Health product allergen risk compliance verification method and system based on multi-modal analysis

Abstract

The application provides a health product allergy risk compliance verification method and system based on multi-modal analysis, relates to the technical field of health product safety detection, and comprises the following steps: acquiring product formula data and label identification data, extracting component entities, performing multi-modal identification on the label identification data, and then performing correlation mapping to obtain product information graph; generating an allergen correlation subgraph through entity linking and knowledge graph extension, obtaining a component risk representation vector through graph embedding coding, and querying an allergen library; coupling the component content parameters and the risk representation vector, combining the allergen cross-reaction relationship to solve a comprehensive risk score and map an allergy risk level; constructing a risk propagation topology graph and labeling the risk level, and matching and verifying according to the compliance rules in the standard product library and combining the allergy risk level to obtain a compliance verification result.

Description

Technical Field

[0001] This invention relates to the field of health product safety testing technology, and in particular to a method and system for compliance verification of allergy risk in health products based on multimodal analysis. Background Technology

[0002] Allergy risk compliance verification before health products are launched on the market is a crucial step in ensuring consumer safety. Current practices typically rely on chemical property analysis of product ingredients and textual verification of labeling. For ingredients, the mainstream approach involves comparing each component against publicly available allergen databases, determining the presence of allergen exposure based on the matching results. For labeling, the standard process relies on manual review or basic optical character recognition (OCR) to extract explicit information from the packaging and manually compare it with the formula data to confirm the inclusion of allergen warnings. These practices have standardized execution procedures and relatively mature database support within the industry, forming the basic technical framework for current compliance verification.

[0003] However, existing verification methods, which rely on static comparison of single components, have blind spots in considering interactions between components. Different components in health products may produce cross-sensitization reactions after being metabolized or by the immune system; the matching results of a single component cannot reflect the true allergic risk of the entire composition. Name-based searches can only reconstruct the physical existence of components and cannot assess their synergistic sensitization intensity when participating in complex biological reactions. This can lead to missed detections of high-risk products and reduce the clinical reference value of compliance verification.

[0004] Furthermore, the existing practice of relying on explicit label information and fixed allergen lists for verification is limited in accuracy by the completeness of the labeling information and the update speed of the allergen knowledge base. Exemption clauses or non-standardized terminology in product labels may cause information loss in the identification process, while item-by-item comparison based on known lists can only cover common allergens already included in the database, lacking the ability to dynamically track newly emerging sensitizing components or hidden derivatives. Static and unrelated processing methods make it difficult to guarantee the consistency and traceability of verification results across different products. Summary of the Invention

[0005] This invention provides a method and system for compliance verification of allergy risk in health products based on multimodal analysis, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of this invention provides a method for compliance verification of allergy risk in health products based on multimodal analysis, comprising:

[0007] The product formula data and label identification data of health products are obtained. Entity extraction is performed on the product formula data to obtain a set of ingredient entities. Multimodal recognition is performed on the label identification data to obtain annotation information and associate it with the set of ingredient entities to obtain a product information map.

[0008] The component entities in the component entity set are linked with the nodes in the pre-constructed domain knowledge graph and expanded along the corresponding knowledge association path to obtain an allergen association subgraph. The allergen association subgraph is then graph-embedded to obtain a component risk representation vector. Based on the component risk representation vector, the allergen database is queried to obtain the target allergen set.

[0009] Based on the product information map, extract the component content parameters and couple them with the component risk characterization vector to obtain the component risk intensity. Obtain the cross-reactivity relationship between different allergens in the target allergen set. Based on the cross-reactivity relationship and the component risk intensity, solve the comprehensive risk score and perform segmented mapping to obtain the allergy risk level.

[0010] Based on the distribution location of the target allergen set in the product information map, a risk propagation topology map is constructed and the allergy risk level is marked. The compliance rules corresponding to the health products are extracted from the preset standard product library and matched with the allergy risk level to obtain the compliance verification result.

[0011] In one alternative implementation,

[0012] The process involves acquiring product formula data and label identification data for health products, performing entity extraction on the product formula data to obtain a set of ingredient entities, and performing multimodal recognition on the label identification data to obtain annotation information. This information is then associated and mapped with the set of ingredient entities to obtain a product information map, including:

[0013] The product formula data is obtained by querying the product database based on the product identifier, and the label identification data is obtained by acquiring the label image data of the product packaging through the image acquisition device. The text description field and the structured formula table are extracted from the product formula data. The text description field is sequence labeled to obtain the text ingredient entity. The ingredient name field and the content field are extracted from the structured formula table to obtain the table ingredient entity. The text ingredient entity and the table ingredient entity are merged to obtain the ingredient entity set.

[0014] Visual and semantic features are extracted from the label data and fused to obtain multimodal annotation features. Allergen warning labels and ingredient labeling labels are identified based on the multimodal annotation features to obtain the annotation information.

[0015] Using the set of ingredient entities as nodes, ingredient reference identifiers are extracted from the annotation information and associated edges are established between different nodes. Supplementary associated edges are established based on the positional distance between the annotation information and the ingredient entities in the structured formula table. The ingredient names and contents in the set of ingredient entities are used as node attributes, and the allergen warning identifiers and ingredient label identifiers in the annotation information are used as edge attributes. The product information graph is constructed by combining the associated edges and the supplementary associated edges.

[0016] In one alternative implementation,

[0017] The component entities in the component entity set are linked with nodes in the pre-constructed domain knowledge graph, and then expanded along the corresponding knowledge association paths to obtain the allergen association subgraph, including:

[0018] Extract the standardized names of each component entity in the component entity set; extract the standardized names of the nodes corresponding to the candidate nodes from the domain knowledge graph and calculate the semantic similarity between them and the standardized names of the components; extract the contextual neighborhood information of each component entity in the product information graph and perform structural matching with the associated node information of the candidate nodes to obtain the structural matching degree; calculate the comprehensive link score based on the semantic similarity and the structural matching degree; and determine the optimal matching node corresponding to each component entity based on the comprehensive link score to obtain the complete entity link mapping.

[0019] Extract the edge relationships associated with the linked nodes in the complete entity link mapping from the domain knowledge graph and identify the paths pointing to the allergen nodes. Extract the attribute labels of the intermediate nodes of the paths. Based on the attribute labels, determine the allergy risk transmission nature of the intermediate nodes and filter the paths with transmission nature to obtain effective knowledge association paths.

[0020] The process extends outward from the linked nodes along the effective knowledge association path and collects the relationships between nodes and edges. The linked nodes are used as starting nodes, and a subgraph structure is constructed by combining the collected node and edge relationships. The importance of nodes is scored based on the path length in the subgraph structure, and nodes and associated edges are selected and retained based on the importance scores to obtain the allergen association subgraph.

[0021] In one alternative implementation,

[0022] Graph embedding encoding is performed on the allergen association subgraph to obtain a component risk characterization vector. Based on the component risk characterization vector, the target allergen set is obtained by querying the allergen database, including:

[0023] Node attributes and edge relationship types are extracted from the allergen association subgraph, and node feature matrix and edge type matrix are constructed. Based on the topological structure of the allergen association subgraph, a neighborhood aggregation matrix is ​​calculated and convolved with the node feature matrix to obtain neighborhood features. Allergy risk transmission edges are extracted from the edge type matrix and a risk transmission weight matrix is ​​constructed. The risk transmission weight matrix and the neighborhood features are weighted and aggregated to obtain risk-oriented features and mapped to obtain node embedding vectors.

