Medical instrument intelligent retrieval system and method based on knowledge graph

Abstract

The invention discloses an intelligent medical instrument retrieval system and method based on a knowledge graph, and relates to the technical field of information retrieval of medical instruments, and the method comprises the following steps: S1, constructing a dynamic term network: extracting terms from registration data, clinical literatures, specifications and industrial standard codes of the medical instruments, and mapping the multi-source terms to a unified vector space through a heterogeneous term dynamic projection module. According to the intelligent medical instrument retrieval system and method based on the knowledge graph, the problems of semantic segmentation and information lag of multi-source heterogeneous terms in the field of medical instruments are effectively solved by constructing a dynamic term network and a dual-drive knowledge graph updating mechanism. The adversarial alignment network realizes accurate semantic mapping of cross-source terms, ensures differential expressions of commodity names, learning names, industry codes and the like, and accords with medical consensus in a unified vector space.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information retrieval of medical devices, in particular to a medical device intelligent retrieval system and method based on a knowledge graph. BACKGROUND

[0002] In the field of information retrieval technology, especially in the highly specialized field of medical devices, an intelligent retrieval system and method based on a knowledge graph is an important direction to improve retrieval efficiency and accuracy. Existing technologies mostly use general knowledge graph construction and information retrieval framework to process related data, but the terminology ecology in the medical device field is complex and unique: data comes from multiple sources such as registration data, clinical literature, etc., and the terminology system varies greatly; there are a large number of highly specialized terms, and the terms are constantly evolving due to standard updates and new product introductions, etc., and there are complex synonymous, near-synonymous, and multi-level relationships between terms. These characteristics pose serious challenges to knowledge graph-based retrieval. Existing general technologies are difficult to cope with: named entity recognition and entity linking tools are difficult to accurately process professional entities and attributes, and cannot achieve accurate mapping and semantic alignment of cross-source terms; knowledge graph construction is based on static data, making it difficult to incorporate new terms and update information and establish correlations in a timely manner; general semantic matching methods are difficult to understand complex professional query intentions, and cannot support deep professional logical association reasoning. This results in existing systems having many noisy retrieval results, missing key information, low precision and recall, and difficulty in meeting the requirements of accuracy, relevance, and timeliness in the field. Therefore, the technical problem to be solved at present is: how to overcome the above challenges in the terminology system, achieve accurate semantic alignment across data sources, and support deep association retrieval based on complex professional logic and dynamic industry standards and regulations. SUMMARY

[0003] To achieve the above purpose, the present application is implemented by the following technical scheme: a medical device intelligent retrieval method based on a knowledge graph, comprising the following steps: S1: Construct a dynamic terminology network: extract terms from medical device registration data, clinical literature, instruction manuals, and industry standard codes, and map multiple source terms to a unified vector space through a heterogeneous term dynamic projection module; S2: Perform dual-driven knowledge graph update: monitor industry standard updates or regulation status change events through a dynamic perception engine, trigger attribute updates of associated graph nodes, and generate a candidate update set of affected historical nodes based on term semantic association strength; S3: Intention-graph joint reasoning retrieval: analyze the medical device professional intention type of the user query, expand the query word set associated by synonymy or hierarchy, perform multi-path traversal retrieval along the attribute path, causal path, and standard path of the knowledge graph, and output the results after preliminary screening and fine screening.

[0004] Preferably, the step S1 comprises: Generating a topological constraint vector of the term by parsing the hierarchical relationship of the industry standard code through the structured channel; Generating a context semantic vector of the term by using a domain pre-training language model through an unstructured channel; Optimizing the vector space by using an adversarial alignment network, so that the vector distance of synonymous terms is less than a preset value.

[0005] Preferably, the operation of the adversarial alignment network comprises: Inputting the topological constraint vector and the context semantic vector into the discriminator to force the output distribution to be consistent; Iterative optimization based on the adversarial loss function until the semantic alignment of cross-source terms is completed.

[0006] Preferably, the double-driven knowledge graph update in the step S2 comprises: Responding to the regulatory state change event to update the attributes of the associated graph node in real time; When the semantic association strength between the new term and the historical term exceeds a preset threshold, automatically backtracking to the associated historical product node and generating a candidate update set.

[0007] Preferably, the medical device professional intent type of the step S3 in the step S3 comprises: Identifying the query type by using a medical device special intent classifier, the type including a clinical problem, a failure mode, or a composite attribute condition; Based on the term relationship of the dynamic term network, the original query word is expanded into a query word set containing synonymous words and hierarchical associated words.

