Multimodal Medical Education Knowledge Graph Construction System Integrating Multi-Source Heterogeneous Data
By constructing a multimodal medical education knowledge graph system, the problems of local evidence failure and relationship chain breakage in the updating of multi-source heterogeneous data were solved, and the high accuracy and continuous updating of the medical education knowledge graph were achieved.
Patent Information
- Application Number
- CN202611619272.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-20
- Publication Date
- 2026-11-17
AI Technical Summary
Existing medical knowledge graphs struggle to identify invalid local evidence and broken relationship chains when updating multi-source heterogeneous data, resulting in residual invalid relationships and insufficient accuracy in graph updates.
A multimodal medical education knowledge graph system is constructed, including a multimodal relation primitive generation module, a same evidence coverage band generation module, a relation break incision identification module, and a relation primitive local write-back module. By generating image evidence units, report evidence units, video evidence units, and courseware evidence units, the system identifies uncovered primitives and disconnected boundaries, suspends uncovered primitives, deletes chain adjacency records at the location of disconnected boundaries, and restores valid relation edges and chain adjacency records.
It improves the accuracy, locality, and continuity of updating medical education knowledge graphs, prevents the splicing of false relationships, reduces the deletion and reconstruction of entire relationship chains, and ensures the traceability and semantic continuity of knowledge graphs.
Smart Images

Figure CN123117293A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of knowledge graph technology, and in particular to a multimodal medical education knowledge graph construction system that integrates multi-source heterogeneous data. Background Technology
[0002] As medical education resources become increasingly digitized, structured, and multimodal, medical images, image reports, medical teaching videos, and teaching courseware are gradually becoming important carriers of medical teaching content. Knowledge graphs can organize scattered medical knowledge through entity nodes and relation edges. By combining natural language processing, image recognition, entity alignment, and relation extraction technologies, they can transform medical education resources from different sources and in different data formats into semantically related data structures, providing a data foundation for medical knowledge retrieval, teaching content organization, and medical education resource management.
[0003] For example, application CN110362690B provides a method and apparatus for constructing a medical knowledge graph. The method includes: acquiring diagnostic evidence information, clinical manifestation information, and disease information; establishing the association between the disease and multi-level clinical manifestations, as well as the association between clinical manifestations and diagnostic evidence at each level, based on the diagnostic evidence information, clinical manifestation information, and disease information, using preset diagnostic criteria, clinical criteria, and diagnostic methods for the disease; and constructing a medical knowledge graph containing the associations between diagnostic evidence entities, clinical manifestation entities, and disease entities, based on the associations between the disease and multi-level clinical manifestations, as well as the associations between clinical manifestations and diagnostic evidence at each level.
[0004] For example, application CN120653782A discloses a method for constructing a biomedical knowledge graph, belonging to the field of knowledge graph technology, including: 1. Data preprocessing: classifying and grading multi-source biomedical data from literature databases, electronic medical records, genomic data, and medical images, and desensitizing sensitive information using differential privacy technology; 2. Knowledge extraction and annotation: extracting entities using a joint model based on BERT-BiLSTM-CRF, identifying relationships between entities using an attention mechanism, and performing semantic annotation using an ontological framework; 3. Cross-source alignment and fusion: calculating entity similarity based on graph embedding algorithms, and integrating entities and relationships from different data sources through a probabilistic soft alignment model; 4. Quality control system: constructing a three-level verification mechanism including a rule engine, statistical analysis, and expert feedback to dynamically monitor the completeness, consistency, and timeliness of the knowledge graph.
[0005] However, existing medical knowledge graph construction technologies mainly focus on medical entity extraction, cross-source entity alignment, relationship fusion, and overall quality verification. When the content of medical education data changes, it is difficult to identify the failure of local evidence and the breakage of relationship chains based on the support range of the same evidence unit for continuous relationships. It is also difficult to update only the invalid relationship primitives and the disconnected boundaries. This can easily lead to invalid relationships being retained, valid relationships being deleted as a whole, and insufficient accuracy in updating the knowledge graph.
[0006] Therefore, in order to address the above issues, there is an urgent need for a multimodal medical education knowledge graph construction system that integrates heterogeneous data from multiple sources. Summary of the Invention
[0007] To address the problem that existing medical education knowledge graphs struggle to identify invalid local evidence and broken relationship chains when updating multi-source heterogeneous data, leading to residual invalid relationships and insufficient accuracy in graph updates, this invention provides a multimodal medical education knowledge graph construction system that integrates multi-source heterogeneous data. The technical solution is as follows:
[0008] A multimodal medical education knowledge graph construction system integrating multi-source heterogeneous data is provided. This system includes:
[0009] The multimodal relation primitive generation module is used to acquire medical image data, image report data, medical teaching video data, and teaching courseware data, and generate image evidence units, report evidence units, video evidence units, and courseware evidence units respectively. It extracts medical entity numbers and relation type encoding values and constructs an ordered relation primitive chain.
[0010] The same evidence coverage band generation module is used to connect the relation primitives that are continuously supported by the same evidence unit according to the intra-chain position value to form the evidence cross-primary primitive coverage band, truncate at the position interruption, generate the same evidence continuous coverage mark, and write the relation primitives that are supported by all corresponding evidence unit types into the valid relation edge table.
[0011] The relationship break incision identification module is used to acquire medical education data that is inconsistent with the content recorded in the original evidence unit, generate updated evidence units and updated evidence cross-element coverage bands, identify uncovered primitives and disconnected boundaries, and generate relationship chain break incisions.
[0012] The relation primitive local write-back module is used to suspend uncovered primitives based on the broken relation chain incision, delete the chain adjacency records at the lost boundary position and generate relation primitive chain fragments, restore valid relation edges and chain adjacency records based on newly added evidence units, and write teaching resource nodes, evidence unit nodes, entity nodes, valid relation edges, and resource inclusion edges into the medical education knowledge graph.
[0013] Further, the specific steps for acquiring medical imaging data, imaging report data, medical teaching video data, and teaching courseware data are as follows: Acquire the medical imaging data, imaging report data, medical teaching video data, teaching courseware data, medical knowledge ontology table, and medical relationship rule table stored in the medical education data management database. All four types of medical education data include resource numbers and case numbers. Medical imaging data includes image frame numbers, lesion annotation numbers, and lesion region coordinate values. Imaging report data includes report sentence numbers and report sentence content. Medical teaching video data includes subtitle sentence numbers, subtitle sentence content, subtitle time intervals, and video frame sequences. Teaching courseware data includes courseware page numbers, courseware annotation numbers, annotation region coordinate values, and annotation text content.
[0014] Further, the specific steps for generating image evidence units, report evidence units, video evidence units, and courseware evidence units are as follows: Extract lesion images according to the coordinate values of the lesion area; extract report sentence content according to the report sentence number; extract subtitle sentence content according to the subtitle sentence number; extract corresponding continuous video frames from the video frame sequence according to the subtitle time interval; extract the courseware annotation area and annotation text content according to the coordinate values of the annotation area; and generate image evidence units, report evidence units, video evidence units, and courseware evidence units respectively. Write the image frame number, lesion annotation number, report sentence number, subtitle sentence number, courseware page number, and courseware annotation number into the content index data of each evidence unit, and write an evidence unit number, resource number, and case number for each evidence unit.
[0015] Further, the specific steps for extracting medical entity numbers and relation type codes to construct an ordered relation primitive chain are as follows: The report sentence content, subtitle sentence content, and annotation text content are matched item by item with entity terms, relation terms, and stage terms in the medical knowledge ontology table to extract lesion entity numbers, anatomical location numbers, medical manifestation numbers, conclusion entity numbers, relation type codes, and treatment stage codes, which are then written into the report evidence unit, video evidence unit, and courseware evidence unit. The boundary contour values, major axis values, minor axis values, and grayscale statistical values of the lesion image are extracted and normalized distances are calculated with the image entity templates in the medical knowledge ontology table. When the minimum normalized distance is not greater than the entity matching distance threshold, the lesion entity number, anatomical location number, and medical manifestation number from the image entity template are written into the image evidence unit. Within the same case number, based on the lesion entity number, anatomical location number, and medical manifestation... The entity ID and conclusion entity ID are used to generate lesion entity nodes, anatomical location nodes, medical manifestation nodes, and conclusion entity nodes, respectively. The diagnosis and treatment stage code value is used as the diagnosis and treatment stage node ID to generate diagnosis and treatment stage nodes. Positional relation primitives are generated based on lesion entity ID and anatomical location ID. Performance relation primitives are generated based on lesion entity ID and medical manifestation ID. Conclusion relation primitives are generated based on medical manifestation ID and conclusion entity ID. Stage relation primitives are generated based on conclusion entity ID and diagnosis and treatment stage node ID. Positional relation primitives sharing the same lesion entity ID are connected to performance relation primitives. Performance relation primitives sharing the same medical manifestation ID are connected to conclusion relation primitives. Conclusion relation primitives sharing the same conclusion entity ID are connected to stage relation primitives to generate an ordered relation primitive chain. For each relation primitive, the relation primitive ID, case ID, head entity ID, tail entity ID, relation type code value, and intra-chain position value are written.
[0016] Furthermore, the specific steps for connecting the continuously supported relational primitives of the same evidence unit into an evidence cross-primary primitive coverage band according to the intra-chain positional value, and truncating at the positional break point, are as follows: Read each evidence unit and the ordered relational primitive chain; when the head entity number, relation type code value, and tail entity number in the report evidence unit, video evidence unit, and courseware evidence unit are consistent with the head entity number, relation type code value, and tail entity number of the relational primitive, respectively, write the evidence unit number into the corresponding relational primitive; when the lesion entity number, anatomical location number, and medical manifestation number in the image evidence unit are consistent with the location relational primitive and manifestation relational primitive, respectively, write the image evidence unit number into the corresponding location relational primitive and manifestation relational primitive; divide the relational primitives supported by the same evidence unit into continuous positional segments according to the intra-chain positional value, generate an evidence cross-primary primitive coverage band based on each continuous positional segment, and write the coverage start positional value, coverage end positional value, and evidence unit number.
[0017] Further, the specific steps for generating continuous coverage markers for the same evidence and writing relational primitives supported by all corresponding evidence unit types into the effective relational edge table are as follows: Determine the relational primitives covered by the evidence cross-primary primitive coverage band and the inter-chain boundaries it crosses based on the coverage start and end position values; when the same evidence cross-primary primitive coverage band simultaneously covers relational primitives on both sides of the inter-chain boundary, write continuous coverage markers for the same evidence for the inter-chain boundary; read the evidence unit types required for each relation type encoding value from the medical relation rule table; when there are evidence cross-primary primitive coverage bands covering all evidence unit types required by the relational primitives, write the relational primitives into the effective relational edge table; when there are continuous coverage markers for the same evidence between adjacent relational primitives, write the connection between adjacent relational primitives into the chain adjacency record table.
[0018] Further, the specific steps for obtaining medical education data that is inconsistent with the original evidence unit's recorded content and generating updated evidence units and updated evidence cross-element coverage bands are as follows: Obtain newly written medical imaging data, image report data, medical teaching video data, and teaching courseware data from the medical education data management database; retrieve the original evidence unit based on the resource number and content index data; when the newly written data is inconsistent with the original evidence unit's recorded content, re-extract the evidence content, perform ontology term comparison on the text content, perform image entity template matching on the lesion image, generate updated evidence units, and determine the relational primitives supported by the updated evidence units; delete all evidence cross-element coverage bands formed by the original evidence units, divide the relational primitives supported by the updated evidence units into continuous positional segments according to the intra-chain positional value, generate updated evidence cross-element coverage bands, and write the coverage start positional value, coverage end positional value, and updated evidence unit number; based on the updated evidence cross-element coverage bands, redetermine the necessary evidence unit type coverage status of each relational primitive and the continuous coverage status of the same evidence at the boundaries of each chain.
