Physical examination report intelligent analysis method and system based on multilevel knowledge graph

Abstract

The invention provides a physical examination report intelligent analysis method and system based on a multi-level knowledge graph, and relates to the technical field of medical data analysis. According to the method, multi-hop retrieval is performed on the individual layer through the user identifier, the abnormal information of the physical examination indexes is quickly positioned and extracted, and the problem of low efficiency of manual part-by-part analysis is solved. And based on cross-layer association, pathological mechanisms and disease knowledge of a public layer and statistical association information of a group layer are automatically and synchronously retrieved, general medical knowledge, individual data and group rules are fused, and the comprehensiveness and objectivity of risk identification are remarkably improved. And finally, integrating retrieval results of each layer and a reasoning chain, carrying out comprehensive reasoning by utilizing a large language model, automatically completing the whole process from anomaly recognition and disease attribution to risk prediction, and generating a structured report. Therefore, the accuracy, interpretability and reliability of an analysis result are enhanced while the processing efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of medical data analysis technology, and in particular to an intelligent analysis method and system for physical examination reports based on a multi-level knowledge graph. Background Technology

[0002] In the healthcare field, health check-up reports serve as a core carrier reflecting an individual's health status, and the quality and efficiency of their analysis directly impact the effectiveness of early disease screening and health management. Traditional health check-up report analysis primarily relies on manual interpretation, which suffers from three major pain points: First, it is inefficient, as manual analysis of each report consumes a significant amount of time, making it difficult to meet the rapid processing needs of large-scale health check-up data; second, risk identification is incomplete, as limitations in physicians' knowledge and experience can easily lead to the omission of potential disease associations or abnormal trends in indicators; and third, the results are poorly interpretable, only providing conclusive judgments without clearly presenting the logical connections between abnormal indicators, pathological mechanisms, and disease risks, which is detrimental to user understanding and subsequent treatment decisions. Summary of the Invention

[0003] This invention provides an intelligent analysis method and system for physical examination reports based on a multi-level knowledge graph, which solves the problems of low efficiency, inaccurate results, and poor interpretability in traditional manual interpretation of physical examination reports.

[0004] In a first aspect, the present invention provides an intelligent analysis method for physical examination reports based on a multi-level knowledge graph. The method includes: receiving query information input by a user; the query information includes a user identifier; based on the user identifier, performing multi-hop retrieval at the individual level of the multi-level knowledge graph to obtain abnormal information about the user's physical examination indicators; the multi-level knowledge graph includes an individual level, a public level, and a group level; based on the abnormal information about the physical examination indicators and the cross-level relationships between the layers in the multi-level knowledge graph, retrieving pathological mechanisms and disease knowledge from the public level and statistical association information from the group level, determining the retrieval results for each layer and the association reasoning chains between the layers; based on the retrieval results for each layer, the association reasoning chains between the layers, and a large language reasoning model, performing indicator anomaly identification, disease attribution analysis, and potential risk prediction, and generating a physical examination analysis report.

[0005] Secondly, this invention provides an intelligent analysis device for physical examination reports based on a multi-level knowledge graph. The intelligent analysis device includes: a communication module for receiving query information input by a user; the query information includes a user identifier; and a processing module for performing multi-hop searches in the individual layer of the multi-level knowledge graph based on the user identifier to obtain abnormal information about the user's physical examination indicators; the multi-level knowledge graph includes an individual layer, a public layer, and a group layer; based on the abnormal information about the physical examination indicators and the cross-layer relationships between the layers in the multi-level knowledge graph, the device retrieves pathological mechanisms and disease knowledge in the public layer and statistical association information in the group layer to determine the retrieval results for each layer and the association reasoning chains between the layers; based on the retrieval results for each layer, the association reasoning chains between the layers, and a large language reasoning model, the device performs indicator anomaly identification, disease attribution analysis, and potential risk prediction to generate a physical examination analysis report.

[0006] Thirdly, embodiments of the present invention provide an intelligent analysis system for physical examination reports based on a multi-level knowledge graph. The system includes an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor is used to call and run the computer program stored in the memory to perform the steps of the method as described in the first aspect and any possible implementation thereof.

[0007] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the method as described in the first aspect and any possible implementation thereof.

[0008] This invention provides an intelligent analysis method and system for physical examination reports based on a multi-level knowledge graph. By receiving query information containing user identifiers, the invention first performs multi-hop searches at the individual level of the multi-level knowledge graph to quickly and accurately locate and extract abnormal information from the user's physical examination indicators, effectively overcoming the efficiency bottleneck of manual review of each report. Subsequently, based on the abnormal information, it automatically and synchronously retrieves pathological mechanisms and disease knowledge in the common layer and statistical correlation information in the group layer using cross-layer association relationships. This integrates general medical knowledge, individual-specific data, and group statistical patterns, making risk identification no longer limited to a single knowledge source and significantly improving the comprehensiveness and objectivity of the analysis. Finally, by integrating the retrieval results and association reasoning chains from each layer, it inputs them into a large language reasoning model for comprehensive reasoning, automating the entire process from indicator anomaly identification and disease attribution to potential risk prediction, and generating a structured physical examination analysis report. This invention solves the problems of low efficiency, inaccurate results, and poor interpretability inherent in traditional manual interpretation of physical examination reports, achieving efficient, comprehensive, and logically interpretable intelligent analysis of physical examination reports. This not only significantly improves processing efficiency but also enhances risk detection capabilities and the reliability of conclusions. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of the architecture of an intelligent analysis system for physical examination reports based on multi-level knowledge graphs and LLM reasoning, provided by an embodiment of the present invention. Figure 2 This is a flowchart illustrating an intelligent analysis method for physical examination reports based on a multi-level knowledge graph, provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of a multi-level knowledge graph architecture provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of an intelligent retrieval process provided by an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of an intelligent analysis device for physical examination reports based on a multi-level knowledge graph, provided in an embodiment of the present invention. Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0011] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0012] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner to facilitate understanding.