[0008] Preferably, the multi-path traversal retrieval comprises: Attribute path: matching device specification parameters and registration status; Causal path: tracing the improved model associated with the failure mode; Standard path: associating the technical category defined by the industry code.

[0009] Preferably, during the initial screening and fine screening of the results in the step S3, comprising: Initial screening: sorting the retrieval results based on the term vector similarity of the dynamic term network; Fine screening: comparing the results with the medical device professional rule library through a logical consistency checker to eliminate the results that violate the rules.

[0010] A medical device intelligent retrieval system based on a knowledge graph, for realizing a medical device intelligent retrieval method based on a knowledge graph, comprising: A dynamic term network construction module is configured to construct a dynamic term network, wherein terms are extracted from registration information of medical devices, clinical literature, instruction manuals and industry standard codes, and the multi-source terms are mapped to a unified vector space through a heterogeneous term dynamic projection module; A double-driven graph updating module is configured to perform double-driven knowledge graph updating, wherein attribute updating of associated graph nodes is triggered by monitoring industry standard updating or regulation state changing events through a dynamic perception engine, and a candidate updating set of affected historical nodes is generated based on term semantic association strength backtracking; An intent-graph joint reasoning module is configured to perform intent-graph joint reasoning retrieval, wherein a medical device professional intent type of a user query is analyzed, a query word set associated by synonymy or hierarchy is generated, multi-path traversal retrieval is performed along attribute paths, causal paths and standard paths of a knowledge graph, and a result is output after preliminary screening and fine screening.

[0011] Preferably, the system further comprises: A user query interface configured to receive a natural language or structured query instruction; A retrieval result output interface configured to return a sorted device information list after screening.

[0012] A computer readable storage medium storing a computer program, wherein the computer program is executed by a processor to implement steps of a medical device intelligent retrieval method based on a knowledge graph.

[0013] The application provides a medical device intelligent retrieval system and method based on a knowledge graph. The medical device intelligent retrieval system and method based on a knowledge graph effectively solves the problems of semantic disconnection and information lag of multi-source heterogeneous terms in the field of medical devices by constructing a dynamic term network and a double-driven knowledge graph updating mechanism. The cross-source term semantic mapping is realized by the adversarial alignment network, ensuring that the different expressions such as trade names, scientific names and industry codes conform to the medical consensus in the unified vector space. The backtracking updating mechanism based on the semantic association threshold dynamically maintains the professional logical association between new terms and historical data, avoiding the invalidation of historical data caused by standard iteration or product evolution. The semantic consistency and timeliness of the retrieval data are ensured from the root.

[0014] The medical device intelligent retrieval system and method based on the knowledge graph improve the retrieval accuracy and safety in complex professional scenarios by means of the intention-graph combined reasoning retrieval and double-layer screening mechanism. The multi-path traversal strategy intelligently selects the attribute path, the cause-effect path or the standard path in combination with the query intention, realizes the deep mining of the device parameter compliance, the defect improvement scheme and the clinical adaptability, the rule library driven fine screening link is embedded with the medical device specific constraints, and the results that violate the core safety criteria are automatically filtered in the output link and the risks are indicated. Finally, the retrieval closed loop with the semantic accuracy, the medical compliance and the dynamic adaptability is formed, and the strict requirements of the medical device field on the professional retrieval are met. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 It is a module interaction schematic diagram of the medical device intelligent retrieval system based on the knowledge graph of the application; Figure 2 It is a flowchart of the medical device intelligent retrieval method based on the knowledge graph of the application. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0017] Please refer to Figure 1 and Figure 2 The application provides a technical solution: a medical device intelligent retrieval method based on a knowledge graph, comprising the following steps: S1: constructing a dynamic term network: extracting terms from the registration data, clinical literature, instruction manuals and industry standard codes of medical devices, mapping the multi-source terms to a unified vector space through a heterogeneous term dynamic projection module; S2: performing double-driven knowledge graph updating: monitoring the industry standard updating or regulation state changing events through a dynamic perception engine, triggering the attribute updating of the associated graph nodes, and generating a candidate updating set of the affected historical nodes based on the term semantic association strength; S3: intention-graph combined reasoning retrieval: analyzing the medical device professional intention type of the user query, expanding the query word set associated with synonyms or levels, performing multi-path traversal retrieval along the attribute path, the cause-effect path and the standard path of the knowledge graph, and outputting the results after the initial screening and fine screening.