[0024] Extract the node embedding vector corresponding to the starting node in the allergen association subgraph as the component representation. Calculate the path weight based on the length of each path and the edge relationship type in the allergen association subgraph, and obtain the path embedding vector by weighted summation of the node embedding vectors corresponding to the nodes on the path. Solve for the component risk representation vector based on the component representation and the path embedding vector.

[0025] Allergen feature vectors are extracted from a pre-defined allergen database and an allergen feature matrix is ​​constructed. The similarity score between the component risk characterization vector and each allergen feature vector in the allergen feature matrix is ​​calculated. Based on the similarity score, a candidate allergen set is extracted and merged with the allergen identifier corresponding to the termination node in the allergen association subgraph to obtain the target allergen set.

[0026] In one alternative implementation,

[0027] Based on the product infographic, ingredient content parameters are extracted and coupled with the ingredient risk characterization vector to obtain ingredient risk intensity. The cross-reactivity relationships between different allergens in the target allergen set are obtained, including:

[0028] The content values ​​corresponding to each component entity are extracted from the product information map and normalized to obtain a standardized content vector. Risk-oriented dimension features are extracted from the component risk characterization vector and the correlation between the risk-oriented dimension features and the standardized content vector is calculated to obtain a content sensitivity coefficient. The standardized content vector is weighted and modulated based on the content sensitivity coefficient and concatenated with the component risk characterization vector to obtain a coupled feature vector. The component risk intensity is obtained by performing a nonlinear transformation on the coupled feature vector.

[0029] The system queries a pre-defined allergen cross-reaction database to retrieve cross-reaction records between any two allergens in the target allergen set. Based on the cross-reaction records, it extracts cross-reaction types and assigns differential weight coefficients based on the differences in sensitization mechanisms of the cross-reaction types. It then extracts reaction intensity identifiers from the cross-reaction records, calculates weighted reaction intensity based on the reaction intensity identifiers and the differential weight coefficients, and constructs a cross-reaction relationship graph. In the cross-reaction relationship graph, it establishes directed edges between vertices with allergens as vertices based on the weighted reaction intensity to obtain the cross-reaction relationship.

[0030] In one alternative implementation,

[0031] Based on the cross-reactivity relationship and the risk intensity of the components, a comprehensive risk score is calculated and segmented to obtain the allergy risk level, including:

[0032] Directed edges are extracted from the cross-reaction graph and a cross-reaction adjacency matrix is ​​constructed. Based on the cross-reaction adjacency matrix, the in-degree and out-degree of each allergen are calculated and strong co-existing allergen nodes are identified. Risk values ​​are extracted from the component risk intensity and a component co-existence matrix is ​​constructed by combining the mapping relationship between the strong co-existing allergen nodes and component entities. The cross-reaction adjacency matrix is ​​iterated exponentially and associated with the component co-existence matrix to obtain a multi-level co-propagation matrix. Based on the multi-level co-propagation matrix, the cumulative contribution of each component in the cross-reaction propagation path is calculated and the co-enhanced risk value is obtained by combining the component risk intensity.

[0033] The basic sensitization intensity of each allergen in the target allergen set is obtained and a sensitization intensity vector is constructed. Bidirectional cross-reaction pairs are extracted from the cross-reaction relationship and the corresponding elements in the sensitization intensity vector are amplified and modulated to obtain a modulated sensitization intensity vector. The vector inner product between the modulated sensitization intensity vector and the co-enhanced risk value is calculated and a hyperbolic tangent transform is performed to obtain a comprehensive risk score.

[0034] Segmentation thresholds are extracted from a preset risk grading standard library and a set of segmented intervals is constructed. The interval to which the individual belongs is determined based on the comprehensive risk score and the interval boundaries in the set of segmented intervals, and the allergy risk level is determined.

[0035] In one alternative implementation,

[0036] A risk propagation topology map is constructed based on the distribution of the target allergen set in the product information map, and the allergy risk level is labeled. Compliance rules corresponding to the health products are extracted from a pre-set standard product library and matched with the allergy risk level to obtain compliance verification results, including:

[0037] The component node positions corresponding to the target allergen set are extracted from the product information graph, and the topological distance between the nodes is calculated to construct a propagation adjacency matrix. Allergen reaction paths are extracted from the cross-reaction relationship graph and mapped to the component nodes of the propagation adjacency matrix to obtain a reaction path mapping matrix. Multi-hop propagation paths are identified based on the reaction path mapping matrix, and the path propagation attenuation coefficient is calculated. The propagation adjacency matrix is ​​iterated exponentially, and the attenuation propagation matrix is ​​solved by combining the path propagation attenuation coefficient. A risk propagation topology graph is constructed based on the attenuation propagation matrix.

[0038] The allergy risk level is labeled to the risk propagation topology map and high-risk level nodes and their corresponding associated nodes are extracted. The risk diffusion range is calculated based on the topological distance between the high-risk level nodes and the associated nodes, and a diffusion region set is constructed. The region boundary nodes are extracted based on the diffusion region set and the region risk density is calculated. The region risk density is then layered and labeled to obtain a layered labeled topology map.

[0039] The set of compliance rules corresponding to the health products is extracted from the preset standard product library. The component content values ​​of the allergen nodes in the hierarchical labeling topology map and the allergen warning labels in the product information map are extracted. Compliance judgments are then performed based on the set of compliance rules to obtain compliance verification results.

[0040] A second aspect of this invention provides a compliance verification system for allergy risk of health products based on multimodal analysis, comprising:

[0041] The data acquisition unit is used to acquire product formula data and label identification data of health products, extract the product formula data to obtain a set of ingredient entities, perform multimodal recognition on the label identification data to obtain annotation information, and associate and map it with the set of ingredient entities to obtain a product information map.

[0042] The graph construction unit is used to link the component entities in the component entity set with the nodes in the pre-constructed domain knowledge graph and expand them along the corresponding knowledge association path to obtain an allergen association subgraph. The allergen association subgraph is then graph-embedded to obtain a component risk representation vector. Based on the component risk representation vector, the allergen database is queried to obtain a target allergen set.

[0043] The risk calculation unit is used to extract the component content parameters based on the product information map and couple them with the component risk characterization vector to obtain the component risk intensity, obtain the cross-reactivity relationship between different allergens in the target allergen set, solve the comprehensive risk score based on the cross-reactivity relationship and the component risk intensity, and perform segmented mapping to obtain the allergy risk level.

[0044] The verification output unit is used to construct a risk propagation topology map based on the distribution location of the target allergen set in the product information map and label the allergy risk level. It also extracts the compliance rules corresponding to the health product from a preset standard product library and performs matching verification in combination with the allergy risk level to obtain the compliance verification result.