[0013] Furthermore, the terms "comprising" and "having," and any variations thereof, used in the description of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or modules is not limited to the steps or modules listed, but may optionally include other steps or modules not listed, or may optionally include other steps or modules inherent to such process, method, product, or device.

[0014] To make the objectives, technical solutions, and advantages of the present invention clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0015] As described in the background section, some current health checkup analysis systems attempt to incorporate knowledge graphs or machine learning models, but significant shortcomings remain. On one hand, most systems employ only a single-level knowledge graph (e.g., based solely on the Universal Medical Terminology Database (UMLS), lacking the integration of individual health checkup data with group statistical data, resulting in low relevance between the analysis results and individual circumstances. On the other hand, retrieval strategies are simple, mostly relying on single-hop keyword matching, failing to construct multi-dimensional reasoning chains, and the model output lacks standardized constraints, making it difficult to guarantee consistency and reliability. Furthermore, cross-level data association often uses fixed rule matching, neglecting semantic similarity of terms, easily leading to association errors and further affecting the accuracy of the analysis.

[0016] To address the issues of low efficiency, incomplete risk identification, and poor interpretability in past medical examination report analysis, as well as the problems of existing systems' single-layer knowledge graph, insufficient retrieval and reasoning, and errors in cross-layer associations, such as... Figure 1As shown in the diagram, this invention provides an architecture diagram of an intelligent analysis system for physical examination reports based on multi-level knowledge graphs and LLM reasoning. This invention achieves efficient, comprehensive, and interpretable analysis of physical examination reports by constructing a "public-individual-group" multi-level knowledge graph architecture, optimizing cross-layer links, designing intelligent retrieval, and standardizing LLM reasoning processes, thus providing more accurate intelligent auxiliary support for medical decision-making.

[0017] like Figure 2 As shown, this embodiment of the invention provides an intelligent analysis method for physical examination reports based on a multi-level knowledge graph. The method includes steps S101-S104.

[0018] S101. Receive query information input by the user.

[0019] In this embodiment of the application, the query information includes the user identifier.

[0020] S102. Based on the user identifier, perform multi-hop retrieval in the individual layer of the multi-level knowledge graph to obtain abnormal information of the user's physical examination indicators.

[0021] In this embodiment of the application, the multi-level knowledge graph includes an individual layer, a public layer, and a group layer.

[0022] As one possible implementation, step S102 can be specifically implemented as steps S1021-S1022.

[0023] S1021. Starting with the user node corresponding to the user identifier, perform multi-hop traversal retrieval at the individual level according to the preset path of user, medical record, test type, and test result.

[0024] S1022. During the retrieval process, different types of nodes are prioritized according to a preset weight allocation strategy to prioritize the acquisition of detection result nodes and detection type nodes.

[0025] S1023. Integrate the detection result nodes on the retrieval path, and determine the abnormal information of the user's physical examination indicators based on the comparison between the values ​​and the reference range.

[0026] S103. Based on abnormal information of physical examination indicators and cross-layer associations between layers in the multi-level knowledge graph, retrieve the pathological mechanisms and disease knowledge of the common layer and the statistical association information of the group layer to determine the retrieval results of each layer and the association reasoning chain between each layer.

[0027] As one possible implementation, step S103 can be specifically implemented as steps S1031-S1037.

[0028] S1031. For each abnormal detection result in the abnormal information of physical examination indicators, based on the first type of relationship in the cross-layer association relationship, retrieve the semantically related medical concept nodes in the public layer.

[0029] In some embodiments, the first type of relationship is the association between individual layer nodes and public layer nodes.

[0030] S1032. Starting from the retrieved medical concept nodes, perform multi-hop extended retrieval along the core semantic relationship in the public layer to obtain pathological mechanism nodes and disease nodes related to the retrieved medical concept nodes, which are used as the retrieval results of the public layer.

[0031] S1033. Establish a public inference subchain between abnormal physical examination indicators and related pathological mechanism nodes and disease nodes.

[0032] In some embodiments, the common inference subchain represents the association between abnormal indicators, pathological mechanisms, and diseases.

[0033] S1034. For combinations of abnormal physical examination indicators, based on the second type of relationship in the cross-layer association relationship, retrieve matching statistical indicator nodes in the group layer.

[0034] In some embodiments, the second type of relationship is the association between individual layer nodes and group layer nodes.

[0035] S1035. Starting from the retrieved statistical indicator nodes, search along statistical relationships in the group layer to obtain statistically significant associated disease nodes and population grouping information, which are used as the search results for the group layer.

[0036] S1036. Establish a group inference sub-chain between abnormal physical examination indicators and related disease nodes and population grouping information.

[0037] In some embodiments, the group inference subchain represents the association between indicators of abnormality, associated diseases, and population groups.

[0038] S1037. Based on abnormal information of physical examination indicators, retrieval results of the common layer, retrieval results of the group layer, common inference subchain, and group inference subchain, integrate and generate retrieval results of each layer, as well as the association inference chain between each layer.

[0039] like Figure 3 As shown in the figure, this embodiment of the invention provides a schematic diagram of a multi-level knowledge graph architecture. Figure 4 This is a schematic diagram of the intelligent search process. Based on... Figure 3 The multi-layered knowledge graph shown below illustrates the intelligent retrieval algorithm provided by this invention. The retrieval algorithm uses the user's core identifier (User_ID) as the entry point to achieve multi-hop retrieval and inference chain construction.

[0040] 1. Core Identifier Recognition: When a user inputs a query (such as "Please help me analyze the medical examination report I just uploaded"), the system extracts the User_ID as 12345, uses this User_ID as the core identifier, and inputs it into the LLM as the search entry point.

