A multi-source knowledge fusion and dynamic knowledge graph construction method and system for rare disease whole-process diagnosis and treatment

Abstract

The application discloses a multi-source knowledge fusion and dynamic knowledge graph construction method and system for rare disease whole-process diagnosis and treatment, and belongs to the technical field of medical information data processing. The method performs SNOMED CT term normalization and CHPO phenotype mapping on Chinese clinical text in parallel, calculates fusion confidence after mapping the two to the same UMLS concept identifier; feeds the merged phenotype set to the previous stage to enhance term disambiguation, and applies three-level gate unlocking and rollback to Chinese expressions not covered by the core vocabulary; then determines the fusion coefficient and conflict resolution priority under the conditions of four stages of screening, auxiliary diagnosis, treatment and follow-up management; finally, the gate cycle unit evolves the edge weight, and only the consistency constraint is applied to the anchor edge defined by the intersection of adjacent stage edge types. The application outputs a knowledge graph data structure, which can improve the coverage rate of Chinese phenotype mapping and reduce the conflict rate of cross-stage edge weight.

Description

Technical Field

[0001] This invention belongs to the field of medical information data processing technology, specifically relating to a method and system for multi-source knowledge fusion and dynamic knowledge graph construction for the whole process of diagnosis and treatment of rare diseases. Background Technology

[0002] Rare diseases are characterized by a large number of diseases, a small number of cases per disease, strong phenotypic heterogeneity, and a long diagnosis cycle. To reuse medical knowledge in screening, assisted diagnosis, treatment, and follow-up management, existing information systems typically rely on multiple sources, including the Human Phenotypic Ontology (HPO), the Chinese Human Phenotyping Standard Terminology (CHPO), the systematic clinical medical terminology set SNOMED CT, and the Integrated Medical Language System (UMLS), and are developed along the following four technical routes: The first approach is based on phenotypic ontology for the organization and matching of rare disease knowledge. CN115482926A discloses a knowledge-driven, visual, question-and-answer-based, assisted differential diagnosis system and method for rare diseases. It is based on a human phenotypic ontology knowledge base and the rare disease knowledge base Orphanet. Utilizing information such as disease prevalence, age of onset and death, phenotypic frequency characteristics, and pathogenic genes, it provides candidate diseases through conversational phenotypic query recommendation services and disease recommendation services. Based on patient self-reports and phenotypic information added during the dialogue, it presents the results using metrics such as posterior probability, phenotypic matching degree, and information entropy. This approach has established an overall framework that combines rare diseases, phenotypic ontology, and disease phenotypic associations.

[0003] The second route is the Chinese phenotypic vocabulary expansion route. CN112687397A discloses a processing method and device for a rare disease knowledge base, as well as a readable storage medium. It extracts clinical phenotypic terms and their frequency and inverse text frequency index, etc., from the electronic medical records of rare disease patients, compares them one by one with the standard phenotypic terms in the authoritative rare disease knowledge base, and optimizes the authoritative knowledge base based on the comparison results. To extract clinical phenotypic terms from electronic medical records, this patent also discloses an implementation method for building a self-constructed Chinese phenotypic terminology database. One implementation method involves taking concepts from the top-level categories of "clinical observations" and "diseases" in an external Chinese medical terminology database, merging their preferred terms, pending preferred terms, and permitted terms to form an external thesaurus. This external thesaurus is then compared with the Chinese human phenotypic standard terminology database. When entries match, all synonyms under the external concept are added as Chinese synonyms to the corresponding Chinese phenotypic concept, and these are also included as hyponyms. This completes the enrichment of synonyms and hyponyms in the human phenotypic ontology terminology set. The enriched terms can be added to the thesaurus after review by a doctor. Based on the enriched terminology database, a reverse maximum matching method and a rule-based algorithm are used to identify phenotypic terms in the electronic medical records. This approach improves the coverage of Chinese phenotypic terms.

[0004] The third approach is the Chinese clinical text entity recognition and terminology normalization approach. This approach uses named entity recognition to identify clinical entities such as disease sites, symptoms, medications, and surgeries from descriptive texts such as chief complaints, present medical history, and physical examinations, and normalizes them into a standard clinical terminology set, providing entity sources for the construction of a medical knowledge graph. It is well known in the field that Chinese electronic medical records lack a unified standardized writing style; the same clinical concept often exists in multiple ways, including colloquial expressions, dialectal expressions, abbreviations, and institutionally defined abbreviations. These situations increase the difficulty of entity recognition and terminology normalization.

[0005] The fourth approach is a cross-ontology fusion and dynamic graph update approach based on unified concept identifiers. This approach first uses a clinical text extraction tool to parse the concepts of an integrated medical language system, assigns a unique concept identifier to each concept as an input node for the graph, and then generates candidate conclusions through multi-hop reasoning. In the field of temporal knowledge graphs, existing methods use relational graph convolutional networks and gated recurrent units to jointly encode node and relation representations, and use global information extracted from static graphs to constrain the relation evolution process.

[0006] The aforementioned existing technology has the following problems: First, cross-ontology concept convergence is treated as the endpoint of unidirectional evidence aggregation. In the approach represented by CN115482926A and in multi-hop inference methods based on unique concept identifiers, after multiple sources converge on the same concept identifier, the result is only used for confidence weighting, candidate ranking, or conflict resolution, and does not flow back to the previous clinical terminology disambiguation stage. As a result, the system cannot distinguish between genuine cross-source semantic convergence and concept identifier collisions caused by homonyms; once the previous disambiguation result is incorrect, subsequent layers cannot correct it, and the error will continue to propagate along the link.

[0007] Second, the expansion of the Chinese phenotypic vocabulary is disconnected from the system's operational status. The enrichment method used in CN112687397A is completed in one go offline, with only doctors reviewing it before writing. Once the expansion results are written, they remain valid indefinitely. However, the applicability of colloquial, dialectal, and abbreviation expressions in Chinese electronic medical records varies across different stages of diagnosis and treatment. Synonyms written in one go may continue to participate in the calculation of disease-phenotype association weights in stages where they are not applicable, thus introducing erroneous associations. Furthermore, this type of offline expansion lacks a mechanism to retract erroneously written synonyms during system operation.

[0008] Third, information about the diagnosis and treatment stages is merely used as a start / stop switch for knowledge sub-repositories. Existing staged knowledge organization schemes divide knowledge scope according to screening, assisted diagnosis, treatment, and follow-up management, addressing only which knowledge to invoke at any given time. Stage information is not included in cross-ontology fusion coefficients, disease-phenotype association weights, or conflict resolution priorities. Therefore, the same phenotype receives the same weight across different diagnosis and treatment stages, which is inconsistent with the actual clinical practice where different stages have different focuses.