[0045] A third aspect of the present invention provides an electronic device, comprising:

[0046] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0047] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0048] In this invention, entity extraction is performed on product formula data, and multimodal recognition is used on label identifiers. Heterogeneous information such as visual and textual data is uniformly associated into a product information graph, effectively solving the problem of ingredient omissions caused by relying on a single data source. By leveraging domain knowledge graphs for entity linking and path expansion, implicit relationships between ingredients can be captured, enabling allergen identification to no longer be limited to isolated ingredients but to cover the complete associated subgraph, thereby accurately locating potential sensitizing sources and reducing the risk of missed detection. By coupling ingredient content parameters with risk characterization vectors, the actual exposure intensity of each allergen can be quantified, and the cross-reactivity relationships between different allergens can be comprehensively analyzed. The risk scoring system overcomes the limitations of independent assessment of single components. The allergy risk levels obtained through segmented mapping are more consistent with the actual clinical patterns of allergic reactions, providing differentiated judgment criteria for different sensitive populations. This improves the precision and reliability of risk classification, avoids misjudgments caused by simple threshold judgments, constructs a risk transmission topology map and marks allergy risk levels, intuitively presenting the distribution and transmission path of allergens in the ingredient system. This makes the verification process interpretable and traceable. Matching compliance rules with risk levels enables rapid identification of whether products meet regulatory requirements, significantly improving verification efficiency and providing clear direction for product improvement. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the compliance verification method for allergy risk of health products based on multimodal analysis, as described in an embodiment of the present invention.

[0050] Figure 2 This is a statistical chart of matching similarity scores for the compliance verification method for allergy risks of health products based on multimodal analysis, as described in an embodiment of the present invention.

[0051] Figure 3This is a risk propagation path attenuation curve of the health product allergy risk compliance verification method based on multimodal analysis, as described in an embodiment of the present invention.

[0052] Figure 4 This is a flowchart illustrating the risk propagation and compliance verification process of the health product allergy risk compliance verification method based on multimodal analysis, as described in an embodiment of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0055] Figure 1 This is a flowchart illustrating the compliance verification method for allergy risk of health products based on multimodal analysis according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0056] The product formula data and label identification data of health products are obtained. Entity extraction is performed on the product formula data to obtain a set of ingredient entities. Multimodal recognition is performed on the label identification data to obtain annotation information and associate it with the set of ingredient entities to obtain a product information map.

[0057] The component entities in the component entity set are linked with the nodes in the pre-constructed domain knowledge graph and expanded along the corresponding knowledge association path to obtain an allergen association subgraph. The allergen association subgraph is then graph-embedded to obtain a component risk representation vector. Based on the component risk representation vector, the allergen database is queried to obtain the target allergen set.

[0058] Based on the product information map, extract the component content parameters and couple them with the component risk characterization vector to obtain the component risk intensity. Obtain the cross-reactivity relationship between different allergens in the target allergen set. Based on the cross-reactivity relationship and the component risk intensity, solve the comprehensive risk score and perform segmented mapping to obtain the allergy risk level.

[0059] Based on the distribution location of the target allergen set in the product information map, a risk propagation topology map is constructed and the allergy risk level is marked. The compliance rules corresponding to the health products are extracted from the preset standard product library and matched with the allergy risk level to obtain the compliance verification result.

[0060] In one alternative implementation,

[0061] The process involves acquiring product formula data and label identification data for health products, performing entity extraction on the product formula data to obtain a set of ingredient entities, and performing multimodal recognition on the label identification data to obtain annotation information. This information is then associated and mapped with the set of ingredient entities to obtain a product information map, including:

[0062] The product formula data is obtained by querying the product database based on the product identifier, and the label identification data is obtained by acquiring the label image data of the product packaging through the image acquisition device. The text description field and the structured formula table are extracted from the product formula data. The text description field is sequence labeled to obtain the text ingredient entity. The ingredient name field and the content field are extracted from the structured formula table to obtain the table ingredient entity. The text ingredient entity and the table ingredient entity are merged to obtain the ingredient entity set.

[0063] Visual and semantic features are extracted from the label data and fused to obtain multimodal annotation features. Allergen warning labels and ingredient labeling labels are identified based on the multimodal annotation features to obtain the annotation information.

[0064] Using the set of ingredient entities as nodes, ingredient reference identifiers are extracted from the annotation information and associated edges are established between different nodes. Supplementary associated edges are established based on the positional distance between the annotation information and the ingredient entities in the structured formula table. The ingredient names and contents in the set of ingredient entities are used as node attributes, and the allergen warning identifiers and ingredient label identifiers in the annotation information are used as edge attributes. The product information graph is constructed by combining the associated edges and the supplementary associated edges.

[0065] Product formula data is retrieved from a product database based on product identifiers (such as barcodes, QR codes, or internal codes). The product database stores structured formula files, with each record containing two types of subfields: a text description field and a structured formula table. The text description field typically records the source of the ingredients, process instructions, and efficacy descriptions in natural language, while the structured formula table stores information such as ingredient names, CAS numbers, content values, and units in a row-column format. Simultaneously, the product packaging is photographed using image acquisition equipment (such as high-resolution industrial cameras or mobile scanning devices) to obtain label image data containing the printed label content, which serves as input for subsequent multimodal recognition.

[0066] For text description fields, a sequence labeling method based on a pre-trained language model is used to label the field content word by word or character by character, identifying the boundaries and categories of constituent entities to obtain text constituent entities. During sequence labeling, the model outputs a corresponding label for each word, using the BIO format, where the B label indicates the starting position of the constituent entity, the I label indicates the interior of the entity, and the O label indicates a non-entity word. For structured recipe tables, the ingredient name field and the content field are directly extracted. The ingredient name field provides the normalized text of the entity, and the content field provides the corresponding numerical value and unit of measurement. Combining these two fields yields the table constituent entities. The text constituent entities and the table constituent entities are deduplicated and merged. For entities with similar names but different origins, identity is determined through string similarity calculation and semantic vector comparison, ultimately resulting in a complete set of constituent entities.

[0067] For label image data, visual and semantic features are extracted separately. Visual feature extraction uses a convolutional neural network to encode multi-scale features in the label image, capturing color distribution, texture patterns, the shape and outline of warning icons, and the spatial layout of printed areas. Semantic feature extraction first performs optical character recognition (OCR) on the label image, converting the printed text in the image into a text sequence. Then, a text encoder is used to semantically vectorize the recognition results, obtaining a high-dimensional feature representation at the semantic level. Visual and semantic features are fused through a cross-modal attention mechanism. This attention mechanism establishes a dynamic weight allocation between the visual and semantic feature sequences, allowing local features of the visual region to mutually enhance the corresponding semantic information of the text, resulting in multimodal annotation features.

[0068] This system classifies and identifies allergen warning labels and ingredient labels based on multimodal annotation features. Allergen warning labels typically appear as specific graphic symbols, bold fonts, or colored borders. By decoding the multimodal annotation features using a trained label classifier, all allergen-related warning areas and their semantic content on the label can be located and extracted. Ingredient labels correspond to the ingredient list area on the label, identifying the text content and arrangement of each ingredient entry. The above identification results together constitute the annotation information, including the category, location coordinates, and corresponding text of the allergen warning labels, as well as the entry content and location information of the ingredient labels.