[0041] 2. Individual-level multi-hop retrieval: (1) Locate the user: Execute a Cypher query: MATCH (p:User) WHERE p.user_ID = "12345" RETURN p, and obtain the user node.

[0042] (2) Multi-hop traversal: Starting from the user node, perform multi-hop retrieval along the path “User→Encounter→Test_Type→Test_Result”. The maximum number of hops is set to 5. Prioritize obtaining Test_Result (weight 40%), Test_Type (weight 30%), and other medical entities (weight 30%).

[0043] 3. Cross-layer association retrieval: (1) Common layer association: Based on the individual layer Test_Result, execute the query: MATCH(r:Test_Result)-[ref:REFERENCE_OF]->(c:Concept)WHERE ref.similarity ≥0.6 RETURN r, c, to obtain the associated common layer concept, and perform multi-hop retrieval (maximum 3 hops) on the common layer concept to construct the link "Test_Result→Concept→Related_Concept".

[0044] (2) Group-level association: Based on the individual-level Test_Result combination, execute the query: MATCH (r_comb:Test_Result_Combination)-[m:MATCHES]->(ic:Indicator) WHERE m.similarity≥0.8RETURN r_comb, ic, to obtain the associated group-level indicator combination and corresponding disease.

[0045] 4. Inference Chain Construction: Integrate search results to form an inference chain of "abnormal individual indicators → pathological mechanism in the common layer → disease association in the population layer". For example: "user's lymphocytes are elevated (5.2×10^9 / L) → REFERENCE_OF → common layer "Lymphocytosis" → manifestation_of → "juvenile idiopathic arthritis"; at the same time, match the population layer "juvenile idiopathic arthritis" → STATISTICALLY_ASSOCIATED → "earlychildhood: 5 - 7 years old".

[0046] For example, the present invention adopts a distributed software architecture, which does not require dedicated hardware equipment and is compatible with mainstream servers and terminal devices. The core hardware requirements are as follows: (1) Server: Supports Neo4j database deployment (recommended configuration: CPU≥8 cores, memory≥32GB, hard disk≥1TB), meets the parallel storage and retrieval requirements of the three-layer knowledge graph; (2) Terminal device: Computer or mobile terminal for uploading reports and viewing analysis results (supports browser access or dedicated client installation); (3) Data storage device: Cloud storage or local storage device for storing physical examination report data, UMLS knowledge base data, group statistical data and analysis results, which must ensure efficient reading and writing of large-scale data at the group level.

[0047] In terms of software platform, the system integrates the following core components: (1) Database: Neo4j (used to store knowledge graph data of public layer, individual layer and group layer), MySQL (used to store structured data of physical examination reports, raw data of group statistics and analysis records); (2) Core algorithm library: SapBERT (used to generate three-layer entity embedding vectors), gemma3:27b (used to LLM inference), PyPDF2 / Jsonpickle (used to parse multi-format reports), Scipy (used to calculate Pearson correlation coefficient of group layer); (3) Development framework: Flask (used to build system interface), Vue.js (used to develop front-end interactive interface); (4) Data processing tools: Pandas (used to standardize data processing), NumPy (used to calculate similarity and statistical analysis).

[0048] For example, the present invention combines a multi-layered knowledge graph architecture, and the data acquisition and preprocessing steps are as follows: (1) Data collection and preprocessing of the common layer: Download biomedical terminology and semantic relationship data from the official UMLS database, and select 24 core semantic relationships that are strongly related to physical examination analysis (such as indicators, manifestation_of, causes, etc.); use Pandas to clean redundant fields and correct format errors, establish relationship mapping according to the logic of "observation indicators-pathological mechanisms-diseases", and generate a standardized common layer dataset.

[0049] (2) Individual-level data collection and preprocessing: The system interface receives the physical examination reports uploaded by users (supporting Notepad, PDF, and JSON formats), uses PyPDF2 to parse PDF files and Jsonpickle to parse JSON files, and extracts key fields such as the user's basic information (age, gender), physical examination indicator data (indicator name, value, unit, reference range), and medical records (date, etc.); the user's basic information and physical examination indicator data are associated through User_ID, and the indicator units and reference range formats are unified to complete the standardization of individual-level data.

[0050] (3) Group-level data collection and preprocessing: Collect raw data of physical examination reports of more than 10,000 people, and classify and organize them according to "population group (age, gender, etc.) - combination of detection indicators - disease type - statistical period"; use Pandas to clean abnormal data and fill in missing fields, calculate the Pearson correlation coefficient of "combination of detection indicators - disease type" through SciPy, retain effective relation pairs with absolute value of Pearson correlation coefficient ≥ 0.3, and generate standardized group-level dataset.

[0051] S104. Based on the retrieval results of each layer, the association reasoning chain between each layer, and the large language reasoning model, perform indicator anomaly identification, disease attribution analysis, and potential risk prediction to generate a physical examination analysis report.

[0052] In some embodiments, the health checkup analysis report includes risk warnings, attribution paths, a list of potential diseases, and reasoning.

[0053] As one possible implementation, step S104 can be specifically implemented as steps S1041-S1044.

[0054] S1041. Construct the retrieval results and related reasoning chains of each layer into structured contextual data suitable for large language reasoning models.

[0055] S1042. Concatenate the structured context data with the preset prompts to obtain the input data.

[0056] The prompts include role definitions and output format constraints.

[0057] S1043. Input the input data into the large language reasoning model, perform chain reasoning operations, and obtain the recognition results of each layer.

[0058] For example, step S1043 can be specifically implemented as steps A1-A5.

[0059] A1. Based on the individual-level retrieval results, the detection values ​​are compared with the reference range to identify anomalies in indicators and generate risk warnings, thereby identifying abnormal indicators and risk warning content.

[0060] A2. Based on the retrieval results and public inference subchains of the public layer, analyze the logical relationship from abnormal indicators to pathological mechanisms and then to related diseases, perform disease attribution analysis and attribution path generation, and generate attribution paths.