[0018] It needs to be further explained that in the specific implementation process, when implementing the medical device intelligent retrieval method based on the knowledge graph, first, a dynamic term network is constructed: multi-source terms are extracted from medical device registration data, clinical literature, instruction manuals and industry standard codes, and are processed through a heterogeneous term dynamic projection module. The module enables the hierarchical relationship of the structured channel to parse the industry standard code, generating a topological constraint vector of the term; at the same time, the unstructured channel is enabled, and the pre-trained language model in the medical device field is used to generate a context semantic vector of the term; then, an adversarial alignment network is used to input the two types of vectors into the discriminator to force the distribution to be consistent, and based on the adversarial loss function, the iteration optimization is performed to reduce the vector distance of synonymous terms to within the preset range, and the semantic alignment of cross-source terms is realized.

[0019] Then, double-drive knowledge graph updating is performed: the industry standard updating or regulation state change event is monitored through a dynamic perception engine. When the regulation state changes, the attributes of the associated graph nodes are updated in real time; when the semantic association strength between the new term and the historical term exceeds the preset threshold, the associated historical product node is automatically traced back, a candidate update set is generated, and attribute synchronization is performed to ensure that the graph dynamically associates historical data.

[0020] Finally, intent-graph joint reasoning retrieval is performed: first, a medical device special intent classifier is used to identify the user query type, if it is a clinical problem, the device type is associated, if it is a failure mode, the product defect is associated, and if it is a composite attribute condition, the multi-dimensional parameters are analyzed; based on the term relationship of the dynamic term network, the original query words are expanded into a query word set containing synonymous words and hierarchical associated words. Then, the three types of paths in the knowledge graph are traversed: in the attribute path, the device specification parameters and the registration state are matched; in the causal path, the improved model associated with the failure mode is traced back; in the standard path, the technical category defined by the industry code is associated. The retrieval results are first sorted and screened based on the term vector similarity, and then compared with the medical device professional rule library through a logical consistency checker, and the results that violate the rules are removed to output the final list.

[0021] Step S1 includes: The hierarchical relationship of the industry standard code is parsed through the structured channel to generate a topological constraint vector of the term; The context semantic vector of the term is generated through the unstructured channel using the field pre-trained language model; The vector space is optimized using the adversarial alignment network, so that the vector distance of synonymous terms is less than a preset value.

[0022] It needs to be further explained that in the specific implementation process, when constructing the dynamic term network, first, the standard terms are extracted from the structured fields of the medical device registration information, and the descriptive terms are extracted from the unstructured text of the clinical literature and the instruction manual. For industry standard coding, the hierarchical relationship is parsed through the structured channel: for example, the tree structure of the Global Medical Device Nomenclature coding is converted into the parent-child dependency relationship between terms, generating a topological constraint vector that carries the hierarchical depth and branch position information.

[0023] For unstructured text terms, a pre-trained language model in the medical device field is input, which generates a context semantic vector containing semantic correlation by analyzing the co-occurrence patterns of terms in the context of clinical literature and the parameter description of the instruction manual.

[0024] Subsequently, the adversarial alignment network is started: the topological constraint vector and the context semantic vector of the same term are input into the discriminator in parallel, and the discriminator generates a feedback signal by comparing the distribution difference of the two types of vectors; based on the feedback signal, the vector space mapping relationship is iteratively adjusted until the discriminator cannot distinguish the source distribution of the two types of vectors.

[0025] In this process, if the vector distance of a pair of synonymous terms is greater than the preset upper limit distance, the vector attraction weight is increased; wherein the pair of synonymous terms includes "blood glucose meter" and "glucose monitor"; on the contrary, if the irrelevant term vector distance is too close, the repulsion weight is increased. Finally, the cross-source terms meet the medical semantic consistency in the unified space.

[0026] The operation of the adversarial alignment network includes: The topological constraint vector and the context semantic vector are input into the discriminator to force the output distribution to be consistent; Iterative optimization based on the adversarial loss function until the semantic alignment of cross-source terms is completed.

[0027] It needs to be further explained that in the specific implementation process, in the running process of the adversarial alignment network, first, the topological constraint vector and the context semantic vector of the same medical device term are input into the discriminator in parallel. The discriminator analyzes the distribution characteristics of the two types of vectors through multiple layers of neural networks: if the discriminator detects that the topological constraint vector has the hierarchical clustering characteristics unique to industry coding, and the context semantic vector presents the continuous distribution characteristics of unstructured text, then a difference signal is output.

[0028] Based on the signal, the generator performs spatial transformation on the two types of vectors: the topological constraint vector is injected with a semantic smoothing operation to weaken its discrete hierarchical boundaries; the context semantic vector is subjected to a structure regularization operation to enhance its association with industry standards.

