A health recommendation system and method based on a medical knowledge graph

Abstract

The present application relates to the technical field of medical health intelligent recommendation, in particular to a health recommendation system and method based on medical knowledge graph, comprising receiving a query request containing a symptom description text and a user historical health record, synchronously extracting a symptom, a body part, a duration entity set, and a multi-dimensional matching with a knowledge graph node to generate a preliminary candidate disease set; calculating a degree of fit score to sort and select highly relevant diseases, extracting a recommended examination path, a treatment option and a lifestyle suggestion to integrate into a structured report; cross-verification of the report with real-time physiological monitoring data of the user to optimize priority and generate a personalized plan. Multi-dimensional entity matching improves the accuracy of candidate diseases, real-time data verification dynamically optimizes the recommendation to adapt to individual conditions. The present application combines multi-dimensional entity recognition and real-time data cross-verification to enhance the personalization and timeliness of health recommendation.

Description

Technical Field

[0001] This invention relates to the field of intelligent recommendation technology in healthcare, and in particular to a health recommendation system and method based on a medical knowledge graph. Background Technology

[0002] Existing health recommendation methods based on medical knowledge graphs typically receive user symptom descriptions and historical health records. They extract single-type entities through medical entity recognition, perform simple matching with knowledge graph nodes to generate candidate diseases, and then extract recommended content to form a report based on the relevance of historical records. However, these techniques neglect crucial dimensions such as the body part where symptoms occur and their duration during entity recognition, resulting in a coarse matching process and the inclusion of weakly related items in the candidate disease set. Furthermore, the generated recommendation reports are mostly static outputs, lacking validation with real-time physiological monitoring data, making it difficult to reflect the dynamic changes in the user's current state, and potentially leading to discrepancies between recommendation priorities and actual needs.

[0003] To address the aforementioned problems, this invention aims to solve how to improve the accuracy of candidate disease sets by simultaneously extracting symptom entities, body part entities, and duration entity sets and performing multi-dimensional matching; and how to cross-validate structured health recommendation reports with users' real-time physiological monitoring data and dynamically adjust recommendation priorities to adapt to real-time status. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a health recommendation system and method based on medical knowledge graphs.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a health recommendation method based on a medical knowledge graph, comprising:

[0006] Receive a user's health query request, which includes symptom description text entered by the user and the user's historical health record data;

[0007] The symptom description text is processed by medical entity recognition to extract the symptom entity set, body part entity set, and duration entity set. The symptom entity set, body part entity set, and duration entity set are then matched with nodes in the medical knowledge graph in multiple dimensions.

[0008] A preliminary candidate disease set is generated based on the matching results, and a score is calculated for each disease in the preliminary candidate disease set to match the user's historical health record data.

[0009] The preliminary candidate disease set is sorted based on the fit score, and diseases with fit scores higher than the threshold are selected to form a highly relevant disease set.

[0010] For each disease in the highly relevant disease set, corresponding recommended examination paths, potential treatment options, and lifestyle intervention suggestions are extracted from the medical knowledge graph.

[0011] Integrate the recommended examination pathways, potential treatment options, and lifestyle intervention suggestions to form a structured health recommendation report;

[0012] The structured health recommendation report is cross-validated with the user's real-time physiological monitoring data. Based on the cross-validation results, the recommendation priority in the structured health recommendation report is adjusted to generate a personalized health recommendation plan.

[0013] As a further aspect of the present invention, the symptom description text is subjected to medical entity recognition processing to extract a set of symptom entities, a set of body part entities, and a set of duration entities, including:

[0014] The symptom description text is input into a pre-trained biomedical named entity recognition model, which is trained based on a medical literature corpus.

[0015] The biomedical named entity recognition model identifies symptom keywords in the text and maps these symptom keywords to standardized terms in a standard medical terminology database to form the symptom entity set.

[0016] Simultaneously, identify human anatomical terms related to the symptoms in the symptom description text, standardize the human anatomical terms into anatomical terms, and form the body part entity set;

[0017] Further identify time-expressing words describing the duration of symptoms in the symptom description text, convert the time-expressing words into a unified time unit representation, and form the duration entity set;

[0018] The confidence level of entities in the symptom entity set, body part entity set, and duration entity set is evaluated, and entities with confidence levels lower than a set value are filtered out.

[0019] As a further aspect of the present invention, the symptom entity set, body part entity set, and duration entity set are matched with nodes in a medical knowledge graph in multiple dimensions, including:

[0020] The medical knowledge graph contains the relationships between disease nodes, symptom nodes, examination item nodes, drug nodes, and treatment plan nodes.

[0021] Calculate the similarity between each symptom entity in the symptom entity set and the symptom nodes in the medical knowledge graph, and obtain the graph symptom node with the highest matching degree with each symptom entity.

[0022] Each body part entity in the body part entity set is associated and matched with the body part attributes recorded in the medical knowledge graph to determine the location information of each body part entity in the graph;

[0023] Each duration entity in the duration entity set is compared and analyzed with the typical symptom duration range recorded in the medical knowledge graph;

[0024] Based on the symptom nodes in the graph, disease nodes directly connected to the symptom nodes are searched along the edge relationships of the medical knowledge graph to form an initial set of matching diseases.

[0025] By combining the location information of the body part entity and the comparison analysis results of the duration entity, the disease nodes in the initial matching disease set are filtered to remove disease nodes that do not meet the location information and duration characteristics, thus obtaining the preliminary candidate disease set.

[0026] As a further aspect of the present invention, a score for matching each disease in the preliminary candidate disease set with the user's historical health record data is calculated, including:

[0027] Extract the user's past medical history data, allergy history data, and family history of genetic diseases from the user's historical health record data;

[0028] For each disease in the preliminary candidate disease set, query the medical knowledge graph for common complication data, inducing factor data, and genetic association data of the disease.

[0029] The matching degree between the user's past medical history data and the common complication data is calculated to obtain a medical history matching score.

[0030] The user's allergy history data and the triggering factor data are correlated to obtain an allergy risk matching score;

[0031] The similarity scores of the family history of genetic diseases and the genetic association data are compared to obtain a genetic risk matching score.

[0032] The matching scores for disease history, allergy risk, and genetic risk are weighted and summed to obtain the matching score for each disease.

[0033] As a further aspect of the present invention, for each disease in the highly relevant disease set, corresponding recommended examination paths, potential treatment options, and lifestyle intervention suggestions are extracted from the medical knowledge graph, including:

[0034] For each disease in the set of highly relevant diseases, the recommended examination path is formed by searching the medical knowledge graph for all examination item nodes that start from the disease node and are connected by examination edges, sorting them according to the examination necessity weight.

[0035] Search the medical knowledge graph for all treatment nodes that originate from the disease node and are connected via treatment edges. The treatment nodes include drug nodes, surgical nodes, and physical therapy nodes, which together constitute the potential treatment options.