[0061] A3. Based on the group-level retrieval results and the group inference sub-chain, potential diseases are sorted according to the statistical association strength, and potential risk prediction and disease list generation are performed to generate a potential disease list.

[0062] A4. Determine the abnormal indicators and risk warning content, the attribution path, and the reasoning basis for the potential disease list.

[0063] In some embodiments, the reasoning basis includes entity nodes and relationships in a multi-level knowledge graph.

[0064] S1044. Based on the recognition results of each layer, generate a physical examination analysis report.

[0065] For example, combining intelligent retrieval and reasoning processes, the steps for constructing multimodal input are as follows: (1) Constructing constant information, integrating the core semantic relationships in the three-layer knowledge graph (pathological association in the public layer, statistical association in the group layer, and indicator association in the individual layer), the preset analysis report template (including risk warning, disease attribution and other module formats), the standardized physical examination indicator data after multimodal data preprocessing, and the effective relationship pairs of group statistics.

[0066] (2) Memory message construction: extract the current user's historical physical examination analysis records, the latest physical examination indicator trend data, and past feedback correction information from the MySQL database, and associate them with the historical statistical data of the same group in the group layer.

[0067] (3) Analyze the demand acquisition. Receive the analysis request (including specific needs such as risk analysis of physical examination report and prediction of potential diseases) input by the user through the front-end interface. The user identification core identifier is obtained from the system.

[0068] (4) Input standardization splicing: splice constant information, memory messages, analysis requirements and three-layer knowledge graph entity association information according to the preset format to generate unified multimodal prompt text that LLM can recognize.

[0069] For example, in the physical examination analysis reasoning and report generation, (1) Reasoning execution: the standardized multimodal prompt text is input into the gemma3:27b model, and the model completes three core reasoning steps based on the preset chain reasoning template: indicator anomaly identification: compare the individual layer Test_Result value with the reference range to generate an indicator over / under warning; disease attribution analysis: based on the relationship between individual layer indicator anomalies and public layer REFERENCE_OF, trace the pathological mechanism and construct the public layer reasoning chain of "indicator anomalies → pathological concepts → associated diseases"; potential risk prediction: combine the individual layer indicator combination and the group layer MATCHES relationship, use the group layer STATISTICALLY_ASSOCIATED statistical correlation data, and combine the user's age, gender and other population attributes to predict potential disease risks.

[0070] (2) Structured output generation: The model outputs a standardized analysis report according to a preset template, which includes five core contents: Risk warning: clearly marking the name, value and degree of deviation from the reference range of abnormal indicators; Disease attribution: presenting the reasoning process of "abnormal individual indicators → pathological mechanism of the public layer" in chain text; Potential diseases: listing potential disease risks in descending order of probability of association with the population layer statistics; Reasoning basis: detailing the nodes, relationships and data sources of the three-layer knowledge graph (including UMLS knowledge base, population statistics data and individual physical examination data); Supplementary information: including disclaimers and data source descriptions (clearly specifying the copyright information of UMLS and population statistics data).

[0071] This invention provides an intelligent analysis method and system for physical examination reports based on a multi-level knowledge graph. By receiving query information containing user identifiers, the invention first performs multi-hop searches at the individual level of the multi-level knowledge graph to quickly and accurately locate and extract abnormal information from the user's physical examination indicators, effectively overcoming the efficiency bottleneck of manual review of each report. Subsequently, based on the abnormal information, it automatically and synchronously retrieves pathological mechanisms and disease knowledge in the common layer and statistical correlation information in the group layer using cross-layer association relationships. This integrates general medical knowledge, individual-specific data, and group statistical patterns, making risk identification no longer limited to a single knowledge source and significantly improving the comprehensiveness and objectivity of the analysis. Finally, by integrating the retrieval results and association reasoning chains from each layer, it inputs them into a large language reasoning model for comprehensive reasoning, automating the entire process from indicator anomaly identification and disease attribution to potential risk prediction, and generating a structured physical examination analysis report. This invention solves the problems of low efficiency, inaccurate results, and poor interpretability inherent in traditional manual interpretation of physical examination reports, achieving efficient, comprehensive, and logically interpretable intelligent analysis of physical examination reports. This not only significantly improves processing efficiency but also enhances risk detection capabilities and the reliability of conclusions.

[0072] Optionally, the intelligent analysis method for physical examination reports based on multi-level knowledge graphs provided in this embodiment of the invention further includes steps S201-S204 before step S102.

[0073] S201. Construct a multi-level knowledge graph, which includes an individual layer, a public layer, and a group layer; As one possible implementation, step S201 can be specifically implemented as steps S2011-S2016.

[0074] S2011. Based on the Unified Medical Language System, core semantic relationships related to physical examination analysis were selected.

[0075] S2012. Based on core semantic relationships, construct a knowledge graph with observation indicators, pathological mechanisms, and diseases as nodes and core semantic relationships as edges, as a common layer.

[0076] S2013. Analyze multiple medical examination reports in the current batch and extract user information, test type and test results for each medical examination report.

[0077] S2014. Based on user information, test types, and test results from each physical examination report, construct an individual health data knowledge graph centered on the user and hierarchically linked to medical records, test types, and test results, as the individual layer.

[0078] S2015. Based on multiple medical examination reports from various batches during historical periods, large-scale medical examination report statistics were obtained through analysis.

[0079] S2016. Based on the statistical data of large-scale physical examination reports, calculate the statistical association between the combination of test indicators and disease types, and construct a knowledge graph reflecting the association between indicators and disease groups as the group layer.

[0080] S202. Use a semantic embedding model to encode entities in the public layer, individual layer and group layer respectively to obtain vector representations of entities in each layer; S203. Based on the vector representations of entities in each layer, calculate the first semantic similarity between individual layer entities and public layer entities, and the second semantic similarity between combinations of individual layer entities and group layer entities. S204. Based on the first semantic similarity and the second semantic similarity, a threshold determination is made to establish cross-layer association relationships between different layers in the multi-level knowledge graph.

