge graph data and preset intelligent detection rule data to obtain candidate primary diagnosis data; S2. Generate diagnostic execution parameters according to the candidate primary diagnostic data to obtain candidate primary diagnostic parameter data; S3, obtaining historical diagnosis request data and historical candidate main diagnosis data, and performing diagnosis linkage processing according to the historical diagnosis request data and the historical candidate main diagnosis data to obtain diagnosis linkage network data; S4. Perform node interaction processing on the candidate main diagnosis data according to the diagnosis linkage network data to obtain candidate main diagnosis weight data.

[0005] In the present invention, by constructing a medical knowledge graph and an intelligent detection rule engine, the intelligent generation of candidate main diagnoses and the structured expression of context parameters are realized; historical diagnosis data are further introduced to establish a diagnosis linkage network, and a multi-dimensional weight evaluation of the candidate main diagnoses is performed through graph structure modeling and node interaction scoring mechanism, which improves the accuracy, explainability and automation level of diagnostic rationality judgment, and effectively enhances the intelligent decision-making assistance capability and clinical data utilization efficiency of the medical information system.

[0006] Optionally, the step of constructing the preset medical knowledge graph data includes the following steps: Acquire structured electronic medical records, medical standard coding data, and unstructured medical literature data to construct a multi-source medical data set; Perform entity recognition and entity relationship extraction based on multi-source medical data sets to obtain medical entity data and initial association relationship data; Construct semantic relationship edges between entities based on medical entity data and initial association relationship data to form medical relationship network data; According to the preset medical knowledge base, the medical entity data is processed with semantic relationship enhancement to obtain the medical entity causal relationship, medical entity adaptation relationship and medical entity concurrent association relationship data; According to the causal relationship, adaptive relationship and concurrent association relationship of medical entities, the edge direction of medical relationship network data is calibrated to obtain medical knowledge graph data.

[0007] The present invention constructs a multi-source heterogeneous medical data set by integrating structured medical record data, standard medical coding and unstructured documents, combines entity recognition and relationship extraction technology to generate a medical relationship network, and further enhances semantic relationships and calibrates edge directionality based on a preset medical knowledge base to construct a directed medical knowledge graph with causal relationships, adaptive relationships and concurrent associations. This graph structure not only reflects the multi-layer semantic logic of medical knowledge, but also has semantic reasoning and upstream and downstream dependency modeling capabilities. Compared with traditional static graphs or rule bases, it improves the systematic nature of knowledge organization, the accuracy of relationship expression and the interpretability of subsequent diagnostic reasoning, and has significant technical advantages and industry applicability.

[0008] Optionally, the preset intelligent detection rule data includes main diagnosis judgment rules, surgical judgment rules and complication handling rules. The main diagnosis judgment rules, surgical judgment rules and complication handling rules include rule identifiers, rule types, trigger conditions, parameter configurations and priority information. The trigger conditions include a combination of inspection indicators, chief complaint information and medical record elements, which are used to realize the linkage judgment and intelligent triggering of diagnosis request data and rules.

[0009] The present invention pre-sets multiple types of intelligent detection rules, and defines rule identifiers, types, trigger conditions, parameter configurations and priority information in the rules to form a structured and configurable rule system, which supports the linkage judgment and intelligent triggering of rules through the combination and matching of inspection indicators, chief complaint information and medical record elements. Compared with traditional fixed rule templates, this rule system has dynamic adaptation capabilities and refined management features, and can quickly match the most relevant rule paths in different clinical scenarios, improve the accuracy and controllability of primary diagnosis generation, and support the continuous optimization and maintenance of the rule base, with stronger scalability and engineering adaptability.

[0010] Optionally, S1 includes: Obtaining diagnosis request data, and using preset medical knowledge graph data to perform entity matching on the diagnosis request data to obtain entity matching data; According to the entity matching data, the medical knowledge graph data is retrieved for an initial diagnosis-related candidate set to obtain initial diagnosis-related candidate set data; The preset intelligent detection rule data is used to match the primary diagnosis candidate items of the initial diagnosis-related candidate set data to obtain candidate primary diagnosis data.

[0011] The present invention achieves rapid retrieval of the initial diagnosis candidate set by performing entity-level matching between the diagnosis request data and the medical knowledge graph, and combining the structural associations in the knowledge graph; the candidate set is then conditionally matched and screened through the preset intelligent detection rule data to obtain highly reliable primary diagnosis candidates. Compared with the traditional diagnosis generation method that relies on manual rule configuration or black box model reasoning, the present invention realizes the interpretability, rule-drivenness and graph structure optimization of the diagnosis reasoning process, effectively improving the accuracy, stability and system scalability of candidate diagnosis generation.

[0012] Optionally, S2 includes: Extract candidate diagnosis context from the diagnosis request data according to the candidate main diagnosis data to obtain candidate diagnosis context data; Initialize rule parameter mapping of candidate main diagnosis data and candidate diagnosis context data according to preset intelligent detection rule data to obtain rule parameter mapping data; The candidate diagnosis parameter structure is constructed according to the rule parameter mapping data to obtain candidate main diagnosis parameter data.

[0013] In the present invention, by extracting the context information of the candidate main diagnosis in the diagnosis request data, and combining it with the preset intelligent detection rule data for parameter mapping and structured construction, the structural expression of the candidate diagnosis in multi-dimensional parameter dimensions such as semantics, indicators, and rule conditions is realized. Compared with the traditional coarse-grained judgment method based on the rule table, this method can construct a candidate diagnosis parameter structure with context perception and rule semantic alignment characteristics, providing clear, traceable, and configurable basic data support for subsequent diagnosis rationality analysis, and significantly improving the interpretation ability and parameterized management ability of the intelligent diagnosis system.

[0014] Optionally, S3 includes: Obtain historical diagnosis request data and historical candidate primary diagnosis data; Extracting diagnostic instance associations from historical diagnostic request data and historical candidate primary diagnostic data to obtain diagnostic instance association data; Perform linkage relationship mining based on the diagnosis instance association data to obtain linkage relationship data; The diagnosis linkage network data is generated according to the diagnosis instance association data and the linkage relationship data to obtain the diagnosis linkage network data.

[0015] In the present invention, by extracting the diagnostic instance-level associations between historical diagnostic request data and historical candidate main diagnostic data, combined with linkage relationship mining methods, diagnostic linkage network data containing diagnostic co-occurrence, sequence and context-dependent features are constructed to achieve semantic relationship modeling and structural expression between diagnostic nodes. Compared with the traditional diagnostic association method based on static co-occurrence statistics or rule settings, the present invention can dynamically reflect the complex diagnostic linkage patterns in real diagnosis and treatment data, significantly enhance the ability to utilize historical experience and the basis of graph structure reasoning in the diagnostic reasoning process, and effectively improve the accuracy and adaptability of subsequent intelligent scoring and rationality judgment.

[0016] Optionally, S4 includes: According to the candidate main diagnosis data, a candidate main diagnosis subgraph is constructed for the diagnosis linkage network data to obtain the candidate main diagnosis subgraph data; Extract node linkage features according to candidate main diagnosis subgraph data and candidate main diagnosis data to obtain node linkage feature data; Interactive scoring is performed based on candidate primary diagnosis data and node linkage feature data to obtain preliminary scoring data; A simplified graph attention mechanism is performed on the candidate main diagnosis data and preliminary scoring data to obtain the candidate main diagnosis weight data.

[0017] In the present invention, a subgraph associated with the candidate primary diagnosis is constructed in the diagnostic linkage network, the linkage features of the candidate diagnosis nodes in the graph structure are extracted, and interactive scoring is performed in combination with context information. The weights between nodes are dynamically adjusted using a simplified graph attention mechanism to generate candidate primary diagnosis weight data with structural interpretability. Compared with traditional static scoring or rule priority methods, the present invention introduces a graph structure semantic propagation and multi-source feature fusion mechanism, which can adaptively evaluate the rationality priority of the primary diagnosis based on the diagnostic context and historical graph relationship, effectively improving the accuracy of diagnostic weight judgment, semantic consistency and system intelligence level.

[0018] Optionally, the interaction score includes: Extracting context semantic features according to the candidate primary diagnosis data to obtain context semantic feature data; Perform bilinear attention calculation based on node linkage feature data and context semantic feature data to obtain matching weight data; The node linkage feature data and context semantic feature data are fused according to the matching weight data to obtain preliminary scoring data.

[0019] In the interactive scoring process of the present invention, the contextual semantic features corresponding to the candidate main diagnosis are first extracted, and combined with the node linkage feature data, the matching weight between the candidate diagnosis and the graph structure is calculated based on the bilinear attention mechanism, and then the preliminary scoring result is generated through weight-driven feature fusion. Compared with the traditional weighted average or rule-based scoring method, the bilinear attention mechanism introduced in the present invention can more accurately model the deep coupling relationship between contextual semantics and graph structure, realize adaptive dynamic adjustment of diagnostic scores and semantic perception enhancement, and significantly improve the distinctiveness, reliability and clinical semantic consistency of the scoring results.