[0009] Fourth, dynamic graph updates lack constraints that match the semantics of each stage. While temporal graph methods, such as relational graph convolutional networks and gated recurrent units, can update relation strength over time and impose constraints on evolution using static global information, their constraints apply to all relations or the global representation, without defining the constraint scope according to stage semantic compatibility. If consistency constraints are imposed on all edges, normal changes in stage-specific relations will be suppressed; if no constraints are imposed, the edge weights of the same disease-phenotype association or test abnormality association will drift during stage switching, resulting in cross-stage semantic conflicts.

[0010] Therefore, there is a need in this field for a dynamic knowledge graph construction scheme that enables cross-ontology convergence results to have a reverse effect on the previous disambiguation and Chinese vocabulary unlocking, enables diagnosis and treatment stage information to enter the fusion coefficient and conflict resolution, and enables consistency constraints to be applied precisely according to the semantic compatibility range of the stage. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for multi-source knowledge fusion and dynamic knowledge graph construction for the whole process of diagnosis and treatment of rare diseases. This method enables the cross-ontology concept fusion results to have a reverse effect on the disambiguation of the preceding clinical terms and the unlocking and rollback of the Chinese phenotypic vocabulary. It allows the information of the diagnosis and treatment stages to enter the cross-ontology fusion coefficient and conflict resolution priority, and applies cross-stage consistency constraints only to the edges of adjacent diagnosis and treatment stages that are semantically compatible. This solves the problems in the prior art such as the lack of backflow of fusion results, the offline expansion of the Chinese vocabulary and the separation of the running state, the stage information being used only as a sub-library switch, and the mismatch between the scope of consistency constraints and the stage semantics.

[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This invention provides a method for multi-source knowledge fusion and dynamic knowledge graph construction for the whole process of diagnosis and treatment of rare diseases. The method is executed by a computer and is used to construct and update the dynamic knowledge graph data structure, including steps S1 to S7.

[0013] A method for multi-source knowledge fusion and dynamic knowledge graph construction for the entire process of diagnosis and treatment of rare diseases includes: S1. Acquire multi-source data and divide the data into segments according to the diagnosis and treatment stage identifiers. The diagnosis and treatment stages include the screening stage, auxiliary diagnosis stage, treatment stage, and follow-up management stage, which are carried out in sequence. S2. Perform SNOMED CT clinical terminology normalization on the Chinese electronic medical record text in the multi-source data to obtain candidate clinical concepts and their channel confidence levels. s Perform CHPO Chinese phenotype mapping on the Chinese self-narrative text in the multi-source data to obtain candidate phenotypes and their channel confidences c. p Candidate clinical concepts and candidate phenotypes are mapped to UMLS concept labels respectively, and the fusion confidence c of candidates merging to the same concept label is calculated according to formula (1). f : c f =1-(1-c s (1-c) p (1) S3, using the HPO phenotype set H obtained in step S2... conv For each candidate clinical concept with the same Chinese expression, the score is recalculated and reordered to obtain the disambiguated clinical concepts. Chinese expressions not covered by the CHPO core thesaurus are placed in a pending confirmation set, and are only written into the valid mapping and participate in subsequent weight calculations after successively meeting the following three levels of gating: The first level is that the Chinese expression has been merged into the same UMLS concept identifier and its fusion confidence c. f The first level is not less than the first threshold; the second level is the set of priority semantic types to which the semantic type of the Chinese expression belongs in the current diagnosis and treatment stage; the third level is that the cross-stage edge weight conflict rate of the anchor edge obtained in the previous diagnosis and treatment stage is not higher than the second threshold, and this level of gating is considered to be satisfied when there is no previous diagnosis and treatment stage. S4. Determine the fusion coefficient α based on the current treatment stage t. h (t), with α h (t) The corrected association weight w is obtained by weighting and summing the publicly known prior weights and the local Chinese clinical statistical weights for the association between disease d and phenotype h. d,h (t); α h (t) consists of the basic fusion coefficient, the stage bias term of t, and c. f To be determined jointly; S5. Based on the set of allowed edge types in the current treatment stage, select nodes and relationships from the pre-built knowledge sub-base to generate a candidate node set and a candidate edge set; S6. The gated loop unit inputs the vector x according to the edge (i,j) in the candidate edge set. ij t The edge weight w in the current treatment stage is obtained by evolving from the hidden state of the edge in the previous treatment stage. ij t Take the edges whose type belongs to the intersection of the allowed edge types of the two adjacent treatment stages from the intersection of the edge sets of those two adjacent stages, and form the anchored edge set E. anchor(t); Only anchor edges are subject to cross-stage consistency constraints that reduce the edge weight difference between two adjacent treatment stages, and edges other than anchor edges are not subject to such constraints. S7. Output a dynamic knowledge graph data structure including a set of nodes, a set of edges, an edge weight matrix, and a stage version identifier.

[0014] Steps S2 to S7 are repeated sequentially for each treatment stage according to the order of the treatment stages. The edge set and edge weight matrix of each treatment stage are saved for use in subsequent treatment stages.

[0015] This invention also provides a multi-source knowledge fusion and dynamic knowledge graph construction system for the whole process of diagnosis and treatment of rare diseases, for implementing the above method. The system includes a multi-source data access module, a dual-channel terminology alignment module, a concept identification closed-loop feedback module, a stage conditional fusion module, a knowledge sub-base stage assembly module, a dynamic edge weight update and cross-stage anchoring module, and a graph storage and stage version output module.

[0016] The multi-source data access module is used to acquire multi-source data and divide the data into fragments according to the diagnosis and treatment stage identifier; The dual-channel terminology alignment module is used to perform SNOMED CT clinical terminology normalization and CHPO Chinese phenotype mapping in parallel, and to calculate fusion confidence on UMLS concept identifiers; The concept identification closed-loop feedback module is used to reorder pre-candidate clinical concepts by merging phenotype sets and to perform three-level gating unlocking and rollback for Chinese expressions not covered by the core vocabulary. The stage-conditional fusion module is used to determine the fusion coefficient, correct the correlation weight, source weighting coefficient, and conflict resolution priority according to the current diagnosis and treatment stage; The knowledge sub-base stage assembly module is used to assemble candidate nodes and candidate edges from the knowledge sub-base according to the set of allowed edge types in the current stage; The dynamic edge weight update and cross-stage anchoring module is used to evolve edge weights, determine the set of anchored edges, and apply cross-stage consistency constraints to the anchored edges. The knowledge graph storage and phase version output module is used to output and store dynamic knowledge graph data structures.