[0019] Further, the specific steps for identifying uncovered primitives and disconnected boundaries to generate relationship chain break cuts are as follows: Record relationship primitives that do not have all corresponding evidence unit types as uncovered primitives and record the missing evidence unit types; record the inter-chain boundaries that originally had continuous coverage markers of the same evidence and do not have the same evidence cross-primary primitive coverage band after the update as disconnected boundaries; merge adjacent uncovered primitives and disconnected boundaries according to the intra-chain positional value to generate relationship chain break cuts, and write the cut start positional value, cut end positional value, uncovered primitive number, missing evidence unit type, and disconnected boundary number; generate relationship chain retention markers when no uncovered primitives and disconnected boundaries are formed; generate relationship primitive suspension markers for uncovered primitives; generate relationship primitive split markers for disconnected primitives when the necessary evidence unit types of the relationship primitives on both sides of the disconnected boundary are all intact.
[0020] Furthermore, based on the relationship chain break slit suspension of the uncovered primitive, the specific steps for deleting the chain adjacency record at the lost boundary position and generating relationship primitive chain fragments are as follows: Read the relationship chain retention flag, maintain valid relationship edges and chain adjacency records, and delete the original evidence unit number from the relationship primitive; when updating the evidence unit still supports the original relationship primitive, write the updated evidence unit number into the relationship primitive, and retain other evidence unit numbers and evidence cross-primary primitive coverage bands; read the relationship primitive suspension flag, remove the uncovered primitive from the valid relationship edge table, delete the chain adjacency record connected to the uncovered primitive, and generate relationship primitive chain fragments from the retained evidence unit. Extract the remaining evidence unit numbers from the evidence cross-element coverage area, and write the relation element number, head entity number, relation type code value, tail entity number, missing evidence unit type, and remaining evidence unit number into the suspended relation element table; the relation elements and chain adjacency records outside the relation chain break remain unchanged; read the relation chain split mark, retain the valid relation edges on both sides of the disconnection boundary, delete the chain adjacency records at the disconnection boundary position, generate relation element chain fragments with each disconnection boundary as the boundary, and write the chain fragment number and the relation element number sequence arranged in ascending order of the intra-chain position value for each relation element chain fragment.
[0021] Furthermore, based on the newly added evidence units, the specific steps for restoring valid relation edges and chain adjacency records, and writing teaching resource nodes, evidence unit nodes, entity nodes, valid relation edges, and resource inclusion edges into the medical education knowledge graph are as follows: Generate teaching resource nodes based on resource numbers, generate evidence unit nodes based on evidence unit numbers, establish resource inclusion edges pointing from teaching resource nodes to evidence unit nodes, write the evidence unit number into the evidence support field of the valid relation edge, and write entity nodes, teaching resource nodes, evidence unit nodes, valid relation edges, and resource inclusion edges into the medical education knowledge graph; read data from the medical education data management database based on resource numbers and content... If the index data does not retrieve the newly written medical education data containing the original evidence unit, a new evidence unit and a new evidence cross-element coverage band are generated. When the new evidence unit and the remaining evidence units together complete all the necessary evidence unit types of the suspended relation primitive, the relation primitive is rewritten into the effective relation edge table, and the restored effective relation edges are updated to the medical education knowledge graph. When the new evidence cross-element coverage band crosses the lost connection boundary, the chain adjacency record at the lost connection boundary position is restored. The new teaching resource node, the new evidence unit node, and the new resource inclusion edge are written into the medical education knowledge graph, and the relation primitives that have not been completed with the necessary evidence unit types are retained in the suspended relation primitive table.
[0022] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:
[0023] (1) By converting medical images, image reports, medical teaching videos and teaching courseware into image evidence units, report evidence units, video evidence units and courseware evidence units respectively, and constructing an ordered relation primitive chain based on medical entity number, relation type encoding value and chain position value, medical knowledge under different data formats can be associated according to a unified data structure, providing a traceable data foundation for the construction of multimodal medical education knowledge graph.
[0024] (2) By generating evidence cross-element coverage bands according to the range of continuous support relation primitives of the same evidence unit, and generating continuous coverage markers of the same evidence only when the same evidence cross-element coverage band covers relation primitives on both sides of the chain boundary, it is possible to prevent multiple independent evidence units from being spliced into a false continuous relationship. At the same time, the completeness of necessary evidence unit types is used to screen effective relation edges, thereby improving the evidence sufficiency and semantic continuity of medical relation edges.
[0025] (3) By regenerating the updated evidence cross-element coverage band after the medical education data is updated, the missing coverage elements lacking the necessary evidence unit type support and the missing connection boundary formed by the disappearance of the original evidence continuous coverage are identified respectively. The relationship chain break cut is generated according to the position value in the chain. It can distinguish between the failure of a single relationship and the failure of the connection of adjacent relationships, and avoid relying solely on the state of the entire relationship chain for coarse-grained judgment.
[0026] (4) By suspending the missing primitives, deleting the chain adjacency records at the missing boundary position and generating relation primitive chain fragments, the effective relation primitives and chain adjacency records outside the broken relation chain remain unchanged; when new evidence is added to supplement the necessary evidence unit type or the missing boundary is crossed again, the effective relation edge and chain adjacency records are restored, thereby reducing the deletion and reconstruction operations of the entire relation chain and improving the accuracy, locality and continuity of the medical education knowledge graph update. Attached Figure Description
[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0028] Figure 1 It is a system structure diagram for constructing a multimodal medical education knowledge graph that integrates heterogeneous data from multiple sources;
[0029] Figure 2 This is a schematic diagram of multimodal evidence coverage update and relationship chain break write-back;
[0030] Figure 3 This is a schematic diagram of the structure for writing nodes and relationships in a medical education knowledge graph. Detailed Implementation
[0031] The technical solution of the present invention will now be described with reference to the accompanying drawings.
[0032] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.
[0033] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.
[0034] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.
[0035] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.
[0036] This invention provides a multimodal medical education knowledge graph construction system that integrates heterogeneous data from multiple sources, such as... Figure 1 As shown, the system includes the following modules:
[0037] The multimodal relation primitive generation module is used to acquire medical image data, image report data, medical teaching video data, and teaching courseware data, and generate image evidence units, report evidence units, video evidence units, and courseware evidence units respectively. It extracts medical entity numbers and relation type encoding values and constructs an ordered relation primitive chain.
[0038] The same evidence coverage band generation module is used to connect the relation primitives that are continuously supported by the same evidence unit according to the intra-chain position value to form the evidence cross-primary primitive coverage band, truncate at the position interruption, generate the same evidence continuous coverage mark, and write the relation primitives that are supported by all corresponding evidence unit types into the valid relation edge table.
[0039] The relationship break incision identification module is used to acquire medical education data that is inconsistent with the content recorded in the original evidence unit, generate updated evidence units and updated evidence cross-element coverage bands, identify uncovered primitives and disconnected boundaries, and generate relationship chain break incisions.
[0040] The relation primitive local write-back module is used to suspend uncovered primitives based on the broken relation chain incision, delete the chain adjacency records at the lost boundary position and generate relation primitive chain fragments, restore valid relation edges and chain adjacency records based on newly added evidence units, and write teaching resource nodes, evidence unit nodes, entity nodes, valid relation edges, and resource inclusion edges into the medical education knowledge graph.
[0041] Optionally, the specific steps for obtaining medical imaging data, imaging report data, medical teaching video data, and teaching courseware data are as follows: using the resource number as the main retrieval field for the same medical education resource and the case number as the aggregation field for the same medical teaching case, an association index between the resource number and the case number is established in the medical education data management database, so that medical imaging data, imaging report data, medical teaching video data, and teaching courseware data can be aggregated to the same case number, and the evidence units generated subsequently can be traced back to the original medical education resource. The system retrieves medical imaging data, imaging report data, medical teaching video data, teaching courseware data, medical knowledge ontology tables, and medical relationship rule tables stored in the medical education data management database. When reading medical imaging data, it retrieves image frames for the same examination item by resource number and case number, and determines the position of each image frame in the image sequence according to the image frame number. The system sets the top-left pixel of the image frame as the origin of the image plane coordinate system, sets the horizontal direction to the right as the positive X-axis, and the vertical direction downwards as the positive Y-axis. It records the coordinate values of the lesion region in the image plane coordinate system, using a sequence of lesion annotation boundary point coordinates. The lesion annotation number is used to distinguish different lesion annotation regions within the same image frame. When reading imaging report data, it locates the imaging report belonging to the same case as the medical imaging data by resource number and case number, and then reads the report sentence content in the order of the report sentence number, preserving the original sentence order in the imaging report. The report sentence number serves as the content retrieval position for subsequent report evidence units. When reading medical teaching video data, the subtitle content, subtitle time interval, and video frame sequence are read according to resource number and case number. The start and end timestamps of the subtitle time interval are uniformly converted to milliseconds. Continuous video frames falling within the subtitle time interval are extracted from the video frame sequence according to the video frame timestamps, ensuring that the subtitle content and the medical images displayed within the subtitle time interval are time-consistent. When reading teaching courseware data, the courseware pages are read according to the courseware page number. The top left corner of the courseware page is set as the origin of the courseware page coordinates. The horizontal direction to the right is set as the positive X-axis, and the vertical direction downwards is set as the positive Y-axis. The position of the annotation area in the courseware page is determined using the coordinate values of the annotation area, and the annotation area coordinate values and annotation text content are associated using the courseware annotation number. All four types of medical education data include resource numbers and case numbers. Medical imaging data includes image frame numbers, lesion annotation numbers, and lesion region coordinate values. Image report data includes report sentence numbers and report sentence content. Medical teaching video data includes subtitle sentence numbers, subtitle sentence content, subtitle time intervals, and video frame sequences. Teaching courseware data includes courseware page numbers, courseware annotation numbers, annotation region coordinate values, and annotation text content.The system reads entity terms, entity synonyms, entity category codes, entity numbers, relation terms, relation type codes, stage terms, treatment stage codes, term priority values, and image entity templates from the medical knowledge ontology table. This enables text content to obtain medical entity numbers, relation type codes, and treatment stage codes through character matching, and enables lesion images to obtain lesion entity numbers, anatomical location numbers, and medical manifestation numbers through image entity template matching. It also reads relation type codes and corresponding evidence unit types from the medical relation rule table, ensuring that each relation type code has a clear correspondence to an evidence unit type, providing a rule-based foundation for subsequent judgments on whether relation primitives have all corresponding evidence unit types. After data reading is complete, check whether the resource number, case number, and each content index field are empty. Check whether the joint index composed of image frame number and lesion annotation number, report sentence number, subtitle sentence number, and courseware page number and courseware annotation number under the same resource number are duplicated. Check whether the lesion area coordinate value is within the boundary of the image frame, whether the annotation area coordinate value is within the boundary of the courseware page, whether the start timestamp of the subtitle time interval is earlier than the end timestamp, and whether the video frame sequence is arranged in ascending order of video frame timestamps. Only medical education data that passes the check is written into the data reading result, thereby ensuring that medical education data of different modalities have a unified case attribution, a clear content retrieval location, and verifiable spatial and temporal boundaries.
[0042] This implementation plan unifies resource numbers, case numbers, content index data, spatial coordinate standards, and temporal boundary standards, enabling medical imaging data, imaging report data, medical teaching video data, and teaching courseware data to be accurately collected within the same case scope. This establishes a stable data correspondence with the medical knowledge ontology table and the medical relationship rule table, providing a consistent data foundation for the generation of evidence units, the construction of relation primitives, and the writing of effective relation edges in the medical education knowledge graph. This improves the traceability, verifiability, and accuracy of knowledge graph construction for multimodal medical education data, and reduces evidence matching bias caused by data attribution errors, index conflicts, spatial out-of-bounds errors, and temporal misalignments.