[0020] Optionally, the simplified graph attention mechanism processing includes: Perform adjacent node similarity calculation based on candidate primary diagnosis data and preliminary scoring data to obtain adjacent node edge weight data; A simplified attention coefficient is calculated based on the adjacent node edge weight data and the preliminary scoring data to obtain simplified attention coefficient data; Perform multi-order aggregation based on the simplified attention coefficient data to obtain multi-order neighbor aggregation data; Residual connection is performed based on multi-order neighbor aggregation data and preliminary scoring data to obtain candidate main diagnosis weight data.

[0021] In the present invention, the edge weight data is generated by calculating the similarity of adjacent nodes for the candidate main diagnosis nodes and their preliminary scoring results, and the attention coefficient is further simplified by combining the edge weight and node score calculation to achieve efficient graph structure attention distribution; the global semantic perception ability is enhanced through a multi-order neighbor aggregation mechanism, and the original score and graph aggregation information are fused under the residual connection to generate the diagnosis weight result. Compared with the traditional static propagation or full-graph attention method, the present invention greatly reduces the computational complexity while ensuring the aggregation accuracy, improves the diagnostic reasoning efficiency and edge deployment adaptability, and retains the original diagnostic semantics, effectively enhancing the robustness, flexibility and interpretability of the diagnostic scoring process. Traditional attention calculations often require multi-layer neural network calculations, with a large number of feature transformations, splicing and activation functions, high computational complexity, and a large reliance on computing resources and training samples during training / deployment, which is not conducive to lightweight embedded deployment or rule collaborative logic expression in clinical systems. In the design of the attention mechanism, the present invention adopts linear combination calculation of structural edge weights and preliminary scoring data to replace the traditional attention scoring method based on deep neural network parameter learning, and constructs a simplified attention coefficient calculation mechanism that does not require training, is easy to deploy, and is interpretable. It not only reduces the system's dependence on computing resources, but also improves the transparency and traceability of the diagnostic scoring process. It is particularly suitable for scenarios where medical assistance systems require high reliability and lightweight deployment capabilities.

[0022] Optionally, the present application further provides an artificial intelligence-based primary diagnosis rationality judgment system, which is used to execute the artificial intelligence-based primary diagnosis rationality judgment method as described above, and the artificial intelligence-based primary diagnosis rationality judgment system includes: A candidate diagnosis generation module is used to obtain diagnosis request data, and use preset medical knowledge graph data and preset intelligent detection rule data to generate candidate main diagnosis for the diagnosis request data to obtain candidate main diagnosis data; A diagnostic parameter extraction module is used to generate diagnostic execution parameters according to the candidate main diagnostic data to obtain candidate main diagnostic parameter data; The diagnosis linkage mapping module is used to obtain historical diagnosis request data and historical candidate main diagnosis data, and perform diagnosis linkage processing according to the historical diagnosis request data and the historical candidate main diagnosis data to obtain diagnosis linkage network data; The diagnosis weight evaluation module is used to perform node interaction processing on the candidate main diagnosis data according to the diagnosis linkage network data to obtain the candidate main diagnosis weight data.

[0023] The purpose of the present invention is to establish a multi-level structural expression model between diagnostic entities and semantic relationship edges by constructing a medical knowledge graph with causal associations, adaptive relationships and concurrent logic, and realize unified semantic mapping of unstructured symptom descriptions, structured examination indicators and chief complaint elements in diagnostic requests; based on the semantic paths between graph nodes and the upstream and downstream dependencies of entities, combined with a configurable intelligent detection rule engine, semantic triggering and rule screening are performed on the preliminary matched candidate main diagnosis set, thereby realizing the intelligent generation and structural traceability of candidate main diagnoses.

[0024] On this basis, the present invention further introduces historical diagnosis request data and candidate main diagnosis instances, and extracts the co-occurrence frequency, path dependency and sequence order relationship between diagnoses by constructing a "diagnosis-context-candidate" triple map, forming diagnostic linkage network data with directional and contextual condition constraints, and providing multi-dimensional historical association support in the real world for candidate main diagnoses. This graph structure not only has the topological propagation capability between entities, but also can serve as a path network for the propagation of scoring information between diagnosis candidate nodes.

[0025] Aiming at the task of judging the rationality of diagnosis, the present invention proposes a bidirectional scoring mechanism that integrates node linkage features and semantic context. By introducing a bilinear attention structure, the graph structural features of candidate nodes and the diagnostic context vector are aligned and calculated, and semantic matching weights are dynamically generated to achieve a refined expression of node-level scoring results. Subsequently, a lightweight graph attention aggregation mechanism is introduced to complete the semantic and structural enhancement of node weights based on multi-order neighbor information, and the residual connection mechanism is combined to maintain the initial semantic stability, and a credible weight score for diagnostic decision-making is output.

[0026] Compared with the traditional main diagnosis generation method based on rule tables or model outputs, the present invention has significant technical advantages such as clear knowledge modeling structure, traceable scoring process, strong context fusion capability, and high graph structure reasoning efficiency. It provides structured, intelligent, and explainable underlying technical support for diagnostic rationality judgment tasks, and greatly improves the integrated utilization efficiency of multi-source data and the automated diagnosis management capabilities of medical decision-making support systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Other features, objects and advantages of the present application will become more apparent by reading the detailed description of non-limiting implementations made with reference to the following drawings: Figure 1 A flowchart showing the steps of a method for determining the rationality of a primary diagnosis based on artificial intelligence according to an embodiment of the present invention is shown; Figure 2 A flowchart showing the steps of a candidate diagnosis generation method according to an embodiment is shown; Figure 3 A flowchart showing the steps of a diagnostic parameter extraction method according to an embodiment is shown; Figure 4 A flowchart of the steps of a diagnosis linkage mapping method according to an embodiment is shown; Figure 5 A flowchart showing the steps of a diagnostic weight evaluation method according to an embodiment is shown; The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0028] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0029] In addition, the drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The functional entities can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0030] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0031] This product is an AI medical reasoning engine with artificial intelligence and medical knowledge graph as the core. It has the ability to analyze medical big texts, clinical diagnosis and treatment reasoning, and efficient computing capabilities. It can accurately and quickly extract key medical descriptions from medical records to determine whether the diagnosis and surgical coding are accurate and whether the charges are reasonable. Through artificial intelligence + medical knowledge graph + scenario rules, the rationality analysis of medical insurance fund use in different scenarios is completed, and artificial intelligence scanning of diagnosis and treatment process data is fully realized, which can deeply empower medical insurance fund supervision and hospital intelligent audit business.

[0032] Product features include: Knowledge graph management covers multi-dimensional data such as symptoms, causes, and drug information, making it easier for medical staff to accurately query and apply knowledge. It has the function of knowledge association, which can connect different knowledge points with each other according to medical logic, such as the association between diseases and corresponding diagnostic standards and therapeutic drugs, to assist intelligent diagnostic reasoning.

[0033] Smart inspection rule management allows professionals to configure and view smart inspection rules. The built-in rule template library covers the main diagnosis judgment rules, surgical judgment rules, etc. It is convenient for the site to quickly understand the judgment logic of the rules and reduce the learning cost. At the same time, it can also support the configuration of rule parameters on site to meet the personalized needs of the site.

[0034] Zhijian execution statistics, comprehensive statistics and Yuzhiyi interface call status, including execution times, interface response time, and the most frequently triggered rules.

[0035] Service log tracking records every operation log of Yuzhi Medical in detail, covering key behaviors such as user login, data query, and intelligent inspection task startup, to ensure that operations are traceable. For abnormal situations, such as system errors, detection interruptions, etc., accurately locate the fault point, record the error details and occurrence time, and assist technical personnel to quickly troubleshoot and repair. Support log screening and retrieval, and can query by multiple dimensions such as user, time, and operation type, so that operation and maintenance personnel can focus on the root cause of the problem and ensure stable system operation.

[0036] The rule-driven matrix builds a visual rule-driven relationship matrix, graphically presents the linkage, priority and mutual influence between intelligent inspection rules, and assists in rule optimization decision-making. It supports interactive operations of matrix nodes. Clicking a single rule node can view its detailed configuration, associated rules and trigger conditions, which facilitates in-depth analysis of the logic of the rule system. According to clinical feedback and intelligent inspection statistics, the matrix layout and rule weights are dynamically adjusted to continuously improve the accuracy and reliability of the system's intelligent diagnosis.

[0037] In one embodiment, the main idea of ​​judging the rationality of the main diagnosis (rotavirus enteritis) in the above case.

[0038] Step 1: Determine whether the main diagnosis (rotavirus enteritis) is reasonable. First, we need to determine whether the patient has rotavirus enteritis. The diagnostic medical knowledge graph will include the basis for judging rotavirus enteritis, such as a positive rotavirus antigen test. AI scans the full text data of the electronic medical record, performs contextual precise semantic recognition analysis, and finds whether the medical record contains relevant descriptions of "positive rotavirus antigen test", and judges whether the patient has rotavirus enteritis based on this.