[0029] The process is iterated until the discriminator cannot distinguish the vector source type according to the distribution characteristics. In the optimization phase, if the cosine distance of the synonymous term pair exceeds the preset upper limit, the attraction gradient update step of its vector space is increased; if the distance of the synonymous term pair is lower than the preset lower limit, the repulsion gradient is applied; wherein the synonymous term pair also includes "vascular stent" and "coronary stent", and the synonymous term pair includes "blood glucose meter" and "cardiac pacemaker".

[0030] Finally, the cross-source terms satisfy the following conditions in the unified vector space: the synonymous term distance tends to zero, the synonymous term distance expands to a reasonable interval, and the mixed distribution of structured and unstructured vectors is indistinguishable.

[0031] The double-driven knowledge graph update in step S2 includes: In response to the regulatory state change event, the attributes of the associated graph nodes are updated in real time; When the semantic association strength between the new term and the historical term exceeds the preset threshold, the history product nodes associated with the new term are automatically traced back, and a candidate update set is generated.

[0036] It should be further explained that, in the specific implementation process, when the double-driven knowledge graph update is performed, the dynamic perception engine continuously monitors the medical device industry standard revision announcements and the regulatory state change files published by the regulatory agencies. When a specific medical device registration state change event is detected, the associated device node in the knowledge graph is immediately located, and its registration state attribute is updated synchronously.

[0033] For the new term event, the similarity between the semantic association vector of the new term in the dynamic term network and the historical term vector is calculated: if the similarity exceeds the preset threshold, for example, the vector similarity between the new term "degradable magnesium alloy coronary stent" and the historical term "bioabsorbable vascular stent" reaches the threshold, the backtracking mechanism is started, including: tracing along the hierarchical relationship and synonymous chain in the term network, locating all associated historical product nodes, such as early metal stent models, and generating a candidate set containing attributes that need to be updated.

[0034] After the candidate set is generated, the attribute synchronization operation is automatically triggered: the historical node is supplemented with new material association attributes, and its iteration relationship with the new model is marked. If the similarity does not reach the threshold, such as the new term "ophthalmic surgical navigation system" has no significant association with "orthopedic implants", only the basic node insertion is performed without triggering backtracking. This mechanism ensures that the knowledge graph dynamically maintains the semantic coherence of historical data when new terms are included.

[0035] The medical device professional intent type of the user query in step S3 includes: The query type is identified by a medical device-specific intent classifier, including clinical problems, failure modes, or composite attribute conditions; Based on the dynamic term network, the original query is expanded to a query set containing synonyms and hierarchical associated terms.

[0036] It needs to be further explained that in the specific implementation process, when analyzing the user query intention, the original query is first input into the medical device special purpose intention classifier. The classifier determines the type by analyzing the keyword combination mode of the query text: if the query contains clinical operation description words such as minimally invasive treatment and intervention, it is identified as a clinical problem type; if it contains device abnormal state words such as fracture and failure, it is determined as a failure mode type; if it contains device parameters and regulatory conditions such as Class III implantable and titanium alloy material, it is classified as a composite attribute condition type.

[0037] After determining the type, the query expansion based on the dynamic term network is started: for clinical problem type queries, extract the device type synonym set strongly associated with the clinical description from the term network, such as “mitral regurgitation treatment” expanded to {“mitral valve clip”, “transcatheter repair system”}; for failure mode type queries, trace the cause-effect chain to obtain the synonymous expressions of associated defects, such as “catheter fracture” expanded to {“pipe body rupture”, “structure separation”}; for composite attribute type queries, expand the parameter conditions by level, such as “III active” associated with specific product categories. During the expansion process, if there are multiple associated branches in the term network, prefer the branch with high semantic similarity; if the size of the expanded term set exceeds the preset upper limit, automatically trim the low-frequency words based on the term usage frequency.

[0038] Multi-path traversal retrieval includes: Attribute path: matching device specification parameters and registration status; Causal path: tracing the improved model associated with the failure mode; Standard path: associating with the technical category defined by industry code.

[0039] It needs to be further explained that in the specific implementation process, when starting multi-path traversal retrieval, select the knowledge graph path according to the expanded query term set: if the query type is a composite attribute condition, prefer to traverse the attribute path, including: matching the specification parameters and registration status in the device node, when there are multiple levels of standards, drill down to the most fine-grained attribute node verification, wherein the specification parameters include material composition, size range, and the registration status includes classification level, approved area; If the query type is a failure mode, activate the causal path, including: locating the problem node along the defect relationship chain preset in the graph, tracing its associated improved model node, and verifying whether the improved model meets the current registration status, wherein the problem node includes catheter fracture, and the improved model node includes a new catheter with the “replacement scheme” attribute; If the query type is a clinical problem, the standard path is traversed, including: comparing the extended term set with the technical category defined by the industry code for semantic coverage, and when there are multiple coding intersections, such as heart intervention instruments associated with both cardiovascular and surgical instrument codes, the main coding path covering the core intent of the clinical description is selected.