[0043] Optionally, the specific steps for generating image evidence units, report evidence units, video evidence units, and courseware evidence units are as follows: Read the resource number, case number, image frame number, lesion annotation number, and lesion region coordinate values from the medical image data. Locate the target image frame from the image sequence associated with the resource number based on the image frame number. Check whether each boundary point in the lesion region coordinate values is within the pixel boundary of the target image frame. Determine the lesion truncation rectangle based on the minimum, maximum, minimum, and maximum abscissa values in the lesion region coordinate values, and retain the original pixel values within the lesion annotation boundary to ensure the truncation result includes the complete lesion region and reduces interference from irrelevant tissues around the lesion on subsequent image entity template matching. Read the resource number, case number, report sentence number, and report sentence content from the image report data. Locate the target report sentence in the image report associated with the resource number based on the report sentence number. Retain the character order and punctuation positions of the target report sentence, and delete the leading and trailing whitespace characters of the target report sentence to ensure the report sentence content is independent and traceable. Textual evidence at the location of the original image report; reading resource numbers, case numbers, subtitle number, subtitle content, subtitle time interval, and video frame sequence from medical teaching video data; locating the target subtitle based on the subtitle number and extracting its content; filtering video frames whose timestamps fall within the subtitle time interval based on the start and end timestamps of the subtitle time interval; arranging the filtered consecutive video frame sequence in ascending order of video frame timestamps; and binding the subtitle content to the consecutive video frames, so that the same video evidence unit simultaneously retains the medical explanation text and the medical images displayed during the explanation period; reading resource numbers, case numbers, courseware page numbers, courseware annotation numbers, annotation area coordinates, and annotation text content from teaching courseware data; locating the target courseware page based on the courseware page number; determining the cutoff boundary of the courseware annotation area based on the annotation area coordinates; extracting the annotation text content associated with the courseware annotation area based on the courseware annotation number; and binding the courseware annotation area to the annotation text content, so that the graphic annotations and text explanations on the courseware page maintain the same evidence attribution.Lesion images are extracted based on lesion region coordinates, report sentence content is extracted based on report sentence number, subtitle sentence content is extracted based on subtitle sentence number, and corresponding consecutive video frames are extracted from the video frame sequence according to subtitle time interval. The courseware annotation area is extracted based on annotation region coordinates, and annotation text content associated with the courseware annotation number is extracted, generating image evidence units, report evidence units, video evidence units, and courseware evidence units respectively. The data payload of the image evidence unit is written with the lesion region coordinates and lesion image, the data payload of the report evidence unit is written with the report sentence content, the data payload of the video evidence unit is written with the subtitle sentence content, the subtitle time interval, and the consecutive video frame sequence arranged according to the video frame timestamp, and the data payload of the courseware evidence unit is written with the annotation region coordinates, the courseware annotation area, and the annotation text content. This ensures that medical education content of different modalities is encapsulated at the smallest evidence granularity that can independently support medical entity recognition and relationship determination. The content index data for image evidence units is written with the image frame number and lesion annotation number; the content index data for report evidence units is written with the report sentence number; the content index data for video evidence units is written with the subtitle sentence number; and the content index data for courseware evidence units is written with the courseware page number and courseware annotation number. An evidence unit number, resource number, and case number are also written for each of the image evidence unit, report evidence unit, video evidence unit, and courseware evidence unit. Evidence unit numbers are generated according to a fixed field order of evidence unit type, resource number, case number, and content index data. This ensures that the evidence unit number is unique within the same medical education data management database. Each evidence unit can be returned to the original medical education resource based on the resource number, assigned to the corresponding medical teaching case based on the case number, and located in the original content position based on the content index data. After generating image evidence units, report evidence units, video evidence units, and courseware evidence units, check whether the evidence unit number, resource number, case number, and content index data are complete. Check whether the lesion image, report sentence content, subtitle sentence content, continuous video frames, courseware annotation area, and annotation text content are empty. Stop writing to the corresponding evidence unit when any required field is empty, thereby ensuring that the evidence units entering the subsequent medical entity number extraction and ordered relation primitive chain construction process all have clear data sources, complete evidence content, and verifiable original locations.
[0044] In this implementation plan, by uniformly encapsulating four types of medical education content into evidence units with evidence unit numbers, resource numbers, case numbers, and content index data, different modal contents have consistent minimum evidence granularity and stable traceability relationships. This provides reliable input for extracting medical entity numbers, constructing ordered relation primitive chains, and writing effective relation edges in the medical education knowledge graph. As a result, it improves the accuracy of evidence attribution, the reliability of cross-modal associations, and the accuracy of subsequent update positioning, and reduces the deviation in knowledge graph construction caused by original content mismatch, duplicate evidence writing, and unclear sources of relation support.
[0045] Optionally, the specific steps for extracting medical entity IDs and relation type codes to construct an ordered relation primitive chain are as follows: Read entity terms, entity synonyms, entity category codes, entity IDs, relation terms, relation type codes, stage terms, treatment stage codes, term priority values, and image entity templates from the medical knowledge ontology table, and establish term retrieval indexes according to lesion entities, anatomical locations, medical manifestations, conclusion entities, and treatment stages; map entity synonyms to corresponding entity IDs, and incorporate entity synonyms that have completed entity ID mapping into the term retrieval index of the corresponding entity category, so that entity terms and entity synonyms can return a unified entity ID using the same character matching method. The report sentence content, subtitle sentence content, and annotation text content are processed by unifying full-width and half-width characters, unifying uppercase and lowercase English letters, deleting consecutive whitespace characters, and expanding medical abbreviations. The text content after character standardization is scanned from front to back according to the character starting position. The continuous character fragments starting from the scan position are matched with the term retrieval index item by item. When multiple terms are matched at the same scan position, the term with the longest character length is retained. When the character lengths are the same, the term with the lowest term priority value in the medical knowledge ontology table is retained to avoid truncating the complete medical expression by terms with shorter character lengths. The standardized report sentences, subtitles, and annotations are matched against entity terms, relation terms, and stage terms in the medical knowledge ontology table. This process extracts the lesion entity number, anatomical location number, medical manifestation number, conclusion entity number, relation type code value, and treatment stage code value, which are then written into the report evidence unit, video evidence unit, and courseware evidence unit. During the writing of each number and code value, the matched term, the starting character sequence number, the ending character sequence number, and the evidence unit number are recorded simultaneously, ensuring that each medical entity number, relation type code value, and treatment stage code value can be traced back to its original character position within the corresponding evidence unit. When the same medical entity number is matched repeatedly within the same evidence unit, only one entity extraction record is retained. When character intervals corresponding to different medical entity numbers overlap, the entity extraction record with the longest character interval is retained, thereby reducing duplicate entity writing caused by repeated and nested terms. Extract the boundary contour values, major axis values, minor axis values, and grayscale statistical values of the lesion image slices, calculate the normalized distance with the image entity template in the medical knowledge ontology table, and write the lesion entity number, anatomical location number, and medical manifestation number in the image entity template into the image evidence unit when the minimum normalized distance is not greater than the entity matching distance threshold.Specifically, the centroid of the lesion region is calculated based on the coordinate sequence of the boundary points of the lesion annotation. Starting from the centroid of the lesion region, the radial distance from the centroid to the lesion boundary is collected at 5° angular intervals. Each radial distance is divided by the maximum value among all radial distances, and arranged in ascending order of the collection angle to form the boundary contour value. The 5° angular interval divides the 360° contour range into 72 collection directions, and the 5° angular interval is determined by the contour feature dimension of the image entity template and the density of the boundary points of the lesion annotation. The maximum Euclidean distance between any two points on the lesion annotation boundary is calculated as the major axis value, and the maximum span distance of the lesion region is calculated along the direction perpendicular to the major axis as the minor axis value. The grayscale mean, grayscale standard deviation, grayscale 10th percentile, and grayscale 90th percentile values of the pixels within the lesion annotation boundary are calculated, and the four values are combined into grayscale statistical values in a fixed field order of grayscale mean, grayscale standard deviation, grayscale 10th percentile, and grayscale 90th percentile. The 72 boundary contour values, major axis values, minor axis values, gray mean values, gray standard deviation values, gray 10th percentile values, and gray 90th percentile values, arranged in ascending order of acquisition angle, are sequentially concatenated to form the image feature sequence of the lesion image patch. The number of feature dimensions of the image feature sequence is denoted as n, which is 78 in this embodiment. The template boundary contour values, template major axis values, template minor axis values, and template gray statistical values are read from the m-th image entity template and concatenated to form the template feature sequence according to the same field order as the image feature sequence of the lesion image patch. From the manually annotated historical lesion image patch samples associated with the m-th image entity template, the feature mean μm,i and feature standard deviation σm,i of the i-th feature dimension are calculated respectively, and the feature mean μm,i and feature standard deviation σm,i are used as the normalization benchmark corresponding to the m-th image entity template. For the image feature sequence of the lesion image patch, the i-th original feature value is denoted as xi, and the i-th template feature value of the m-th image entity template is denoted as tm,i. The dimensionless feature value of the lesion image patch relative to the m-th image entity template is calculated as x′m,i=(xi-μm,i) / max(σm,i,10^-6), and the dimensionless template feature value of the m-th image entity template is calculated as t′m,i=(tm,i-μm,i) / max(σm,i,10^-6); where 10^-6 is the lower limit of numerical stability to prevent the feature standard deviation from being zero, which would lead to division by zero. The normalized distance Dm corresponding to the m-th image entity template is calculated as Dm=√[(1 / n)Σi=1n(x′m,i-t′m,i)²], where n represents the number of feature dimensions involved in the normalized distance calculation, and (1 / n) is used to eliminate the influence of the number of feature dimensions on the distance numerical scale. The normalized distance between each lesion image and each image entity template is calculated. The minimum value among all normalized distances is taken as the minimum normalized distance, and the image entity template that produces the minimum normalized distance is taken as the candidate image entity template.The entity matching distance threshold is determined by the 95th percentile of the normalized distance distribution calculated according to the same normalization rule for manually annotated historical lesion image samples associated with the candidate image entity template. This ensures that the entity matching distance threshold covers the normal morphological differences of lesions of the same type and limits the false hits of cross-category image entity templates. When the minimum normalized distance is not greater than the entity matching distance threshold, the lesion entity number, anatomical location number, and medical manifestation number in the candidate image entity template are written into the image evidence unit. When the minimum normalized distance is greater than the entity matching distance threshold, the image evidence unit is retained, and the lesion entity number, anatomical location number, and medical manifestation number are not written into the image evidence unit. Within the same case number, lesion entity nodes, anatomical location nodes, medical manifestation nodes, and conclusion entity nodes are generated based on the lesion entity number, anatomical location node, medical manifestation node, and conclusion entity number, respectively. The treatment stage code value is used as the treatment stage node number to generate the treatment stage node. Before generating entity nodes, existing entity nodes are retrieved according to the case number and entity node number. Specifically, the entity node numbers for lesion entity nodes, anatomical location nodes, medical manifestation nodes, and conclusion entity nodes are taken from their corresponding entity numbers, while the entity node number for treatment stage nodes is taken from the treatment stage node number. If the same case number and entity node number are found, the existing entity node is reused; otherwise, a new entity node is generated, and the case number and entity node number are written into the new entity node. For lesion entity nodes, anatomical location nodes, medical manifestation nodes, and conclusion entity nodes, the corresponding entity category code value and entity term are written into the new entity node. For treatment stage nodes, the treatment stage code value and stage term are written into the new entity node to avoid duplicate creation of the same medical entity in the same case. A positional relation primitive is generated based on the lesion entity number and anatomical location number; a manifestation relation primitive is generated based on the lesion entity number and medical manifestation number; a conclusion relation primitive is generated based on the medical manifestation number and conclusion entity number; and a stage relation primitive is generated based on the conclusion entity number and treatment stage node number. The head entity number of the positional relation primitive is written to the lesion entity number, and the tail entity number of the positional relation primitive is written to the anatomical location number. The head entity number of the manifestation relation primitive is written to the lesion entity number, and the tail entity number of the manifestation relation primitive is written to the medical manifestation number. The head entity number of the conclusion relation primitive is written to the medical manifestation number, and the tail entity number of the conclusion relation primitive is written to the conclusion entity number. The head entity number of the stage relation primitive is written to the conclusion entity number, and the tail entity number of the stage relation primitive is written to the treatment stage node number. These are then written into the relation type code value in the medical knowledge ontology table that matches the positional relation, manifestation relation, conclusion relation, and stage relation, respectively.Connect the positional relation primitives of the shared lesion entity number to the manifestation relation primitives, connect the manifestation relation primitives of the shared medical manifestation number to the conclusion relation primitives, and connect the conclusion relation primitives of the shared conclusion entity number to the stage relation primitives to generate an ordered relation primitive chain. Write the relation primitive number, case number, head entity number, tail entity number, relation type code value, and intra-chain position value for each relation primitive. Specifically, only adjacent relation primitives with the same case number and the same shared medical entity number are connected. Relation primitives in the same ordered relation primitive chain are arranged according to a fixed medical semantic order of position relation primitives, manifestation relation primitives, conclusion relation primitives, and stage relation primitives. The intra-chain position value of position relation primitives is written as 1, the intra-chain position value of manifestation relation primitives is written as 2, the intra-chain position value of conclusion relation primitives is written as 3, and the intra-chain position value of stage relation primitives is written as 4. The intra-chain position value is used to determine the coverage start position value and coverage end position value of the evidence cross-primary primitive coverage band, the incision start position value and incision end position value of the relation chain break incision, and the arrangement order of relation primitive numbers within the relation primitive chain segment. The intra-chain position value of the same relation primitive remains unchanged during evidence unit updates, evidence cross-primary primitive coverage band replacements, and relation primitive local write-back processes. Relational primitive numbers are generated by concatenating fixed fields such as case number, head entity number, relation type code value, and tail entity number. When the same relational primitive number is generated repeatedly, only one relational primitive record is retained. When there are multiple lesion entity numbers within the same case number, a shared medical entity number connection is performed with each lesion entity number as the chain starting point to prevent cross-connection of anatomical location, medical manifestation, conclusion entity, and treatment stage corresponding to different lesion entities. This ensures that the ordered relational primitive chain accurately represents the progressive medical semantics of a single lesion entity within the same case, from anatomical location, medical manifestation, conclusion entity to treatment stage.