[0039] Step 2: Determine whether the main diagnosis (rotavirus enteritis) consumes the most medical resources. Combined with the monitoring model and scenario rules set by artificial intelligence, analyze the detailed cost data of the patient's entire hospitalization process, and classify the costs according to the diagnosis based on the medical knowledge map. If rotavirus enteritis is the diagnosis that consumes the most medical resources, it is directly determined that the main diagnosis (rotavirus enteritis) is a reasonable choice; otherwise, the next step of judgment is required.

[0040] Step 3: Determine whether the purpose of hospitalization is consistent with the main diagnosis (rotavirus enteritis). If the resource consumption of the main diagnosis (rotavirus enteritis) is not the highest, but is close to that of the diagnosis with the highest medical resource consumption, further analysis will be conducted in combination with the patient's chief complaint. The chief complaint is: chest tightness, wheezing, and cough for 3 months, which is quite different from the clinical findings of rotavirus enteritis (rotavirus enteritis patients generally have symptoms such as watery diarrhea, vomiting, and low fever), so it can be determined that the purpose of hospitalization is inconsistent with the main diagnosis (rotavirus enteritis), and the choice of the main diagnosis is unreasonable.

[0041] See also Figure 1 , the present application provides a method for judging the rationality of a primary diagnosis based on artificial intelligence, the method comprising: S1. Obtain diagnosis request data, and generate candidate primary diagnosis for the diagnosis request data using preset medical knowledge graph data and preset intelligent detection rule data to obtain candidate primary diagnosis data; In one embodiment, a diagnosis request data of an outpatient is obtained from the hospital HIS system, including: patient's main complaint (such as "cough for 2 weeks, with low fever"), examination results (such as blood routine, chest CT, etc.). The above data is input into the natural language processing module, key medical entities (such as symptoms, examinations, past history, etc.) are extracted, and entity alignment is performed with the disease nodes in the medical knowledge graph. The knowledge graph contains a structured causal relationship path between disease-symptom-examination, and a number of diseases with high matching degree (such as tuberculosis, bronchitis, etc.) are identified through graph search (for example, based on the path Rank algorithm) as candidate main diagnoses. At the same time, the system scores and sorts the candidate diagnoses according to intelligent detection rules (such as "if the CT results show nodules + chronic cough, it is prompted to exclude tuberculosis") to obtain candidate main diagnosis data.

[0042] S2. Generate diagnostic execution parameters according to the candidate primary diagnostic data to obtain candidate primary diagnostic parameter data; In one embodiment, for the candidate primary diagnosis of "pulmonary tuberculosis", the system automatically matches the standard diagnosis and treatment parameters that need to be performed, such as recommended further examinations (sputum acid-fast staining, tuberculin test), ICD code of the disease (such as A15.0), treatment cycle recommendation (6 months of anti-tuberculosis drug treatment), medical insurance review risk level (medium and high). These parameters are automatically extracted through the knowledge base template generation system or AI model (such as BERT fine-tuned ICD encoder, and semantic editing model with similar functions) to form structured candidate primary diagnosis parameter data.

[0043] S3, obtaining historical diagnosis request data and historical candidate main diagnosis data, and performing diagnosis linkage processing according to the historical diagnosis request data and the historical candidate main diagnosis data to obtain diagnosis linkage network data; In one embodiment, the system retrieves diagnostic requests and candidate main diagnostic records of similar medical records (matched by similarity models such as SimCSE) within the past year from the data platform of local or alliance hospitals. Construct a linkage network diagram between "diagnosis-symptom-parameters", in which the nodes include diseases, symptoms, examinations, diagnosis and treatment pathways, etc. The weight of the edge represents the co-occurrence frequency or temporal relationship. For example, in the past 100 cases, the proportion of "cough + abnormal chest CT" diagnosed as "bronchitis" was 60%, and "tuberculosis" was 30%, so this part of historical linkage information formed the diagnosis linkage network data.

[0044] In one embodiment, the diagnostic linkage network data is constructed based on historical diagnostic request data and historical candidate main diagnostic data. Considering that the process relies on a large amount of historical records and involves complex relationship extraction and graph structure generation operations, in order to improve the system response efficiency, the construction process of the diagnostic linkage network is designed as a periodic offline preprocessing task, which can be run regularly on a daily, weekly or monthly basis, and the constructed network structure is stored in the graph database for call. In the actual diagnostic process, the system does not generate a diagnostic linkage graph in real time for the current diagnostic request received, but directly queries the existing linkage network structure, and locates the subgraph area and its adjacent structure in the graph according to the candidate main diagnosis, which is used for subsequent node interaction modeling and weight evaluation; During the offline phase, the system periodically retrieves historical diagnosis request data and corresponding candidate primary diagnosis records from local or alliance medical institution platforms, and screens similar instances based on semantic similarity models (such as SimCSE); for these historical data, the system extracts co-occurrence relationships and temporal conversion trends between entities such as "diagnosis-symptoms-examination-treatment pathway", and constructs a linkage graph structure; the nodes in the graph include medical entities such as diseases, symptoms, examination results, and treatment pathways, the edges represent semantic or temporal relationships, and the edge weights represent co-occurrence frequencies, conversion probabilities, or clinical weights; the constructed diagnostic linkage network is stored in the form of a graph database.

[0045] The system only selects diagnostic request records and diagnostic result data within a preset time range (such as the past three years) to build the basis of the map, avoiding the introduction of outdated information caused by factors such as clinical pathway updates and diagnostic standard adjustments. This time range can be flexibly adjusted according to the actual deployment scenario (such as the frequency of hospital database updates). Perform field integrity checks on diagnostic records in historical data and remove abnormal records that are missing key fields such as chief complaints, test results, diagnosis information, and diagnostic codes. It is preferred to retain only samples with a high degree of structure that can be used for entity recognition and standardization. The system performs coding legitimacy verification on diagnostic and operation fields, giving priority to retaining diagnostic items that comply with national standard terminology (such as ICD-10, SNOMED CT) or the hospital's standard coding system, and removing spelling errors, unknown categories, or manually supplemented abnormal data. For diagnostic nodes that appear very rarely but are not rare disease labels, they are checked in combination with the corresponding treatment departments and diagnosis processes. If it is found that they are likely to be caused by diagnostic errors or unreasonable pathways, they can be marked as "low-confidence samples" or removed. Basic statistical indicators (such as gender, age group, department source, etc.) of cases are introduced to evaluate the balance of sample distribution and ensure that the coverage of different diseases and their characteristic nodes in the diagnostic linkage diagram is not affected by the extreme bias of a certain group.

[0046] S4. Perform node interaction processing on the candidate main diagnosis data according to the diagnosis linkage network data to obtain candidate main diagnosis weight data.

[0047] In one embodiment, a graph neural network (such as GraphSAGE, GAT, or a graph neural network with similar functions) is used to aggregate and update the candidate main diagnosis nodes in the diagnosis linkage network. The characteristics of the node include the historical diagnosis rate of the candidate diagnosis, the matching degree with the symptoms in the current request, the historical medical insurance return rate, etc. The graph neural network calculates the weight value of each candidate diagnosis node to form the candidate main diagnosis weight data. For example: tuberculosis 0.62, bronchitis 0.35, and the remaining diagnoses are less than 0.1. The system will recommend "tuberculosis" as the main diagnosis of the current diagnosis request, and sort the remaining main diagnosis weights for visual display.

[0048] Optionally, the step of constructing the preset medical knowledge graph data includes the following steps: Acquire structured electronic medical records, medical standard coding data, and unstructured medical literature data to construct a multi-source medical data set; In one embodiment, the structured electronic medical record data refers to electronic medical visit data with a standard field structure collected through medical information platforms such as hospital information systems (HIS), laboratory information systems (LIS), and picture archiving and communication systems (PACS). The structured electronic medical record data includes but is not limited to the following fields: patient_id: coded information used to uniquely identify the patient's identity; visit_date: date-type data indicating the patient's visit time; chief_complaint: patient's chief complaint information, used to describe the main symptoms of the visit; diagnosis: the doctor's preliminary or specific diagnosis information of the patient's current disease; prescription: a list of drug prescriptions issued for this visit; check_result: examination or test result data related to this visit, including laboratory test indicators, imaging descriptions, etc.

[0049] The medical standard coding data refers to the medical information coding system used to standardize the identification of medical entities and their relationships, including but not limited to the following categories: ICD-10 is used to represent the classification standard of diseases and related health problems, and is widely used in clinical diagnosis and medical insurance settlement scenarios; SNOMED-CT is used to describe clinical terms such as diseases, symptoms, signs, procedures, equipment, etc., and has strong semantic expression capabilities; LOINC is mainly used for standardized coding of laboratory test items and clinical observation items. The above coding system is used to give unified structured coding to medical entities such as diagnosis names, examination indicators, surgical operations, and medication behaviors, which facilitates entity alignment and semantic fusion in the process of graph modeling.