[0036] The medical knowledge graph is searched for all lifestyle nodes that start from the disease node and are connected via prevention edges. These lifestyle nodes include dietary advice nodes, exercise advice nodes, and daily routine advice nodes. These are then integrated to form the lifestyle intervention recommendations.

[0037] Based on the fit scores of diseases in the highly relevant disease set, priority weights are assigned to the recommended examination pathways, potential treatment options, and lifestyle intervention recommendations for each disease.

[0038] As a further aspect of the present invention, the recommended examination pathways, potential treatment options, and lifestyle intervention suggestions are integrated to form a structured health recommendation report, including:

[0039] Create a multi-level reporting structure, which includes a disease overview layer, an examination suggestion layer, a treatment option layer, and a lifestyle layer;

[0040] The diseases in the highly relevant disease set are sorted in descending order of their matching scores and placed in the disease overview layer. Each disease is accompanied by its matching score and main symptom matching information.

[0041] After sorting the recommended examination paths for each disease according to the weight of examination necessity, they are placed under the corresponding disease sub-item in the examination suggestion layer;

[0042] After sorting the potential treatment options for each disease according to treatment routines, they are placed under the corresponding disease sub-item in the treatment option layer;

[0043] After sorting the lifestyle intervention recommendations for each disease according to their implementation difficulty, place them under the corresponding disease sub-item in the lifestyle layer;

[0044] Add time sensitivity and execution urgency tags to each entry in the report structure to form the complete structured health recommendation report.

[0045] As a further aspect of the present invention, the structured health recommendation report is cross-validated with the user's real-time physiological monitoring data, including:

[0046] The real-time physiological monitoring data includes heart rate variability data, sleep quality index, and daily activity level statistics;

[0047] Abnormal physiological indicators are extracted from the real-time physiological monitoring data. These abnormal physiological indicators include heart rate variability data that deviates from the normal range, sleep quality index that is below the standard value, and statistics on abnormally fluctuating daily activity levels.

[0048] The abnormal physiological indicators are matched with the physiological parameters targeted by the examination items recommended in the structured health recommendation report.

[0049] For each matching pair, the correlation between the degree of deviation of the abnormal physiological indicators and the recommendation strength of the corresponding examination items in the structured health recommendation report is calculated;

[0050] Based on the aforementioned relevance, the real-time matching degree of the recommended content in the structured health recommendation report is evaluated, and a matching degree evaluation result is generated.

[0051] For recommendations with low matching scores, a verification flag is added to the structured health recommendation report.

[0052] As a further aspect of the present invention, the step of adjusting the recommendation priority in the structured health recommendation report based on the cross-validation results to generate a personalized health recommendation scheme includes:

[0053] Based on the matching degree assessment results, increase the priority of examination items and lifestyle recommendations that are highly correlated with the abnormal physiological indicators in the structured health recommendation report;

[0054] Reduce or flag recommendations that are not related to the abnormal physiological indicators;

[0055] By combining the user's real-time medication records, the potential treatment options in the structured health recommendation report are screened for contraindications, and conflicting treatment options are removed;

[0056] Based on the user's geographic location information, supplement the structured health recommendation report with information on locally available medical resources, including hospital specialty information and drug inventory information;

[0057] The adjusted structured health recommendation report is reorganized according to urgency and implementation order to form an executable personalized health recommendation plan.

[0058] As a further aspect of the present invention, the construction and updating of the medical knowledge graph includes:

[0059] Extract entities related to diseases, symptoms, examinations, drugs, and treatment plans, as well as their relationships, from authoritative medical databases and clinical guidelines to construct an initial medical knowledge graph;

[0060] Regularly crawl the latest medical research literature and clinical trial reports to extract new medical discoveries and treatments;

[0061] Align and integrate new medical discoveries and treatments with existing nodes in the initial medical knowledge graph;

[0062] Based on disease co-occurrence patterns and symptom association patterns discovered in real-world medical data, the edge weights in the initial medical knowledge graph are dynamically adjusted.

[0063] Maintain the version history of the medical knowledge graph to ensure that the version of the knowledge graph used by the recommendation system at different points in time is traceable.

[0064] As a further aspect of the present invention, the present invention also includes a health recommendation system based on a medical knowledge graph, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the health recommendation method based on a medical knowledge graph as described above.

[0065] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0066] This technology simultaneously extracts symptom entity sets, body part entity sets, and duration entity sets from symptom description text, and uses these three as multi-dimensional features for joint matching with nodes in a medical knowledge graph. This approach overcomes the limitations of single-entity dimension matching, enabling precise correspondence between the combined features of symptoms, location, and time and the associated attributes of disease nodes in the knowledge graph. It reduces mismatches caused by ignoring locations or times, making the initial candidate disease set more closely match the user's actual symptom context and minimizing interference from irrelevant diseases.

[0067] The system integrates recommended examination pathways, potential treatment options, and lifestyle intervention suggestions into a structured health recommendation report, which is then cross-validated with the user's real-time physiological monitoring data. By comparing the report's recommendations with real-time physiological indicators and identifying any discrepancies, the priority of each recommendation is dynamically adjusted. This technology changes the static recommendation model, allowing recommendations to dynamically adapt to the user's real-time physiological state, avoiding a disconnect between recommendations and current vital signs, and enhancing the timeliness and relevance of the recommendations. Attached Figure Description

[0068] Figure 1This is a flowchart of the health recommendation method based on medical knowledge graph described in this invention;

[0069] Figure 2 The flowchart for calculating the fit score. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0071] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0072] See Figure 1 The system receives a user's health query request, which includes the user's symptom description text and historical health records. The symptom description text is then processed using medical entity recognition, extracting sets of symptom entities, body parts entities, and duration entities. These three sets are then matched against nodes in a medical knowledge graph across multiple dimensions. Based on the matching results, a preliminary candidate disease set is generated, and a score is calculated for each disease in the preliminary candidate disease set to align with the user's historical health records. The preliminary candidate disease set is then sorted based on the alignment scores, and diseases with alignment scores above a threshold are selected to form a highly relevant disease set. For each disease in the highly relevant disease set, corresponding recommended examination paths, potential treatment options, and lifestyle intervention suggestions are extracted from the medical knowledge graph. These recommendations are integrated to form a structured health recommendation report. The structured health recommendation report is cross-validated with the user's real-time physiological monitoring data. Based on the cross-validation results, the recommendation priorities in the structured health recommendation report are adjusted to generate the final personalized health recommendation plan.