[0046] In this implementation plan, report evidence units, video evidence units, courseware evidence units, and image evidence units are uniformly mapped to lesion entity numbers, anatomical location numbers, medical manifestation numbers, conclusion entity numbers, relation type codes, and treatment stage codes. An ordered relation primitive chain is constructed according to the shared medical entity numbers, so that the medical entities, relation types, and treatment stages within the same case form a stable and traceable progressive semantic structure. This improves the matching accuracy of entity nodes and relation primitives in the medical education knowledge graph, avoids cross-connections between different lesion entities, and provides a reliable foundation for subsequent evidence cross-primary coverage generation, effective relation edge determination, and local update of relation chains.
[0047] Optionally, the relationship primitives continuously supported by the same evidence unit are connected according to the intra-chain positional value to form an evidence cross-primary primitive coverage band. The specific steps for truncating at the positional interruption are as follows: Read each evidence unit and the ordered relationship primitive chain; read the evidence unit number, resource number, case number, and the written medical entity number and relationship type code value from each evidence unit; read the relationship primitive number, case number, head entity number, tail entity number, relationship type code value, and intra-chain positional value from the ordered relationship primitive chain; first, limit the matching range between evidence units and relationship primitives according to the case number; then, within each ordered relationship primitive chain corresponding to the same case number, establish the support mapping between evidence units and relationship primitives one by one according to the evidence unit number to prevent cross-matching of evidence units and relationship primitives with different case numbers. For report evidence units, video evidence units, and courseware evidence units, the head entity number, relation type code value, and tail entity number in the same evidence unit are compared item by item with the head entity number, relation type code value, and tail entity number of each relation primitive under the same case number. Only when all three items are consistent is the corresponding evidence unit confirmed to support the corresponding relation primitive, and the evidence unit number is written into the corresponding relation primitive. When there are multiple sets of head entity numbers, relation type code values, and tail entity numbers in the same evidence unit, each set of numbers is compared with each relation primitive one by one. Relation primitives that meet the condition of consistency in all three items are all supported by a support mapping. A single relation primitive is not selected from multiple relation primitives that meet the condition. If the same evidence unit number has already been written into the same relation primitive, it will not be written again, so that the same evidence unit forms only one support record for the same relation primitive. For each image evidence unit, the lesion entity number and anatomical location number in the image evidence unit are compared with the head entity number and tail entity number of each positional relationship primitive under the same case number. The relationship type code value of the positional relationship primitive is confirmed to be the relationship type code value corresponding to the positional relationship. When the three items match, the image evidence unit number is written into the corresponding positional relationship primitive. The lesion entity number and medical manifestation number in the image evidence unit are compared with the head entity number and tail entity number of each manifestation relationship primitive under the same case number. The relationship type code value of the manifestation relationship primitive is confirmed to be the relationship type code value corresponding to the manifestation relationship. When the three items match, the image evidence unit number is written into the corresponding manifestation relationship primitive. Thus, the support of the same lesion image for the lesion location and lesion manifestation is written into the corresponding relationship primitive. After the evidence unit number is written, the same ordered relation primitive chain corresponding to the same case number is used as the processing scope, and the evidence unit number is used as the grouping field to extract all relation primitives that have been written for the same evidence unit number, delete duplicate relation primitive numbers, and arrange them in ascending order according to the position value in the chain; the same continuous position segment is composed only of relation primitives with the same evidence unit number and adjacent position values in the chain.The current continuous position segment is established with the first relation primitive after the arrangement, and the intra-chain position value of the current relation primitive is compared with the intra-chain position value of the previous relation primitive. If the evidence unit number corresponding to the current relation primitive and the previous relation primitive are the same, and the difference between the current intra-chain position value and the previous intra-chain position value is 1, the current relation primitive is merged into the current continuous position segment. If the difference between the current intra-chain position value and the previous intra-chain position value is greater than 1, it indicates that there is an intra-chain position between the two relation primitives that is not supported by the same evidence unit, the current continuous position segment ends, and a new continuous position segment is established with the current relation primitive. If the evidence unit number corresponding to the current relation primitive and the previous relation primitive are different, the two relation primitives are not classified into the same continuous position segment. When two relation primitives with intra-chain position values j and j+2 are both supported by the same evidence unit, but a relation primitive with intra-chain position value j+1 is not supported by the same evidence unit, the current continuous position segment ends at the intra-chain position value j, and a new continuous position segment is established starting from the relation primitive with intra-chain position value j+2. No other evidence units are used to supplement the two continuous position segments for the supporting record with intra-chain position value j+1. A single relation primitive also constitutes a continuous position segment, and the starting position value and ending position value of the continuous position segment are both taken from the intra-chain position value of the single relation primitive. Based on the intra-chain positional value, the relational primitives supported by the same evidence unit are divided into consecutive positional segments. Evidence cross-primary primitive coverage bands are generated based on each consecutive positional segment, and the coverage start positional value, coverage end positional value, and evidence unit number are written into them. The coverage start positional value is taken from the intra-chain positional value of the first relational primitive in the consecutive positional segment, and the coverage end positional value is taken from the intra-chain positional value of the last relational primitive in the consecutive positional segment. When the same evidence unit forms multiple non-adjacent consecutive positional segments in the same ordered relational primitive chain, multiple evidence cross-primary primitive coverage bands are generated respectively. When the same evidence unit forms consecutive positional segments in different ordered relational primitive chains, evidence cross-primary primitive coverage bands are generated separately according to each ordered relational primitive chain. Supporting records formed by different evidence unit numbers are not merged into the same evidence cross-primary primitive coverage band. The evidence cross-element coverage band is used to characterize the continuous support range of the same evidence unit in the same ordered relational element chain, so that subsequent continuous coverage marks of the same evidence can be generated based on the real support results with the same evidence unit number and continuous positional value in the chain, preventing the discrete support of different evidence units and the discrete support formed by the same evidence unit across missing positions from being incorrectly spliced into a continuous relation.
[0048] In this implementation plan, by using the same evidence unit as a continuous support source and utilizing the evidence cross-element coverage band to solidify the true support range of the same evidence unit in the ordered relational element chain, the medical education knowledge graph can accurately distinguish between continuous evidence support and discrete evidence splicing, avoiding false continuous relationships between different case numbers and different evidence units, thereby improving the credibility of the generation of continuous coverage markers of the same evidence, the accuracy of effective relational edge determination, and the stability of subsequent relational chain break identification.
[0049] Optionally, the specific steps for generating continuous coverage markers for the same evidence and writing the relation primitives supported by all corresponding evidence unit types into the effective relation edge table are as follows: Read the evidence unit number, coverage start position value, and coverage end position value in each evidence cross-primary coverage band; read the case number from the corresponding evidence unit according to the evidence unit number; determine the relation primitives covered by the evidence cross-primary coverage band and the inter-chain boundaries crossed within the ordered relation primitive chain to which the evidence cross-primary coverage band belongs; read the relation primitive number, case number, head entity number, tail entity number, relation type code value, and intra-chain position value of each relation primitive in the ordered relation primitive chain; determine the relation primitives whose intra-chain position values are within the range of coverage start position value to coverage end position value as the relation primitives covered by the evidence cross-primary coverage band; and determine the position between two relation primitives with adjacent intra-chain position values in the same ordered relation primitive chain as the inter-chain boundary; generate the inter-chain boundary number according to the fixed field order of case number, left relation primitive number, and right relation primitive number. The inter-chain boundary crossed by the evidence cross-element coverage band is determined based on the coverage start position value and the coverage end position value. For a left relation element with an intra-chain position value of j and a right relation element with an intra-chain position value of j+1, when the coverage start position value is not greater than j and the coverage end position value is not less than j+1, it is determined that the same evidence cross-element coverage band simultaneously covers relation elements on both sides of the inter-chain boundary, and the evidence cross-element coverage band is determined to cross the corresponding inter-chain boundary. When the evidence cross-element coverage band only covers relation elements on one side of the inter-chain boundary, it is not determined that the evidence cross-element coverage band crosses the corresponding inter-chain boundary. This ensures that the continuous support of the inter-chain boundary comes from a single evidence cross-element coverage band formed by the same evidence unit, preventing adjacent relation elements from being judged as continuously supported by the same evidence when they are supported by different evidence units. When the same evidence cross-element coverage band simultaneously covers relational primitives on both sides of the inter-chain boundary, a continuous coverage marker for the same evidence is written to the inter-chain boundary, and the inter-chain boundary number, left relational primitive number, right relational primitive number, evidence unit number, coverage start position value, and coverage end position value are written simultaneously. When the same inter-chain boundary is crossed by multiple evidence cross-element coverage bands, the corresponding continuous coverage records are retained according to the evidence unit number. When the same continuous coverage record corresponding to the same evidence unit number is generated repeatedly, only one record is retained. The medical relationship rule table is used to record the mapping relationship between relationship type encoding values and corresponding evidence unit types. Each rule record includes a relationship type encoding value and a corresponding evidence unit type. When the same relationship type encoding value corresponds to multiple evidence unit types, multiple rule records are established respectively, and they are grouped into a set of corresponding evidence unit types according to the same relationship type encoding value. In this embodiment, the position relationship type encoding value corresponds to the image evidence unit and the report evidence unit, the presentation relationship type encoding value corresponds to the image evidence unit and the video evidence unit, the conclusion relationship type encoding value corresponds to the report evidence unit and the courseware evidence unit, and the stage relationship type encoding value corresponds to the report evidence unit and the courseware evidence unit.When reading the medical relationship rule table, the relationship type code value of the relationship primitive is used as the retrieval field. All rule records matching the relationship type code value are extracted. The corresponding evidence unit type in each rule record is read. After deleting duplicate corresponding evidence unit types, all corresponding evidence unit types of the current relationship primitive are formed. Evidence cross-primary primitive coverage bands covering the current relationship primitive are collected according to the relationship primitive number. The evidence unit type of the corresponding evidence unit is read according to the evidence unit number in each evidence cross-primary primitive coverage band. After deleting duplicate evidence unit types, a set of evidence unit types that have been supported by the relationship primitive is formed. All corresponding evidence unit types of the current relationship primitive are compared item by item with the set of evidence unit types that have been supported by the relationship primitive. If each corresponding evidence unit type can be retrieved in the set of evidence unit types that have been supported by the relationship primitive, it is determined that the relationship primitive has been supported by all corresponding evidence unit types. If there is a corresponding evidence unit type that cannot be retrieved in the set of evidence unit types that have been supported by the relationship primitive, the missing evidence unit type is recorded, and it is determined that the relationship primitive has not been supported by all corresponding evidence unit types. When all evidence unit types corresponding to a relation primitive have cross-primary coverage bands that cover the current relation primitive, the relation primitive is written into the valid relation edge table. The relation primitive number, case number, head entity number, tail entity number, relation type code value, intra-chain position value, and evidence unit number set are written into the same valid relation edge record. The evidence unit number set consists of the evidence unit numbers corresponding to each cross-primary coverage band that actually covers the current relation primitive, and duplicate evidence unit numbers are deleted. This ensures that the valid relation edge simultaneously records entity connection relationships, intra-chain positions, and sources of evidence support. Relation primitives that do not receive support from all corresponding evidence unit types remain in the ordered relation primitive chain, are not written into the valid relation edge table, and retain the missing evidence unit types to provide a basis for subsequent identification of uncovered primitives. When adjacent relation primitives have continuous coverage markers of the same evidence, the connection between adjacent relation primitives is written into the chain adjacency record table. When writing into the chain adjacency record table, the case number, the inter-chain boundary number, the left relation primitive number, the right relation primitive number, the continuous coverage marker of the same evidence, and the set of evidence unit numbers forming continuous coverage are written into the same chain adjacency record. When multiple evidence units form continuous coverage markers of the same evidence at the same inter-chain boundary, the corresponding evidence unit number is written into the set of evidence unit numbers forming continuous coverage. Thus, the effective relation edges and chain adjacency records are determined according to the support status of all corresponding evidence unit types and the continuous coverage status of the same evidence, respectively, providing a verifiable data foundation for subsequent identification of missing primitives, identification of missing boundaries, and local rewriting of relation primitives.