[0069] The product infographic uses each component entity in the component entity set as a node. Node attributes include two types of information: component name and content. The content attribute stores the numerical value and unit extracted from the structured recipe table. Component reference identifiers are extracted from the labeling information; these are labels that explicitly reference or point to a specific component entity. Association edges are established between the corresponding component entity nodes, with the direction of the association edge pointing from the source of the reference to the referenced component. For component entities in the labeling information that are geographically close but do not have an explicit reference relationship, the positional distance between them is calculated based on the coordinates of each identifier in the labeling information and the row order position of the component entity in the structured recipe table. ,when When the value is below a preset threshold, supplementary edges are established between the corresponding nodes to compensate for the missing graph structure caused by format differences or implicit associations.

[0070] In the product infographic, related edges and supplementary related edges play different semantic roles. Related edges represent explicit ingredient references, while supplementary related edges represent potential relationships inferred from proximity. Both types of edges are stored in a directed format. Edge attributes are jointly assigned by allergen warning labels and ingredient label labels in the annotation information. The allergen warning label serves as the risk marker attribute of the edge, recording the allergen warning categories present in the association path; the ingredient label label serves as the annotation source attribute of the edge, recording the label entry content corresponding to the association path. By integrating node attributes, edge attributes, the set of related edges, and the set of supplementary related edges, a product infographic is constructed. This infographic, in a structured graph data format, fully expresses the multidimensional relationships between product formulation ingredients and label labels, providing a unified data foundation for subsequent allergen risk analysis and compliance verification.

[0071] The node size and edge density of a product infographic depend on the complexity of the product formula. For complex health products with many ingredients, the set of ingredient entities may contain dozens of nodes, each connected to multiple label information entries via associative and supplementary associative edges, forming a relatively dense graph structure. For single-function products with simpler ingredients, the graph structure is relatively sparse, but still retains the complete correspondence between ingredients and label identifiers. By normalizing the content field in the node attributes, content data from different unit systems can be uniformly converted into standard percentages or milligrams per gram, ensuring consistent data caliber in subsequent risk intensity calculations. The graph structure design of the product infographic takes into account both the structured characteristics of formula data and the visual semantic characteristics of label data, enabling deep integration of these two heterogeneous data sources within a unified graph representation framework.

[0072] In one alternative implementation,

[0073] The component entities in the component entity set are linked with nodes in the pre-constructed domain knowledge graph, and then expanded along the corresponding knowledge association paths to obtain the allergen association subgraph, including:

[0074] Extract the standardized names of each component entity in the component entity set; extract the standardized names of the nodes corresponding to the candidate nodes from the domain knowledge graph and calculate the semantic similarity between them and the standardized names of the components; extract the contextual neighborhood information of each component entity in the product information graph and perform structural matching with the associated node information of the candidate nodes to obtain the structural matching degree; calculate the comprehensive link score based on the semantic similarity and the structural matching degree; and determine the optimal matching node corresponding to each component entity based on the comprehensive link score to obtain the complete entity link mapping.

[0075] Extract the edge relationships associated with the linked nodes in the complete entity link mapping from the domain knowledge graph and identify the paths pointing to the allergen nodes. Extract the attribute labels of the intermediate nodes of the paths. Based on the attribute labels, determine the allergy risk transmission nature of the intermediate nodes and filter the paths with transmission nature to obtain effective knowledge association paths.

[0076] The process extends outward from the linked nodes along the effective knowledge association path and collects the relationships between nodes and edges. The linked nodes are used as starting nodes, and a subgraph structure is constructed by combining the collected node and edge relationships. The importance of nodes is scored based on the path length in the subgraph structure, and nodes and associated edges are selected and retained based on the importance scores to obtain the allergen association subgraph.

[0077] For each component entity in the component entity set, its corresponding standardized name is extracted. The extraction of standardized names relies on the standardization of component descriptions in the original formula text, including removing modifiers, standardizing chemical nomenclature, and handling synonyms and abbreviations. For example, "Vitamin C," "Ascorbic Acid," and "L-Ascorbic Acid" are all mapped to the same standardized name after standardization. Simultaneously, standardized names for candidate nodes are extracted from the domain knowledge graph, and the semantic similarity between each standardized component name and the standardized names of each candidate node is calculated. Semantic similarity is calculated by using a pre-trained language model to encode the two names separately, and then measuring the directional consistency between the two encoded vectors using cosine similarity to obtain a numerical value reflecting the semantic closeness.

[0078] Building upon semantic similarity calculation, we further extract the contextual neighborhood information of each component entity in the product information graph. This contextual neighborhood information includes the component entity's direct neighbors in the product information graph, the types of connecting edges, and the attribute labels of neighboring nodes. This information collectively characterizes the local topological features of the component within the product formula structure. We then extract the associated node information of candidate nodes from the domain knowledge graph, including the category distribution of their first-order neighbor nodes, edge relationship types, and attribute label sets. Finally, we perform structural matching between the contextual neighborhood information and the associated node information, and measure the structural matching degree. It reflects the degree of similarity between the two at the graph structure level, and is calculated by comparing the overlap ratio of the neighboring node attribute sets and the degree of matching of edge relationship types.

[0079] Overall link rating Based on semantic similarity Structural matching degree The weighted fusion yields, i.e. ,in This is a weighting coefficient for semantic similarity, ranging from 0 to 1, and can be adjusted according to the specific characteristics of the domain's data. In practical applications, for chemical components with relatively uniform naming conventions, Larger values ​​are chosen to emphasize semantic matching; for components with significant naming differences but clear structural relationships... The smaller value is selected to emphasize structural matching. Based on the comprehensive link score, the candidate node with the highest score is selected as the optimal matching node for each component entity, thereby constructing a complete entity link mapping and establishing a clear correspondence between each component entity in the component entity set and the corresponding node in the domain knowledge graph.

[0080] After completing entity linking, all edge relationships associated with the linked nodes in the complete entity linking map are extracted from the domain knowledge graph, and paths pointing to allergen nodes are identified. Path identification is achieved through a graph traversal algorithm, starting from the linked nodes and searching hop-by-hop along the edge relationships until a node labeled as an allergen type is reached. During this process, attribute labels of intermediate nodes in the path are extracted. These attribute labels contain multi-dimensional information such as the node's functional classification, bioactivity description, and immunogenicity markers. Based on these attribute labels, it is determined whether the intermediate nodes possess allergy risk transmission properties. The criteria for determining allergy risk transmission properties include: whether the intermediate node contains known sensitizing domain labels, whether it is labeled as a cross-reaction-related substance, and whether it plays a mediating role in the allergic reaction mechanism path. Only paths with intermediate nodes possessing transmission properties are retained and identified as valid knowledge association paths, thereby excluding invalid paths generated solely due to weak associations in the graph, ensuring the quality of subsequent subgraph construction.