[0073] In one embodiment of the present invention, a user-provided symptom description text, such as "persistent headache for three days, accompanied by neck muscle stiffness," is input into a pre-trained biomedical named entity recognition model. The pre-trained biomedical named entity recognition model is trained based on a large-scale medical literature corpus. The model identifies symptom keywords such as "pain" and "stiffness" in the text and maps these keywords to standard terms in a standard medical terminology lexicon, such as "headache" and "neck stiffness," forming a symptom entity set containing "headache" and "neck stiffness." In some embodiments, the symptom description text may contain non-standard expressions such as "headache," which the pre-trained biomedical named entity recognition model maps to the standard term "headache," ensuring the consistency of the symptom entity set. It is understood that the mapping process depends on the coverage of the standard medical terminology lexicon, which is updated regularly to include new terms.

[0074] In practice, the pre-trained biomedical named entity recognition model simultaneously identifies anatomical terms related to symptoms, such as "head" and "neck," in the symptom description text. It standardizes these anatomical terms into anatomical terms such as "skull" and "cervical spine region," forming a body part entity set that includes "skull" and "cervical spine region." Optionally, the standardization process uses an anatomical coordinate system to convert colloquial descriptions into precise location information, such as converting "neck" into "cervical spine region." In some embodiments, the body part entity set may contain nested structures, such as "head" being subdivided into "forehead" and "occipital region." The pre-trained biomedical named entity recognition model outputs hierarchical part information.

[0075] In practical implementation, the pre-trained biomedical named entity recognition model further identifies time-related words describing the duration of symptoms in the symptom description text, such as "three days". It converts these time-related words into a unified time unit representation, such as "72 hours", forming a duration entity set containing "72 hours". It can be understood that the time unit standardization process supports conversions to hours, days, weeks, etc., ensuring consistency in subsequent comparative analysis. For example, "one week" is converted into "168 hours". Data comparison shows that the original text "three days" becomes the standardized duration entity "72 hours" after conversion, which is convenient for alignment with the typical symptom duration range in the medical knowledge graph.

[0076] In practice, confidence scores are assessed for entities in the symptom entity set, body part entity set, and duration entity set. Entities with confidence scores below a set threshold are filtered out. The confidence score assessment is based on the probability score output by the pre-trained biomedical named entity recognition model and the degree of matching of the entity in a standard medical terminology lexicon. For example, the confidence score for the symptom entity "headache" is 0.95, the confidence score for the body part entity "skull" is 0.90, and the confidence score for the duration entity "72 hours" is 0.88. If the set threshold is 0.85, all entities are retained. If an entity's confidence score is below the set threshold, such as 0.80, it is filtered out. The confidence score calculation formula is as follows:

[0077]

[0078] in: Indicates the confidence level of an entity. This represents the probability score of the pre-trained biomedical named entity recognition model for recognizing this entity. This represents the semantic similarity score between the entity and terms in the standard medical terminology lexicon. It is a weighting coefficient, with a value ranging from 0 to 1. In specific implementation, the weighting coefficient... Setting it to 0.7 emphasizes the importance of the output of the pre-trained biomedical named entity recognition model; optional, semantic similarity score. By calculating the cosine similarity of word vectors, for example, the similarity score between the terms "pain" and "pain" is 0.85. Data comparison shows that after introducing confidence assessment, low-quality entities such as the misspelled word "toutong" are filtered out, improving the accuracy of subsequent matching.

[0079] In one embodiment of the present invention, the medical knowledge graph includes the relationships between disease nodes, symptom nodes, examination item nodes, drug nodes, and treatment plan nodes. The generation process from the symptom entity set, body part entity set, and duration entity set to the preliminary candidate disease set is as follows: For example, the symptom entity set includes "headache" and "neck stiffness", the body part entity set includes "skull" and "cervical spine region", and the duration entity set includes "72 hours". The similarity between each symptom entity in the symptom entity set and the symptom nodes in the medical knowledge graph is calculated to obtain the graph symptom node with the highest matching degree for each symptom entity. The calculation uses the cosine similarity method based on word vectors. The similarity between the symptom entity "headache" and the symptom node "headache" in the medical knowledge graph is 0.98, and the similarity between the symptom entity "neck stiffness" and the symptom node "neck stiffness" in the medical knowledge graph is 0.95. Therefore, the matched graph symptom nodes are determined to be "headache" and "neck stiffness". It can be understood that a threshold needs to be set for the similarity calculation, and matches below the threshold will be considered invalid.

[0080] Specifically, the multi-dimensional matching process includes the following technical steps: First, the standardized symptom entities, body part entities, and duration entities are transformed into high-dimensional feature vectors using a pre-trained word embedding model. For example, the symptom entity "headache" is mapped to a vector... The body part "skull" is mapped as a vector. The duration entity "72 hours" is quantized as a vector. These feature vectors together constitute the feature vector set for this query. In the medical knowledge graph, each node, such as a disease node or symptom node, is associated with an attribute vector learned through a graph neural network, and each edge between nodes is also associated with a weight tensor representing the strength of the relationship. During matching, the system first selects the symptom feature vector from the feature vector set. The system calculates cosine similarity with the attribute vectors of all symptom nodes in the graph to quickly locate the Top-K most similar graph symptom nodes, such as the node "headache". This process is called vectorized retrieval. Starting from these matched graph symptom nodes, the system performs multi-hop path retrieval: the first hop is along the "manifests as" edge to retrieve all directly connected disease nodes, forming an initial matching set; based on the body part vectors in the feature vector group... and duration vector Calculate the association attribute vectors of these disease nodes and , The similarity scores are used to weight and sort the disease nodes in the initial set. The weight calculation here combines the weight tensors of the edges and the vector similarity scores, ultimately resulting in a preliminary set of candidate diseases.

[0081] In specific implementation, each body part entity in the body part entity set is associated and matched with the body part attributes recorded in the medical knowledge graph to determine the location information of each body part entity in the graph. The body part entity "skull" is associated with the higher-level anatomical system of the brain, blood vessels and other tissues recorded in the medical knowledge graph, and the body part entity "cervical spine region" is associated with the cervical spine, nerves, muscles and other structures recorded in the medical knowledge graph. In some embodiments, the location information is represented by anatomical codes, such as "A01.1" for skull and "A02.2.01" for cervical spine. Each duration entity in the duration entity set is compared and analyzed with the typical symptom duration range recorded in the medical knowledge graph. The duration entity "72 hours" is compared with the typical duration range of "4 to 72 hours" for the "migraine" node in the medical knowledge graph, and also with the typical duration range of "several days to several weeks" for the "cervical myofascitis" node. The data comparison shows that "72 hours" falls within the typical range of "migraine" and also within the typical range of "cervical myofascitis".