[0040] During the traversal process, if the same instrument node is hit in multiple paths, the path weight is superimposed; if the path retrieval result is empty, cross-path association backtracking is triggered, such as jumping from the failure mode path to the attribute path to screen similar material instruments.

[0041] During the initial screening and fine screening of the results in step S3, the following is included: Initial screening: term vector similarity sorting of retrieval results based on dynamic term network; Fine screening: through logical consistency checker and medical instrument professional rule library comparison, eliminate results that violate rules.

[0042] It needs to be further explained that in the specific implementation process, after completing the multi-path traversal retrieval, the result set is executed for double-layer screening. In the initial screening stage, the term vector similarity calculation based on the dynamic term network: compare the instrument description terms in the retrieval results with the extended query term set for vector distance comparison, the smaller the distance value, the higher the ranking; when multiple results have similar distance values, the instrument node with higher industry standard coding level is preferred.

[0043] In the fine screening stage, the logical consistency checker is called, and the constraint conditions in the medical instrument professional rule library are loaded: if the retrieval result contains a composite attribute condition class instrument, check whether the parameter combination meets the regulatory requirements, for example, class III implant instruments must have a biocompatibility report attribute; if it is an improved model associated with the failure mode, verify whether it is marked as valid registration status and not canceled; if the result involves clinical indication conflict, such as orthopedic instruments appearing in cardiovascular treatment scenarios, trigger the contraindication items in the rule library for interception.

[0044] During the screening process, if the initial screening result has passed the fine screening rules, it is directly retained; if the fine screening finds that it violates the core rules, such as registration status failure, it is immediately eliminated; if it violates the secondary rules, such as material report version not updated, it is sorted and output with a warning identifier. Finally, ensure that the output list meets both semantic relevance and medical compliance.

[0045] A medical instrument intelligent retrieval system based on a knowledge graph, for realizing a medical instrument intelligent retrieval method based on a knowledge graph, comprising: A dynamic term network construction module for constructing a dynamic term network: extracting terms from medical instrument registration data, clinical literature, instructions and industry standard codes, and mapping multiple source terms to a unified vector space through a heterogeneous term dynamic projection module; A dual-driven knowledge graph updating module is configured to perform dual-driven knowledge graph updating: monitoring industry standard updates or regulatory status change events through a dynamic perception engine, triggering attribute updates of associated graph nodes, and generating a candidate update set of affected historical nodes based on the semantic association strength of the terms; An intent-graph joint reasoning module is configured to perform intent-graph joint reasoning retrieval: analyzing the medical device professional intent type of the user query, expanding the query word set associated by synonym or hierarchy, performing multi-path traversal retrieval along the attribute path, causal path, and standard path of the knowledge graph, and outputting the results after preliminary screening and fine screening.

[0046] It should be further explained that, in the specific implementation process, during system operation, the dynamic term network construction module first interfaces with the medical device multi-source database, extracts the structured fields of the registration data and the unstructured text of the clinical literature, generates a unified vector space through heterogeneous term dynamic projection, and outputs the cross-source aligned term network to the knowledge graph library.

[0047] The dual-driven graph updating module listens to the industry standard update events of the regulatory information publishing interface, and when detecting a regulatory status change, directly calls the graph library API to update the node attributes; when the dynamic perception engine captures new terms, it triggers semantic association strength calculation, and if it exceeds the set threshold, it sends a backtracking update instruction to the graph library.

[0048] After the intent-graph joint reasoning module receives the instructions of the user query interface, it analyzes the intent type through the term network data, generates an expanded query word set to drive multi-path traversal retrieval, and sends the retrieval results to the output interface after double-layer screening.

[0049] If the user query contains complex clinical conditions, the joint reasoning module synchronously requests the rule library for logical pre-checking; if the graph updating module performs large-scale backtracking, it automatically delays low-priority retrieval tasks to ensure system response stability.

[0050] It also includes: A user query interface that receives natural language or structured query instructions; A retrieval result output interface that returns a sorted list of device information.