[0081] For example, the core algorithm of this invention is a multi-level knowledge graph construction algorithm. This algorithm constructs a public layer, an individual layer, and a group layer graph in three steps. Each layer graph forms a complete medical knowledge network through structured data parsing and semantic relationship extraction.

[0082] 1. Construction of the common layer graph.

[0083] (1) Data filtering: From the 54 semantic relations in UMLS, 24 core relations that are strongly related to physical examination analysis (such as indicators (symptoms point to the state), manifestation_of (symptoms belong to the disease), causes (cause chain) etc.) were filtered out to eliminate redundant relations and reduce the complexity of the graph.

[0084] (2) Relationship mapping: Classify the filtered relationships by function and establish the association link of "observation index-pathological mechanism-disease", such as "lymphocyte count elevated (Observation) → manifestation of → Lymphocytosis (Pathology) → causes → virus infection (Disease)".

[0085] (3) Storage architecture: Neo4j graph database is used for storage. The node type is set to Concept, which includes CUI (UMLS concept unique identifier), definition, semantic type and other attributes; the relation type is 24 kinds of semantic relations after filtering, and the attributes include relation description and confidence.

[0086] 2. Individual-level map construction (based on medical examination report parsing).

[0087] (1) Data parsing: Input the physical examination report (supports JSON / PDF / Notepad format), and extract the core entities and relationships through a Python parsing script. Entities include User (user, including attributes such as age and gender), Test_Type (test type, such as "white blood cell count"), Test_Result (test result, including attributes such as value, unit, and reference range), and Encounter (medical record, including date attribute); relationships include has_encountered (user-medical record), has_test (medical record-test type), and has_result (test type-test result).

[0088] (2) Construction process: With the user as the core node, the relationship is established according to the hierarchy of "User→Encounter→Test_Type→Test_Result", as shown in the following example (JSON format):

[0089] 3. Construction of population-level maps (based on large-scale physical examination report statistics).

[0090] (1) Data source: Collect data from physical examination reports of more than 10,000 people and conduct statistical analysis according to "test indicator combination - disease type".

[0091] (2) Node and relationship types: Indicator (single detection indicator), Abnormality (abnormality type of indicator), Disease (disease type), Population_Group (population group), Statistical_Period (statistical period); Relationship type: has_statistical_period (has...statistical period), Prevalent_In (high incidence in...population); Cross-graph relationship type: relationship between individual layer and group layer STATISTICALLY_ASSOCIATED.

[0092] (3) Association calculation: The Pearson correlation coefficient is used to calculate the association strength between the combination of indicators and the disease. The relationship is maintained if the absolute value of the Pearson correlation coefficient is ≥0.3 to ensure statistical significance.

[0093] 4. Cross-layer linking algorithm: This algorithm establishes connections between three layers of the graph through semantic embedding and similarity calculation. The specific steps are as follows: (1) Embedding vector generation: Using the SapBERT model, the names of the common layer Concept (such as "Lymphocytosis" in UMLS), the individual layer Test_Result (such as "lymphocyte count elevated"), and the population layer Indicator (such as "Lymphocytosis") are encoded to generate 768-dimensional embedding vectors.

[0094] (2) Similarity calculation: Calculate the cosine similarity between the individual layer Test_Result and the public layer Concept, and the cosine similarity between the combination of individual layer Test_Result and the group layer Indicator.

[0095] (3) Association establishment: When the cosine similarity is ≥0.6, a cross-layer relationship is established: the relationship between individual layer and public layer is REFERENCE_OF; when the cosine similarity is ≥0.8, a cross-layer relationship is established: the relationship between individual layer and group layer is MATCHES.

[0096] An example of three-layer knowledge graph construction and cross-layer linking is shown below.

[0097] (1) Construction of public layer knowledge graph: import the preprocessed UMLS data into the Neo4j database, create Concept nodes (including CUI, definition, semantic type and other attributes), establish the relationship between nodes based on the 24 core semantic relationships after screening; check invalid relationships through graph traversal tools, and finally form a clear “observation index-pathological mechanism-disease” association network.

[0098] (2) Construction of individual-level knowledge graph: Import the standardized individual physical examination data into the Neo4j database, create four types of entity nodes: User, Encounter, Test_Type, and Test_Result, and supplement attribute information; establish the association of has_encountered, has_test, and has_result according to the hierarchical relationship of “User→Encounter→Test_Type→Test_Result”, and form an individual-level knowledge network with users as the core.

[0099] (3) Construction of the knowledge graph of the group layer: The standardized group statistical data is imported into the Neo4j database, and five types of entity nodes are created: Indicator (single detection indicator), Abnormality (abnormality type of indicator), Disease (disease type), Population_Group (population grouping), and Statistical_Period (statistical period). Relationships such as Prevalent_In (high incidence in... population) and STATISTICALLY_ASSOCIATED (statistical association) are established. Among them, the STATISTICALLY_ASSOCIATED relationship associates the group layer indicator combination with the corresponding disease to ensure the statistical validity of the group layer data.

[0100] (4) Cross-layer linking is implemented by calling the SapBERT model to encode the names of the public layer Concept, the individual layer Test_Result, and the group layer Indicator, respectively, and generating a 768-dimensional embedding vector; the cosine similarity between the individual layer Test_Result and the public layer Concept is calculated, and a REFERENCE_OF relationship (individual layer → public layer) is established when the similarity is ≥0.6; the cosine similarity between the individual layer Test_Result combination and the group layer Indicator is calculated, and a MATCHES relationship (individual layer → group layer) is established when the similarity is ≥0.8, thus completing the fusion of the three-layer knowledge graph.