[0050] The unstructured medical literature data refers to literature resources in the medical field that have not been structured and exist in the form of natural language. Its sources include but are not limited to public papers in medical literature databases, such as PubMed, Embase, etc.; various authoritative medical guidelines, clinical pathway specification documents; Chinese core medical journals, specialist papers, disease knowledge base texts, etc. This type of data is captured and downloaded through web crawler tools or open interfaces, and the plain text corpus of medical knowledge content is extracted through subsequent steps such as PDF document conversion, HTML structure analysis, and text cleaning, providing original semantic materials for entity recognition and relationship extraction.

[0051] Perform entity recognition and entity relationship extraction based on multi-source medical data sets to obtain medical entity data and initial association relationship data; In one embodiment, in the medical entity recognition process, predefined entity categories include, but are not limited to, diseases, symptoms, tests, drug names, surgical procedures, human organs or anatomical parts, etc. This classification system is used to guide the subsequent entity labeling and model training process.

[0052] In the entity recognition stage, a named entity recognition model based on deep learning is used to model medical texts. The model can use a bidirectional long short-term memory network combined with a conditional random field (BiLSTM-CRF) structure, or a pre-trained language model BERT combined with a CRF output layer structure, or a semantic editing model with similar functions. The training corpus uses the Chinese clinical medical entity annotation corpus (such as the CMeEE dataset), which covers more than six categories of medical entity labels and has been manually verified to ensure semantic accuracy and label consistency.

[0053] Taking "The patient had fever and cough, CT showed lung shadows, and ceftriaxone was given for treatment" as the original text sample, the entity recognition model can extract the following medical entities: "fever" and "cough" are identified as symptom entities (Symptom); "lung shadows" are identified as check finding entities (Check Finding); and "cefotaxime" is identified as a drug entity (Drug).

[0054] On the basis of completing entity recognition, a relation extraction model is further used to identify the semantic relationship between entity pairs. The model can adopt a pre-trained structure based on sentence pair modeling (such as BERT-PAIR or a semantic editing model with similar functions), or a model based on a convolutional neural network (CNN) combined with an attention mechanism. The output form of relation extraction is a triple structure, namely (entity 1, relation type, entity 2), such as ("lung shadow", representation, "pneumonia"), ("cough", symptoms, "pneumonia"), ("cefotaxime", treatment, "pneumonia").

[0055] Construct semantic relationship edges between entities based on medical entity data and initial association relationship data to form medical relationship network data; In one embodiment, the medical knowledge graph is a graph data structure with structured semantic expression capabilities, wherein each graph node represents a standardized medical entity, which may be a disease, symptom, drug, examination item, examination result or surgery, etc.; each graph edge represents a semantic association relationship between entities, and the relationships include but are not limited to the following initial relationship types: symptom-manifestation-disease: indicating that a symptom is a typical manifestation of a disease; drug-treatment-disease: indicating that a drug is suitable for the treatment of a disease; examination result-support-disease: indicating that a certain examination finding can be used to support the diagnosis of a disease.

[0056] Each edge not only contains the relationship type, but also comes with multiple edge attribute information, including co-occurrence frequency, which indicates the frequency of the relationship appearing in historical medical records or documents; text source identifier, which records the original text number or data source of the source of the relationship; semantic similarity score, which calculates the contextual semantic similarity between entities based on a pre-trained language model (such as BERT or a semantic editing model with similar functions), and the value range is 0 to 1.

[0057] To quantify the importance of edges, a numerical weight is assigned to each edge during the composition process. The weight value is calculated according to the following formula: .in, and is the weighting coefficient, set to =0.6, =0.4; the co-occurrence frequency is normalized by Min-Max, and the semantic similarity is calculated by the BERT model or a semantic editing model with similar functions to calculate the sentence vector cosine value. The constructed knowledge graph is stored and managed in the form of a graph database, preferably using a graph database system that supports attribute graph modeling and efficient graph query, such as Neo4j or OrientDB.

[0058] According to the preset medical knowledge base, the medical entity data is processed with semantic relationship enhancement to obtain the medical entity causal relationship, medical entity adaptation relationship and medical entity concurrent association relationship data; In one embodiment, in order to avoid redundancy or semantic conflict with the initial entity relationship extracted by the natural language processing method mentioned above, when the system performs semantic enhancement of the knowledge base, it gives priority to the entity relationship already in the knowledge graph as the main structure, and introduces enhanced edges only in the following two situations: 1. When the relationship type does not exist in the graph, that is, when there is no semantic edge such as causality, adaptation, or concurrency between two entities, if there is a clearly defined relationship in the knowledge base, a new supplement is added; 2. The confidence of the existing relationship is insufficient. For the low-confidence edge automatically extracted from the graph (such as a weak expression from unstructured text), if there is a strong semantic relationship of the same entity pair in the knowledge base, the knowledge base edge is used to replace the original edge, or the edge weight is increased. Before performing the enhancement, the system will deduplicate and check the consistency of the candidate edges to ensure that only a single semantic edge is retained for the same entity pair to avoid relationship conflicts or reasoning ambiguity. Therefore, the role of the preset knowledge base is to supplement, correct and enhance the automatically extracted relationship.

[0059] The medical knowledge base used for enhanced processing includes but is not limited to the following data resources, such as national standard medical terminology libraries such as SNOMED CT and ICD coding systems; official clinical pathway guidelines include clinical pathway documents for various diseases issued by the National Health Commission; drug-related knowledge bases such as the National Pharmacopoeia, drug instructions, and adverse drug reaction databases; medical insurance rule bases include medical insurance restricted drug use and reasonable diagnosis and treatment judgment rules issued by local medical insurance bureaus; drug indication instructions are derived from drug use guidelines filed with or publicly released by drug regulatory authorities, which are used to identify the indication relationship between drugs and diseases.

[0060] The above-mentioned knowledge base may contain structured (such as coding tables, databases) and semi-structured data (such as PDF format guidelines, HTML web page format rule descriptions). In the actual construction process, the following parsing methods are used to extract semantic relationships for enhancing the graph, map the entries in the knowledge base to the medical entity nodes in the graph, and use name exact matching, medical term synonym expansion, ICD coding alignment, and semantic similarity calculation based on word vector models (such as BERT or semantic editing models with similar functions) to perform entity alignment and standardization. For structured knowledge tables, field relationships are directly identified, such as the "disease→recommended drug" field is parsed as the "drug-indication-disease" relationship; for semi-structured text data, rule templates or natural language-based relationship extraction models are used to identify semantic sentence patterns between entity pairs, such as "suitable for treatment...", "commonly complicated by...", "mainly caused by..." and other expression patterns. By parsing the structured or semi-structured information in the above knowledge base, semantically enhanced relationships are extracted to supplement the graph, mainly including causal relationships, indicating that a medical entity is the cause or inducing factor of another entity, such as "inhaled pathogens" → "pneumonia"; adaptive relationships, indicating that a certain treatment method is suitable for a specific disease, such as "cefotaxime" → "bacterial pneumonia"; concurrent relationships, indicating that two diseases have a high probability of co-morbidity in clinical practice, such as "hypertension" associated with "diabetes".

[0061] After the knowledge graph is constructed, the following semantic enhancement strategy is further introduced: for entity pairs with clear relationship definitions in the knowledge base, a new graph edge is automatically added as an "enhanced edge"; the enhanced edge has an independent relationship type identifier and is given a higher trust weight, preferably with a weight value of 0.3 to increase its priority in graph reasoning; before edge construction, the edit distance (Levenshtein distance) is used to Distance), Jaccard similarity, etc., if the similarity is greater than a threshold (such as 0.85), they are directly matched as the same node, or the knowledge base entity and the graph node name are encoded as vector representations (semantic editing models with similar functions such as BERT, SimCSE, ESimCSE, etc. can be used), and the cosine similarity of the two entity vectors is calculated. If the semantic similarity is greater than 0.9, they are mapped to the same graph node; if there is a unique high-confidence match (such as the highest semantic similarity and exceeding the threshold), the knowledge base entity is bound to the node in the graph; if there are multiple candidate nodes, the previous item with the highest similarity is retained and marked as "partial match"; if there is no match, the original entity is retained for manual review or heterogeneous graph expansion, so that there is a one-to-one correspondence between the knowledge base entity and the graph node; if there are entity aliases, abbreviations or language differences, the matching results with word vector similarity greater than the set threshold (such as 0.85) are preferentially used as the basis for alignment.

[0062] According to the causal relationship, adaptive relationship and concurrent association relationship of medical entities, the edge direction of medical relationship network data is calibrated to obtain medical knowledge graph data.