[0051] It should be further explained that, in the specific implementation process, the user query interface supports natural language input and structured form submission. When receiving a natural language query, it automatically identifies the medical device professional parameters implied therein: if the description contains clinical scenario terms such as "cardiac intervention surgery", it extracts the core operation verbs and anatomical site keywords; if it contains product defect descriptions such as "catheter multiple breakage", it marks the failure frequency modifier; the structured query directly analyzes the preset fields, such as the material drop-down menu options and the registration classification check boxes.

[0052] The input content is standardized and passed to the intent-mapping joint inference module. The retrieval result output interface receives the sorted list after double-layer screening, each result carrying the standard code of the instrument, the registration number, and the key attribute summary; if the result is down-weighted due to the precision screening rule, such as a lagging material report version, an additional warning icon and revision prompt are attached. When the user selects a result, the causal path information in the knowledge graph is automatically associated, such as clicking on the defective instrument to display the improved model link, and simultaneously linking the supervision interface to verify the effectiveness of the current registration status. If the user submits a composite clinical condition query, the output list automatically folds the non-indicated instrument branch.

[0053] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a knowledge graph-based intelligent retrieval method for medical instruments. It needs to be further explained that, in the specific implementation process, when the computer program in the storage medium is executed by the processor, the registration data of the medical instrument multi-source database and the industry standard code are first loaded, and the dynamic term network construction module is driven to start the heterogeneous term projection: the hierarchical dependency relationship of the structured code is analyzed to generate a topological constraint vector, and the semantic vector is generated by executing the field pre-training on the clinical literature text, and the medical semantic distance of the cross-source terms is converged through the iterative optimization of the adversarial alignment network.

[0054] Then the dynamic perception engine listens to the supervision data interface, and if a regulatory status change event is detected, the double-driven update module is called to synchronize the graph node attributes; if the semantic association strength between the new term and the historical term exceeds the set threshold, the backtracking mechanism is triggered to update the associated historical nodes.

[0055] When the user query instruction is input, intent classification and term expansion are performed, and the retrieval result is traversed along the attribute, causal, and standard paths of the knowledge graph. After the retrieval result is sorted and screened by the term vector similarity, the rule library is called to perform precision screening: the parameter compliance of the composite attribute instrument is verified, the registration status effectiveness of the improved model is checked, and the contraindication rule is applied to the clinical conflict result to intercept. The final output is a sorted list carrying the standard code and risk identification.

[0056] It needs to be further explained that in the specific implementation process, when implementing the medical device intelligent retrieval method based on knowledge graph, first, the terms are extracted from the structured fields of medical device registration data, the text description of clinical literature, the technical parameter chapter of the instruction manual and the standard coding library of global medical device terminology system. For industry standard coding, the internal hierarchical relationship is analyzed through the structured channel, the tree branch structure of the coding is converted into the parent-child dependent topology between terms, and the vector carrying hierarchical position information is generated. For unstructured terms in clinical literature and instruction manual, input the language model pre-trained by massive medical device text. The model generates vectors containing semantic correlation by analyzing the co-occurrence rules of terms in surgical operation description, material performance parameters and other professional contexts.

[0057] Subsequently, the adversarial alignment network is started: the structured topology vector and the unstructured semantic vector of the same term are input into the multi-layer neural network discriminator in parallel, and the discriminator outputs a difference signal by comparing the distribution patterns of the two types of vectors; based on the signal, the generator applies semantic smoothing operation to weaken the discrete boundary of the structured vector, and performs structure regularization operation to enhance the standard correlation of the unstructured vector. This process is iterated until the discriminator cannot distinguish the source of the vector, and the vector distance of the synonymous term pair is stable within the range recognized by medical consensus.

[0058] The dynamic perception engine continuously monitors the industry standard revision documents and registration status change announcements released by regulatory agencies. When a specific device registration status update is detected, the corresponding node attributes in the knowledge graph are immediately located. For new terms, calculate their similarity with historical terms in the unified vector space: if the similarity exceeds the preset threshold, locate the associated historical device nodes along the hierarchical relationship chain and the synonymous relationship chain of the term network, generate an attribute update candidate set and supplement iterative relationship annotation; if the similarity does not reach the threshold, only perform new node insertion. This mechanism ensures the dynamic association of historical product data and new terms.

[0059] When the user inputs the query, the medical device dedicated intent classifier analyzes the text features: if the query contains surgical action words and anatomical site words, it is identified as a clinical problem type; if it contains abnormal state words of the device, it is determined as a failure mode type; if both technical parameters and regulatory conditions are present, it is classified as a complex attribute type. Query expansion based on dynamic term network: clinical problem class query is expanded to the set of synonymous terms of associated devices, failure mode class query is traced back to the chain of defect synonymous expressions, and complex attribute class query is drilled down to parameter sub-conditions. During the expansion process, preferentially select branches with high semantic similarity, and automatically trim low-frequency terms when the expansion size is too large.