[0050] In this implementation plan, by solidifying the integrity of the corresponding evidence unit type and the continuous coverage status of the same evidence into effective relation edges and chain adjacency records respectively, the credibility of individual relations and the continuity of relation chains in the medical education knowledge graph have clear evidentiary basis. This can reduce relation distortion caused by splicing discrete evidence, missing evidence types, and misjudgment of inter-chain connections, and improve the accuracy and traceability of effective relation edge screening, relation chain break identification, and subsequent local write-back.
[0051] Optionally, the specific steps for obtaining medical education data that is inconsistent with the original evidence unit record and generating updated evidence units and updated evidence cross-element coverage bands are as follows: Read the newly written records in the medical education data management database according to the data writing timestamp, and extract the resource number, case number, and content index data from the newly written records; use the resource number and content index data to form a joint search field, and search for the original evidence unit in the medical education data management database. When an original evidence unit with consistent resource number and content index data is found, the newly written record and the original evidence unit are included in the update verification process. When no original evidence unit is found, the newly written record is retained until the new evidence unit generation process. After retrieving the original evidence unit, the original evidence unit number, resource number, case number, content index data, evidence unit type, and data payload are read, and the comparison fields are determined according to the evidence unit type. The comparison fields for the image evidence unit include the coordinate values of the lesion area represented by the coordinate sequence of the lesion annotation boundary points and the pixel values of the lesion image. The comparison fields for the report evidence unit include the report sentence content after character standardization. The comparison fields for the video evidence unit include the subtitle sentence content after character standardization, the subtitle time interval, and the video frame sequence arranged in ascending order of video frame timestamps. The comparison fields for the courseware evidence unit include the coordinate values of the annotation area, the pixel values of the courseware annotation area, and the annotation text content after character standardization. Character standardization includes unifying full-width and half-width characters, unifying uppercase and lowercase English letters, deleting consecutive whitespace characters, and expanding medical abbreviations. The coordinate comparison fields in the original evidence unit and the newly written record are unified to the same coordinate system, and compared item by item according to the order of coordinate points. The text comparison fields are compared item by item according to the character sequence number. The image comparison fields are compared item by item according to the order of pixel row number and pixel column number within the same pixel row. The video frame sequence is compared item by item according to the order of video frame timestamp, comparing the number of video frames, video frame timestamp, number of pixel rows and columns, and video frame pixel value in ascending order of video frame timestamp. When the number of data and the data value at the corresponding position of the corresponding comparison field are consistent, it is determined that the newly written record is consistent with the content of the original evidence unit record, the original evidence unit and the original evidence cross-element coverage band are maintained, and no update writing is performed. When the number of data or the data value of any corresponding comparison field are inconsistent, it is determined that the newly written record is inconsistent with the content of the original evidence unit record, and the process of updating the evidence unit generation begins. When the newly written record is inconsistent with the original evidence unit record, the evidence content is re-extracted according to the newly written record, and the resource number, case number and content index data of the original evidence unit are used. An updated evidence unit number is generated according to the original evidence unit number and the data timestamp of the newly written record. According to the evidence unit type, the new lesion image is written into the updated imaging evidence unit, the new report sentence content is written into the updated report evidence unit, the new subtitle sentence content and the new video frame sequence are written into the updated video evidence unit, and the new courseware annotation area and the new annotation text content are written into the updated courseware evidence unit.Based on the character matching methods corresponding to the report evidence unit, video evidence unit, and courseware evidence unit, the lesion entity number, anatomical location number, medical manifestation number, conclusion entity number, relationship type code value, and treatment stage code value are re-extracted; based on the image entity template matching method corresponding to the image evidence unit, the boundary contour value, major diameter value, minor diameter value, and grayscale statistical value are re-extracted, and the lesion entity number, anatomical location number, and medical manifestation number are determined based on the minimum normalized distance. The head entity number, relation type code value, and tail entity number in the updated report evidence unit, updated video evidence unit, and updated courseware evidence unit are compared item by item with the head entity number, relation type code value, and tail entity number of each relation primitive under the same case number. If all three data are consistent, the updated evidence unit is confirmed to support the corresponding relation primitive. The lesion entity number and anatomical location number in the updated imaging evidence unit are compared with the head entity number and tail entity number of the positional relation primitive, and the lesion entity number and medical manifestation number in the updated imaging evidence unit are compared with the head entity number and tail entity number of the manifestation relation primitive. If the numbers are consistent, the updated imaging evidence unit is confirmed to support the corresponding positional relation primitive and manifestation relation primitive. Before deleting the evidence cross-primary primitive coverage band formed by the original evidence unit, the inter-chain boundaries crossed by the evidence cross-primary primitive coverage band corresponding to the original evidence unit are read. The current continuous coverage markers of the same evidence at each inter-chain boundary are saved as the continuous coverage markers of the same evidence before the update. The inter-chain boundary number, the left relation primitive number, the right relation primitive number, and the original evidence unit number are written into the corresponding continuous coverage record before the update. Delete all cross-element coverage bands formed by the original evidence unit. During deletion, the original evidence unit number is used as the deletion condition; only cross-element coverage bands with the same evidence unit number as the original evidence unit number are deleted, while cross-element coverage bands formed by other evidence units on the same relational element are retained. The relational elements supported by the updated evidence unit are arranged in ascending order of their intra-chain positional values. When the updated evidence unit numbers are the same and the difference in intra-chain positional values between adjacent chains is 1, the corresponding relational elements are grouped into the same continuous positional segment. When the difference in intra-chain positional values between adjacent chains is greater than 1, the current continuous positional segment ends, and a new continuous positional segment is established starting from the next relational element. When the updated evidence unit only supports a single relational element, the single relational element is determined as a continuous positional segment. When the updated evidence unit does not support any relational elements, no updated evidence cross-element coverage bands are generated. Based on each consecutive positional segment, an updated evidence cross-element coverage band is generated. The intra-chain positional value of the first relational element in the consecutive positional segment is written into the coverage start positional value, the intra-chain positional value of the last relational element in the consecutive positional segment is written into the coverage end positional value, and the updated evidence unit number is written in.The updated evidence cross-element coverage band and the retained evidence cross-element coverage band are used together as the current evidence coverage data. Based on the current evidence coverage data, the necessary evidence unit type coverage status of each relational element and the continuous coverage status of the same evidence at each chain boundary are re-determined. The current evidence cross-element coverage bands covering the corresponding relational elements are collected according to the relational element number. The evidence unit type is read according to the evidence unit number in the current evidence cross-element coverage band and compared item by item with the evidence unit type corresponding to the relational type code value in the medical relational rule table. When all corresponding evidence unit types are supported, the necessary evidence unit type coverage status is recorded as complete. When there is a corresponding evidence unit type that is not supported, the necessary evidence unit type coverage status is recorded as incomplete, and the missing evidence unit type is recorded. According to the chain boundary number, it is checked whether there is a single evidence cross-element coverage band in the current evidence coverage data that simultaneously covers the relational elements on both sides of the chain boundary. If it exists, the continuous coverage status of the same evidence is retained. If it does not exist, the continuous coverage status of the same evidence is canceled. Thus, the evidence support range formed by the updated evidence unit replaces the original evidence unit, while keeping the evidence cross-element coverage bands formed by other evidence units unchanged.
[0052] In this implementation plan, by replacing the local evidence support of the original evidence unit with the updated evidence unit and simultaneously reconstructing the cross-element coverage of the updated evidence, the medical education knowledge graph can accurately reflect the actual impact of changes in medical education data content on relational primitives and inter-chain boundaries, while keeping the cross-element coverage of unaffected evidence unchanged. This improves the accuracy of updating the coverage status of necessary evidence unit types and the continuous coverage status of the same evidence, and reduces the risk of large-scale relation reconstruction and accidental deletion of valid evidence caused by local content changes.