[0082] In practice, based on the symptom nodes "headache" and "neck stiffness" in the medical knowledge graph, disease nodes directly connected to the symptom nodes are searched along the edge relationships of the medical knowledge graph to form an initial matching disease set. For example, starting from the symptom node "headache," disease nodes "migraine," "tension headache," and "intracranial hypertension" are found through the "manifestations" edge. Similarly, starting from the symptom node "neck stiffness," disease nodes "cervical myofascitis," "meningitis," and "cervical spondylosis" are found through the "manifestations" edge. The initial matching disease set includes "migraine," "tension headache," "intracranial hypertension," "cervical myofascitis," "meningitis," and "cervical spondylosis." The disease nodes in the initial matching disease set are then filtered based on the location information of body part entities and the comparative analysis results of duration entities. The filtering process removes disease nodes that do not meet the location and duration criteria, resulting in a preliminary set of candidate diseases. The location filtering logic requires the disease to be associated with either the "skull" or "cervical spine region." The duration filtering logic requires that "72 hours" fall within the typical symptom duration range of the disease. For example, the disease "migraine" is associated with the location "skull" and its duration matches; the disease "meningitis" is associated with both the "skull" and "meninges" and its duration matches; the disease "cervical spondylosis" is associated with the location "cervical spine region" and its duration matches; and the disease "tension headache" is associated with the location "skull," but the medical knowledge graph records its typical duration as "30 minutes to 7 days," which also meets the criteria. Optionally, a weighted scoring model can be used for the filtering process, where location matching and duration matching each contribute to the score. The total score formula is:

[0083]

[0084] in: Indicates the disease node screening score, This represents the score for body part matching. A score of 1.0 is given when the body part attribute associated with the disease node completely contains or belongs to the user-provided body part entity, a score of 0.5 is given when it is partially relevant, and a score of 0 is given when it is irrelevant. The duration matching score is 1.0 when the duration entity provided by the user falls exactly within the typical duration range of the disease node, 0.5 when it falls on the boundary, and 0 otherwise. and It is a weighting coefficient and A threshold of 0.7 was set, and disease nodes with scores below the threshold were filtered out. After calculation, "intracranial hypertension" was filtered out because its typical duration is often long-term, while "cervical myofascitis" was retained due to its high score. The final preliminary candidate disease set included "migraine", "tension headache", "cervical myofascitis", "meningitis" and "cervical spondylosis".

[0085] In practice, the construction and updating of the medical knowledge graph involves extracting entities and their relationships related to diseases, symptoms, examinations, drugs, and treatment plans from authoritative medical databases and clinical guidelines to build an initial medical knowledge graph. The authoritative medical databases include the ICD-10 disease classification database, the SNOMEDCT medical terminology set, and the UpToDate clinical decision support database. The extracted relationships include "disease-manifestation-symptoms", "disease-recommended examinations-examination items", and "disease-available treatments-treatment plans". The initial medical knowledge graph contains tens of thousands of nodes and hundreds of thousands of edges. Understandably, the extraction of entities and relationships uses natural language processing technology, including syntactic analysis and relationship extraction models.

[0086] In practice, the latest medical research literature and clinical trial reports are crawled regularly to extract new medical discoveries and treatments. Crawling sources include academic databases such as PubMed and the Cochrane Library. New discoveries extracted include "Drug A has a new therapeutic effect on disease B," and new methods include "A new imaging examination C is used to diagnose disease D." The new medical discoveries and treatments are then aligned and integrated with existing nodes in the initial medical knowledge graph. For example, the new discovery "Drug A has a new therapeutic effect on disease B" is aligned with the existing "Disease B" and "Drug A" nodes in the initial medical knowledge graph, and a "Disease B - Available Treatment - Drug A" relationship edge is added. If the "Drug A" node does not exist, a new node is created and a relationship is established with the "Disease B" node. In some embodiments, relationship alignment is based on entity names, synonyms, and contextual semantic similarity. Data comparison shows that the number of new nodes and edges added to the medical knowledge graph after each update cycle can be quantified.

[0087] In practice, the edge weights in the initial medical knowledge graph are dynamically adjusted based on disease co-occurrence patterns and symptom association patterns discovered in real-world medical data. This real-world medical data comes from an anonymous electronic health record database. Analysis shows that disease co-occurrence patterns, such as "hypertension" and "diabetes," often occur simultaneously, and symptom association patterns, such as "headache" and "vomiting," show high association strength under specific diseases. The specific method for dynamically adjusting edge weights is as follows: each relation edge in the initial medical knowledge graph has an initial weight. These weights are periodically updated based on the co-occurrence frequency or association strength statistically analyzed in the real-world medical data. For example, the weight of the "disease-manifestation-symptom" edge is adjusted based on the frequency of the symptom appearing in the disease's medical records. The version history of the medical knowledge graph is maintained to ensure that the knowledge graph versions used by the recommendation system at different points in time are traceable. Each update generates a new knowledge graph version number and records a change log. For example, version "KG-2023-10" records the complete graph status in October 2023. When processing historical queries, the system can call the corresponding historical version of the knowledge graph for retrospective analysis.

[0088] See Figure 2 In one embodiment of the present invention, a score is calculated for each disease in the preliminary candidate disease set to match the user's historical health record data. This process begins by extracting the user's past medical history data, allergy history data, and family history of genetic diseases from the user's historical health record data. The user's historical health record data comes from the user's authorized electronic health record. The extracted past medical history data includes, for example, "hypertension" and "gastric ulcer". The extracted allergy history data includes "penicillin allergy". The extracted family history of genetic diseases includes "mother has type II diabetes". It is understood that the data extraction process must follow a privacy protection agreement and anonymize sensitive information.

[0089] In practice, for each disease in the preliminary candidate disease set, the medical knowledge graph is queried to obtain data on common complications, triggering factors, and genetic associations. For example, the preliminary candidate disease set includes "migraine," "tension headache," and "cervical spondylosis." For the disease "migraine," the medical knowledge graph shows common complications including "anxiety" and "depression," triggering factors including "alcohol," "stress," and "cervical food additives," and genetic associations showing "familial aggregation." For the disease "cervical spondylosis," the medical knowledge graph shows common complications including "radiculopathy" and "spinal cord disease," triggering factors including "neck strain" and "trauma," and genetic associations showing "no obvious single-gene genetic association." Data comparison shows that there are differences in the complications, triggering factors, and genetic association data recorded in the medical knowledge graph for different diseases.

[0090] In practice, the matching degree of the user's past medical history data and the common complications data obtained from the query is calculated to obtain a medical history matching score. The matching degree calculation is based on the correlation strength between the user's past diseases and common complications in the medical knowledge graph. The correlation strength is quantified by the weight of the edge relationship in the medical knowledge graph. For example, "hypertension" in the user's past medical history data and "anxiety" as a common complication of the disease "migraine" have a "comorbidity" relationship in the medical knowledge graph, with a weight of 0.6. "Gastric ulcer" in the user's past medical history data has no direct correlation with the common complications of the disease "migraine", with a weight of 0. The medical history matching score for the disease "migraine" is calculated to be 0.6. For the disease "cervical spondylosis", its common complication "radiculopathy" has no direct correlation with the user's past diseases "hypertension" and "gastric ulcer", so the medical history matching score for the disease "cervical spondylosis" is 0. In some embodiments, the matching degree calculation adopts the method of set intersection and weighted summation.