[0060] The extended query word set drives the multi-path traversal of the knowledge graph: for composite attribute class queries, match the device parameters and registration status along the attribute path, and verify the parameters layer by layer to the finest granularity when there are multiple levels of standards; for failure mode class queries, locate the defect node along the causal path and trace back to the improved model, and verify the current registration validity of the improved model at the same time; for clinical problem class queries, compare the technical categories coded by the industry along the standard path, and select the main path that covers the core clinical intent when multiple codes intersect. If there is no result in a single path, switch to the complementary path to re-search.

[0061] The search results are first sorted based on the similarity of the term vector, and when the similarity is the same, the device with a high standard coding level is preferentially displayed. Then, the medical device professional rule library is called for fine screening: the composite attribute class result verifies whether the parameter combination meets the regulatory whitelist; the improved model class result verifies the validity of the registration status and the defect solution evidence chain; the clinical class result triggers the contraindication rule to intercept the conflict items of indications. Results that violate core safety rules are immediately excluded, and secondary defect items are output with reduced weight and additional warning explanations. The final list carries the device standard code, registration number, and key attribute summary, and the user can click on the result to verify the real-time validity of the registration status through the regulatory interface.

[0062] During system operation, if the double-drive update module performs large-scale historical data backtracking, low-priority retrieval requests are automatically delayed; when the user submits a complex clinical condition query, the joint reasoning module preloads the rule library to accelerate the fine screening process. All operations rely on computer programs for fixed execution: the termination condition of the adversarial alignment is determined by the convergence state of the medical semantic distance, the backtracking mechanism of the double-drive update is strictly controlled by the semantic association threshold, and the selection of the retrieval path and the application of the fine screening rule are both embedded with medical device field logic that cannot be replaced.

[0063] During the implementation of the adversarial alignment network, the convergence of the medical semantic distance is based on the distance range of the synonym term pairs annotated by clinical experts as the convergence benchmark; the structural regularization operation is essentially: mapping unstructured terms to the nearest parent node of the industry standard code; in the double-drive update priority rules, the regulatory status change response level is higher than the new term processing, and the historical node backtracking is only executed in real time when the system load is below the threshold; The multi-path conflict resolution mechanism includes: when the attribute path and the standard path conflict, the technical category of the standard code is used as the reference; the improved model retrieval of the causal path must pass the current registration status verification; the fine screening rule library is constructed from the following sources: the regulatory whitelist is derived from the appendix of the medical device supervision regulations; the contraindication rule library is generated by structuring the clinical diagnosis and treatment guidelines; in the system resource scheduling principle, the delay time of low-priority retrieval tasks is linearly related to the amount of backtracking data, and the rule preloading range of complex queries is determined by the depth of the intent type; By constructing a dynamic term network and a double-driven knowledge graph updating mechanism, the semantic fragmentation and information lag of multi-source heterogeneous terms in the medical device field are effectively solved. The adversarial alignment network realizes precise semantic mapping across source terms, ensuring that trade names, scientific names, industry codes, and other different expressions conform to medical consensus in a unified vector space. The backtracking updating mechanism based on semantic association thresholds dynamically maintains the professional logical association between new terms and historical data, avoiding the invalidation of historical data due to standard iteration or product evolution. The semantic consistency and timeliness of the search data are ensured from the root.

[0064] With the help of intent-graph joint reasoning retrieval and double-layer screening mechanism, the retrieval accuracy and safety in complex professional scenarios are improved. The multi-path traversal strategy intelligently selects attribute paths, causal paths, or standard paths according to the query intent, realizing deep mining of device parameter compliance, defect improvement schemes, and clinical adaptability. The rule library-driven fine screening process embeds medical device-specific constraints, automatically filtering results that violate core safety standards and clearly indicating risks. Ultimately, a retrieval closed loop with semantic precision, medical compliance, and dynamic adaptability is formed, meeting the stringent requirements of the medical device field for professional retrieval.

[0065] Efficiency of medical device R&D is improved: integrating product design documents, process parameters, material lists, and other unstructured data, a "product-component-process" knowledge graph is constructed. Engineers can quickly retrieve similar cases (such as "guide wire cleaning process parameters") through intelligent question answering, reducing the R&D cycle by 40%. According to the case of DataVisor, cardiovascular device companies achieve cross-department knowledge sharing (such as heart and coronary departments collaborating to optimize guide wire design), reducing repeated trial and error.