[0017] The system also includes a storage and configuration unit for storing the knowledge sub-base, stage mode configuration, set of statements to be confirmed and their unlock level identifiers, and atlas versions for each diagnosis and treatment stage.

[0018] The above modules form two feedback paths: the concept identification closed-loop feedback module sends the reordering results back to the dual-channel term alignment module, so that the merging results have a reverse effect on the previous disambiguation; the dynamic edge weight update and cross-stage anchoring module sends the cross-stage edge weight conflict rate of the anchored edges back to the concept identification closed-loop feedback module as the judgment basis for the cross-stage anchoring confirmation gate, thereby enabling the graph consistency status to control the unlocking and rollback of the Chinese vocabulary in reverse.

[0019] The present invention has the following beneficial effects: 1. This invention transforms the conceptual identifier aggregation result from a one-way aggregation endpoint to a two-way feedback starting point. The HPO phenotype set obtained from the aggregation is fed back to the pre-clinical terminology disambiguation according to equation (2), so that true cross-source aggregation and homonymous collisions can be distinguished, and pre-disambiguation errors can be corrected within the same processing cycle. In the following embodiments, the coverage of Chinese phenotype mapping is improved from 35 / 45 of the tandem baseline to 39 / 45, and the top three hits in the disease-phenotype association ranking are improved from 19 / 24 to 22 / 24.

[0020] 2. This invention determines anchor edges based on the intersection of the sets of allowed edge types in adjacent stages, imposing consistency constraints only on anchor edges and no constraints on stage-specific edges outside the intersection. This design simultaneously considers cross-stage semantic stability and stage-specific evolution. In the following embodiments, the cross-stage edge weight conflict rate of anchor edges decreases from 12.5% ​​to 0.

[0021] 3. This invention couples the expansion of the Chinese phenotypic vocabulary with the system's operational status. Chinese expressions not covered by the core vocabulary must undergo three levels of gating—convergence confirmation, stage semantic type confirmation, and cross-stage anchoring confirmation—before being written into a valid mapping. When the gating conditions are no longer met, the expression reverts to the pending confirmation set according to the unlocking level. Compared to offline one-time enrichment plus doctor review, this mechanism improves coverage while suppressing erroneous synonyms from participating in weight calculations during inapplicable stages.

[0022] 4. This invention incorporates information from the diagnosis and treatment stages into the fusion coefficient, source weighting coefficient, and conflict resolution priority, rather than simply serving as a knowledge sub-base switch. For the same phenotype, the screening stage relies more on publicly available prior weights, while the follow-up management stage relies more on local follow-up statistical weights. In the following embodiments, the top three hits in the disease-phenotype association ranking are further improved from 22 / 24 to 24 / 24.

[0023] 5. In this invention, during phase advancement, only edges outside the anchored edge set undergo complete edge weight evolution. Anchored edges inherit the hidden state of the previous stage and use the edge weights of the previous stage as the initial values ​​of the current stage's edge weights, eliminating the need for full graph reconstruction. In the following embodiments, the total number of edge operations when reconstructing in the full manner for each treatment stage is 93; according to the incremental method of this invention, the number of complete edge weight evolutions saved is equal to the total number of anchored edges in each phase advancement.

[0024] 6. The processing objects of this invention are medical terminology data and graph edge weight data, directly outputting a knowledge graph data structure containing a set of nodes, a set of edges, an edge weight matrix, and a stage version identifier. The disease and phenotype association strength in the edge weight matrix is ​​stored and displayed as edge attributes of the graph for medical personnel to consult and refer to. This invention does not directly provide disease diagnosis conclusions for living individuals. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a structural block diagram of the system of the present invention. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0027] Example 1 like Figure 1 As shown, this embodiment provides a method for multi-source knowledge fusion and dynamic knowledge graph construction for the whole process of diagnosis and treatment of rare diseases. The method is executed by a computer and includes steps S1 to S7.

[0028] Step S1: Obtain multi-source data and divide the data into segments according to the diagnosis and treatment stage identifier.

[0029] The multi-source data includes Chinese electronic medical record text, Chinese patient-reported text, and structured medical codes. The Chinese electronic medical record text includes the chief complaint, present illness history, physical examination, medical orders, and nursing records; the Chinese patient-reported text includes colloquial, dialectal, and generalized descriptions provided by the patient and their family; and the structured medical codes include disease codes, laboratory test codes, and drug codes. The diagnosis and treatment stages include a sequentially progressive screening stage, auxiliary diagnostic stage, treatment stage, and follow-up management stage, with each segment carrying a stage identifier and timestamp.

[0030] Step S2: First-round alignment of dual channels.

[0031] For Chinese electronic medical record text, a rule-based channel and a learning channel are used in parallel to perform medical entity recognition and SNOMEDCT clinical terminology normalization. The rule-based channel is used for standard medical terms with good writing and high confidence, while the learning channel is used for text fragments with strong context dependence or poor writing. The results from the two channels are combined according to the maximum value to obtain candidate clinical concepts and their channel confidence c. s For Chinese self-narrative texts, they are first mapped to Chinese phenotypic entries using the CHPO core vocabulary, and then mapped to HPO phenotypic identifiers to obtain candidate phenotypes and their channel confidence scores c. p .

[0032] Among them, the channel confidence level c s The value selection rule is as follows: 0.95 is taken when the rule channel hits the standard medical terminology, and the normalized score of the model output is taken when the learning channel is taken. Channel confidence level c p The value selection rules are as follows: 1.00 is used when it completely matches the CHPO core vocabulary entries, 0.90 is used when it matches the thesaurus, and the similarity value is used when it matches the character similarity. Candidates with a similarity value lower than 0.60 are not retained.

[0033] In this invention, being covered by the CHPO core thesaurus means that the Chinese expression completely matches an entry in the CHPO core thesaurus, or is matched by the thesaurus attached to the core thesaurus; Chinese expressions that are only matched by character similarity or do not match any entries are considered not covered by the CHPO core thesaurus and are placed into the set to be confirmed in step S3.

[0034] Candidate clinical concepts and candidate phenotypes are mapped to UMLS concept identifiers, thus forming a four-hop alignment chain: Chinese expression to CHPO entry, CHPO entry to HPO identifier, HPO identifier to UMLS concept identifier, and SNOMED CT concept to UMLS concept identifier. For candidates mapped to the same UMLS concept identifier, the fusion confidence c is calculated according to equation (1). f : c f =1-(1-c s (1-c) p (1) Equation (1) adopts the form of probability complement product, which means that the product of the probabilities of both channels not confirming the concept is the complement. Therefore, the fusion confidence is low only when both channels give low confidence at the same time. Unlike conventional weighted summation, the result of Equation (1) is not the fusion endpoint in this invention, but the trigger signal and condition quantity of steps S3 and S4.