[0101] Thus, this invention significantly improves the systematicness and accuracy of health checkup report analysis by constructing a three-layer knowledge graph of "public-individual-group" and achieving cross-layer links based on a semantic embedding model. First, the three-layer structure integrates general medical knowledge, individual real-time data, and group statistical patterns, overcoming the shortcomings of incomplete coverage in a single-layer graph and making risk identification more comprehensive. Second, by using semantic similarity calculation (such as cosine similarity) instead of fixed rules for cross-layer association, it can more accurately discover the deep semantic connections between abnormal indicators and pathological mechanisms and disease risks, effectively reducing false associations and missed associations, thereby enhancing the reliability and interpretability of the entire analysis and reasoning process.

[0102] Optionally, the intelligent analysis method for physical examination reports based on multi-level knowledge graphs provided in this embodiment of the invention further includes steps S301-S304 before step S104.

[0103] S301. Based on preset conditions, select a large language model.

[0104] In some embodiments, the preset conditions include having medical expertise and understanding capabilities, supporting complex prompt word engineering constraints, having multi-hop reasoning capabilities, meeting medical data privacy and security requirements, supporting localized deployment, and having compliant medical data processing qualifications.

[0105] S302. Based on the selected large language model and combined with the physical examination analysis scenario, type selection and combination are performed to determine the customized model.

[0106] These include open-source medical-adaptive large language models, commercially compliant large language models, or models that are fine-tuned for specific domains based on a general large language model foundation.

[0107] S303. For customized models, construct standardized prompt word templates that include role definitions, task requirements, context structure, and output format constraints to obtain a large language reasoning model.

[0108] In some embodiments, the task requirements include indicator risk warning, disease attribution analysis, potential disease prediction, and explanation of the reasoning basis.

[0109] S304. Deploy the large language reasoning model in a local or cloud environment, and fine-tune the parameters or scenario adaptability according to the actual reasoning efficiency and accuracy requirements to obtain the optimized large language reasoning model.

[0110] In some embodiments, the Large Language Inference Model (LLM) is built upon a large language model that meets the requirements for adaptation to medical scenarios, and standardized report output is achieved through Prompt constraints, as detailed below: 1. Model Selection Principles: The selected large language model must meet the following core requirements to ensure the accuracy, logic, and standardization of medical examination report analysis: It should possess strong medical expertise and be able to accurately interpret medical terminology, pathological mechanisms, and the logical relationship between indicators and diseases; support complex Prompt engineering constraints and output structured content according to preset task requirements, avoiding redundant non-standardized information; possess stable multi-hop inference capabilities and be able to build a complete inference chain based on multi-layered knowledge graph-related data; adapt to the privacy and security requirements of medical data, prioritizing models that support local deployment or have compliant medical data processing qualifications; and possess a certain degree of scalability, enabling parameter tuning or model iteration according to the needs of actual application scenarios (such as analysis efficiency and inference depth).

[0111] 2. Optional Model Range: Models that meet the above selection principles include, but are not limited to: Open-source medical-adaptive large language models: Llama2 / 3 (7B-70B parameters, fine-tuned with a medical knowledge base), Mistral8x7B / 8x22B (supports medical domain plugin extensions), Qwen-Med (an open-source model optimized for medical scenarios), etc.; Commercial compliant large language models: GPT series models, Claude3 series models, Wenxin Yiyan Medical Edition, etc., which have medical data processing qualifications (requires API calls and ensures secure data transmission); Customized training models: Customized models based on a general large language model foundation, combined with a dedicated knowledge base for physical examination analysis scenarios (such as UMLS terminology database, local population statistics data), fine-tuned.

[0112] 3. Prompt design, including role definition (professional medical examination report analysis physician), task requirements (indicator risk warning, disease attribution, potential disease prediction, reasoning basis), context (retrieved reasoning chain + three-layer graph data), and output format (fixed 5 parts: risk warning, disease attribution, potential disease, reasoning basis, and disclaimer).

[0113] 4. Reasoning Logic: LLM, based on context, first determines whether the indicator is abnormal (compared to the reference range) and generates an over / under warning; then, it combines the common layer reasoning chain to analyze the pathological mechanism and generate disease attribution; finally, it combines the population layer statistical association to rank potential disease risks (from high to low according to the association probability); and finally outputs a standardized report containing the reasoning basis.

[0114] Thus, by setting strict model selection criteria and a customized process, this invention ensures that the adopted large language reasoning model possesses medical professionalism, logical reasoning ability, and data compliance and security. Furthermore, by combining scenario-adaptive prompt word engineering and deployment optimization, the general model can accurately adapt to the task of analyzing physical examination reports, ultimately ensuring that the system output analysis results have high accuracy, strong interpretability, and reliability in practical applications.

[0115] Optionally, the intelligent analysis method for physical examination reports based on multi-level knowledge graphs provided in this embodiment of the invention further includes steps S401-S403 after step S104.

[0116] S401. Collect feedback data from the current batch of physical examination analysis reports.

[0117] In some embodiments, feedback data includes corrective confirmation information from healthcare professionals or statistical association data of newly included groups.

[0118] S402. Based on the feedback data, verify the cross-layer correlations involved in the current batch of physical examination analysis reports and determine the verification results.

[0119] In some embodiments, the verification result includes the inference result being confirmed as valid and the inference result having a deviation. If the inference result is confirmed as valid, the confidence weight is increased; if the inference result has a deviation, backtracking calibration is triggered. Backtracking calibration includes recalculating the semantic similarity or adjusting the similarity threshold based on the feedback data.

[0120] S403. Based on the verification results, incrementally update the multi-level knowledge graph.

[0121] In some embodiments, incremental updates include: updating the long-term health record nodes and relationships of users in the individual layer, updating the association strength and significance information of statistical indicators in the group layer, and optimizing cross-layer association relationships.