[0053] Optionally, the specific steps for identifying missing primitives and disconnected boundaries to generate relationship chain break points are as follows: Read the updated necessary evidence unit type coverage status, continuous coverage status of the same evidence, ordered relationship primitive chains, and the previous continuous coverage markers of the same evidence, and process each ordered relationship primitive chain according to the case number. For each relationship primitive, read the relationship primitive number, relationship type code value, and intra-chain position value. Based on the relationship type code value, read the corresponding evidence unit type from the medical relationship rule table. Then, read the evidence cross-primary primitive coverage band currently covering the relationship primitive and the evidence unit types associated with each evidence cross-primary primitive coverage band. Compare each corresponding evidence unit type with the evidence unit type currently covering the relationship primitive item by item. If all corresponding evidence unit types are supported, the relationship primitive is not recorded as a missing primitive. If not all corresponding evidence unit types are supported, the relationship primitive is recorded as a missing primitive, and the relationship primitive number, intra-chain position value, and unsupported evidence unit types are written into the missing primitive record. If the same relationship primitive lacks multiple corresponding evidence unit types, record each missing evidence unit type separately. For each inter-chain boundary, read the inter-chain boundary number, the left relation primitive number, the right relation primitive number, the intra-chain positional value of the left relation primitive, the intra-chain positional value of the right relation primitive, the continuous coverage mark of the same evidence before the update, and the cross-primary primitive coverage band of the evidence after the update. Check whether there is a single cross-primary primitive coverage band of evidence in the updated cross-primary primitive coverage band where the starting positional value of the coverage is not greater than the intra-chain positional value of the left relation primitive and the ending positional value of the coverage is not less than the intra-chain positional value of the right relation primitive. If the inter-chain boundary had a continuous coverage mark of the same evidence before the update, and there is no single cross-primary primitive coverage band of evidence covering both relation primitives on both sides of the inter-chain boundary after the update, record the inter-chain boundary as a lost boundary; if the inter-chain boundary did not have a continuous coverage mark of the same evidence before the update, do not record the inter-chain boundary as a lost boundary; if there is still a single cross-primary primitive coverage band of evidence covering both relation primitives on both sides of the inter-chain boundary after the update, retain the continuous coverage status of the same evidence of the inter-chain boundary. After identifying the missing primitives and disconnected boundaries, each missing primitive is mapped to a primitive anomaly interval. The starting and ending position values of the primitive anomaly interval are both taken from the position values within the chain of the missing primitive. Each disconnected boundary is mapped to a boundary anomaly interval. The starting position value of the boundary anomaly interval is taken from the position value within the chain of the left relation primitive, and the ending position value of the boundary anomaly interval is taken from the position value within the chain of the right relation primitive.Arrange the primitive abnormal intervals and boundary abnormal intervals in ascending order of their starting position values. Establish the current relationship chain break incision with the first abnormal interval after the arrangement. If the starting position value of the next abnormal interval is not greater than the termination position value of the current relationship chain break incision plus 1, the next abnormal interval is merged into the current relationship chain break incision, and the termination position value is updated to the larger of the termination position value of the current incision and the termination position value of the next abnormal interval. If the starting position value of the next abnormal interval is greater than the termination position value of the current relationship chain break incision plus 1, the current relationship chain break incision ends, and a new relationship chain break incision is established with the next abnormal interval. Adjacent uncovered primitives and lost-connection boundaries are merged according to their intra-chain positional values to generate a relationship chain break. The break's starting positional value, ending positional value, uncovered primitive number, missing evidence unit type, and lost-connection boundary number are then written into the break. The starting positional value is the minimum starting positional value among all abnormal intervals included in the break, and the ending positional value is the maximum ending positional value among all abnormal intervals. When the same relationship chain break contains multiple uncovered primitives and lost-connection boundaries, each uncovered primitive number, each missing evidence unit type, and each lost-connection boundary number is retained, ensuring that the break defines the continuous intra-chain intervals where the evidence support status changes. If no uncovered primitives or lost-connection boundaries are formed, a relationship chain retention marker is generated. The case number and the relationship primitive number sequence arranged in ascending order of intra-chain positional values are written into the relationship chain retention marker, which is then written into the corresponding ordered relationship primitive chain record for the relationship primitive local write-back module to read and enter the retention processing path. For uncovered primitives, a relation primitive suspension marker is generated. The uncovered primitive number, missing evidence unit type, and corresponding relation chain break cut are written into the relation primitive suspension marker, and the relation primitive suspension marker is written into the corresponding relation primitive record for the relation primitive local write-back module to read and enter the suspension processing path. When the necessary evidence unit types of relation primitives on both sides of the lost boundary are complete, a relation chain split marker is generated for the lost boundary. The lost boundary number, left relation primitive number, right relation primitive number, and corresponding relation chain break cut are written into the relation chain split marker, and the relation chain split marker is written into the corresponding inter-chain boundary record for the relation primitive local write-back module to read and enter the split processing path. When a relation primitive on either side of the lost boundary is recorded as an uncovered primitive, no relation chain split marker is generated. The uncovered primitive and its adjacent records are processed by the suspension processing path corresponding to the relation primitive suspension marker, thus establishing a correspondence between the relation chain retention marker, relation primitive suspension marker, and relation chain split marker with the subsequent retention processing path, suspension processing path, and split processing path, respectively.
[0054] In this implementation plan, by establishing the missing primitives, disconnected boundaries, and broken relationship chains as corresponding breakage state records, the medical education knowledge graph can accurately distinguish between the failure of relationship primitive evidence and the failure of continuous support between chains, avoiding the amplification of local evidence changes into the failure of the entire relationship chain. This improves the precision of relationship chain anomaly localization, the accuracy of breakage cause determination, and the reliability of subsequent relationship primitive suspension, chain adjacency record deletion, and relationship primitive chain fragment generation.
[0055] Optionally, the specific steps for deleting the chain adjacency record at the lost connection boundary position and generating a relationship primitive chain fragment based on the relationship chain break incision suspension of the lost coverage primitive are as follows: read the relationship chain retention mark, relationship primitive suspension mark and relationship chain split mark, and when there is a relationship chain break incision, read the incision start position value, incision end position value, lost coverage primitive number, missing evidence unit type and lost connection boundary number; locate the effective relationship edge, chain adjacency record, evidence cross-primary coverage band and relationship primitive in the same ordered relationship primitive chain according to the case number, and enter the retention processing path, suspension processing path and split processing path according to the three types of marks respectively. Among them, the suspension processing path and split processing path only operate on the intra-chain position range defined by the relationship chain break incision. The process reads the relationship chain retention flag, maintains valid relationship edges and chain adjacency records, and deletes the original evidence unit number from the relationship primitive. When deleting the original evidence unit number, the original evidence unit is determined based on the resource number and content index data. Only the evidence supporting record with the same original evidence unit number is deleted; other evidence unit numbers are not deleted, and the relationship primitive number, head entity number, relationship type encoding value, tail entity number, and intra-chain position value are not changed. When the updated evidence unit still supports the original relationship primitive, the updated evidence unit number is written to the relationship primitive, and other evidence unit numbers and evidence cross-primary overlay bands are retained. Specifically, the process checks whether the updated evidence unit number has been written to the relationship primitive based on the relationship primitive number supported by the updated evidence unit. If it has not been written, a write operation is performed; if it has been written, the original record is maintained. When the updated evidence unit no longer supports the original relationship primitive, but other evidence units still supplement all necessary evidence unit types, the updated evidence unit number is not written to the relationship primitive, and valid relationship edges and chain adjacency records are retained, thereby avoiding the impact of changes in the content of a single evidence unit on the valid relationship maintained by other evidence units. Read the suspended markers of relation primitives, remove the uncovered primitives from the effective relation edge table, delete the chain adjacency records connected to the uncovered primitives, extract the remaining evidence unit numbers from the evidence cross-primary primitive coverage bands retained by the uncovered primitives, and write the relation primitive number, head entity number, relation type code value, tail entity number, missing evidence unit type, and remaining evidence unit number into the suspended relation primitive table. When removing from the effective relation edge table, delete the corresponding effective relation edge record according to the uncovered primitive number, but retain the original relation primitive records and entity association information in the ordered relation primitive chain. When extracting the remaining evidence unit numbers, exclude the original evidence unit numbers and retain the evidence unit numbers associated with the evidence cross-primary primitive coverage bands that still cover the uncovered primitives, so that the suspended relation primitive table can record the existing evidence support and the missing evidence unit types that have not yet been supplemented, providing a directly searchable data foundation for restoring effective relation edges with newly added evidence.Relationship primitives and chain adjacency records outside the break in the relationship chain remain unchanged. Specifically, only relationship primitives with intra-chain position values within the range of the break in the starting position value to the ending position value of the break, and chain adjacency records pointed to by the loss boundary numbers, are processed. Valid relationship edges outside the break in the relationship chain are not recalculated, and the evidence unit numbers and evidence cross-primary coverage zones outside the break in the relationship chain are not changed, thus limiting relationship updates to the actual failure location. The relationship chain splitting marker is read, valid relationship edges on both sides of the loss boundary are retained, chain adjacency records at the loss boundary position are deleted, and relationship primitive chain fragments are generated with each loss boundary as a demarcation. For each relationship primitive chain fragment, a chain fragment number and a sequence of relationship primitive numbers arranged in ascending order of intra-chain position values are written. When generating relation primitive chain fragments, relation primitives still in the valid relation edge table are scanned according to their intra-chain positional values from smallest to largest. If there is a chain adjacency record between the current relation primitive and the next relation primitive, the next relation primitive is merged into the current relation primitive chain fragment. If there is no chain adjacency record between the current relation primitive and the next relation primitive, the current relation primitive chain fragment is terminated and a new relation primitive chain fragment is generated starting from the next relation primitive. Chain fragment numbers are generated according to the case number, the chain fragment start positional value, and the chain fragment end positional value. The numbers of each relation primitive are written into the relation primitive number sequence according to their intra-chain positional values from smallest to largest, so that each relation primitive chain fragment maintains continuous positional order, complete valid relation edges, and traceable evidence support sources. By performing local processing on relation chain preservation, relation primitive suspension, and relation chain splitting respectively, it is possible to isolate invalid relations, cut off lost connections, and retain valid chain segments without reconstructing the entire ordered relation primitive chain. This provides a clear local write-back boundary for subsequent recovery of valid relation edges and chain adjacency records based on newly added evidence units.
[0056] In this implementation scheme, by defining the local processing scope with the break incision of the relationship chain, and by adopting differentiated write-back methods for relationship chain preservation, relationship primitive suspension, and relationship chain splitting, the medical education knowledge graph update process can retain the effective relationship edges outside the break incision of the relationship chain, and keep the chain adjacency records in the chain adjacency record table and the evidence cross-primary coverage band in the evidence coverage data unchanged. At the same time, the remaining evidence unit numbers and missing evidence unit types of the uncovered primitives are completely preserved, thereby reducing the diffusion impact of local evidence failure on the entire ordered relationship primitive chain, improving the accuracy of effective relationship preservation, the reliability of relationship primitive chain fragment generation, and the efficiency of subsequent new evidence recovery.
[0057] In this embodiment, the ordered relation primitive chain includes position relation primitive K1, performance relation primitive K2, conclusion relation primitive K3, and stage relation primitive K4, with position values of 1, 2, 3, and 4 in sequence. Here, K1 to K4 only represent relation primitive numbers. The image evidence unit, report evidence unit, video evidence unit, and courseware evidence unit are numbered E-IMG-01, E-REP-01, E-VID-01, and E-PPT-01, respectively. The updated image evidence unit and updated report evidence unit generated after the content recorded by the original evidence unit is inconsistent with the original evidence unit are numbered E-IMG-02 and E-REP-02, respectively.
[0058] like Figure 2 As shown, before the update, E-IMG-01 forms an evidence cross-element coverage band, covering K1 to K2; E-REP-01 forms an evidence cross-element coverage band, covering K1 to K4; E-VID-01 supports K2; E-PPT-01 forms an evidence cross-element coverage band, covering K3 to K4; K1 to K4 remain continuously connected. After the medical education data was updated, E-IMG-02 only supports K2, and E-REP-02 forms two consecutive positional segments, one of which covers K1 and the other covers K3 to K4. E-VID-01 continues to support K2, and E-PPT-01 continues to cover K3 to K4. K1 is identified as a missing primitive due to the lack of corresponding image evidence unit support and is written into the suspended relation primitive table. At the same time, the chain adjacency record between K1 and K2 is deleted. The cross-primary coverage band of report evidence that spanned K2 and K3 before the update disappeared after the update, forming a missing boundary between K2 and K3. After deleting the chain adjacency record at the missing boundary position, a relation primitive chain fragment P1 containing only K2 is generated. K3 and K4 are still simultaneously covered by the continuous coverage formed by E-REP-02 and E-PPT-01. The same evidence continuous coverage mark between K3 and K4 is retained, and a relation primitive chain fragment P2 is generated.