[0081] The process expands outwards from the already linked nodes along effective knowledge association paths, gradually collecting the relationships between nodes and edges along the path. The expansion process follows a bounded depth strategy, setting a maximum expansion depth. To avoid subgraph bloat due to excessively large graph size, within the extended depth range, all nodes reachable along effective knowledge association paths and their associated edge relationships are collected, forming the node set and edge set of the original subgraph. Linked nodes in the complete entity link mapping are used as the starting nodes of the subgraph, and the collected node and edge relationships are combined to construct the subgraph structure. The subgraph structure preserves the complete association links from component entities to allergen nodes, while also including intermediate nodes with transitive properties in the path, forming a directed graph structure that reflects the allergy risk propagation mechanism.

[0082] Node importance is scored based on path length within the subgraph structure. The path length from this node to the starting node There is an inverse relationship; that is, nodes with shorter paths are more directly related to the transmission of allergy risks and are therefore more important. Specifically, ,in This represents the shortest path length from the starting node to this node. Adding 1 avoids division by zero when the path length to the starting node itself is zero. For nodes reachable by multiple valid paths, the maximum importance score across all paths is taken as the node's final importance score. An importance score threshold is set. Nodes with importance scores above the threshold are retained, along with the associated edges between these nodes. Nodes with importance scores below the threshold and their corresponding isolated edges are deleted, resulting in a simplified allergen association subgraph.

[0083] The constructed allergen association subgraph contains a complete knowledge association structure starting from health product ingredient entities, passing through intermediate nodes with allergy risk transmission properties, and finally reaching the allergen node. This subgraph not only retains highly important directly associated nodes but also ensures the biological significance and rationality of the allergy mechanism of each association link through an effective knowledge association path filtering mechanism. In subsequent graph embedding encoding of this subgraph, its topology and node attributes will be encoded as ingredient risk representation vectors, providing a high-quality feature representation foundation for allergen database queries and risk intensity calculations.

[0084] In one alternative implementation,

[0085] Graph embedding encoding is performed on the allergen association subgraph to obtain a component risk characterization vector. Based on the component risk characterization vector, the target allergen set is obtained by querying the allergen database, including:

[0086] Node attributes and edge relationship types are extracted from the allergen association subgraph, and node feature matrix and edge type matrix are constructed. Based on the topological structure of the allergen association subgraph, a neighborhood aggregation matrix is ​​calculated and convolved with the node feature matrix to obtain neighborhood features. Allergy risk transmission edges are extracted from the edge type matrix and a risk transmission weight matrix is ​​constructed. The risk transmission weight matrix and the neighborhood features are weighted and aggregated to obtain risk-oriented features and mapped to obtain node embedding vectors.

[0087] Extract the node embedding vector corresponding to the starting node in the allergen association subgraph as the component representation. Calculate the path weight based on the length of each path and the edge relationship type in the allergen association subgraph, and obtain the path embedding vector by weighted summation of the node embedding vectors corresponding to the nodes on the path. Solve for the component risk representation vector based on the component representation and the path embedding vector.

[0088] Allergen feature vectors are extracted from a pre-defined allergen database and an allergen feature matrix is ​​constructed. The similarity score between the component risk characterization vector and each allergen feature vector in the allergen feature matrix is ​​calculated. Based on the similarity score, a candidate allergen set is extracted and merged with the allergen identifier corresponding to the termination node in the allergen association subgraph to obtain the target allergen set.

[0089] After the allergen association subgraph is constructed, graph embedding encoding is required to transform the topological relationships and semantic information contained in the graph structure into a computable vector representation. Attribute information for each node is extracted from the allergen association subgraph, including multi-dimensional attribute fields such as component type, chemical classification, bioactivity markers, and allergen classification codes. Simultaneously, the relationship types of each edge in the graph are extracted, such as "containment relationship," "cross-reactivity relationship," "homology relationship," and "risk transmission relationship." The node attributes are then concatenated into a node feature matrix. ,in Each row corresponds to the attribute vector of a node in the graph; the edge type information is organized into an edge type matrix. Record the start and end points of each directed edge and the corresponding relation type code.

[0090] Calculate the neighborhood aggregation matrix based on the topological structure of the allergen association subgraph. This matrix reflects the neighborhood connection structure of each node in the graph. It is typically built upon a normalized adjacency matrix by introducing a self-loop correction, allowing nodes to retain their own characteristics while aggregating neighborhood information. The neighborhood aggregation matrix... With node feature matrix Perform graph convolution operation to obtain the neighborhood feature matrix. ,Right now ,in The learnable weight matrix of the graph convolutional layer. This is a non-linear activation function. This convolution operation enables each node to aggregate the attribute information of all its first-order neighbors, thereby capturing local structural features.

[0091] From the edge type matrix Edges belonging to the "risk transmission" category were selected. These edges specifically describe pathways through which allergy risk is transmitted between components via biochemical mechanisms, such as cross-reaction transmission caused by protein homologous sequences or cascade allergic responses triggered by metabolites. A risk transmission weight matrix was constructed based on the selected risk transmission edges. In the matrix, each non-zero element corresponds to a risk transmission edge, and the value is assigned based on the risk transmission strength marked on the knowledge graph for that edge. Unmarked edges have a weight of 0. The risk transmission weight matrix is ​​then used. With neighborhood feature matrix Weighted aggregation is performed to obtain the risk-oriented feature matrix. ,Right now Risk-oriented feature matrix This highlights the characteristic contributions of nodes along the risk transmission path in the diagram and suppresses the influence of background nodes unrelated to allergy risk. By projecting the vector onto a uniform-dimensional embedding space through a linear mapping layer, we obtain the node embedding vector corresponding to each node. ,in For node indexing.

[0092] After obtaining all node embedding vectors, extract the node embedding vector corresponding to the starting node (i.e., the node corresponding to the original component entity) in the allergen association subgraph, and use it as the component representation vector. The component representation vector encodes the attribute information of the component entity itself and its local neighborhood features in the graph structure. The allergen association subgraph contains multiple paths starting from the initial node, passing through different intermediate nodes, and finally reaching the allergen termination node. The lengths and edge relationships of these paths vary, and their contributions to risk assessment also differ. For each path in the graph... Based on path length Calculate path weights based on the relationship types of the edges along the path. The shorter the path and the more risk propagation edges it passes through, the higher the path weight. Specifically, ,in This is the path length attenuation coefficient. For the edge The risk correlation coefficient corresponding to the relationship type. For the path The node embedding vectors corresponding to all nodes are arranged according to path weights. Perform a weighted summation to obtain the path embedding vector. . Component characterization vector With all path embedding vectors The process involves fusing the vectors by assigning dynamic weights to the embedding vectors of each path using an attention mechanism, summing the weights, and then combining the results with the summation of the vectors. The concatenation and mapping through a fully connected layer yield the final component risk representation vector. This vector comprehensively encodes the attribute features of component entities, graph structure neighborhood information, and aggregate semantics of multiple risk transmission paths, possessing a strong ability to express allergy risk.