[0122] Thus, this invention, by introducing a verification and incremental update mechanism based on feedback data, endows the system with the ability to continuously learn and optimize. This mechanism utilizes confirmation from medical professionals or new statistical data to dynamically verify and calibrate the accuracy of cross-layer associations, thereby improving the credibility and adaptability of the inference results. Through incremental updates to the knowledge graph, new knowledge can be continuously integrated and biases corrected, enabling the analytical model to self-evolve and achieve long-term performance improvement, enhancing its practicality, reliability, and adaptability in dynamic medical environments.

[0123] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0124] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0125] Figure 5This diagram illustrates the structure of an intelligent analysis device for physical examination reports based on a multi-level knowledge graph, according to an embodiment of the present invention. The intelligent analysis device 500 includes a communication module 501 and a processing module 502.

[0126] The communication module 501 is used to receive query information input by the user; the query information includes the user identifier.

[0127] The processing module 502 is used to perform multi-hop retrieval in the individual layer of a multi-level knowledge graph based on user identifiers to obtain abnormal information of the user's physical examination indicators. The multi-level knowledge graph includes an individual layer, a public layer, and a group layer. Based on the abnormal information of the physical examination indicators and the cross-layer association relationships between the layers in the multi-level knowledge graph, the module retrieves the pathological mechanisms and disease knowledge of the public layer and the statistical association information of the group layer to determine the retrieval results of each layer and the association reasoning chain between each layer. Based on the retrieval results of each layer, the association reasoning chain between each layer, and the large language reasoning model, the module performs indicator anomaly identification, disease attribution analysis, and potential risk prediction to generate a physical examination analysis report.

[0128] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device 600 includes: a processor 601, a memory 602, and a computer program 603 stored in the memory 602 and executable on the processor 601. When the processor 601 executes the computer program 603, it implements the steps in the above-described method embodiments. Alternatively, when the processor 601 executes the computer program 603, it implements the functions of each module / unit in the above-described device embodiments.

[0129] For example, the computer program 603 may be divided into one or more modules / units, which are stored in the memory 602 and executed by the processor 601 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program 603 in the electronic device 600.

[0130] The processor 601 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0131] The memory 602 can be an internal storage unit of the electronic device 600, such as a hard disk or memory of the electronic device 600. The memory 602 can also be an external storage device of the electronic device 600, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD) card, flash card, etc., equipped on the electronic device 600. Furthermore, the memory 602 can include both internal and external storage units of the electronic device 600. The memory 602 is used to store the computer program and other programs and data required by the terminal. The memory 602 can also be used to temporarily store data that has been output or will be output.

[0132] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for intelligent analysis of physical examination reports based on multi-level knowledge graphs, characterized in that, include: Receive query information input by the user; The query information includes a user identifier; Based on the user identifier, multi-hop retrieval is performed in the individual layer of the multi-level knowledge graph to obtain the user's abnormal physical examination indicators; the multi-level knowledge graph includes an individual layer, a public layer, and a group layer; Based on the abnormal information of the physical examination indicators and the cross-layer association relationship between the layers in the multi-level knowledge graph, the pathological mechanism and disease knowledge of the public layer and the statistical association information of the group layer are retrieved to determine the retrieval results of each layer and the association reasoning chain between each layer. Based on the retrieval results at each layer, the inference chains between layers, and the large language inference model, abnormal indicator identification, disease attribution analysis, and potential risk prediction are performed to generate a physical examination analysis report.

2. The intelligent analysis method for physical examination reports based on multi-level knowledge graphs according to claim 1, characterized in that, Before obtaining relevant abnormal health check indicators by performing multi-hop retrieval in the individual-level knowledge graph based on user identifiers, the process also includes: Construct the multi-level knowledge graph, which includes an individual layer, a public layer, and a group layer; The entities in the public layer, individual layer and group layer are encoded using a semantic embedding model to obtain vector representations of the entities in each layer. Based on the vector representations of entities in each layer, the first semantic similarity between individual layer entities and public layer entities, and the second semantic similarity between combinations of individual layer entities and group layer entities are calculated. Based on the first semantic similarity and the second semantic similarity, a threshold determination is made to establish cross-layer association relationships between the layers in the multi-level knowledge graph.

3. The intelligent analysis method for physical examination reports based on multi-level knowledge graphs according to claim 2, characterized in that, The construction of the multi-level knowledge graph includes: Based on the unified medical language system, core semantic relationships related to physical examination analysis were selected. Based on the core semantic relationships, a knowledge graph is constructed with observation indicators, pathological mechanisms, and diseases as nodes and the core semantic relationships as edges, serving as a common layer. Analyze multiple medical examination reports in the current batch to extract user information, test type, and test results for each report; Based on user information, test types, and test results from each physical examination report, an individual health data knowledge graph is constructed with the user at its core and hierarchically linked to medical records, test types, and test results, serving as the individual layer; Based on multiple medical examination reports from various batches during historical periods, large-scale medical examination report statistics were obtained through analysis. Based on the statistical data of the large-scale physical examination reports, the statistical association between the combination of test indicators and the disease type is calculated, and a knowledge graph reflecting the association between indicators and disease groups is constructed as the group layer.

4. The intelligent analysis method for physical examination reports based on multi-level knowledge graphs according to claim 1, characterized in that, The step of performing multi-hop retrieval at the individual level of a multi-level knowledge graph based on the user identifier to obtain abnormal information about the user's physical examination indicators includes: Starting with the user node corresponding to the user identifier, a multi-hop traversal search is performed at the individual level according to the preset path of user, medical record, test type, and test result; During the retrieval process, different types of nodes are prioritized according to a preset weight allocation strategy in order to prioritize the acquisition of detection result nodes and detection type nodes. By integrating the detection result nodes along the retrieval path and comparing the values ​​with the reference range, abnormal information in the user's physical examination indicators is determined.