[0035] Step S3: Concept Identification and Closed-Loop Feedback. This step includes feedback in two directions.

[0036] The first approach is to disambiguate feedback to higher-level clinical terminology. When the same Chinese expression corresponds to multiple candidate SNOMEDCT concepts, the HPO phenotype set H obtained by merging the same UMLS concept identifiers in step S2 is taken. conv The scores of each candidate concept are recalculated and reordered according to formula (2): r k =min(1,c k +η·II(h k ∈H conv (2) In equation (2), r kFor the recalculated score of the k-th candidate clinical concept, c k h represents the original confidence level of the k-th candidate clinical concept. k The HPO identifier is obtained by mapping the candidate concept using UMLS concept identifiers, where II is the indicator function and η is the feedback gain coefficient. In this embodiment, η = 0.15. Wherein, H... conv The HPO identifier is defined only for aggregates that simultaneously meet the following two conditions: first, the confidence levels c of the two channels in the aggregate. s With c p All are not lower than the lower limit of channel confidence, which is set to 0.60 in this embodiment; secondly, the fusion confidence level c of the merged term. f Not less than the first threshold. The basis for setting the lower limit of channel confidence is that equation (1) is in the form of probability complement product, and any channel alone can give a high confidence level so that c f Close to 1, only with c f The determination cannot distinguish between a true cross-source convergence where both channels provide moderate or higher evidence and a conceptual marker collision where one channel is strong and the other is weak.

[0037] When the HPO identifier corresponding to a candidate concept already appears in the cross-source pooling set, its score is increased and reordered, thereby causing the pooling result to have a reverse impact on the preceding disambiguation. If two channel candidates of a certain Chinese expression are mapped to the same UMLS concept identifier, but the confidence of either channel is lower than the lower channel confidence limit, or the fusion confidence of the pooling term is lower than the first threshold, then the pooling is judged as a homonymous collision, and its HPO identifier is not included in the HPO count. conv The corresponding candidate clinical concepts are not enhanced.

[0038] The second approach is to provide feedback to the Chinese phenotypic mapping. For Chinese expressions not covered by the CHPO core vocabulary, they are not directly written into the valid mapping; instead, they are placed in a pending confirmation set and unlocked step-by-step using the following three-level gating system. The first level is the convergence confirmation gating, which requires that the Chinese expression converge on the same UMLS concept identifier in step S2, and that the fusion confidence level c... fThe threshold is not less than the first threshold, which is set to 0.70 in this embodiment. The second level is the stage semantic type confirmation gate, which requires that the semantic type of the Chinese expression belongs to the priority semantic type set of the current diagnosis and treatment stage; the semantic type of the Chinese expression is determined according to the semantic type of the UMLS concept identifier obtained by its aggregation in the UMLS semantic network, and according to the merging relationship listed in Table 3; the semantic type of Chinese expressions that have not been aggregated to the UMLS concept identifier is classified as symptoms and signs. The third level is the cross-stage anchoring confirmation gate, which requires that the cross-stage edge weight conflict rate of the current adjacent stage anchoring edge is not higher than the second threshold, which is set to 0.10 in this embodiment. In the screening stage, since there is no preceding adjacent diagnosis and treatment stage, the anchoring edge set is empty, and the cross-stage edge weight conflict rate is recorded as 0, and this level of gate is considered to be satisfied; from the auxiliary diagnosis stage onwards, this level of gate is judged according to the cross-stage edge weight conflict rate obtained by the previous stage advancement.

[0039] Each Chinese expression in the pending confirmation set carries an unlock level identifier, which records the highest gating level passed by the Chinese expression in the most recent judgment, with a value of level one, two, or three. Only Chinese expressions with an unlock level identifier of level three are written into the valid mapping and participate in the calculation of the corrected association weight. Chinese expressions with unlock level identifiers of level one or two remain in the pending confirmation set, waiting for other gating conditions to be met. When a certain level of gating is no longer met in a subsequent diagnosis stage, the unlock level identifier of Chinese expressions with an unlock level identifier not lower than that level is downgraded to the level before that level. Chinese expressions with a level lower than three after downgrading are returned to the pending confirmation set and do not participate in the calculation of the corrected association weight in the stage after the downgrade. This downgrade mechanism makes the effective range of the Chinese vocabulary dynamically change with the running status, which is different from the method of offline one-time enrichment that is effective for a long time.

[0040] Step S4: Stage-based conditional fusion.

[0041] The four stages of diagnosis and treatment have different focuses: the screening stage emphasizes symptoms, signs, and disease risk; the auxiliary diagnostic stage emphasizes laboratory tests, imaging, and differential diagnosis; the treatment stage emphasizes medications, procedures, and treatment plans; and the follow-up management stage emphasizes repeat visits, follow-ups, and long-term risks. This step expands the stage information from a knowledge sub-base switch to conditional quantities such as fusion coefficients, source weighting, and conflict resolution priorities. Fusion coefficient α h (t) is calculated according to formula (3): α h (t)=clip [0.05,0.95] (α0+β t -γ·c f (3) In equation (3), α0 is the basic fusion coefficient, which is taken as 0.40 in this embodiment; β tThe phase bias term is 0.10, 0, -0.10, and -0.05 for screening, assisted diagnosis, treatment, and follow-up management phases, respectively; γ is the fusion confidence influence coefficient, which is 0.20 in this embodiment; clip is the interval truncation operation, limiting the result to between 0.05 and 0.95. The fusion confidence c in equation (3) is... f Take the fusion confidence score obtained by merging phenotype h on the same UMLS concept identifier in step S2; when phenotype h does not undergo cross-channel merging in step S2, c f Take the channel confidence level c of this phenotype. p The values ​​for the stage bias terms are based on the following: In the screening stage, the local sample size is the smallest, so more reliance should be placed on the publicly available prior weights; therefore, a positive bias is used to increase the proportion of prior weights. In the follow-up management stage, local follow-up records have been accumulated, so more reliance should be placed on local statistical weights; therefore, a negative bias is used. The negative sign of the fusion confidence impact coefficient is based on the following: the stronger the cross-source confirmation, the higher the consistency between local evidence and publicly available priors; in this case, the proportion of local statistical weights can be appropriately increased without sacrificing stability.