[0063] In one embodiment, in order to ensure the logical consistency and reasoning validity of the medical knowledge graph relationship expression, the directionality and type of the graph edge are standardized and defined and coded. The directionality of the graph edge is set according to the causal logic or co-occurrence characteristics of different semantic relationship types. For example, a causal edge is used to indicate that a medical entity is a pathogenic factor or antecedent event of another entity, and the direction of the edge is from the cause entity to the result entity, for example: "bacterial infection" → "pneumonia"; an adaptation edge is used to indicate that a certain treatment method or drug is suitable for the treatment of a specific disease, and the direction of the edge is from the treatment plan entity to the target disease entity, for example: "cefotaxime" → "pneumonia"; a concurrent edge is used to indicate that two diseases or medical conditions have a high probability of co-morbidity or simultaneous occurrence in clinical practice. This type of edge is a bidirectional edge without causal order, for example: "hypertension" "coronary heart disease". In order to achieve the manageability and computer recognizability of the graph structure, a unique identifier code is configured for each type of semantic relationship: causal relationship: encoded as R1; adaptation relationship: encoded as R2; concurrency relationship: encoded as R3; the code can be used as the attribute field of the edge in the graph database, and used for relationship screening, graph path calculation or type identification in the graph neural network training process.

[0064] Optionally, the preset intelligent detection rule data includes main diagnosis judgment rules, surgical judgment rules and complication handling rules. The main diagnosis judgment rules, surgical judgment rules and complication handling rules include rule identifiers, rule types, trigger conditions, parameter configurations and priority information. The trigger conditions include a combination of inspection indicators, chief complaint information and medical record elements, which are used to realize the linkage judgment and intelligent triggering of diagnosis request data and rules.

[0065] In one embodiment, the preset intelligent detection rule data includes main diagnosis judgment rules, surgical judgment rules and complication handling rules. The three types of rules all contain rule identifiers, rule types, trigger conditions, parameter configurations and priority information; the main diagnosis judgment rules are used to judge candidate main diagnoses based on fields such as chief complaint information, examination indicators and medical imaging results, and the trigger conditions include multi-field combination matching, and after triggering, the candidate diagnosis name and confidence score are output; the surgical judgment rules are used to recommend surgical plans based on image features and laboratory indicators in specific diagnostic scenarios, and output execution parameters such as surgical coding and anesthesia methods; the complication handling rules are used to intelligently prompt the risk of concurrent diseases when there is a high-risk main diagnosis or abnormal indicator combination, and recommend corresponding screening examinations; the trigger conditions of the rules are combined and matched using Boolean logic expressions. After receiving the diagnosis request data, the system automatically traverses the rule base, matches and intelligently triggers in sequence based on the rule priority, and realizes the logical linkage and decision interpretation of the diagnostic reasoning process. For example, if the chief complaint includes "cough", the WBC index in the routine blood test is greater than 10, and the chest CT shows lung shadows, the main diagnosis judgment rule "D001" is triggered, and the candidate diagnosis "bacterial pneumonia" is output, and the confidence score is 0.85. The parameter configuration fields include information such as ICD code (such as J15.9), recommended drugs (such as ceftriaxone), priority (such as P1), etc. The system uses a rule engine to execute rule traversal and conditional judgment, supports parallel matching of multiple rules, and outputs candidate results according to the rule priority for diagnostic reasoning and recommendation process calls.

[0066] Optionally, see Figure 2 , S1 includes: S11, obtaining diagnosis request data, and performing entity matching on the diagnosis request data using preset medical knowledge graph data to obtain entity matching data; In one embodiment, for text fields in the diagnosis request data, such as the description of the chief complaint, the text of the imaging examination report, etc., a pre-trained language model trained on medical field corpus (such as Chinese-Clinical-BERT or a semantic editing model with similar functions) is used for named entity recognition to extract medical entities such as diseases, symptoms, examination items, and drugs. The model can automatically identify non-standard terms in clinical expressions and output structured entity labeling results. For the identified text entities, the system performs entity standardization by comparing medical terminology dictionaries (such as ICD term phrase dictionaries and SNOMED-CT term libraries) to achieve semantic alignment of multi-source texts in the medical atlas. For example, non-standard expressions such as "shadows in the lungs" and "flaky density shadows found on CT" are uniformly mapped to the "lung shadow" entity node. For structured numerical fields in the diagnosis request data (such as blood routine test results, imaging conclusion parameters, etc.), the system uses a rule matching mechanism for classification. For example, if the WBC value is greater than 10.0, it can be matched to the "elevated white blood cell count" entity in the atlas according to the preset rules; if the imaging report contains the keyword "flaky shadows", it will be matched to the "lung shadows" entity.

[0067] S12, searching the medical knowledge graph data for an initial diagnosis-related candidate set according to the entity matching data to obtain initial diagnosis-related candidate set data; In one embodiment, in order to retrieve the candidate diseases most relevant to the current diagnosis request from the knowledge graph, the system uses the "disease" node as the target node, and based on the semantic association relationship between the entities in the graph that match the current diagnosis request, performs a reverse graph retrieval operation to construct an initial set of diagnosis candidates. Specifically, the system first searches the knowledge graph for disease nodes that have semantic relationships such as "manifested as" and "supported" with the structured medical entities (such as "cough", "fever", "lung shadows", "elevated white blood cells", etc.) extracted from the current diagnosis request data. Each disease node is considered an alternative to the candidate main diagnosis, and a relevance score is assigned based on the hits of the associated entities. The scoring process uses a weighted co-occurrence scoring model, which is denoted as: ,in For the matching scores of candidate disease nodes, different weight coefficients are set for different types of entities to reflect their importance in diagnostic reasoning. For example, the weight of symptom entities is set to 1.0, the weight of imaging entities is set to 1.5, and the weight of laboratory indicator entities is set to 1.2. The system sorts the matching scores of each candidate disease node and selects the top N high-scoring nodes (such as Top10) as the initial diagnosis-related candidate set corresponding to the current diagnosis request.

[0068] S13. Use the preset intelligent detection rule data to match the primary diagnosis candidate items to the initial diagnosis related candidate set data to obtain candidate primary diagnosis data.

[0069] In one embodiment, based on preset intelligent detection rule data, a trigger condition matching judgment is performed on each candidate diagnosis in the initial diagnosis-related candidate set, and the trigger condition includes the combination relationship of the examination indicators, the chief complaint information and the graph matching entity in the diagnosis request; if the rule is triggered, the candidate diagnosis is identified as a candidate main diagnosis to form candidate main diagnosis data.

[0070] Optionally, see Figure 3 , S2 includes: S21, extracting candidate diagnosis context from the diagnosis request data according to the candidate main diagnosis data to obtain candidate diagnosis context data; In one embodiment, according to the candidate main diagnosis data, context information semantically related to each candidate diagnosis is extracted from the diagnosis request data, and the context information includes symptom description, laboratory indicators, imaging features, etc., and the candidate diagnosis context data is obtained by reverse matching of the knowledge graph and semantic association analysis. Based on the established entity relationship path in the medical knowledge graph, the system takes the candidate diagnosis as the central node, reversely retrieves its adjacent typical symptoms, examination indicators and other supporting entities, and determines whether these entities exist in the current diagnosis request data. If there is a match, it is identified as the context feature of the candidate diagnosis. For example, when the candidate diagnosis is "bacterial pneumonia", its adjacent nodes in the graph include "fever", "cough", "lung shadow", "elevated white blood cells", etc.; if the diagnosis request contains the above information, it can be extracted as the context data of "bacterial pneumonia". For unstructured text information in the diagnosis request (such as the description of the chief complaint, imaging report, etc.), the system uses a natural language processing module based on a pre-trained language model to extract keyword vectors and perform similarity calculations with the semantic vectors of the candidate diagnosis entities. If the similarity is higher than a set threshold (e.g., 0.75), the field corresponding to the keyword is extracted as a contextual supplementary feature of the candidate diagnosis. This method can be used to identify information matching of implicit expressions or synonymous expressions.

[0071] S22, initializing rule parameter mapping of candidate primary diagnosis data and candidate diagnosis context data according to preset intelligent detection rule data to obtain rule parameter mapping data; In one embodiment, the preset intelligent detection rule data is called to perform rule matching on the candidate main diagnosis data and its context data. If the diagnosis context meets the triggering conditions of the corresponding rules, the standardized diagnostic parameters corresponding to the rules are extracted, including ICD codes, medical insurance risk levels, recommended treatment pathway IDs, recommended drugs, etc., to form rule parameter mapping data.