[0091] In practice, the user's allergy history data is correlated with the retrieved triggering factor data to obtain an allergy risk matching score. The correlation analysis checks whether the user's allergens are directly or indirectly triggering factors of the disease. For example, if the user's allergy history data is "penicillin allergy", the triggering factors for the disease "migraine" include "alcohol", "stress", and "cervical food additives", but not "penicillin", so the allergy risk matching score is 0. Similarly, the triggering factors for the disease "cervical spondylosis" include "cervical strain" and "trauma", but also not "penicillin", so the allergy risk matching score is also 0. Optionally, if the user's allergy history is "aspirin", and the triggering factors for a certain disease include "nonsteroidal anti-inflammatory drugs", then "aspirin" is a type of "nonsteroidal anti-inflammatory drug", and the allergy risk matching score is not 0, indicating that there is a triggering risk.

[0092] In practice, the similarity score is compared between family history of genetic diseases and the genetic association data obtained from the query to obtain a genetic risk matching score. The similarity comparison assesses the closeness of the genetic association between family history and disease. For example, if the family history data shows "mother has type II diabetes," the genetic association data for the disease "migraine" shows "familial clustering," and the medical knowledge graph records that the heritability of migraine is approximately 50%, while the genetic association data for the disease "cervical spondylosis" shows "no obvious single-gene genetic association," indicating a low heritability, then the calculation of the genetic risk matching score requires quantitative comparison. The calculation formula is:

[0093]

[0094] in: Indicates the genetic risk matching score. This represents the heritability value of the disease as recorded in the medical knowledge graph (ranging from 0 to 1). It is an indicator function that responds to a user's family history of genetic diseases. Diseases in the current assessment In the genetic disease family lineage of the medical knowledge graph, a value of 1 is assigned when there is a direct association, a value of 0.5 when there is an indirect association, and a value of 0 when there is no association. The adjustment coefficient is used for normalization. In specific implementation, it is set... The heritability of the disease "migraine" is 1.0. The value was 0.5. A family history of type II diabetes and migraine are indirectly associated in common polygenic inheritance backgrounds. Taking 0.5, the genetic risk matching score is calculated. Regarding the disease "cervical spondylosis", its heritability is... The lower limit was set at 0.2, and there was no clear genetic association with "type II diabetes," therefore... The genetic risk matching score is calculated to be 0. The data comparison showed the differences in genetic risk matching scores for different diseases due to different genetic backgrounds.

[0095] In practice, the scores for matching medical history, allergy risk, and genetic risk are weighted and summed to obtain a fit score for each disease. The weighted sum reflects the overall contribution weight of different historical factors to disease risk. For example, a weight is set for the medical history matching score. The weight of the allergy risk matching score is 0.5. The weight of the genetic risk matching score is 0.2. For the disease "migraine," the matching score is 0.3, the matching score for medical history is 0.6, the matching score for allergy risk is 0, and the matching score for genetic risk is 0.25. Therefore, the compatibility score is... For the disease "tension headache", assuming a medical history matching score of 0.2, an allergy risk matching score of 0, and a genetic risk matching score of 0.1, then the compatibility score is... For the disease "cervical spondylosis", the matching score for medical history is 0, the matching score for allergy risk is 0, and the matching score for genetic risk is 0. Therefore, the compatibility score is... In some embodiments, the weighting coefficient , , The scores can be adjusted based on clinical importance or epidemiological data. It is understandable that the fit score obtained after weighted summation is used for subsequent ranking and screening of the preliminary candidate disease set.

[0096] In one embodiment of the present invention, for each disease in the highly relevant disease set, corresponding recommended examination paths, potential treatment options, and lifestyle intervention suggestions are extracted from the medical knowledge graph. The highly relevant disease set consists of diseases with a fit score higher than a threshold. For example, the fit score threshold is 0.3. After calculation, the fit score of "migraine" is 0.375, the fit score of "tension headache" is 0.13, and the fit score of "cervical spondylosis" is 0. Therefore, the highly relevant disease set includes the disease "migraine". It can be understood that the fit score threshold can be adjusted according to the actual application scenario to control the size of the set.

[0097] In specific implementation, for the disease "migraine" in the highly relevant disease set, the medical knowledge graph searches for all examination item nodes connected by examination edges starting from the "migraine" disease node. The examination edges have a "necessity weight" attribute, the value of which is defined by the medical knowledge graph according to clinical guidelines and expert consensus. The queried examination item nodes include "neurological physical examination", "head CT scan", and "head MRI scan". The necessity weight of "neurological physical examination" is 0.9, the necessity weight of "head CT scan" is 0.6, and the necessity weight of "head MRI scan" is 0.7. After sorting the examinations in descending order of necessity weight, the recommended examination path is formed as ["neurological physical examination", "head MRI scan", "head CT scan"]. In some embodiments, the examination necessity weights can be dynamically updated to reflect the latest medical evidence.

[0098] In practice, the medical knowledge graph searches for all treatment nodes connected via treatment edges starting from the "migraine" disease node. Treatment nodes include drug nodes, surgical nodes, and physical therapy nodes. The retrieved drug nodes include "ibuprofen," "zolmitriptan," and "propranolol," surgical nodes include "nerve block," and physical therapy nodes include "relaxation training" and "biofeedback therapy." These nodes form a set of potential treatment options {"ibuprofen," "zolmitriptan," "propranolol," "nerve block," "relaxation training," and "biofeedback therapy"}. It can be understood that treatment edges also contain attributes such as "treatment routineness," which are used to identify the degree of routine application of the treatment.

[0099] In practice, the medical knowledge graph is searched for all lifestyle nodes connected via prevention edges, starting from the "migraine" disease node. These lifestyle nodes include dietary advice nodes, exercise advice nodes, and sleep advice nodes. The dietary advice nodes found include "avoid foods high in tyramine" and "drink water regularly," the exercise advice nodes include "regular aerobic exercise," and the sleep advice nodes include "maintain regular sleep" and "manage stress." These nodes are integrated to form a set of lifestyle intervention recommendations {"avoid foods high in tyramine," "drink water regularly," "regular aerobic exercise," "maintain regular sleep," and "manage stress"}. Optionally, the prevention edge can have an "evidence level" attribute, representing the strength of the medical evidence for the recommendation.