[0042] The adjusted association weight w between disease d and phenotype h d,h (t) is calculated according to formula (4): w d,h (t)=α h (t)·w d,h prior +(1-α h (t))·w d,h local (4) In equation (4), w d,h prior To disclose the prior weights of the association between disease and phenotype, w d,h local The weight is the local Chinese clinical statistical weight. This weight only participates in the ranking within the set of preferred semantic types at the current stage. The fusion coefficient of equation (3) varies with the fusion confidence c of phenotype h. f The difference, for the sake of brevity, is denoted as α. h (t).

[0043] When the same disease and phenotype pair are associated with weight values ​​from two or more sources, and the difference between these values ​​is greater than the source conflict determination threshold, a source conflict is determined to exist between the disease and phenotype pair. In this embodiment, the source conflict determination threshold is set to 0.10. This source conflict determination threshold is used to determine the value discrepancies between different sources within the same stage, and it is an independent threshold compared to the cross-stage conflict determination threshold τ used in step S6 to determine cross-stage conflicts of anchor edges. Furthermore, for sources whose semantic types do not belong to the current stage's priority semantic type set, their confidence is multiplied by a weighting factor before participating in conflict resolution. In this embodiment, the weighting factor is set to 0.30. The priority of conflict resolution is determined according to the order of priority semantic types in the current stage, i.e., the values ​​given by sources with semantic types ranked higher are adopted first. Thus, sources that are not applicable to the stage are automatically weighted down in conflict resolution, rather than being hidden only at the sub-database layer.

[0044] Step S5: Four-stage assembly of the seven knowledge sub-bases. The composition of the seven knowledge sub-bases is shown in Table 1, and the definition of edge types is shown in Table 2.

[0045] Table 1. Composition Record of the Seven Types of Knowledge Sub-Bases Table 2. Definition Record of Edge Types The set of preferred semantic types and the set of allowed edge types for the four diagnosis and treatment stages are shown in Table 3. The set of preferred semantic types is arranged in order of priority, and those that are listed earlier are adopted first in conflict resolution.

[0046] Table 3. Record of Priority Semantic Type Sets and Allowed Edge Type Sets for the Four Diagnosis and Treatment Stages In Table 3, symptoms and signs are included in the priority semantic type set for the four treatment stages. This is based on the fact that the phenotypic matching edges are the common intersection of the allowed edge type sets for the four treatment stages, and Chinese phenotypic descriptions must participate in graph construction in all four stages. In the auxiliary diagnosis stage, treatment stage, and follow-up management stage, symptoms and signs are ranked after the stage-specific semantic types, and their priority in conflict resolution is lower than that of the stage-specific semantic types. The priority semantic types listed in Table 3 are categories obtained by merging UMLS semantic types according to the focus of treatment.

[0047] Table 3 shows the intersection of the allowed edge types for adjacent stages: the screening stage and the auxiliary diagnosis stage have phenotypic matching edges and gene pathogenesis edges; the auxiliary diagnosis stage and the treatment stage have phenotypic matching edges and abnormal test edges; the treatment stage and the follow-up management stage have phenotypic matching edges and treatment plan edges. This intersection is the range of edge types for the anchored edge set in step S6.

[0048] Step S6: Dynamic edge weight update and cross-stage anchoring constraints.

[0049] Construct an edge input vector x for each edge (i,j) in the candidate edge set. ij t Its components include node type embedding, edge type embedding, and fusion confidence c. f The treatment stage code and the edge weight w of the edge in the previous treatment stage. ij t-1 If the edge did not exist in the previous treatment stage, this component is set to 0. Edge input vector x ij t Enter the gated loop unit according to equation (5); optionally, the staged representation e of the edge can be obtained first by linear mapping according to equation (5a). ij t , and then e ij t Substitute x in equation (5) ij t Substitute: h ij t =GRU(x ij t ,h ij t-1 ), w ij t =σ(u T h ij t (5) e ij t =φ(W e x ij t +b e (5a) In equations (5) and (5a), φ is a nonlinear activation function, and W e and b e h is the linear mapping parameter. ij t Let be the hidden state of edge (i,j) in stage t, u be the output head parameter, and σ be the activation function that constrains the edge weight to the interval between 0 and 1. The linear mapping in equation (5a) is an optional step, and when this step is used, e ij t Substitute x in equation (5) ij t As the input to the gated loop unit, if this step is omitted, x is used directly. ij t This serves as the input to the gated recurrent unit. In this embodiment, the embedding dimension is 128, the hidden dimension is 256, the learning rate is 0.001, the number of training epochs is 80, and the random seed is 42.

[0050] The anchored edge set is the set of edges whose type belongs to the intersection of the edge sets of two adjacent stages and the intersection of the allowed edge types sets of the two adjacent stages.

[0051] The above edge type conditions have a substantial limiting effect in the following situations: when there are nodes in the candidate edge set that carry different edge types in two adjacent stages, the intersection of the node pairs is taken, and it is required that the edge types carried by the node pairs in the two adjacent stages belong to the intersection of the allowed edge type sets of the two adjacent stages. Edges with the same node pairs but different edge types are not included in the anchored edge set. Cross-stage consistency constraints are applied only to edges in the anchored edge set according to equation (6): L cons =Σ (i,j)∈Eanchor(t) (w ij t -w ij t-1 ) 2 (6) This constraint is not applied to edges outside the anchored edge set, allowing stage-specific edges to evolve freely as the stage progresses. After training, each anchored edge is first determined to be a cross-stage conflict edge based on whether the absolute value of the change in edge weight is greater than the cross-stage conflict determination threshold τ, and the conflict rate is calculated. Then, an alignment operation is performed only on the anchored edges that are not determined to be conflict edges, adjusting the edge weights of the same node pair in adjacent stages to the average of the edge weights in the two stages. Thus, the conflict rate is calculated before the alignment operation, and the alignment operation does not cover up cross-stage conflicts that have already occurred.

[0052] The total loss L for model training is determined according to equation (7): L=(1+λ1)L dpa +λ2L cons (7) In equation (7), L dpa For the ranking loss of disease-phenotype association, L cons The cross-stage consistency loss is given by equation (6), where λ1 and λ2 are the loss weights. In this embodiment, λ1 = 0.30 and λ2 = 0.10.

[0053] The cross-stage conflict rate of anchored edges is the ratio of the number of edges in the anchored edge set that are judged as cross-stage conflicting edges to the total number of edges in the anchored edge set. The cross-stage conflict judgment threshold τ is set to 0.15 in this embodiment, that is, when the change in edge weight of an anchored edge in an adjacent stage exceeds 0.15, it is judged as a cross-stage conflicting edge. This conflict rate also serves as the judgment basis for the cross-stage anchoring confirmation gate, i.e., the third-level gate, in step S3, thereby enabling the graph consistency state to control the unlocking and rollback of the Chinese vocabulary in reverse, forming a closed loop. If there is no preceding adjacent stage in the screening stage, the anchored edge set is an empty set, and the conflict rate is recorded as 0, and the third-level gate is considered to be satisfied.