[0093] Feature vectors of all known allergens are extracted from a pre-defined allergen database. Each allergen entry in the database includes its standardized name, protein family, description of the biological mechanism triggering the allergy, and statistics of recorded clinical cases. The feature vectors of each allergen are pre-calculated using the same embedding and encoding process as the component risk characterization vector. The feature vectors of all allergens in the database are then arranged row-wise to construct an allergen feature matrix. Each row corresponds to a feature vector of one allergen. Calculate the component risk characterization vector. With allergen feature matrix The cosine similarity between each row vector in the vector is used to obtain the similarity score vector. , of which element The component risk representation vector and the first The degree of similarity between the feature vectors of various allergens Index allergens. Set similarity filtering thresholds. ,extract The allergen entries are used to form a candidate allergen set.

[0094] The candidate allergen set is obtained through vector similarity retrieval, reflecting the degree of association between components and known allergens in the semantic embedding space. However, it may miss some allergens that have clear links in the knowledge graph but have low embedding similarity. To compensate for this deficiency, allergen identifiers corresponding to all terminal nodes are extracted from the allergen association subgraph. Terminal nodes are allergen nodes directly reached through knowledge association paths during the subgraph expansion process, and have clear graph structure support. The candidate allergen set and the allergen identifiers corresponding to the terminal nodes are merged, and duplicate allergen entries are deduplicated to obtain the target allergen set. The target allergen set includes both implicitly associated allergens discovered based on embedding similarity retrieval and explicit allergens directly associated based on knowledge graph paths, providing more comprehensive coverage and effectively reducing the risk of missed allergen detection, thus providing reliable input data for subsequent comprehensive risk score calculation.

[0095] Figure 2This is a statistical chart showing the matching similarity scores of the health product allergy risk compliance verification method based on multimodal analysis in this invention. It illustrates the cosine similarity distribution between the target ingredient in the health product, after graph embedding encoding, and the feature vectors of five known allergens. The analysis involves calculating the item-by-item similarity between the 128-dimensional risk representation vector of the starting node in the allergen association subgraph and the allergen feature matrix stored in a pre-defined allergen database. The similarity score ranges from 0 to 1; a higher score indicates a stronger association between the component and the corresponding allergen's risk features in the semantic embedding space.

[0096] from Figure 2 It can be observed that the component risk characterization vector has the highest similarity score (0.89) with peanut allergen Arah1, indicating a high correlation between this component and peanut allergen in terms of protein sequence homology and IgE binding site characteristics, classifying it as a high-risk match. The similarity score with milk allergen Bosd6 is 0.78, and with soybean allergen Glym5 is 0.74; both scores exceed the preset screening threshold of 0.72, forming the core members of the target allergen set. In contrast, the similarity score with Alta1 (Alternaria allergen) is 0.40, and with Derp1 (house dust mite allergen) is 0.21, both below the screening threshold, indicating a low risk of cross-reactivity between this component and inhaled allergens. Through the above similarity distribution, the types of allergens requiring special attention can be quickly identified, providing clear data support for subsequent comprehensive risk scoring and compliance verification.

[0097] In one alternative implementation,

[0098] Based on the product infographic, ingredient content parameters are extracted and coupled with the ingredient risk characterization vector to obtain ingredient risk intensity. The cross-reactivity relationships between different allergens in the target allergen set are obtained, including:

[0099] The content values ​​corresponding to each component entity are extracted from the product information map and normalized to obtain a standardized content vector. Risk-oriented dimension features are extracted from the component risk characterization vector and the correlation between the risk-oriented dimension features and the standardized content vector is calculated to obtain a content sensitivity coefficient. The standardized content vector is weighted and modulated based on the content sensitivity coefficient and concatenated with the component risk characterization vector to obtain a coupled feature vector. The component risk intensity is obtained by performing a nonlinear transformation on the coupled feature vector.

[0100] The system queries a pre-defined allergen cross-reaction database to retrieve cross-reaction records between any two allergens in the target allergen set. Based on the cross-reaction records, it extracts cross-reaction types and assigns differential weight coefficients based on the differences in sensitization mechanisms of the cross-reaction types. It then extracts reaction intensity identifiers from the cross-reaction records, calculates weighted reaction intensity based on the reaction intensity identifiers and the differential weight coefficients, and constructs a cross-reaction relationship graph. In the cross-reaction relationship graph, it establishes directed edges between vertices with allergens as vertices based on the weighted reaction intensity to obtain the cross-reaction relationship.

[0101] Extracting the content values ​​of each ingredient entity from the product infographic requires traversing the attribute fields attached to each ingredient node in the infographic, including various forms of content expression such as mass fraction, volume ratio, and milligrams per gram. Since the units used to label the content of different ingredients differ, direct numerical comparisons or vector operations will introduce errors due to inconsistencies in dimensions. Therefore, unit normalization is required for all content values. Specifically, unit normalization involves converting all contents to mass fraction (g / 100g). For liquid components labeled as volume ratios, the conversion is performed using the corresponding density parameters; for vitamin components labeled as International Units (IU), the conversion factor for each vitamin is used to map them to mass units. After the conversion, the standardized content values ​​of each ingredient are organized into a standardized content vector according to the order of the ingredient entities in the product infographic. ,in The dimension is equal to the total number of ingredient entities in the product formula.

[0102] In obtaining the standardized content vector Next, it is necessary to identify the dimensional features related to content sensitivity from the component risk representation vector. Component Risk Representation Vector This is a high-dimensional vector obtained during the pre-graph embedding and encoding stage. Each dimension encodes risk semantic information at different levels, and not all dimensions are directly related to changes in concentration. The risk-oriented dimension refers to the risk-transfer weight matrix used in the allergen association subgraph encoding process. The subset of dimensions with higher activation intensity after application. When extracting risk-oriented dimension features, [the following is considered]. Sort the dimensions by activation magnitude and select the top-ranked dimensions to form risk-oriented dimensional feature subvectors. Then calculate With standardized content vector The correlation between the two was measured using the Pearson correlation coefficient to determine the degree of linear association in their numerical trends, resulting in the content sensitivity coefficient. . The range of values ​​is ,when When the value approaches 1, it indicates that the higher the content of the component, the stronger its risk-oriented characteristic response; when When the concentration approaches 0, it indicates a weak correlation between the concentration and risk characteristics, and the concentration has a relatively small modulating effect on risk assessment.

[0103] Based on content sensitivity coefficient For standardized content vectors Weighted modulation is performed to obtain the modulated content vector. The calculation method is as follows The modulated content vector With component risk characterization vector Perform vector concatenation to obtain coupled feature vectors. ,Right now ,in This represents a vector concatenation operation. It involves coupling eigenvectors. Simultaneously carrying risk representation information from the semantic structure of the knowledge graph and content modulation information from the formulation annotation, it can more comprehensively reflect the actual risk level of a certain ingredient in the context of a specific product formulation. This involves coupling the feature vector... To perform the nonlinear transformation, a two-layer fully connected network structure is used. The first layer uses the ReLU activation function to perform a nonlinear mapping on the features, and the second layer uses the Sigmoid activation function to compress the output. The interval is used to output scalar values ​​as the component risk intensity. . The closer it is to 1, the higher the risk of sensitization of the ingredient under the current formulation concentration conditions.