[0072] In one embodiment, the function of the rule parameter mapping is to correspond the diagnosis-related information to multiple standardized fields, including but not limited to ICD codes (for disease identification and classification), medical insurance review classification (for rationality assessment), preoperative examination recommendations (for preoperative safety assessment), recommended treatment pathway ID (for clinical pathway execution docking), etc. The system uses a matching mechanism to call and trigger parameter rules. For example, for rule retrieval condition matching, each rule sets an applicable diagnosis name and a trigger condition combination. The system performs matching judgment based on the current candidate main diagnosis name and its extracted context keywords, examination index results, etc. The trigger condition supports Boolean logic combination expression, such as: "cough + WBC>10 and there is a 'shadow' in the image". For example, for the candidate diagnosis "bacterial pneumonia", if the system detects that the laboratory indicator WBC value is greater than 10, and there is a description of "lung shadow" in the imaging report, the following parameter rules can be triggered: Rule_ID: P001, ICD10 code: J15.9, medical insurance risk level: medium risk, recommended pathway ID: PATH_03, treatment recommendation: ceftriaxone is the first choice of antibiotics. If a candidate diagnosis matches multiple parameter rules at the same time, the system will select the rule with the highest priority as the basis for parameter output based on the rule priority field; if some parameters in the rule are missing or the context information does not meet the complete trigger conditions, the parameter item will be marked as "pending manual confirmation" and its context source will be retained for manual review or system secondary reasoning.

[0073] S23. Construct a candidate diagnosis parameter structure according to the rule parameter mapping data to obtain candidate main diagnosis parameter data.

[0074] In one embodiment, structured parameter information of candidate primary diagnoses is constructed based on the rule parameter mapping data, and the parameter structure includes diagnosis name, coding information, confidence score, clinical pathway ID, recommended drug information and whether a review field is required to generate candidate primary diagnosis parameter data.

[0075] Optionally, see Figure 4 , S3 includes: S31, obtaining historical diagnosis request data and historical candidate primary diagnosis data; In one embodiment, historical diagnosis request data and corresponding historical candidate primary diagnosis data are obtained, the data coming from historical medical records of medical institutions and output records of candidate diagnosis systems, and undergoing integrity screening and structured cleaning processing.

[0076] S32, extracting diagnostic instance associations from historical diagnostic request data and historical candidate primary diagnostic data to obtain diagnostic instance association data; In one embodiment, in order to construct a diagnostic linkage structure based on historical data, the system performs instance-level association analysis on historical diagnostic request records and their corresponding candidate main diagnostic information. Specifically, the system determines whether there are semantic commonalities, similar diagnostic structures, or disease evolution trends between two historical diagnostic records to identify pairs of diagnostic instances with significant linkage relationships. The diagnostic instance association judgment is based on the following three strategies: symptom or feature semantic similarity analysis, the system vectorizes the text fields such as the chief complaint, symptom description, and examination information in each diagnostic request data, and uses sentence vector modeling methods (such as SimCSE, BERT-CLS, or semantic editing models with similar functions) to generate corresponding semantic vectors, and calculates the semantic similarity between historical instances. When the similarity is higher than a set threshold (such as 0.85), the two records are considered to have high semantic consistency in subjective symptoms or objective descriptions and are marked as "semantically similar". To determine the overlap of candidate diagnoses, the system compares the overlap of the candidate main diagnosis sets in the two historical instances. If the ratio of the number of intersections to the number of unions is greater than or equal to 0.5 (i.e., the overlap of candidate diagnoses reaches 50%), the two are considered to have candidate neighbor characteristics at the structural level and are marked as "diagnostic candidate similarity". To evaluate the similarity of evolutionary paths after diagnosis, for cases that have been diagnosed and have follow-up records, the system determines whether there is consistency in the complications that occur within 7 days after diagnosis or the clinical pathways they enter. If two cases show similar complications or receive the same treatment pathway in a short period of time, they are considered to have a "clinical evolution similarity" relationship. The results of the above association analysis are output in a structured form. Each historical diagnosis record is associated with a set of similar diagnosis instance identifiers (Instance_ID), and is accompanied by similarity type labels (such as semantic similarity, diagnosis candidate similarity, evolutionary path similarity), corresponding similarity scores, and time difference information (such as within 3 days, 7 days, etc.).

[0077] S33, performing linkage relationship mining according to the diagnosis instance association data to obtain linkage relationship data; In one embodiment, the system performs linkage relationship identification and modeling operations based on historical instance association data and knowledge graph structure. The linkage relationship refers to the co-occurrence, causal, evolutionary or structural relationship between the candidate diagnosis and other diagnoses, symptoms or examination features. Specifically, the mining of linkage relationships includes the following three methods: co-occurrence relationship statistical analysis, the system traverses all historical instances and their identified context entities, counts the frequency of candidate diagnosis A and other entities B (which can be diseases, symptoms or examination indicators) appearing in the same instance, and calculates their co-occurrence rate. If the co-occurrence frequency exceeds 10%, it is considered that there is a weak linkage relationship between the two; if it exceeds 30%, it is defined as a strong linkage relationship, which is used to construct a high-confidence co-occurrence edge. Time series transformation relationship identification, the system performs sequence analysis on historical cases with time tags to determine whether diagnosis A significantly precedes diagnosis or symptom B in time. If there is a stable sequence, record its average transformation time interval and direction (such as A→B), and construct evolutionary relationship edges based on this, indicating the disease development path or clinical status change trend. The graph path relationship is strengthened. Combined with the structure of the medical knowledge graph, the system searches whether the candidate diagnosis A and the target entity B have a direct path connection or share the same upper-level semantic node (i.e., a common parent node). If there is a co-occurrence path, causal chain, or co-disease spectrum relationship between A and B, the credibility of their linkage is improved to enhance the structural consistency and reasoning reliability in the graph. The linkage relationship is represented in the form of a structured triple, with the format of: <Entity A, linkage type, entity B>, where the linkage type includes type identifiers such as "co-occurrence", "transformation", "evolution", and "graph structure"; each relationship edge is accompanied by multiple attribute fields, including linkage weight value, number of supported instances, whether it is a temporal relationship, average time interval, etc.

[0078] S34. Generate diagnosis linkage network data according to the diagnosis instance association data and linkage relationship data to obtain diagnosis linkage network data.

[0079] In one embodiment, based on the diagnosis instance association data and linkage relationship data, diagnosis linkage network data including multiple relationship types, multilateral weight features and time directionality is constructed with candidate diagnoses, symptoms, examination items, etc. as nodes.

[0080] Optionally, see Figure 5 , S4 includes: S41, constructing a candidate main diagnosis subgraph for the diagnosis linkage network data according to the candidate main diagnosis data to obtain candidate main diagnosis subgraph data; In one embodiment, the system takes each candidate main diagnosis node as the center, extracts a subgraph within its K-hop range from the diagnosis linkage graph, and preferably sets it to a 2-hop adjacency range to cover the direct features of the diagnosis (1 hop) and its context-related nodes (2 hops), forming a diagnosis context subgraph with a complete local structure. The nodes in the subgraph include, but are not limited to, symptom entities that are directly related to the candidate diagnosis (such as "cough", "fever"); examination or test indicators (such as "elevated white blood cell count", "lung shadows"); concurrent diseases or pre- and post-evolution diagnoses (such as "lung abscess", "chronic bronchitis"), etc. The subgraph edge retains the relationship type and its weight information in the original graph, and the relationship types include "manifested as", "supported", "concurrent with", "co-occurred with", "evolved into", etc. When there are multiple paths between a candidate diagnosis and a certain associated node, the system prioritizes the path with the largest edge weight to ensure that only the most representative semantic path is retained in the subgraph structure; to avoid feature overfitting due to an overly large graph structure, an upper limit N is set for the number of subgraph nodes, preferably within 30 nodes; for edges with low edge weight or insufficient semantic confidence, a threshold filtering strategy (such as setting a threshold of 0.85) can be used to remove them. Through the above subgraph construction method, the system generates a contextual semantic subgraph with a complete structure and clear relationships for each candidate main diagnosis.

[0081] S42, extracting node linkage features according to the candidate main diagnosis subgraph data and the candidate main diagnosis data to obtain node linkage feature data; In one embodiment, for each node in the candidate diagnosis subgraph, the system extracts multiple types of features of the following dimensions, such as the structural feature dimension including node degree, which indicates the number of edges directly connected to the current node in the graph, and is used to measure the connection strength of the entity in the local structure; hop distance, which indicates the number of shortest path hops between the node and the candidate diagnosis center node (such as 1 hop, 2 hops), reflecting the closeness of its association; frequency of occurrence, which indicates the number of times or frequency of the node appearing in the historical instance graph, and is used to measure its global representativeness or importance. The semantic feature dimension includes semantic similarity, which is the cosine similarity between the node word vector generated based on the pre-trained language model (such as BERT or a semantic editing model with similar functions) and the current diagnosis request main complaint vector, and is used to measure the degree of fit between the entity expression and the main complaint semantics; node type, which uses a one-hot method to indicate the category to which the node belongs, such as symptoms, examination items, diagnostic entities, etc.; edge type and edge weight features, which indicate the connection relationship type between the current node and the candidate diagnosis node (such as "manifested as", "co-occurs in") and the weight value of the edge, and are used to convey the relationship semantics and structural importance. The context matching feature dimension includes the context hit status, a Boolean value field that indicates whether the medical entity corresponding to the node exists in the current diagnosis request data (for example, if the chief complaint contains "fever", the value of the "fever" node is True), which is used to reflect the association between the entity and the current request context.