[0054] When the stage progresses from t-1 to t, edge input vectors are constructed for edges outside the anchored edge set, and the complete edge weight evolution is performed according to equation (5). This type of edge includes the associated edge of the newly added node, the newly added edge connecting the existing node, and the edge with the same node pair but different edge types. The anchored edge does not reconstruct the edge input vector, nor does it re-evolve from the initial hidden state, but inherits its hidden state h from the previous stage. ij t-1 And based on its edge rights in the previous stage w ij t-1 As the initial value of the edge weights for the current stage, these edge weights are updated only with the training of the current stage under the cross-stage consistency constraint described in Equation (6), without performing a full reconstruction of the graph. Thus, the number of edges required to fully execute edge weight evolution for one stage advancement is reduced from the size of the edge set for that stage to the number of edges other than the anchored edges in the edge set for that stage, and the number of attempts saved is equal to the number of anchored edges in that advancement. The edge weights of the anchored edges can still change with training in the current stage, and when the absolute value of the change is greater than the cross-stage conflict determination threshold τ, it is included in the cross-stage edge weight conflict rate; the role of the cross-stage consistency constraint described in Equation (6) is to compress this change of the anchored edges to within τ, thereby reducing the cross-stage edge weight conflict rate of the anchored edges.

[0055] Step S7: Output the dynamic knowledge graph data structure, including a node set, an edge set, an edge weight matrix, and a stage version identifier. The association strength between disease nodes and phenotype nodes in the edge weight matrix is ​​sorted in descending order of value and stored and displayed as the edge attributes of the graph for medical personnel to consult and reference.

[0056] Example 2 To verify the technical functions of the three mechanisms in steps S3, S4 and S6, this embodiment performs step-by-step ablation verification.

[0057] It should be noted that in this embodiment, each control group only disables the corresponding mechanism; the remaining processes, hyperparameters, training rounds, and random seeds remain consistent. No artificial drift is injected into any edge, and the loss term is not amplified for the control group without consistency constraints. Therefore, the differences between groups arise entirely from the opening and closing of the mechanism itself. In this embodiment, the five-layer concatenation refers to the unidirectional concatenation of five steps: data access, dual-channel term alignment, cross-ontology fusion, knowledge subbase assembly, and edge weight update. There are no feedback paths between the steps; that is, the closed-loop feedback and level 3 gating in step S3, the stage-conditional fusion in step S4, and the anchoring consistency constraints in step S6 are all disabled simultaneously.

[0058] The basic sample set includes 10 rare diseases, 45 Chinese phenotypic descriptions, 24 disease-phenotype control examples, 25 terminology identification and annotation examples, and 3 case workflow examples. The ablation results are shown in Table 4. In this embodiment, the Chinese phenotypic mapping coverage rate is the ratio of the number of Chinese phenotypic descriptions correctly mapped to HPO identifiers to the total number of Chinese phenotypic descriptions; the top three association ranking hit rate is the ratio of the number of examples containing the gold standard disease among the top three diseases ranked by corrected association weights to the total number of examples; the terminology recognition F1 score is the F1 value of the Chinese clinical text entity recognition and terminology normalization results relative to manual annotation.

[0059] Table 4. Record of stepwise ablation results on the basic sample set Configuration Chinese phenotypic mapping coverage Top three hit rates in association sorting Anchored edge cross-stage edge weight conflict rate Term recognition F1 Five layers in series, all three mechanisms are turned off. 77.8% 79.2% 12.5% 0.884 Enable only the closed-loop feedback in step S3. 86.7% 91.7% 12.5% 0.913 Enable steps S3 and S4 95.6% 100% 12.5% 0.913 Enable anchoring constraints in steps S3, S4, and S6 100% 100% 0% 0.913 As shown in Table 4, the closed-loop feedback in step S3 mainly improves the coverage of Chinese phenotypes and the quality of the preceding disambiguation, while the stage-conditional fusion in step S4 mainly improves the quality of disease-phenotype association ranking. However, neither of these can reduce the cross-stage edge weight conflict rate of the anchoring edge. Only after enabling the anchoring consistency constraint in step S6 does the cross-stage edge weight conflict rate of the anchoring edge decrease from 12.5% ​​to 0. The three mechanisms perform different and non-substitutable technical functions.

[0060] Example 3 In the individual examination during the screening phase, the patient's self-reported progressive muscle weakness and dysphagia, as well as the muscle atrophy recorded in the physical examination, were mapped through the SNOMED CT channel and the CHPO channel, respectively, and merged on the same UMLS concept identifier. The fusion confidence was calculated according to formula (1) and was not lower than the first threshold, triggering the pre-stage disambiguation reordering and unlocking of the pending statement. The atlas at this stage contains 7 nodes and 6 edges, with an average edge weight of 0.8686. After advancing to the auxiliary diagnosis stage, the atlas expanded from 7 nodes and 6 edges to 10 nodes and 19 edges, with a cross-stage edge weight conflict rate of 0 and a final training loss of 0.0041. The total number of edge operations when reconstructing in the full-scale manner in each diagnosis and treatment stage was 93. According to the incremental method of this invention, the anchored edges inherit the hidden state and initial edge weight values ​​of the previous stage without repeating the complete edge weight evolution. The number of complete edge weight evolutions reduced by one stage advancement is equal to the number of anchored edges in that advancement.

[0061] In a cohort of 6 cases of amyotrophic lateral sclerosis (ALS), cross-source conceptual marker convergence was 5 / 6, Chinese phenotype hits via four-hop aligned chains were 13 / 13, and hits via HPO Chinese direct translation alone were 8 / 13. It should be noted that the sample size in this cohort is very small, and the above results are only used to illustrate the feasibility of the procedure and do not constitute a conclusion on clinical performance.

[0062] To verify the method's ability to access real medical texts and structured population variables, this embodiment also performed an offline data access check. This check read 58,550 inpatient records and 21,641 patient discharge texts derived from a publicly available intensive care unit database. Of these, 32,907 discharge records matched the current SNOMED CT English label, resulting in a label matching rate of 56.20%. It also read 111,797 questionnaire records from 11 periods of a publicly available population health survey database, completing field parsing and variable alignment. The above data pertains to common disease populations and is limited to verifying the engineering feasibility of data reading, field parsing, and terminology coverage. It does not constitute performance verification for rare disease scenarios, nor is it equivalent to gold standard verification for rare disease-specific electronic medical records. Verification of rare disease-specific electronic medical records requires separate procedures after obtaining ethical approval, data anonymization, and data authorization.