[0104] When obtaining cross-reaction relationships between different allergens in the target allergen set, the system queries a pre-defined allergen cross-reaction database for cross-reaction records between any two allergens within the target allergen set. This database contains multi-source cross-reaction data from clinical research literature, food safety regulatory databases, and allergen protein sequence homology analysis. Each cross-reaction record includes an identifier for the reacting allergen pair, a cross-reaction type field, a reaction intensity identifier field, and a data source reliability rating. All allergens in the target allergen set are queried pairwise. If a cross-reaction record for a particular allergen pair does not exist in the database, it is assumed that there is no known cross-reaction between that pair of allergens, and no corresponding edge is created in the subsequent graph construction.

[0105] Cross-reaction types were extracted from the retrieved cross-reaction records. Common cross-reaction types include protein sequence homology, co-epitope, and pan-allergenic prototypes. Protein sequence homology refers to two allergens having high homology at the amino acid sequence level, with highly similar IgE binding sites, and the sensitization mechanism is cross-recognition of B cell epitopes. Co-epitope refers to two allergens sharing specific antigenic determinants, but with low overall sequence homology. Pan-allergenic prototypes refer to broad-spectrum cross-reactions triggered by conserved protein families widely found in plants or animals (such as the pollen protein PR-10 family). Based on the differences in the sensitization mechanisms of these three cross-reaction types, differential weighting coefficients were assigned to each. Among them, the weight coefficient of the pan-allergic prototype is the highest because it has the widest range of influence and sufficient clinical evidence; the protein sequence homology is the second highest; the weight coefficient of the coepitope is relatively low because the clinical manifestations of cross-reactivity are usually milder.

[0106] Extract reaction intensity markers from cross-reaction records. This label is typically represented by three levels: strong, medium, and weak, or by a continuous numerical value between 0 and 1. For the three-level labeling, strong, medium, and weak are mapped to the values ​​1.0, 0.6, and 0.3, respectively. (Based on reaction intensity labeling) and differential weighting coefficient Solving for weighted reaction intensities The calculation method is as follows Weighted reaction strength It comprehensively reflects the type mechanism weights of cross-reactions and the measured reaction intensity, and can more accurately characterize the cross-sensitization potential between different allergen pairs.

[0107] When constructing the cross-reactivity graph, each allergen in the target allergen set is used as a vertex, and the vertex attributes include the allergen's identifier and component risk intensity. Directed edges are established between allergen vertex pairs with cross-reaction records. The direction of the directed edges is determined by the sensitization direction of the cross-reaction, i.e., from the primary allergen to the cross-activated allergen; if the cross-reaction is bidirectional, a bidirectional directed edge is established. The weights of the directed edges are set to the corresponding weighted response intensities. The larger the weight, the higher the risk of cross-sensitization between the two allergens. The resulting cross-reaction relationship graph is stored in the form of a directed weighted graph. The set of vertices in the graph corresponds to the set of target allergens, and the set of edges corresponds to all known weighted cross-reaction relationships, providing a complete topological basis for solving the subsequent comprehensive risk score.

[0108] In one alternative implementation,

[0109] Based on the cross-reactivity relationship and the risk intensity of the components, a comprehensive risk score is calculated and segmented to obtain the allergy risk level, including:

[0110] Directed edges are extracted from the cross-reaction graph and a cross-reaction adjacency matrix is ​​constructed. Based on the cross-reaction adjacency matrix, the in-degree and out-degree of each allergen are calculated and strong co-existing allergen nodes are identified. Risk values ​​are extracted from the component risk intensity and a component co-existence matrix is ​​constructed by combining the mapping relationship between the strong co-existing allergen nodes and component entities. The cross-reaction adjacency matrix is ​​iterated exponentially and associated with the component co-existence matrix to obtain a multi-level co-propagation matrix. Based on the multi-level co-propagation matrix, the cumulative contribution of each component in the cross-reaction propagation path is calculated and the co-enhanced risk value is obtained by combining the component risk intensity.

[0111] The basic sensitization intensity of each allergen in the target allergen set is obtained and a sensitization intensity vector is constructed. Bidirectional cross-reaction pairs are extracted from the cross-reaction relationship and the corresponding elements in the sensitization intensity vector are amplified and modulated to obtain a modulated sensitization intensity vector. The vector inner product between the modulated sensitization intensity vector and the co-enhanced risk value is calculated and a hyperbolic tangent transform is performed to obtain a comprehensive risk score.

[0112] Segmentation thresholds are extracted from a preset risk grading standard library and a set of segmented intervals is constructed. The interval to which the individual belongs is determined based on the comprehensive risk score and the interval boundaries in the set of segmented intervals, and the allergy risk level is determined.

[0113] Extract all directed edges from the cross-reactivity graph, and represent the cross-reactivity relationship between the source allergen node and the target allergen node corresponding to each directed edge in a structured matrix to construct the cross-reactivity adjacency matrix. The first in the matrix Line 1 Column elements Indicates the first Allergens to the first When there is a cross-reaction triggering relationship between the various allergens, the corresponding reaction weight value is used; otherwise, 0 is used. Based on Calculate the in-degree for each allergen node. With out The in-degree reflects the frequency with which this allergen is triggered by other allergens, while the out-degree reflects the ability of this allergen to trigger other allergens. Allergens whose in-degree and out-degree both exceed a preset synergistic threshold are considered allergens. The allergen nodes are identified as strongly synergistic allergen nodes. These nodes play a dual role as both receptors and triggers in the allergic response network and have a strong potential for cascading amplification.

[0114] Component risk intensity obtained from the preceding steps The risk values ​​corresponding to each component are extracted, and combined with the mapping relationship between strongly synergistic allergen nodes and component entities, the risk values ​​of each component are filled into the corresponding matrix positions according to the mapping relationship to construct the component synergy matrix. The row indices of the component synergy matrix correspond to component entities, and the column indices correspond to allergen nodes. Matrix elements reflect the initial risk contribution of that component in the synergistic response through its corresponding allergen. For the cross-reaction adjacency matrix... Perform exponential iterations and calculate sequentially. , Until the The result of each exponentiation is multiplied by the corresponding decay factor. (in The current iteration order is... The multi-level cooperative propagation matrix is ​​obtained by summing the propagation attenuation coefficients. This matrix comprehensively characterizes the cumulative synergistic effect of allergens in multi-hop propagation paths. Long-distance indirect triggering relationships are given lower weights due to the presence of attenuation factors, thus avoiding excessive influence of distant low-confidence propagation paths on the scores.

[0115] Multi-level collaborative propagation matrix With component synergy matrix By performing correlation calculations, the multi-level cumulative contribution matrix of each component in the entire cross-reaction propagation network is obtained. ,right The cumulative contribution vector for each component is obtained by summing the rows. The th element in the cumulative contribution vector element Indicates the first The combined impact of each component on the overall allergic response network via all accessible pathways. The cumulative contribution vector is then used. The synergistic enhancement risk value is calculated by performing element-wise weighted fusion with the component risk intensity vector. The fusion weight is determined by the degree of strong synergistic allergen association between the components. The higher the degree of association, the greater the weight coefficient of the component during fusion, so that the synergistic enhancement risk value can accurately reflect the joint sensitization amplification effect between multiple components.