[0082] S43, interactively scoring the candidate primary diagnosis data and the node linkage feature data to obtain preliminary scoring data; In one embodiment, the method calculates the support degree of each adjacent node in the candidate diagnosis subgraph for the main diagnosis based on the predefined feature importance rule. The specific calculation formula is as follows: (node ​​edge weight × similarity between node and main complaint × context hit flag), where It is the preliminary scoring data. The edge weight indicates the strength of the structural relationship between the diagnosis node and the adjacent nodes. The semantic similarity is calculated based on the cosine value between the entity word vector and the main complaint semantic vector. The context hit flag is a Boolean value indicating whether the node is mentioned in the current diagnosis request. The contribution of all supporting nodes is accumulated by weighted summation to form the context confidence score of the candidate diagnosis.

[0083] In one embodiment, the system may use a shallow graph neural network model (such as GCN, GraphSAGE or a graph neural network with similar functions) to aggregate node features of the diagnostic subgraph structure. In the specific operation, the candidate diagnostic node is used as the target node, and the embedded representation of its adjacent nodes in the structural and semantic feature space is received. The multi-order neighbor feature fusion is completed through the graph convolution operation, and the representation vector of the diagnostic node is output. A score value is generated through a linear mapping layer as the credibility score of the diagnosis in the current context. For example, a candidate diagnosis scores 0.78 after aggregation in its context graph structure.

[0084] S44. Perform simplified graph attention mechanism processing on the candidate primary diagnosis data and preliminary scoring data to obtain candidate primary diagnosis weight data.

[0085] In one embodiment, for each candidate main diagnosis node, the system traverses all adjacent nodes in its subgraph (such as symptoms, examinations, concurrent diagnoses, etc.), and comprehensively calculates the attention weight based on the following factors: node importance, including the frequency of occurrence of the node in the global graph structure, type (such as typical symptoms vs. secondary symptoms), and structural connectivity with the main diagnosis (such as whether it is a 1-hop adjacency); semantic similarity, the semantic matching between the word vector corresponding to the node and the diagnostic complaint vector, which measures the degree of fit between the entity expression and the current context; structural edge weight, the edge weight between the candidate diagnosis node and the adjacent node in the original graph, reflects the strength of its structural association. After the above factors are standardized and fused, the attention weight of each adjacent node is calculated and used to weight the feature vector of the node. Subsequently, the system aggregates the weighted features of all adjacent nodes into a representation of the candidate diagnosis node and outputs the diagnostic credibility weight value; this mechanism can achieve the reinforcement of key information and the suppression of interference information in the diagnosis subgraph. For example, when typical symptoms such as "fever" and "cough" exist in the subgraph corresponding to the candidate diagnosis "bacterial pneumonia", these nodes will be given higher attention weights due to their high importance and semantic similarity; while for non-specific symptoms such as "mild fatigue", the attention is lower and the contribution to the weight is smaller. The system realizes the fusion modeling of structural information and semantic information through the adaptive attention mechanism, and outputs the credibility score of the candidate diagnosis node as the basis for decision-making.

[0086] Optionally, the interaction score includes: Extracting context semantic features according to the candidate primary diagnosis data to obtain context semantic feature data; In one embodiment, the text of the chief complaint, symptoms, examination fields, etc. is input into a pre-trained language model dedicated to the medical field (such as Chinese-Clinical-BERT, and semantic editing models with similar functions); the sentence vector representation of the entire context is extracted. The typical symptoms and examination items of the candidate primary diagnosis in the knowledge graph are extracted; it is determined whether these elements are hit in the current request, and a set of Boolean flags or numerical enhancement dimensions are set. The context vector schematic dimensions are obtained, such as 128 dimensions, which contain subfields such as text semantic dimensions, hit enhancement dimensions, disease type identifiers, and feature missing identifiers.

[0087] Perform bilinear attention calculation based on node linkage feature data and context semantic feature data to obtain matching weight data; In one embodiment, the bilinear attention mechanism constructs a context when calculating the matching weight between the context and the graph node. The bidirectional interaction representation of node features measures the degree of matching in the multidimensional semantic space through a learnable weight matrix or a static weight function. For example, when the diagnosis request context contains "CT suggests lung shadows", the bilinear interaction score between the "lung shadows" node and the context in the figure is high and is given a higher attention weight; conversely, if a node (such as "headache") does not appear in the current context or has no semantic association, its attention score will be lowered.

[0088] The system extracts the semantic representation of the diagnosis request text and records it as the context vector , and the linkage feature vectors of each node in the candidate diagnosis subgraph , and construct the following bilinear interaction scoring function: , is the raw attention score, context and node The matching score between is the transposed representation of the context vector, the semantic representation of the current request (chief complaint, medical history, etc.), is a bilinear weight matrix used to capture the interaction between the context and the graph nodes (a learnable parameter matrix, that is, automatically updated by gradient backpropagation during model training, or a diagonal matrix of specified dimensional weights), is the node linkage feature data, indicating the adjacent nodes in the subgraph Multi-dimensional semantic and structural features; Softmax calculation is performed on all adjacent node sets of the candidate diagnosis node to obtain the attention weight of each adjacent node , For Node The attention weight, is the raw attention score, is an exponential function, Think of it as node order, Candidate diagnosis node The first-order neighbor set of For Node ; Use attention weights to perform weighted aggregation on adjacent node features to form an enhanced representation of the candidate diagnosis node in the current context: , To match the weight data, is the node order item, Candidate diagnosis node The first-order neighbor set of For Node The attention weight, is the node linkage feature data, indicating the adjacent nodes in the subgraph The multidimensional semantic and structural characteristics of

[0089] The node linkage feature data and context semantic feature data are fused according to the matching weight data to obtain preliminary scoring data.

[0090] In one embodiment, the linkage features of all graph nodes are aggregated according to the attention weights and fused with the contextual semantic vector to form a preliminary credible score for the candidate diagnosis. The feature vectors of all graph nodes are weighted averaged according to the corresponding attention weights to obtain a graph feature aggregation vector. The contextual semantic vector and the graph aggregation vector are concatenated, nonlinearly mapped, or vector inner-producted to generate a scalar score, i.e., matching weight data; this score is the semantic rationality score of the candidate diagnosis in the current request context.

[0091] Optionally, the simplified graph attention mechanism processing includes: Perform adjacent node similarity calculation based on candidate primary diagnosis data and preliminary scoring data to obtain adjacent node edge weight data; In one embodiment, for each candidate primary diagnosis node, the system extracts the multi-dimensional feature representation of all first-order adjacent nodes in the corresponding subgraph (such as symptoms, examinations, concurrent diagnoses, etc.), including but not limited to structural linkage features (such as degree value, hop distance, node type in the graph), contextual relevance features (such as hit status, semantic similarity score), etc., to form a unified node feature vector representation. By calculating the cosine similarity between the candidate diagnosis node feature vector and the adjacent node feature vector, the basic similarity between the two in the structural and semantic space is measured; the context score of the adjacent node in the current diagnosis request context (from the previous module) is introduced as a semantic importance weight enhancement factor; , is the edge weight data of adjacent nodes, is the weight of similarity, is the similarity, is the weight of the preliminary scoring data, For preliminary scoring data, and is an adjustable weighting factor, preferably set to =0.6, =0.4, to ensure that the weight distribution strategy is based on structural similarity and supplemented by semantic signals.

[0092] A simplified attention coefficient is calculated based on the adjacent node edge weight data and the preliminary scoring data to obtain simplified attention coefficient data; In one embodiment, for each candidate primary diagnosis node, all its first-order adjacent edges have been weighted by structural similarity and context semantic score to obtain edge weights. The system performs a Softmax operation on all adjacent edges of the node to generate a simplified attention coefficient, which is calculated as follows: , For Node For candidate diagnosis nodes The attention score, is the original edge weight, is the candidate primary diagnosis node With adjacent nodes The original edge weight between is calculated by structural similarity and context score. is an exponential function, is the adjacent node index, Candidate diagnosis node The set of all adjacent nodes of For Node With The simplified attention mechanism has the following characteristics: no trainable parameters are used to avoid introducing additional model burden; no multi-head mechanism and nonlinear activation are introduced, and only static attention scores based on graph structure and semantic feature distribution are retained; the attention score reflects the relative contribution of adjacent nodes and is used for weighted aggregation processing of node features. Through this simplified attention normalization method, the system can achieve an efficient and controllable feature propagation mechanism while maintaining the integrity of the graph structure.

[0093] Perform multi-order aggregation based on the simplified attention coefficient data to obtain multi-order neighbor aggregation data; In one embodiment, the attention coefficient is used to transfer information between nodes to generate an enhanced representation of the candidate primary diagnosis node, which represents not only the aggregation of first-order neighbors, but also includes multi-order neighbor information (such as second-order neighbors) to increase semantic coverage. The features of each first-order neighbor are weighted and summed using the attention coefficient to form a direct context aggregation vector of the candidate diagnosis, i.e., the first-order aggregation. The above aggregation method is repeated for the neighbors of the first-order neighbors (i.e., the second-order neighbors); information is allowed to affect the primary diagnosis score through indirect paths (such as "cough" → "bronchitis" → "bacterial pneumonia"). For each outward jump, the weight decay coefficient is multiplied by (e.g. 0.5); avoid excessive amplification of the influence of distant neighboring nodes. The multi-order aggregation result is a new diagnostic feature vector that integrates structural + semantic information.