[0063] Example 4 like Figure 2 As shown, this embodiment provides a multi-source knowledge fusion and dynamic knowledge graph construction system for the entire process of rare disease diagnosis and treatment. It includes a multi-source data access module, a dual-channel terminology alignment module, a concept identification closed-loop feedback module, a stage-based conditional fusion module, a knowledge sub-base stage assembly module, a dynamic edge weight update and cross-stage anchoring module, and a graph storage and stage version output module. Each module is used to execute steps S1 to S7 in Embodiment 1. The system also includes a storage and configuration unit for storing seven types of knowledge sub-bases, stage mode configurations, sets of expressions to be confirmed and their unlocking level identifiers, and graph versions for each stage.

[0064] The data flow between modules is as follows: the multi-source data access module sends data fragments with stage identifiers to the dual-channel terminology alignment module; the dual-channel terminology alignment module sends candidate clinical concepts, candidate phenotypes, and fusion confidence scores to the concept identifier closed-loop feedback module; the concept identifier closed-loop feedback module sends the re-ranking results back to the dual-channel terminology alignment module and sends the unlocked valid mappings to the stage conditional fusion module; the stage conditional fusion module sends the corrected association weights and conflict resolution results to the knowledge sub-base stage assembly module; the knowledge sub-base stage assembly module sends the candidate node set and candidate edge set to the dynamic edge weight update and cross-stage anchoring module; the dynamic edge weight update and cross-stage anchoring module sends the cross-stage edge weight conflict rate of the anchored edges back to the concept identifier closed-loop feedback module as the judgment basis for cross-stage anchoring confirmation gating, and sends the edge weight matrix to the graph storage and stage version output module.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for multi-source knowledge fusion and dynamic knowledge graph construction for the entire process of diagnosis and treatment of rare diseases, characterized in that, include: S1. Acquire multi-source data and divide the data into segments according to the diagnosis and treatment stage identifiers. The diagnosis and treatment stages include the screening stage, auxiliary diagnosis stage, treatment stage, and follow-up management stage, which are carried out in sequence. S2. Perform SNOMED CT clinical terminology normalization on the Chinese electronic medical record text in the multi-source data to obtain candidate clinical concepts and their channel confidence levels. s Perform CHPO Chinese phenotype mapping on the Chinese self-narrative text in the multi-source data to obtain candidate phenotypes and their channel confidence scores c. p Candidate clinical concepts and candidate phenotypes are mapped to UMLS concept labels respectively, and the fusion confidence c of candidates merging to the same concept label is calculated according to formula (1). f : c f =1-(1-c s )(1-c p ) (1) S3, using the HPO phenotype set H obtained in step S2... conv For each candidate clinical concept with the same Chinese expression, the score is recalculated and reordered to obtain the disambiguated clinical concepts. Chinese expressions not covered by the CHPO core thesaurus are placed in a pending confirmation set, and are only written into the valid mapping and participate in subsequent weight calculations after successively meeting the following three levels of gating: The first level is that the Chinese expression has been merged into the same UMLS concept identifier and its fusion confidence c. f Not less than the first threshold; The second level is the set of priority semantic types to which the semantic type of the Chinese expression belongs in the current diagnosis and treatment stage; the third level is considered to be satisfied when the cross-stage edge weight conflict rate of the anchor edge obtained in the previous diagnosis and treatment stage is not higher than the second threshold and there is no previous diagnosis and treatment stage. S4. Determine the fusion coefficient α based on the current treatment stage t. h (t), with α h (t) The corrected association weight w is obtained by weighting and summing the publicly known prior weights and the local Chinese clinical statistical weights for the association between disease d and phenotype h. d,h (t); α h (t) consists of the basic fusion coefficient, the stage bias term of t, and c. f To be determined jointly; S5. Based on the set of allowed edge types in the current treatment stage, select nodes and relationships from the pre-built knowledge sub-base to generate a candidate node set and a candidate edge set; S6. The gated loop unit inputs the vector x according to the edge (i,j) in the candidate edge set. ij t The edge weight w in the current treatment stage is derived from the hidden state of the edge in the previous treatment stage. ij t Take the edges whose type belongs to the intersection of the allowed edge types of the two adjacent treatment stages from the intersection of the edge sets of those two adjacent stages, and form the anchored edge set E. anchor (t); Only anchor edges are subject to cross-stage consistency constraints that reduce the edge weight difference between two adjacent treatment stages; no such constraints are applied to edges other than anchor edges. S7. Output a dynamic knowledge graph data structure including a set of nodes, a set of edges, an edge weight matrix, and a stage version identifier; wherein, steps S2 to S7 are repeated sequentially for each treatment stage according to the order of treatment stage progression, and the edge set and edge weight matrix of each treatment stage are saved for subsequent treatment stages to call.

2. The construction method according to claim 1, characterized in that, In step S1, the multi-source data includes Chinese electronic medical record text, Chinese self-report text, and structured medical codes; the nodes of the dynamic knowledge graph include disease nodes, phenotype nodes, gene nodes, test nodes, drug nodes, operation nodes, and follow-up nodes. In step S5, the knowledge sub-base consists of seven categories: rare disease seed database, phenotype database, pathogenic gene and variant database, testing and examination database, drug and treatment plan database, diagnosis and treatment pathway and follow-up management database, and terminology mapping database.

3. The construction method according to claim 1, characterized in that, In step S3, the scores of candidate clinical concepts are recalculated according to equation (2): r k =min(1,c k +η·II(h k ∈H conv )) (2) In equation (2), r k For the recalculated score of the k-th candidate clinical concept, c k h represents the original confidence level of the k-th candidate clinical concept. k Let I be the HPO identifier obtained by mapping the k-th candidate clinical concept using UMLS concept identifiers, where I is the indicator function and η is the feedback gain coefficient ranging from 0.10 to 0.20; according to r k Reorder from largest to smallest, and take the one that appears first after disambiguation as the clinical concept; H conv Only the HPO identifier corresponding to the following aggregate term will be included: the confidence level c of the two channels in the aggregate term. s With c p All are not lower than the lower limit of the channel confidence level, and the fusion confidence level c of the convergence term is also not lower than the lower limit of the channel confidence level. f Not less than the first threshold.