5. The intelligent analysis method for physical examination reports based on multi-level knowledge graphs according to claim 1, characterized in that, Based on the abnormal information of the physical examination indicators and the cross-layer associations between the layers in the multi-level knowledge graph, the method of retrieving pathological mechanisms and disease knowledge of the common layer and statistical association information of the population layer, and determining the retrieval results of each layer and the association reasoning chain between each layer, includes: For each abnormal detection result in the abnormal information of the physical examination indicators, based on the first type of relationship in the cross-layer association relationship, the semantically related medical concept nodes are retrieved in the public layer; the first type of relationship is the association relationship between individual layer nodes and public layer nodes; Starting with the retrieved medical concept nodes, a multi-hop extended search is performed along the core semantic relationship in the public layer to obtain pathological mechanism nodes and disease nodes related to the retrieved medical concept nodes, which are used as the search results of the public layer. Establish a common inference sub-chain between the abnormal information of the physical examination indicators and the related pathological mechanism nodes and disease nodes. The common inference sub-chain is the association between abnormal indicators, pathological mechanisms, and diseases. For the combination of abnormal physical examination indicators, based on the second type of relationship in the cross-layer association, matching statistical indicator nodes are retrieved in the group layer; the second type of relationship is the association between individual layer nodes and group layer nodes. Starting with the retrieved statistical indicator nodes, a search is conducted along statistical relationships within the group layer to obtain statistically significant associated disease nodes and population grouping information, which serve as the search results for the group layer. Establish a group inference sub-chain between the abnormal physical examination indicators and the associated disease nodes and population grouping information. The group inference sub-chain represents the association relationship between abnormal indicators, associated diseases, and population groups. Based on the abnormal information of the physical examination indicators, the retrieval results of the public layer, the retrieval results of the group layer, the public inference subchain, and the group inference subchain, the retrieval results of each layer and the association inference chain between each layer are integrated to generate the retrieval results of each layer.

6. The intelligent analysis method for physical examination reports based on multi-level knowledge graphs according to claim 1, characterized in that, The physical examination analysis report includes risk warnings, attribution paths, a list of potential diseases, and the basis for reasoning. Based on the retrieval results at each layer, the correlation and reasoning chains between layers, and the large language reasoning model, the system performs anomaly identification, disease attribution analysis, and potential risk prediction to generate a physical examination analysis report, including: The retrieval results of each layer and the associated reasoning chain are constructed into structured context data suitable for large language reasoning models; The structured context data is concatenated with preset prompts to obtain input data, wherein the prompts include role definitions and output format constraints; The input data is input into the large language reasoning model, and chain reasoning operation is performed to obtain the recognition results of each layer; Based on the recognition results of each layer, a physical examination analysis report is generated.

7. The intelligent analysis method for physical examination reports based on multi-level knowledge graphs according to claim 6, characterized in that, The input data is fed into the large language reasoning model, and chain reasoning operations are performed to obtain the recognition results of each layer, including... Based on the individual-level retrieval results, the detection values ​​are compared with the reference range to identify anomalies in indicators and generate risk warnings, thereby identifying abnormal indicators and risk warning content. Based on the retrieval results of the public layer and the public inference subchain, the logical relationship from abnormal indicators to pathological mechanisms and then to related diseases is analyzed, and disease attribution analysis and attribution path generation are performed to generate attribution paths. Based on the group-level retrieval results and the group inference subchain, potential diseases are sorted according to the statistical association strength, and potential risk prediction and disease list generation are performed to generate a potential disease list. The abnormal indicators and risk warning content are determined, as well as the reasoning basis for the attribution path and the potential disease list. The reasoning basis includes entity nodes and relationships in a multi-level knowledge graph.

8. The intelligent analysis method for physical examination reports based on multi-level knowledge graphs according to any one of claims 1 to 7, characterized in that, Before generating a health checkup analysis report, the process includes: based on the retrieval results at each layer, the inference chains between layers, and the large language inference model, it performs anomaly identification, disease attribution analysis, and potential risk prediction. Based on preset conditions, a large language model is selected. These preset conditions include having the ability to understand medical professional knowledge, supporting complex prompt word engineering constraints, having multi-hop reasoning ability, meeting medical data privacy and security requirements, supporting local deployment, and having compliant medical data processing qualifications. Based on the selected large language model, combined with the physical examination analysis scenario, type selection and combination are carried out to determine the customized model. The types include open source medical-adaptive large language models, commercial compliant large language models, or models based on general large language model foundations with domain fine-tuning. For the customized model, a standardized prompt word template containing role definition, task requirements, context structure and output format constraints is constructed to obtain a large language reasoning model. The task requirements include indicator risk warning, disease attribution analysis, potential disease prediction and explanation of reasoning basis. Deploy the large language reasoning model in a local or cloud environment, and fine-tune the parameters or scenario adaptability according to the actual reasoning efficiency and accuracy requirements to obtain the optimized large language reasoning model.

9. The intelligent analysis method for physical examination reports based on multi-level knowledge graphs according to any one of claims 1 to 7, characterized in that, The method also includes Collect feedback data from the current batch of physical examination analysis reports, including correction and confirmation information from medical professionals or statistical correlation data of newly included groups; Based on the feedback data, the cross-layer relationships involved in the current batch of physical examination analysis reports are verified, and the verification results are determined. The verification results include whether the inference results are valid or whether the inference results have deviations. If the inference results are valid, the confidence weight is increased; if the inference results have deviations, backtracking calibration is triggered. The backtracking calibration includes recalculating the semantic similarity or adjusting the similarity threshold based on the feedback data. Based on the verification results, the multi-level knowledge graph is incrementally updated; the incremental update includes: updating the long-term health record nodes and relationships of users in the individual layer, updating the association strength and significance information of statistical indicators in the group layer, and optimizing cross-layer association relationships.

10. A smart analysis system for physical examination reports based on a multi-level knowledge graph, characterized in that, The system includes an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor is configured to invoke and run the computer program stored in the memory to perform the method as described in any one of claims 1 to 9.

Images