[0100] In practice, based on the fit scores of diseases in the highly relevant disease set, priority weights are assigned to the recommended examination pathways, potential treatment options, and lifestyle intervention recommendations for each disease. The calculation formula is:

[0101]

[0102] in: This indicates the priority weight assigned to a specific recommended content item under the current disease. This represents the fit score for the current disease "i". This represents the sum of the fit scores for all "n" diseases in the highly correlated disease set. This represents the importance coefficient of the recommended content within the medical knowledge graph, such as the necessity weight of examination items or the routine coefficient of treatment options. For example, the relevance score for "migraine," the only disease in a highly relevant disease set, is... The score is 0.375, and the total score is 0.375, so the ratio is 1. Assuming that priority weight is assigned to the examination item "Neurological Physical Examination," its importance coefficient in the medical knowledge graph... That is, if the necessity weight is 0.9, then the calculation is as follows: Data comparison shows that when a highly relevant disease set contains multiple diseases, the recommended content under the disease with the higher relevance score will receive a higher base weight.

[0103] In practice, recommended examination pathways, potential treatment options, and lifestyle intervention suggestions are integrated to form a structured health recommendation report. This report structure is multi-level, including a disease overview layer, an examination suggestion layer, a treatment option layer, and a lifestyle layer. Diseases in the highly relevant disease set are sorted in descending order of their fit scores and placed in the disease overview layer. Each disease is accompanied by its fit score and main symptom matching information. For example, the disease overview layer contains a record: disease name "migraine", fit score 0.375, and main symptom matching information "headache, stiff neck". It can be understood that the disease overview layer provides a top-level disease summary.

[0104] In practice, the recommended examination paths for each disease are sorted according to their necessity weight and then placed under the corresponding disease sub-item in the examination suggestion layer. Each examination item in the recommended examination path is accompanied by a necessity weight and an assigned priority weight, as shown in Table 1.

[0105] Table 1: Details of Recommended Examination Pathways for the Disease "Migraine"

[0106] Check item name Check the necessity weight Priority weights assigned Neurological physical examination 0.9 0.9 Head MRI scan 0.7 0.7 Head CT scan 0.6 0.6

[0107] In practice, the potential treatment options for each disease are sorted by their treatment conventionality and then placed under the corresponding disease sub-item in the treatment option layer. The treatment conventionality is obtained from the treatment edge attribute of the medical knowledge graph. For example, the treatment conventionality of the drug "ibuprofen" is 0.9 (very high), the treatment conventionality of the drug "zomitriptan" is 0.8, the treatment conventionality of the surgery "nerve block" is 0.4, and the treatment conventionality of the physical therapy "relaxation training" is 0.7. After being sorted in descending order of treatment conventionality, they are placed under the "migraine" sub-item in the treatment option layer. Each treatment option is accompanied by a treatment conventionality value. In some embodiments, the treatment conventionality can be calculated by combining the treatment guideline recommendation level and the frequency of clinical application.

[0108] In practice, lifestyle intervention recommendations for each disease are sorted by implementation difficulty and placed under the corresponding disease sub-item in the lifestyle layer. Implementation difficulty is a preset quantitative value that represents the ease or difficulty of following the recommendation. For example, the implementation difficulty of the lifestyle recommendation "maintaining regular sleep" is 0.3 (low), "managing stress" is 0.7 (high), and "regular aerobic exercise" is 0.6. After being sorted in ascending order of implementation difficulty, they are placed under the "migraine" sub-item in the lifestyle layer. Time sensitivity tags and execution urgency tags are added to each item in the report structure to form a complete structured health recommendation report. Time sensitivity tags include "long-term" and "acute phase," and execution urgency tags include "high," "medium," and "low." For example, the examination item "head MRI scan" is labeled with the time sensitivity tag "non-acute phase" and the execution urgency tag "medium." Optionally, the tags are generated based on predefined rules and disease states in the medical knowledge graph.

[0109] In one embodiment of the present invention, a structured health recommendation report is cross-validated with the user's real-time physiological monitoring data. The user's real-time physiological monitoring data includes heart rate variability data, sleep quality index, and daily activity level statistics. This data is continuously collected and uploaded from a wearable device worn by the user. For example, heart rate variability data represents the variation between adjacent heartbeats in milliseconds; the sleep quality index is a comprehensive score; and daily activity level statistics are measured by "daily steps" and "minutes of activity." Abnormal physiological indicators are parsed from the real-time physiological monitoring data. The parsing process is based on a preset normal range, where the normal heart rate variability data range is set as follows: The range is set to [20, 100] milliseconds. The normal sleep quality index is set to greater than or equal to 70 points. The normal daily activity level is set to more than 5000 steps per day. In the user data, the heart rate variability is 15 milliseconds, which is lower than the lower limit of the normal range. The sleep quality index is 65 points, which is lower than the standard value. The daily activity level is 3000 steps per day, which is lower than the normal range. Therefore, the abnormal physiological indicators obtained by parsing include "heart rate variability lower than 15 milliseconds", "sleep quality index lower than 65 points", and "daily activity level lower than 3000 steps". It can be understood that the normal value range can be personalized according to the user's age, gender and other basic information.

[0110] In practice, abnormal physiological indicators are matched with the physiological parameters targeted by the examination items recommended in the structured health recommendation report. The recommended examination path for the disease "migraine" in the structured health recommendation report includes "neurological physical examination", "head MRI scan", and "head CT scan". In the medical knowledge graph, the physiological parameters associated with the examination item "neurological physical examination" include "neural reflexes" and "sensory function", which do not directly correspond to the abnormal physiological indicators "low heart rate variability" and "low sleep quality index". The examination items "head MRI scan" and "head CT scan" mainly target intracranial structures and do not directly correspond to the parsed abnormal physiological indicators. In some embodiments, the matching relationship is established through the predefined "examination-monitoring-parameter" edge in the medical knowledge graph. For example, the "polysomnography" examination directly corresponds to the "sleep quality index" parameter. If this examination item is in the report, it will be matched. Data comparison shows that the direct correspondence between the examination items in the current report and the abnormal physiological indicators is weak.

[0111] In practice, for each matching pair, the correlation between the deviation of the abnormal physiological indicator and the recommendation strength of the corresponding examination item in the structured health recommendation report is calculated. A matching pair consists of the abnormal physiological indicator and the examination item directly related to it. Since there are few direct matching pairs in the current example, the correlation between lifestyle intervention recommendations and physiological parameters can be considered. The lifestyle intervention recommendations in the structured health recommendation report include "maintaining regular sleep" and "managing stress." The target physiological parameters for these recommendations include "sleep quality index" and "heart rate variability," thus forming matching pairs. For example, the abnormal physiological indicator "sleep quality index lower than 65 points" matches the recommendation "maintaining regular sleep," and the abnormal physiological indicator "heart rate variability lower than 15 milliseconds" matches the recommendation "managing stress." The recommendation strength is reflected by the priority weight of the recommendation in the report. Assuming the priority weight of "maintaining regular sleep" is 0.8 and the priority weight of "managing stress" is 0.6, calculating the correlation requires quantifying the degree of deviation. The calculation formula is:

[0112]

[0113] in: Indicates the degree of deviation. This indicates the lower limit of the normal value for this physiological parameter (e.g., a sleep quality index of 70). This indicates the actual monitored value (e.g., a sleep quality index of 65). For "sleep quality index below 65," the degree of deviation is indicated. Regarding "heart rate variability is 15 milliseconds lower", the degree of deviation Correlation Calculated as the degree of deviation The Pearson product-moment correlation coefficient between the priority weights of the corresponding recommendations is calculated based on data sequences of multiple matching pairs. Used to evaluate the overall fit. This correlation serves not only as a quantitative indicator of fit but also as a core calibration parameter in the dynamic feedback loop. The system compares the correlation with preset thresholds and generates verification instructions of different levels. For example, when When the value is greater than 0.7, a "strong verification" instruction is generated, and the system will allocate higher computation and push priority to such highly relevant recommendations.