4. The construction method according to claim 1, characterized in that, In step S3, the first threshold is between 0.60 and 0.80, and the second threshold is between 0.05 and 0.

15. Each Chinese expression in the set to be confirmed carries an unlocking level identifier, which records the highest gating level passed by the Chinese expression in the most recent judgment. The value is level one, level two, or level three. Only Chinese expressions with an unlocking level identifier of level three are written into the valid mapping and participate in the calculation of the corrected association weight. When a certain level of gating is no longer satisfied in the subsequent diagnosis and treatment stage, the unlocking level identifier of Chinese expressions with an unlocking level identifier not lower than that level is downgraded to the level before that level. Chinese expressions that are lower than level three after downgrading are returned to the set to be confirmed and do not participate in the calculation of the corrected association weight in the diagnosis and treatment stage after the downgrade. In the screening stage, there is no preceding adjacent diagnosis and treatment stage, the anchored edge set is an empty set, the cross-stage edge weight conflict rate is recorded as 0, and the level three gating is considered satisfied.

5. The construction method according to claim 1, characterized in that, The set of permissible edge types is defined as follows: For the screening phase, the permissible edge types are screening risk edge, phenotypic matching edge, and gene pathogenicity edge; for the auxiliary diagnosis phase, the permissible edge types are phenotypic matching edge, abnormal test result edge, differential association edge, and gene pathogenicity edge; for the treatment phase, the permissible edge types are phenotypic matching edge, abnormal test result edge, treatment plan edge, indication edge, and contraindication association edge; and for the follow-up management phase, the permissible edge types are phenotypic matching edge, treatment plan edge, follow-up edge, and management plan edge. Therefore, the anchoring edge types between the screening phase and the auxiliary diagnosis phase are phenotypic matching edge and gene pathogenicity edge; the anchoring edge types between the auxiliary diagnosis phase and the treatment phase are phenotypic matching edge and abnormal test result edge; and the anchoring edge types between the treatment phase and the follow-up management phase are phenotypic matching edge and treatment plan edge.

6. The construction method according to claim 1, characterized in that, In step S4, the fusion coefficient α h (t) is calculated according to equation (3), and the corrected correlation weight w is calculated. d,h (t) is calculated according to formula (4): a h (t)=clip [0.05,0.95] (α0+β t -γ·c f ) (3) w d,h (t)=α h (t)·w d,h prior +(1-α h (t))·w d,h local (4) In equations (3) and (4), clip is the interval truncation operation, the basic fusion coefficient α0 is 0.35 to 0.45, the fusion confidence influence coefficient γ is 0.15 to 0.25, and w d,h prior To disclose the prior weights of the association between disease and phenotype, w d,h local Local Chinese clinical statistical weights; stage bias term β t Positive values ​​are used in the screening stage, zero values ​​are used in the auxiliary diagnosis stage, and negative values ​​are used in the treatment and follow-up management stages, so that the prior weight has a higher proportion in the screening stage than in the treatment and follow-up management stages. In step S4, the source weight reduction coefficient and conflict resolution priority are also determined according to the priority semantic type set of the current diagnosis and treatment stage: for sources whose semantic types do not belong to the priority semantic type set of the current diagnosis and treatment stage, their confidence is multiplied by the weight reduction coefficient before participating in conflict resolution. The priority of conflict resolution is determined according to the order of the priority semantic type set of the current diagnosis and treatment stage, and the value given by the semantic type source that ranks first is adopted first.

7. The construction method according to claim 1, characterized in that, In step S6, the anchored edge set is the set of edges whose type belongs to the intersection of the edge sets of two adjacent treatment stages and the intersection of the allowed edge type sets of the two adjacent treatment stages; the cross-stage consistency constraint is applied in the form of cross-stage consistency loss according to equation (6): L cons =S (i,j)∈Eanchor(t) (w ij t -w ij t-1 ) 2 (6) For an edge (i,j) in the anchored edge set, if the absolute value of the change in its edge weight is greater than the cross-stage conflict determination threshold τ, then the edge is determined to be a cross-stage conflict edge. τ takes a value of 0.10 to 0.

20. The cross-stage edge weight conflict rate of the anchored edge is the ratio of the number of edges in the anchored edge set that are determined to be cross-stage conflict edges to the total number of edges in the anchored edge set. When the anchored edge set is empty, the cross-stage edge weight conflict rate is recorded as 0; the cross-stage edge weight conflict rate serves as the basis for the third-level gating in step S3.

8. The construction method according to claim 7, characterized in that, In step S6, the input vector x is... ij t Including node type embedding, edge type embedding, and fusion confidence c f The treatment stage coding and the edge weight of the edge in the previous treatment stage; the edge weight evolves according to formula (5): h ij t =GRU(x ij t ,h ij t-1 ),w ij t =σ(u T h ij t ) (5) In equation (5), h ij t and h ij t-1 Let be the hidden states of edge (i,j) at treatment stage t and treatment stage t-1, respectively; u be the output head parameter; and σ be the activation function that constrains the edge weights to the interval between 0 and 1. The total loss L of the model training is determined according to equation (7): L=(1+λ1)L dpa +λ2L cons (7) In equation (7), L dpa For the disease-phenotype association ranking loss, L cons Let λ1 and λ2 be the cross-stage consistency loss in equation (6), where λ1 is 0.20 to 0.40 and λ2 is 0.05 to 0.

15.

9. The construction method according to claim 8, characterized in that, When the diagnosis and treatment stage progresses from t-1 to t, the edge input vector is constructed for the edges outside the anchored edge set and the complete edge weight evolution is performed according to Equation (5); the edge input vector is not reconstructed for the edges within the anchored edge set, but its hidden state in the previous diagnosis and treatment stage is inherited, and its edge weight in the previous diagnosis and treatment stage is used as the initial value of the edge weight in the current diagnosis and treatment stage. The edge weight is still updated under the cross-stage consistency constraint of Equation (6), and the full reconstruction of the dynamic knowledge graph is not performed.

10. A system for multi-source knowledge fusion and dynamic knowledge graph construction for the entire process of diagnosis and treatment of rare diseases, characterized in that: include: A multi-source data access module, used to perform step S1 as described in claim 1; A dual-channel term alignment module for performing step S2 as described in claim 1; A concept identification closed-loop feedback module is used to execute step S3 as described in claim 1; A stage-conditional fusion module is used to execute step S4 as described in claim 1; The knowledge sub-base stage assembly module is used to execute step S5 as described in claim 1; The dynamic edge weight update and cross-stage anchoring module is used to perform step S6 as described in claim 1; The map storage and stage version output module is used to perform step S7 as described in claim 1.

Images