[0059] Optionally, based on the newly added evidence units, the specific steps for restoring valid relation edges and chain adjacency records, and writing teaching resource nodes, evidence unit nodes, entity nodes, valid relation edges, and resource inclusion edges into the medical education knowledge graph are as follows: Read the valid relation edge table, chain adjacency record table, suspended relation primitive table, entity node records, each evidence unit, and each evidence cross-primary overlay band. Establish a retrieval index according to case number, resource number, and relation primitive number, enabling teaching resources, evidence units, and relation primitives to be associated within the same case number range. Generate teaching resource nodes based on resource numbers, generate evidence unit nodes based on evidence unit numbers, establish resource inclusion edges pointing from teaching resource nodes to evidence unit nodes, write the evidence unit number into the evidence support field of the valid relation edge, and write entity nodes, teaching resource nodes, evidence unit nodes, valid relation edges, and resource inclusion edges into the medical education knowledge graph. When generating teaching resource nodes, the resource number is used as the unique identifier. If a teaching resource node with the same resource number exists in the medical education knowledge graph, the existing node is reused; otherwise, a new node is created, and the case number and medical education data type are written into it. When generating evidence unit nodes, the evidence unit number is used as the unique identifier, and the resource number, case number, evidence unit type, and content index data are written into it. When establishing resource inclusion edges, the teaching resource node is used as the starting point, and the evidence unit node as the ending point, with both the resource number and the evidence unit number written into it, enabling the evidence unit node to return the original medical education resource based on the resource inclusion edge. When writing the evidence support field for valid relation edges, the evidence cross-element coverage bands covering the current relation primitives are collected according to the relation primitive number. Evidence unit numbers are read from each evidence cross-element coverage band, duplicate evidence unit numbers are deleted, and then arranged and written into the evidence support field according to the evidence unit type and evidence unit number, ensuring that valid relation edges simultaneously record the entity connection result and the source of evidence support. When writing to the medical education knowledge graph, the joint field composed of case number and entity node number is used as the entity node retrieval key, and the relation primitive number is used as the valid relation edge retrieval key. When the same entity node retrieval key exists, the existing entity node is reused. When the same valid relation edge retrieval key exists, the evidence support field is updated. When the same retrieval key does not exist, new entity node records and valid relation edge records are created to prevent entity nodes under different case numbers from being merged incorrectly and to prevent duplicate nodes and edges from being created.Newly written medical education data is read from the medical education data management database. Resource IDs, case IDs, and content index data are extracted from the newly written medical education data. The resource IDs and content index data are used to form a joint search key. If the original evidence unit is not found using the joint search key, it is confirmed that the newly written medical education data belongs to new evidence data. According to the data type of medical education, lesion images, report sentence content, subtitle sentence content, video frame sequences, courseware annotation areas, and annotation text content are extracted to generate new imaging evidence units, new report evidence units, new video evidence units, and new courseware evidence units, respectively. According to the character matching method corresponding to the report evidence unit, video evidence unit, and courseware evidence unit, ontology term matching is performed on the new report evidence unit, new video evidence unit, and new courseware evidence unit to extract lesion entity ID, anatomical location ID, medical manifestation ID, conclusion entity ID, relation type code value, and treatment stage code value. According to the image entity template matching method corresponding to the imaging evidence unit, image entity template matching is performed on the new imaging evidence unit to extract lesion entity ID, anatomical location ID, and medical manifestation ID. For newly added report evidence units, newly added video evidence units, and newly added courseware evidence units, the case number, head entity number, relation type code value, and tail entity number generated are compared item by item with the case number, head entity number, relation type code value, and tail entity number of each relation primitive. If the case number, head entity number, relation type code value, and tail entity number are all consistent, the newly added evidence unit is confirmed to support the corresponding relation primitive. For newly added imaging evidence units, the case number, lesion entity number, and anatomical location number in the newly added imaging evidence unit are compared with the case number, head entity number, and tail entity number of the positional relation primitive, respectively, and the relation type code value of the positional relation primitive is confirmed to be the relation type code value corresponding to the positional relation. If all data are consistent, the newly added imaging evidence unit is confirmed to support the corresponding positional relation primitive. The case number, lesion entity number, and medical manifestation number in the newly added imaging evidence unit are compared with the case number, head entity number, and tail entity number of the manifestation relation primitive, respectively, and the relation type code value of the manifestation relation primitive is confirmed to be the relation type code value corresponding to the manifestation relation. If all data are consistent, the newly added imaging evidence unit is confirmed to support the corresponding manifestation relation primitive.Relational primitives supported by the same newly added evidence unit are arranged in ascending order of their intra-chain positional values. Relational primitives with the same newly added evidence unit number and adjacent intra-chain positional values are grouped into the same continuous positional segment. When the intra-chain positional value is interrupted, the current continuous positional segment ends and a new continuous positional segment is established. A new evidence cross-primary primitive coverage band is generated based on each continuous positional segment. The intra-chain positional value of the first relational primitive in the continuous positional segment is written into the coverage start positional value, and the intra-chain positional value of the last relational primitive in the continuous positional segment is written into the coverage end positional value, along with the newly added evidence unit number. When the newly added evidence unit supports only a single relational primitive, both the coverage start positional value and the coverage end positional value are written into the intra-chain positional value of the single relational primitive. When the newly added evidence unit and the remaining evidence units together complete all the necessary evidence unit types for the suspended relational primitive, the suspended relational primitive is rewritten into the valid relational edge table, and the restored valid relational edges are updated to the medical education knowledge graph. Specifically, the head entity number, relation type code value, tail entity number, missing evidence unit type, and remaining evidence unit number are read according to the relation primitive number in the suspended relation primitive table. The case number and intra-chain position value are then read from the original relation primitive record retained in the ordered relation primitive chain based on the relation primitive number. The case number, head entity number, relation type code value, and tail entity number of the newly added evidence unit are compared item by item with the case number, head entity number, relation type code value, and tail entity number of the suspended relation primitive. If all four data are consistent, it is confirmed that the newly added evidence unit has the same entity connection relationship and relation type as the suspended relation primitive. The type of the newly added evidence unit is then compared with the type of the missing evidence unit. If the type of the newly added evidence unit belongs to the type of the missing evidence unit, the number of the newly added evidence unit is added to the set of restored evidence unit numbers of the suspended relation primitive. Based on the remaining evidence unit numbers, the evidence cross-element coverage band associated with the remaining evidence units is read. Only the remaining evidence unit numbers whose case numbers match the suspended relationship element and actually cover the suspended relationship element are retained. The remaining evidence unit numbers that meet the conditions are merged with the newly added evidence unit numbers to form a restored evidence unit number set. Based on the restored evidence unit number set, the evidence unit type of each evidence unit is read, and all necessary evidence unit types are read from the medical relationship rule table based on the relationship type code value of the suspended relationship element. When the evidence unit type corresponding to the restored evidence unit number set covers all necessary evidence unit types, it is determined that the newly added evidence unit and the remaining evidence unit together supplement all necessary evidence unit types of the suspended relationship element.After completing the supplementary determination, the relation primitive number, head entity number, relation type code value, and tail entity number are read from the suspended relation primitive table. Based on the relation primitive number, the case number and intra-chain position value are read from the original relation primitive records retained in the ordered relation primitive chain. The suspended relation primitive is then rewritten into the effective relation edge table. The restored evidence unit number set is written into the evidence support field of the effective relation edge, and the corresponding record in the suspended relation primitive table is deleted, allowing the restored effective relation edge to re-participate in entity associations in the medical education knowledge graph. When new evidence crosses the boundary of lost connection across primitive coverage, the continuous coverage marker of the same evidence corresponding to the boundary of lost connection is regenerated, and the chain adjacency record at the boundary of lost connection is restored. When the relation primitives on both sides of the boundary of lost connection are located in the effective relation edge table, the relation primitive chain fragments on both sides of the boundary of lost connection are merged based on the restored chain adjacency records. Specifically, the system reads the lost boundary number, left relation primitive number, right relation primitive number, and the positional value within the left relation primitive chain and the right relation primitive chain. It then checks whether the starting positional value of the coverage of the same newly added evidence cross-primary primitive coverage band is not greater than the positional value within the left relation primitive chain, and checks whether the ending positional value is not less than the positional value within the right relation primitive chain. If both conditions are met, it determines that a single newly added evidence cross-primary primitive coverage band simultaneously covers the relation primitives on both sides of the lost boundary, and regenerates the corresponding continuous coverage marker for the same evidence at the lost boundary. It then checks whether both the left and right relation primitives have been written into the valid relation edge table. When the newly added evidence crosses the lost boundary, the continuous coverage marker for the same evidence has been regenerated, and the relation primitives on both sides of the lost boundary are located in the valid relation edge table, the case number, lost boundary number, left relation primitive number, right relation primitive number, continuous coverage marker for the same evidence, and newly added evidence unit number are written into the chain adjacency record table, restoring the chain adjacency record at the lost boundary position. When newly added evidence does not cross the lost boundary or re-establish continuous coverage with the same evidence across the primitive coverage band, the chain adjacency record at the lost boundary position is not restored. When any relation primitive on either side of the lost boundary is not located in the valid relation edge table, the restored chain adjacency record is retained, but relation primitive chain fragment merging is not performed. After restoring the chain adjacency record at the lost boundary position, the relation primitive chain fragments on both sides of the lost boundary are read in ascending order of intra-chain position value. The relation primitive number sequence in the left relation primitive chain fragment is merged with the relation primitive number sequence in the right relation primitive chain fragment, and the merged relation primitive numbers are arranged in ascending order of intra-chain position value to generate a new relation primitive number sequence. The new relation primitive number sequence is then written into the merged relation primitive chain fragment.New teaching resource nodes, new evidence unit nodes, and new resource inclusion edges are written into the medical education knowledge graph. If a teaching resource node associated with a new evidence unit exists in the medical education knowledge graph, the existing teaching resource node is reused; if no teaching resource node with the same resource number exists, a new teaching resource node is created. New evidence unit nodes are created according to their numbers, and new teaching resource nodes are connected to new evidence unit nodes via new resource inclusion edges. If the new evidence unit and remaining evidence units do not complete all necessary evidence unit types for the suspended relation primitives, the new evidence unit number is written into the remaining evidence unit number in the suspended relation primitive table, and the missing evidence unit type is re-determined according to the medical relation rule table; valid relation edges are not restored. If all conditions are not met—new evidence crossing primitive coverage bands across lost boundaries, continuous coverage of the same evidence re-formed, and relation primitives on both sides of the lost boundary being located in the valid relation edge table—the chain adjacency record is not restored. Thus, the partial restoration of valid relation edges and chain adjacency records is completed based on the completeness of evidence unit types and the continuous coverage of the same evidence, respectively.
[0060] In this implementation plan, by using newly added evidence units and newly added evidence cross-element coverage zones as the basis for local restoration, the medical education knowledge graph can restore valid relation edges when the necessary evidence unit types are restored, and restore chain adjacency records when the continuous support of the same evidence is re-established. At the same time, relation primitives with insufficient evidence conditions are kept in a suspended state. This improves the integrity of the association between teaching resource nodes, evidence unit nodes, entity nodes, valid relation edges, and resource inclusion edges, reduces the risk of erroneous restoration caused by insufficient evidence, and enhances the accuracy, continuity, and traceability of incremental updates to the medical education knowledge graph.
[0061] like Figure 3As shown, the medical education knowledge graph includes teaching resource nodes, evidence unit nodes, entity nodes, effective relation edges, and resource inclusion edges. Purple nodes represent teaching resource nodes, corresponding to medical teaching video data, image report data, teaching courseware data, and medical image data, respectively. Gray arrows represent resource inclusion edges, with each teaching resource node pointing to its corresponding video evidence unit node, report evidence unit node, courseware evidence unit node, and image evidence unit node through these edges. Blue nodes represent evidence unit nodes, each containing evidence unit number, resource number, case number, evidence unit type, and content index data. Green nodes represent entity nodes, which, in medical semantic order, include lesion entity nodes, anatomical location nodes, medical manifestation nodes, conclusion entity nodes, and diagnosis / treatment stage nodes. Orange solid lines represent effective relation edges, where the lesion entity node and the anatomical location node are linked. Nodes are connected by valid relation edges written using position relation primitives. Pathology entity nodes and medical manifestation nodes are connected by valid relation edges written using manifestation relation primitives. Medical manifestation nodes and conclusion entity nodes are connected by valid relation edges written using conclusion relation primitives. Conclusion entity nodes and treatment stage nodes are connected by valid relation edges written using stage relation primitives. Orange dashed lines indicate the writing of evidence support fields, representing the writing of corresponding evidence unit numbers into the evidence support fields of the corresponding valid relation edges. Evidence unit nodes do not establish medical relation edges with entity nodes. Multiple evidence unit numbers can be written into the same valid relation edge, allowing the valid relation edge to trace back to the corresponding evidence unit node and teaching resource node, thus forming a medical education knowledge graph structure with clear source and evidence support relationships among teaching resources, evidence units, entity nodes, and valid relation edges.
[0062] The following points need to be explained:
[0063] (1) The accompanying drawings of the embodiments of the present invention only involve the structures involved in the embodiments of the present invention. Other structures can refer to the general design.