[0114] In practical implementation, based on correlation The system assesses the real-time relevance of recommendations in structured health recommendation reports, generating a relevance assessment result. This result is a quantified score or rating, for example, by setting an assessment score. (when ),like A negative value indicates an opposite trend and a low degree of matching. In the example, the calculated relevance... If the value is positive, such as 0.5, the matching degree assessment result is "moderate matching". For recommendations with low matching degree, a verification mark is added to the structured health recommendation report. Low matching degree means that the recommendations are not related to abnormal physiological indicators or have a very weak correlation. For example, the examination item "head CT scan" has no obvious correlation with all current abnormal physiological indicators. Therefore, a verification mark of "[requires clinical evaluation]" is added to the "head CT scan" item in the report. It can be understood that the verification mark is used to remind users or doctors that this recommendation is not directly based on current real-time physiological data.

[0115] In practice, the recommendation priorities in the structured health recommendation report are adjusted based on the cross-validation results to generate personalized health recommendation schemes. Based on the matching degree evaluation results, the priority of examination items and lifestyle recommendations that are highly correlated with abnormal physiological indicators is increased in the structured health recommendation report. For example, the abnormal physiological indicator "heart rate variability is low by 15 milliseconds" is highly correlated with the lifestyle recommendation "manage stress", so the priority weight of "manage stress" is increased from 0.6 to 0.9. The abnormal physiological indicator "sleep quality index is low by 65 points" is correlated with the recommendation "maintain regular sleep", so its priority weight is increased from 0.8 to 0.95. Recommendations that are not related to abnormal physiological indicators are reduced or marked. For example, the examination item "head CT scan" is not related to the current abnormal physiological indicator, so its recommendation strength (or priority) is marked as "low" or placed in the middle or later in the ranking. In some embodiments, priority adjustment is achieved by modifying the original weights and reordering.

[0116] The system resource allocation module dynamically adjusts subsequent processing strategies based on this verification flag and corresponding verification instructions. For recommended items marked with "strong verification," the system prioritizes using a more detailed explanatory model to generate explanatory text and places it at the top of the user's terminal push queue, ensuring that highly timely information reaches users first. For items with low matching scores but added verification flags, such as "head CT scan," their internal processing priority will be lowered. During peak concurrent requests, the generation and push of related content may be postponed or allocated less computing bandwidth, thereby optimizing the overall system resource utilization efficiency and ensuring efficient processing of high-value information streams. This dynamic resource allocation mechanism based on real-time data cross-validation results constitutes a technical closed loop from data input, analysis, verification to resource scheduling feedback.

[0117] In practice, by combining the user's real-time medication records, contraindication screening is performed on potential treatment options in the structured health recommendation report. Conflicting treatment options are removed. For example, if the user is currently taking aspirin for antiplatelet therapy, and the potential treatment option in the structured health recommendation report includes ibuprofen, then ibuprofen and aspirin are both nonsteroidal anti-inflammatory drugs (NSAIDs). Combined use may increase the risk of bleeding. Since this conflict is recorded in the medical knowledge graph under the "drug-interaction-contraindication" section, ibuprofen is removed from the potential treatment options, while non-conflicting options such as zolmitriptan are retained. This contraindication screening requires real-time access to medication records and the latest drug interaction knowledge base.

[0118] In practice, based on the user's geographic location information, the structured health recommendation report is supplemented with information on available local medical resources, including hospital specialty information and drug inventory information. The user's geographic location information is obtained through device GPS or user settings. For example, if the location is "Haidian District, Beijing", the local medical resource database is queried, and information such as "Haidian Hospital's Neurology Department has specialist outpatient clinics on Thursday mornings" is added to the hospital specialty information. Similarly, information such as "Pharmacy A has zolmitriptan in stock" is added to the drug inventory information. The adjusted structured health recommendation report is then reorganized according to urgency and implementation order to form an executable personalized health recommendation plan. The urgency level is determined comprehensively based on the severity of symptoms, the degree of abnormality of physiological indicators, and the time sensitivity of the recommended content. The implementation order is categorized as "Immediate Execution", "Schedule for the Near Future", and "Long-Term Management". For example, "Managing Stress" and "Maintaining Regular Sleep" after being given higher priority are listed as "Immediate Execution" items, while "Head CT Scan" marked as "Low Priority" is listed as "Long-Term Management" or "Follow Doctor's Orders". The final generated plan is a list containing specific action items, implementation time, and local resources.

[0119] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A health recommendation method based on medical knowledge graph, characterized in that, The method includes: Receive a user's health query request, which includes symptom description text entered by the user and the user's historical health record data; The symptom description text is processed by medical entity recognition to extract the symptom entity set, body part entity set, and duration entity set. The symptom entity set, body part entity set, and duration entity set are then matched with nodes in the medical knowledge graph in multiple dimensions. A preliminary candidate disease set is generated based on the matching results, and a score is calculated for each disease in the preliminary candidate disease set to match the user's historical health record data. The preliminary candidate disease set is sorted based on the fit score, and diseases with fit scores higher than the threshold are selected to form a highly relevant disease set. For each disease in the highly relevant disease set, corresponding recommended examination paths, potential treatment options, and lifestyle intervention suggestions are extracted from the medical knowledge graph. Integrate the recommended examination pathways, potential treatment options, and lifestyle intervention suggestions to form a structured health recommendation report; The structured health recommendation report is cross-validated with the user's real-time physiological monitoring data. Based on the cross-validation results, the recommendation priority in the structured health recommendation report is adjusted to generate a personalized health recommendation plan. The structured health recommendation report is cross-validated with the user's real-time physiological monitoring data, including: The real-time physiological monitoring data includes heart rate variability data, sleep quality index, and daily activity level statistics; Abnormal physiological indicators are extracted from the real-time physiological monitoring data. These abnormal physiological indicators include heart rate variability data that deviates from the normal range, sleep quality index that is below the standard value, and statistics on abnormally fluctuating daily activity levels. The abnormal physiological indicators are matched with the physiological parameters targeted by the examination items recommended in the structured health recommendation report. For each matching pair, the correlation between the degree of deviation of the abnormal physiological indicators and the recommendation strength of the corresponding examination items in the structured health recommendation report is calculated; Based on the aforementioned relevance, the real-time matching degree of the recommended content in the structured health recommendation report is evaluated, and a matching degree evaluation result is generated. For recommendations with low matching scores, a verification flag is added to the structured health recommendation report.