[0066] It should be noted that, in this text, relational terms such as first and second are used merely to distinguish one entity or action from another, without necessarily requiring or implying any actual such relationship or order between or among the entities or actions. Also, the terms "comprises," "comprising," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without more limitations, an element defined by the phrase "comprising a..." does not exclude the existence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0067] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A knowledge graph-based intelligent retrieval method for medical devices, characterized in that, Includes the following steps: S1: Constructing a dynamic terminology network: Extracting terms from medical device registration data, clinical literature, instructions for use, and industry standard codes, and mapping multi-source terms to a unified vector space through a heterogeneous terminology dynamic projection module; S2: Perform dual-drive knowledge graph updates: monitor industry standard updates or regulatory status changes through the dynamic perception engine, trigger attribute updates of related graph nodes, and generate a candidate update set of affected historical nodes based on the semantic association strength of terms. S3: Intent-Graph Joint Reasoning Retrieval: Parses the medical device professional intent type of the user query, expands and generates a set of synonyms or hierarchical related query terms, performs multi-path traversal retrieval along the attribute path, causal path and standard path of the knowledge graph, and outputs the results after initial screening and fine screening.

2. The intelligent medical device retrieval method based on knowledge graph according to claim 1, characterized in that: Step S1 includes: The hierarchical relationship of industry standard codes is parsed through structured channels to generate topological constraint vectors for terms; Contextual semantic vectors of terms are generated using domain-pretrained language models through unstructured channels; An adversarial alignment network is used to optimize the vector space, making the vector distance between synonyms less than a preset value.

3. The intelligent medical device retrieval method based on knowledge graph according to claim 2, characterized in that: The operations of the adversarial alignment network include: Input the topological constraint vector and the context semantic vector into the discriminator, and force its output distribution to be consistent; The process involves iterative optimization based on the adversarial loss function until semantic alignment of cross-source terms is completed.

4. The intelligent medical device retrieval method based on knowledge graph according to claim 1, characterized in that: The dual-driven knowledge graph update in step S2 includes: In response to regulatory status change events, update the attributes of associated graph nodes in real time; When the semantic association between a new term and a historical term exceeds a preset threshold, the system automatically backtracks to the associated historical product nodes and generates a candidate update set.

5. The intelligent medical device retrieval method based on knowledge graph according to claim 1, characterized in that: The medical device professional intent types parsed by the user in step S3 include: The query type is identified by a medical device-specific intent classifier, and the type includes clinical questions, failure modes, or composite attribute conditions. Based on the term relationships of the dynamic term network, the original query terms are expanded into a query term set that includes synonyms and hierarchical related terms.

6. The intelligent medical device retrieval method based on knowledge graph according to claim 5, characterized in that: The multi-path traversal retrieval includes: Attribute path: Matches device specifications and registration status; Causal path: Tracing the improved model based on the association of failure modes; Standard path: The technical scope defined by industry-related coding.

7. The intelligent medical device retrieval method based on knowledge graph according to claim 1, characterized in that: During the initial and fine screening of the results in step S3, the following steps are included: Initial screening: retrieval results ranked by term vector similarity based on a dynamic terminology network; Fine screening: The results that violate the rules are eliminated by comparing them with the medical device professional rule base through a logical consistency checker.

8. A knowledge graph-based intelligent medical device retrieval system, used to implement the method described in any one of claims 1-7, characterized in that, include: The dynamic terminology network construction module is used to build a dynamic terminology network: extracting terms from medical device registration data, clinical literature, instructions for use and industry standard codes, and mapping multi-source terms to a unified vector space through the heterogeneous terminology dynamic projection module; The dual-drive graph update module is used to perform dual-drive knowledge graph updates: by monitoring industry standard updates or regulatory status changes through the dynamic perception engine, it triggers attribute updates of associated graph nodes and backtracks to generate a candidate update set of affected historical nodes based on the semantic association strength of terms. The Intent-Graph Joint Reasoning module is used to perform intent-graph joint reasoning retrieval: it parses the medical device professional intent type of the user query, expands and generates a set of query terms with synonyms or hierarchical associations, performs multi-path traversal retrieval along the attribute path, causal path and standard path of the knowledge graph, and outputs the results after initial screening and fine screening.

9. The intelligent medical device retrieval system based on knowledge graph according to claim 8, characterized in that: Also includes: The user query interface receives natural language or structured query commands. The search results output interface returns a list of screened and sorted medical device information.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-7.

Images