[0064] (2) For clarity, the thickness of layers or regions is enlarged or reduced in the drawings used to describe embodiments of the invention, i.e., these drawings are not drawn to scale. It is understood that when an element such as a layer, film, region or substrate is referred to as being “above” or “below” another element, the element may be “directly” located “above” or “below” the other element or there may be intermediate elements.
[0065] (3) Where there is no conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other to obtain new embodiments.
[0066] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A multimodal medical education knowledge graph construction system integrating multi-source heterogeneous data, characterized in that, include: The multimodal relation primitive generation module is used to acquire medical image data, image report data, medical teaching video data, and teaching courseware data, and generate image evidence units, report evidence units, video evidence units, and courseware evidence units respectively. It extracts medical entity numbers and relation type encoding values and constructs an ordered relation primitive chain. The same evidence coverage band generation module is used to connect the relation primitives that are continuously supported by the same evidence unit according to the intra-chain position value to form the evidence cross-primary primitive coverage band, truncate at the position interruption, generate the same evidence continuous coverage mark, and write the relation primitives that are supported by all corresponding evidence unit types into the valid relation edge table. The relationship break incision identification module is used to acquire medical education data that is inconsistent with the content recorded in the original evidence unit, generate updated evidence units and updated evidence cross-element coverage bands, identify uncovered primitives and disconnected boundaries, and generate relationship chain break incisions. The relation primitive local write-back module is used to suspend uncovered primitives based on the broken relation chain incision, delete the chain adjacency records at the lost boundary position and generate relation primitive chain fragments, restore valid relation edges and chain adjacency records based on newly added evidence units, and write teaching resource nodes, evidence unit nodes, entity nodes, valid relation edges, and resource inclusion edges into the medical education knowledge graph.
2. The multimodal medical education knowledge graph construction system integrating multi-source heterogeneous data according to claim 1, characterized in that, The specific steps for acquiring medical imaging data, imaging report data, medical teaching video data, and teaching courseware data are as follows: Retrieve medical imaging data, image report data, medical teaching video data, teaching courseware data, medical knowledge ontology table, and medical relationship rule table stored in the medical education data management database. All four types of medical education data include resource ID and case ID. Medical imaging data includes image frame ID, lesion annotation ID, and lesion region coordinate values. Image report data includes report sentence ID and report sentence content. Medical teaching video data includes subtitle sentence ID, subtitle sentence content, subtitle time interval, and video frame sequence. Teaching courseware data includes courseware page ID, courseware annotation ID, annotation region coordinate values, and annotation text content.
3. The multimodal medical education knowledge graph construction system integrating multi-source heterogeneous data according to claim 2, characterized in that, The specific steps for generating the image evidence unit, report evidence unit, video evidence unit, and courseware evidence unit are as follows: Lesion images are extracted based on the coordinates of the lesion area. Report sentence content is extracted based on the report sentence number, and subtitle sentence content is extracted based on the subtitle sentence number. Corresponding continuous video frames are extracted from the video frame sequence according to the subtitle time interval. The courseware annotation area and annotation text content are extracted based on the coordinates of the annotation area, generating image evidence units, report evidence units, video evidence units, and courseware evidence units respectively. The image frame number, lesion annotation number, report sentence number, subtitle sentence number, courseware page number, and courseware annotation number are written into the content index data of each evidence unit, and an evidence unit number, resource number, and case number are written for each evidence unit.
4. The multimodal medical education knowledge graph construction system integrating multi-source heterogeneous data according to claim 3, characterized in that, The specific steps for extracting medical entity IDs and relation type codes to construct an ordered relation primitive chain are as follows: The report sentence content, subtitle sentence content, and annotation text content are matched with the entity terms, relation terms, and stage terms in the medical knowledge ontology table one by one to extract the lesion entity number, anatomical location number, medical manifestation number, conclusion entity number, relation type code value, and treatment stage code value, and written into the report evidence unit, video evidence unit, and courseware evidence unit; the boundary contour value, major axis value, minor axis value, and grayscale statistical value of the lesion image are extracted, and the normalized distance is calculated with the image entity template in the medical knowledge ontology table. When the minimum normalized distance is not greater than the entity matching distance threshold, the lesion entity number, anatomical location number, and medical manifestation number in the image entity template are written into the image evidence unit; Within the same case number, lesion entity nodes, anatomical location nodes, medical manifestation nodes, and conclusion entity nodes are generated based on the lesion entity number, anatomical location number, medical manifestation number, and conclusion entity number, respectively. Treatment stage nodes are generated using the treatment stage code value as the treatment stage node number. Positional relation primitives are generated based on the lesion entity number and anatomical location number; manifestation relation primitives are generated based on the lesion entity number and medical manifestation number; conclusion relation primitives are generated based on the medical manifestation number and conclusion entity number; and stage relation primitives are generated based on the conclusion entity number and treatment stage node number. Positional relation primitives sharing the same lesion entity number are connected to manifestation relation primitives; manifestation relation primitives sharing the same medical manifestation number are connected to conclusion relation primitives; and conclusion relation primitives sharing the same conclusion entity number are connected to stage relation primitives, generating an ordered relation primitive chain. For each relation primitive, a relation primitive number, case number, head entity number, tail entity number, relation type code value, and intra-chain position value are written.
5. The multimodal medical education knowledge graph construction system integrating multi-source heterogeneous data according to claim 4, characterized in that, The specific steps for connecting consecutively supported relational primitives of the same evidence unit into an evidence cross-primary primitive coverage band according to the intra-chain positional value, and truncating at the positional break point, are as follows: Read each evidence unit and the ordered relation primitive chain; when the head entity number, relation type code value, and tail entity number in the report evidence unit, video evidence unit, and courseware evidence unit are consistent with the head entity number, relation type code value, and tail entity number of the relation primitive, respectively, write the evidence unit number into the corresponding relation primitive; when the lesion entity number, anatomical location number, and medical manifestation number in the image evidence unit are consistent with the location relation primitive and manifestation relation primitive, respectively, write the image evidence unit number into the corresponding location relation primitive and manifestation relation primitive. The relation primitives supported by the same evidence unit are divided into continuous positional segments according to the positional value within the chain. An evidence cross-primary primitive coverage band is generated based on each continuous positional segment, and the coverage start positional value, coverage end positional value, and evidence unit number are written in.
6. The multimodal medical education knowledge graph construction system integrating multi-source heterogeneous data according to claim 5, characterized in that, The specific steps for generating continuous overlay markers for the same evidence and writing the relation primitives supported by all corresponding evidence unit types into the effective relation edge table are as follows: The relationship primitives covered by the evidence cross-primary primitive coverage band and the inter-chain boundary crossed are determined based on the coverage start position value and the coverage end position value; when the same evidence cross-primary primitive coverage band covers the relationship primitives on both sides of the inter-chain boundary at the same time, the inter-chain boundary is written with the same evidence continuous coverage mark. Read the required evidence unit type for each relationship type encoding value from the medical relationship rule table; when there is an evidence cross-element coverage band covering the relationship primitive for each required evidence unit type, write the relationship primitive into the valid relationship edge table; when there is a continuous evidence coverage mark between adjacent relationship primitives, write the connection between adjacent relationship primitives into the chain adjacency record table.
7. The multimodal medical education knowledge graph construction system integrating multi-source heterogeneous data according to claim 6, characterized in that, The specific steps for obtaining medical education data that is inconsistent with the original evidence unit record, and generating updated evidence units and updated evidence cross-element coverage bands are as follows: The system retrieves newly written medical imaging data, imaging report data, medical teaching video data, and teaching courseware data from the medical education data management database. It then searches for the original evidence unit based on the resource number and content index data. When the newly written data is inconsistent with the content recorded in the original evidence unit, the system re-extracts the evidence content, performs ontology term comparison on the text content, performs image entity template matching on the lesion image, generates an updated evidence unit, and determines the relation primitives supported by the updated evidence unit. Delete all evidence cross-element coverage bands formed by the original evidence unit, divide the relational primitives supported by the updated evidence unit into continuous positional segments according to the positional value within the chain, generate the updated evidence cross-element coverage band, and write the coverage start positional value, coverage end positional value and updated evidence unit number. Based on the updated evidence cross-element coverage band, the necessary evidence unit type coverage status of each relational element and the continuous coverage status of the same evidence at the boundaries of each chain are redefined.
8. The multimodal medical education knowledge graph construction system integrating multi-source heterogeneous data according to claim 7, characterized in that, The specific steps for identifying uncovered primitives and disconnected boundaries to generate relationship chain break points are as follows: Record the relation primitives that do not have all the corresponding evidence unit types as uncovered primitives and record the missing evidence unit types; record the inter-chain boundaries that originally had continuous coverage of the same evidence and do not have the same evidence cross-primary primitive coverage band after the update as disconnected boundaries; merge adjacent uncovered primitives and disconnected boundaries according to the intra-chain position value to generate relation chain break cuts, and write the cut start position value, cut end position value, uncovered primitive number, missing evidence unit type and disconnected boundary number; When no missing primitives or missing boundaries are formed, a relation chain retention mark is generated; a relation primitive suspension mark is generated for missing primitives; when the necessary evidence unit types of relation primitives on both sides of the missing boundary are all intact, a relation chain split mark is generated for the missing boundary.
9. The multimodal medical education knowledge graph construction system integrating multi-source heterogeneous data according to claim 8, characterized in that, The specific steps for suspending uncovered primitives based on relationship chain break cuts, deleting chain adjacency records at the lost-connection boundary positions, and generating relationship primitive chain fragments are as follows: Read the relationship chain retention flag, maintain valid relationship edges and chain adjacency records, and delete the original evidence unit number from the relationship primitive; when the updated evidence unit still supports the original relationship primitive, write the updated evidence unit number into the relationship primitive, and retain other evidence unit numbers and evidence cross-primary overlay bands; Read the suspended marker of the relation primitive, remove the uncovered primitive from the effective relation edge table, delete the chain adjacency record connected to the uncovered primitive, extract the remaining evidence unit number from the evidence cross-primary primitive coverage band retained by the uncovered primitive, and write the relation primitive number, head entity number, relation type code value, tail entity number, missing evidence unit type, and remaining evidence unit number into the suspended relation primitive table; the relation primitives and chain adjacency records outside the relation chain break remain unchanged; Read the relationship chain splitting marker, retain the valid relationship edges on both sides of the lost connection boundary, delete the chain adjacency records at the lost connection boundary position, generate relationship primitive chain fragments with each lost connection boundary as the boundary, and write the chain fragment number and the relationship primitive number sequence arranged in ascending order of the intra-chain position value for each relationship primitive chain fragment.
10. The multimodal medical education knowledge graph construction system integrating multi-source heterogeneous data according to claim 9, characterized in that, The specific steps for restoring valid relation edges and chain adjacency records based on newly added evidence units, and writing teaching resource nodes, evidence unit nodes, entity nodes, valid relation edges, and resource inclusion edges into the medical education knowledge graph are as follows: A teaching resource node is generated based on the resource number, an evidence unit node is generated based on the evidence unit number, a resource inclusion edge is established pointing from the teaching resource node to the evidence unit node, the evidence unit number is written into the evidence support field of the valid relation edge, and the entity node, teaching resource node, evidence unit node, valid relation edge and resource inclusion edge are written into the medical education knowledge graph. Read newly written medical education data from the medical education data management database where the original evidence unit could not be found based on the resource number and content index data. Generate new evidence units and new evidence cross-element coverage bands. When the new evidence units and the remaining evidence units together complete all the necessary evidence unit types of the suspended relation primitives, rewrite the relation primitives into the effective relation edge table and update the restored effective relation edges to the medical education knowledge graph. When the new evidence cross-element coverage band crosses the lost connection boundary, restore the chain adjacency record at the lost connection boundary position. Write the new teaching resource node, the new evidence unit node, and the new resource inclusion edge into the medical education knowledge graph, and retain the relation primitives that have not been completed with the necessary evidence unit types in the suspended relation primitive table.
Citation Information
Patent Citations
A method and apparatus for constructing a medical knowledge graph
CN110362690B
Construction method of biomedical knowledge graph
CN120653782A