[0094] Residual connection is performed based on multi-order neighbor aggregation data and preliminary scoring data to obtain candidate main diagnosis weight data.

[0095] In one embodiment, for a candidate diagnosis whose score is in scalar form, its score is defined as: ,in is the candidate primary diagnosis weight data, is the fusion weight coefficient, preferably set to 0.6, to maintain the dominance of semantic scoring while introducing a certain proportion of structural enhancement compensation. For preliminary scoring data, Aggregate data for multi-order neighbors. When there are high-confidence adjacent nodes in the graph structure, even if the initial score is insufficient, the structural enhancement part can effectively improve the candidate diagnosis score, reflecting its potential contextual rationality; when the graph structure signal is weak and the aggregation path noise is high, the residual mechanism can retain the initial score result to prevent the uncertainty introduced by the structure from misleading the diagnosis score.

[0096] Optionally, the present application further provides an artificial intelligence-based primary diagnosis rationality judgment system, which is used to execute the artificial intelligence-based primary diagnosis rationality judgment method as described above, and the artificial intelligence-based primary diagnosis rationality judgment system includes: A candidate diagnosis generation module is used to obtain diagnosis request data, and use preset medical knowledge graph data and preset intelligent detection rule data to generate candidate main diagnosis for the diagnosis request data to obtain candidate main diagnosis data; A diagnostic parameter extraction module is used to generate diagnostic execution parameters according to the candidate main diagnostic data to obtain candidate main diagnostic parameter data; The diagnosis linkage mapping module is used to obtain historical diagnosis request data and historical candidate main diagnosis data, and perform diagnosis linkage processing according to the historical diagnosis request data and the historical candidate main diagnosis data to obtain diagnosis linkage network data; The diagnosis weight evaluation module is used to perform node interaction processing on the candidate main diagnosis data according to the diagnosis linkage network data to obtain the candidate main diagnosis weight data.

[0097] Therefore, from any point of view, the embodiments should be regarded as illustrative and non-restrictive, and the scope of the present invention is limited by the attached application documents rather than the above description, and it is intended that all changes falling within the meaning and scope of equivalent elements of the application documents are included in the present invention.

[0098] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A method for judging the rationality of primary diagnosis based on artificial intelligence, characterized in that: The method comprises: S1. Obtain diagnosis request data, and generate candidate primary diagnosis for the diagnosis request data using preset medical knowledge graph data and preset intelligent detection rule data to obtain candidate primary diagnosis data; S2. Generate diagnostic execution parameters according to the candidate primary diagnostic data to obtain candidate primary diagnostic parameter data; S3, obtaining historical diagnosis request data and historical candidate main diagnosis data, and performing diagnosis linkage processing according to the historical diagnosis request data and the historical candidate main diagnosis data to obtain diagnosis linkage network data, wherein the diagnosis linkage processing is a process in which the system extracts diagnosis association information corresponding to diseases, symptoms, examinations, and treatment pathways based on the historical diagnosis request data and the historical candidate main diagnosis data, constructs a diagnosis-symptom-parameter ternary graph through matching with a preset similarity model, and calculates the co-occurrence frequency between each node as the edge weight; S4. Perform node interaction processing on the candidate main diagnosis data according to the diagnosis linkage network data to obtain candidate main diagnosis weight data, and perform a main diagnosis recommendation operation on the diagnosis request data according to the candidate main diagnosis weight data, and output a recommended main diagnosis result.

2. The method according to claim 1, characterized in that The steps of constructing the preset medical knowledge graph data include the following steps: Acquire structured electronic medical records, medical standard coding data, and unstructured medical literature data to construct a multi-source medical data set; Perform entity recognition and entity relationship extraction based on multi-source medical data sets to obtain medical entity data and initial association relationship data; Construct semantic relationship edges between entities based on medical entity data and initial association relationship data to form medical relationship network data; According to the preset medical knowledge base, the medical entity data is processed with semantic relationship enhancement to obtain the medical entity causal relationship, medical entity adaptation relationship and medical entity concurrent association relationship data; According to the causal relationship, adaptive relationship and concurrent association relationship of medical entities, the edge direction of medical relationship network data is calibrated to obtain medical knowledge graph data.

3. The method according to claim 1, characterized in that The preset intelligent detection rule data includes the main diagnosis judgment rules, surgical judgment rules and complication handling rules. The main diagnosis judgment rules, surgical judgment rules and complication handling rules include rule identifiers, rule types, trigger conditions, parameter configurations and priority information. The trigger conditions include a combination of inspection indicators, chief complaint information and medical record elements, which are used to realize the linkage judgment and intelligent triggering of diagnosis request data and rules.

4. The method according to claim 1, characterized in that: S1 includes: Obtaining diagnosis request data, and using preset medical knowledge graph data to perform entity matching on the diagnosis request data to obtain entity matching data; According to the entity matching data, the medical knowledge graph data is retrieved for an initial diagnosis-related candidate set to obtain initial diagnosis-related candidate set data; The preset intelligent detection rule data is used to match the primary diagnosis candidate items of the initial diagnosis-related candidate set data to obtain candidate primary diagnosis data.

5. The method according to claim 1, characterized in that S2 include: Extract candidate diagnosis context from the diagnosis request data according to the candidate main diagnosis data to obtain candidate diagnosis context data; Initialize rule parameter mapping of candidate main diagnosis data and candidate diagnosis context data according to preset intelligent detection rule data to obtain rule parameter mapping data; The candidate diagnosis parameter structure is constructed according to the rule parameter mapping data to obtain candidate main diagnosis parameter data.

6. The method according to claim 1, characterized in that S3 include: Obtain historical diagnosis request data and historical candidate primary diagnosis data; Extracting diagnostic instance associations from historical diagnostic request data and historical candidate primary diagnostic data to obtain diagnostic instance association data; Perform linkage relationship mining based on the diagnosis instance association data to obtain linkage relationship data; The diagnosis linkage network data is generated according to the diagnosis instance association data and the linkage relationship data to obtain the diagnosis linkage network data.

7. The method according to claim 1, characterized in that S4 includes: According to the candidate main diagnosis data, a candidate main diagnosis subgraph is constructed for the diagnosis linkage network data to obtain the candidate main diagnosis subgraph data; Extract node linkage features according to candidate main diagnosis subgraph data and candidate main diagnosis data to obtain node linkage feature data; Interactive scoring is performed based on candidate primary diagnosis data and node linkage feature data to obtain preliminary scoring data; A simplified graph attention mechanism is performed on the candidate main diagnosis data and preliminary scoring data to obtain the candidate main diagnosis weight data.

8. The method according to claim 7, characterized in that The interaction scores include: Extracting context semantic features according to the candidate primary diagnosis data to obtain context semantic feature data; Perform bilinear attention calculation based on node linkage feature data and context semantic feature data to obtain matching weight data; The node linkage feature data and context semantic feature data are fused according to the matching weight data to obtain preliminary scoring data.

9. The method according to claim 7, characterized in that: The simplified graph attention mechanism processing includes: Perform adjacent node similarity calculation based on candidate primary diagnosis data and preliminary scoring data to obtain adjacent node edge weight data; A simplified attention coefficient is calculated based on the adjacent node edge weight data and the preliminary scoring data to obtain simplified attention coefficient data; Perform multi-order aggregation based on the simplified attention coefficient data to obtain multi-order neighbor aggregation data; Residual connection is performed based on multi-order neighbor aggregation data and preliminary scoring data to obtain candidate main diagnosis weight data.

10. A primary diagnosis rationality judgment system based on artificial intelligence, characterized in that: Used to execute the artificial intelligence-based primary diagnosis rationality judgment method as claimed in claim 1, the artificial intelligence-based primary diagnosis rationality judgment system comprises: A candidate diagnosis generation module is used to obtain diagnosis request data, and use preset medical knowledge graph data and preset intelligent detection rule data to generate candidate main diagnosis for the diagnosis request data to obtain candidate main diagnosis data; A diagnostic parameter extraction module is used to generate diagnostic execution parameters according to the candidate main diagnostic data to obtain candidate main diagnostic parameter data; The diagnosis linkage mapping module is used to obtain historical diagnosis request data and historical candidate main diagnosis data, and perform diagnosis linkage processing according to the historical diagnosis request data and the historical candidate main diagnosis data to obtain diagnosis linkage network data; The diagnosis weight evaluation module is used to perform node interaction processing on the candidate main diagnosis data according to the diagnosis linkage network data to obtain the candidate main diagnosis weight data.

Images