2. The health recommendation method based on medical knowledge graph according to claim 1, characterized in that, The symptom description text is processed using medical entity recognition to extract a set of symptom entities, a set of body part entities, and a set of duration entities, including: The symptom description text is input into a pre-trained biomedical named entity recognition model, which is trained based on a medical literature corpus. The biomedical named entity recognition model identifies symptom keywords in the text and maps these symptom keywords to standardized terms in a standard medical terminology database to form the symptom entity set. Simultaneously, identify human anatomical terms related to the symptoms in the symptom description text, standardize the human anatomical terms into anatomical terms, and form the body part entity set; Further identify time-expressing words describing the duration of symptoms in the symptom description text, convert the time-expressing words into a unified time unit representation, and form the duration entity set; The confidence level of entities in the symptom entity set, body part entity set, and duration entity set is evaluated, and entities with confidence levels lower than a set value are filtered out.

3. The health recommendation method based on medical knowledge graph according to claim 2, characterized in that, The symptom entity set, body part entity set, and duration entity set are matched with nodes in the medical knowledge graph in multiple dimensions, including: The medical knowledge graph contains the relationships between disease nodes, symptom nodes, examination item nodes, drug nodes, and treatment plan nodes. Calculate the similarity between each symptom entity in the symptom entity set and the symptom nodes in the medical knowledge graph, and obtain the graph symptom node with the highest matching degree with each symptom entity. Each body part entity in the body part entity set is associated and matched with the body part attributes recorded in the medical knowledge graph to determine the location information of each body part entity in the graph; Each duration entity in the duration entity set is compared and analyzed with the typical symptom duration range recorded in the medical knowledge graph; Based on the symptom nodes in the graph, disease nodes directly connected to the symptom nodes are searched along the edge relationships of the medical knowledge graph to form an initial set of matching diseases. By combining the location information of the body part entity and the comparison analysis results of the duration entity, the disease nodes in the initial matching disease set are filtered to remove disease nodes that do not meet the location information and duration characteristics, thus obtaining the preliminary candidate disease set.

4. The health recommendation method based on medical knowledge graph according to claim 3, characterized in that, For each disease in the preliminary candidate disease set, a score is calculated to match the user's historical health record data, including: Extract the user's past medical history data, allergy history data, and family history of genetic diseases from the user's historical health record data; For each disease in the preliminary candidate disease set, query the medical knowledge graph for common complication data, inducing factor data, and genetic association data of the disease. The matching degree between the user's past medical history data and the common complication data is calculated to obtain a medical history matching score. The user's allergy history data and the triggering factor data are correlated to obtain an allergy risk matching score; The similarity scores of the family history of genetic diseases and the genetic association data are compared to obtain a genetic risk matching score. The matching scores for disease history, allergy risk, and genetic risk are weighted and summed to obtain the matching score for each disease.

5. The health recommendation method based on medical knowledge graph according to claim 4, characterized in that, For each disease in the highly relevant disease set, corresponding recommended examination paths, potential treatment options, and lifestyle intervention suggestions are extracted from the medical knowledge graph, including: For each disease in the set of highly relevant diseases, the recommended examination path is formed by searching the medical knowledge graph for all examination item nodes that start from the disease node and are connected by examination edges, sorting them according to the examination necessity weight. Search the medical knowledge graph for all treatment nodes that originate from the disease node and are connected via treatment edges. The treatment nodes include drug nodes, surgical nodes, and physical therapy nodes, which together constitute the potential treatment options. The medical knowledge graph is searched for all lifestyle nodes that start from the disease node and are connected via prevention edges. These lifestyle nodes include dietary advice nodes, exercise advice nodes, and daily routine advice nodes. These are then integrated to form the lifestyle intervention recommendations. Based on the fit scores of diseases in the highly relevant disease set, priority weights are assigned to the recommended examination pathways, potential treatment options, and lifestyle intervention recommendations for each disease.

6. The health recommendation method based on medical knowledge graph according to claim 5, characterized in that, Integrating the recommended examination pathways, potential treatment options, and lifestyle intervention suggestions, a structured health recommendation report is generated, including: Create a multi-level reporting structure, which includes a disease overview layer, an examination suggestion layer, a treatment option layer, and a lifestyle layer; The diseases in the highly relevant disease set are sorted in descending order of their matching scores and placed in the disease overview layer. Each disease is accompanied by its matching score and main symptom matching information. After sorting the recommended examination paths for each disease according to the weight of examination necessity, they are placed under the corresponding disease sub-item in the examination suggestion layer; After sorting the potential treatment options for each disease according to treatment routines, they are placed under the corresponding disease sub-item in the treatment option layer; After sorting the lifestyle intervention recommendations for each disease according to their implementation difficulty, place them under the corresponding disease sub-item in the lifestyle layer; Add time sensitivity and execution urgency tags to each entry in the report structure to form the complete structured health recommendation report.

7. A health recommendation method based on medical knowledge graph according to claim 6, characterized in that, The step of adjusting the recommendation priority in the structured health recommendation report based on the cross-validation results to generate a personalized health recommendation plan includes: Based on the matching degree assessment results, increase the priority of examination items and lifestyle recommendations that are highly correlated with the abnormal physiological indicators in the structured health recommendation report; Reduce or flag recommendations that are not related to the abnormal physiological indicators; By combining the user's real-time medication records, the potential treatment options in the structured health recommendation report are screened for contraindications, and conflicting treatment options are removed; Based on the user's geographic location information, supplement the structured health recommendation report with information on locally available medical resources, including hospital specialty information and drug inventory information; The adjusted structured health recommendation report is reorganized according to urgency and implementation order to form an executable personalized health recommendation plan.

8. A health recommendation method based on medical knowledge graph according to claim 7, characterized in that, The construction and updating of the medical knowledge graph includes: Extract entities related to diseases, symptoms, examinations, drugs, and treatment plans, as well as their relationships, from authoritative medical databases and clinical guidelines to construct an initial medical knowledge graph; Regularly crawl the latest medical research literature and clinical trial reports to extract new medical discoveries and treatments; Align and integrate new medical discoveries and treatments with existing nodes in the initial medical knowledge graph; Based on disease co-occurrence patterns and symptom association patterns discovered in real-world medical data, the edge weights in the initial medical knowledge graph are dynamically adjusted. Maintain the version history of the medical knowledge graph to ensure that the version of the knowledge graph used by the recommendation system at different points in time is traceable.

9. A health recommendation system based on a medical knowledge graph, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the health recommendation method based on medical knowledge graph as described in any one of claims 1 to 8.

Images