[0116] The baseline sensitization intensity of each allergen in the target allergen set is obtained. This baseline sensitization intensity is derived from clinical data and literature records stored in the allergen database, reflecting the probability and severity of an allergic reaction induced by the allergen at a standard exposure dose. The baseline sensitization intensities of each allergen are then arranged in index order to construct a sensitization intensity vector. Extract all bidirectional cross-reactivity pairs from the cross-reactivity relationships, i.e., those that simultaneously satisfy the allergen requirement. Triggering allergens And allergens Triggering allergens The pairwise relationship. For each pair of bidirectional cross-reaction pairs, in the sensitization intensity vector Find the two corresponding elements and multiply them by the magnification factor. ,in This indicates that bidirectional cross-reactions have a stronger synergistic sensitization effect compared to unidirectional cross-reactions. After amplification and modulation operations on all bidirectional cross-reaction pairs, the modulation sensitization intensity vector is obtained. The modulated sensitization intensity vector already includes the contribution of the bidirectional synergistic amplification effect, which can more accurately characterize the actual sensitization ability when multiple allergens coexist.

[0117] Calculate the modulation sensitization intensity vector Enhanced risk value through collaboration The dot product of the corresponding extended vectors yields the raw comprehensive risk score. The vector inner product operation aligns and fuses the directional distribution of sensitization intensity with the magnitude distribution of co-enhancement risk, ensuring that the final score is influenced by both the sensitization intensity distribution characteristics and the co-propagation intensity. Apply the hyperbolic tangent transform, i.e. Mapping the raw comprehensive risk score to Within the range, a comprehensive risk score is obtained. The introduction of hyperbolic tangent transform can effectively suppress the problem of infinite expansion of the original score in extremely high-risk scenarios, while preserving the linear resolution of the score in the medium and low-risk range, so that the comprehensive risk score has good interpretability and stability throughout the entire value range.

[0118] The segmented thresholds are extracted from a pre-defined risk grading standard library, which stores multi-level risk classification standards based on industry regulations, clinical data, and regulatory requirements, and includes several segmented thresholds. The range of values ​​for the comprehensive risk score is divided into: Each segment corresponds to an allergy risk level label, such as very low risk, low risk, moderate risk, high risk, and very high risk. The overall risk score will be calculated. The risk level of a product is determined by comparing it sequentially with the boundaries of each interval. The product is then assigned a risk level label corresponding to that interval to arrive at its final allergy risk level. When the comprehensive risk score falls exactly at an interval boundary, it is assigned to the higher risk level interval according to the principle of choosing the highest risk. This ensures that the compliance verification results remain conservative in boundary situations and avoids underestimating high-risk products due to misjudgment of the critical score.

[0119] Once the allergy risk level is determined, the level label, along with the corresponding comprehensive risk score, synergistic enhancement risk value, and list of strongly synergistic allergen nodes, are recorded to form a traceable risk assessment result for subsequent compliance verification steps. Through the complete process described above—from constructing the cross-reaction adjacency matrix, calculating the multi-level synergistic propagation matrix, modulating sensitization intensity, to mapping the comprehensive risk score—a quantitative characterization of the complex synergistic sensitization mechanisms among multiple components and allergens in health products is achieved, providing sufficient data support and methodological basis for determining the allergy risk level.

[0120] In one alternative implementation,

[0121] A risk propagation topology map is constructed based on the distribution of the target allergen set in the product information map, and the allergy risk level is labeled. Compliance rules corresponding to the health products are extracted from a pre-set standard product library and matched with the allergy risk level to obtain compliance verification results, including:

[0122] The component node positions corresponding to the target allergen set are extracted from the product information graph, and the topological distance between the nodes is calculated to construct a propagation adjacency matrix. Allergen reaction paths are extracted from the cross-reaction relationship graph and mapped to the component nodes of the propagation adjacency matrix to obtain a reaction path mapping matrix. Multi-hop propagation paths are identified based on the reaction path mapping matrix, and the path propagation attenuation coefficient is calculated. The propagation adjacency matrix is ​​iterated exponentially, and the attenuation propagation matrix is ​​solved by combining the path propagation attenuation coefficient. A risk propagation topology graph is constructed based on the attenuation propagation matrix.

[0123] The allergy risk level is labeled to the risk propagation topology map and high-risk level nodes and their corresponding associated nodes are extracted. The risk diffusion range is calculated based on the topological distance between the high-risk level nodes and the associated nodes, and a diffusion region set is constructed. The region boundary nodes are extracted based on the diffusion region set and the region risk density is calculated. The region risk density is then layered and labeled to obtain a layered labeled topology map.

[0124] The set of compliance rules corresponding to the health products is extracted from the preset standard product library. The component content values ​​of the allergen nodes in the hierarchical labeling topology map and the allergen warning labels in the product information map are extracted. Compliance judgments are then performed based on the set of compliance rules to obtain compliance verification results.

[0125] The component node positions corresponding to the target allergen set are extracted from the product infographic. Each component node carries coordinate information and adjacency relationships in the graph structure. By traversing the node set of the product infographic, nodes belonging to the target allergen set are selected, and the in-graph position of each node is recorded. For any two component nodes, the shortest path steps in the graph structure are calculated as the topological distance. ,in , These are the identifiers of the two component nodes. The topological distances between all node pairs are then filled into the propagation adjacency matrix. In the matrix, elements Represents a node With nodes The topological distance between nodes is set; if two nodes are unreachable, the corresponding element is set to the preset maximum distance limit. The adjacency matrix is ​​then propagated. The dimension corresponds to the number of component nodes in the target allergen set, forming a symmetric or directed sparse matrix structure, which provides a basis for the calculation of subsequent propagation paths.

[0126] Extract allergen response paths from the cross-reaction graph by traversing the directed edges in the graph, where each edge represents a response association between one allergen and another. Map the allergen nodes involved in the extracted response paths to a propagation adjacency matrix. The corresponding component node positions form a reaction path mapping matrix. Its elements At the node With nodes If there are cross-reaction paths, the corresponding reaction strength value is used; otherwise, zero is used. This is based on the reaction path mapping matrix. This identifies multi-hop propagation paths, i.e., the sequence of paths that start from a certain allergen node, propagate through intermediate nodes, and then to other nodes. For a number of hops... The propagation path, path propagation attenuation coefficient Calculated according to the attenuation rule, specifically, ,in It is the single-hop decay factor (a constant with a value between 0 and 1). This represents the hop count of the current path. The attenuation coefficient of multi-hop propagation paths decreases exponentially with the number of hops, ensuring a reasonable reduction in the risk contribution of long-distance propagation.

[0127] propagation of adjacency matrix Perform power-law iterations, and let the maximum iteration order be . In the Calculate in the next iteration And the matrix of this order is compared with the corresponding path propagation attenuation coefficient. After multiplying and summing, we obtain the attenuation propagation matrix. The calculation process is as follows Attenuation propagation matrix The elements in the matrix comprehensively reflect the cumulative risk contribution of each node pair propagating through paths with different hop counts; the larger the value, the stronger the risk propagation correlation between the two nodes. This is represented by the attenuation propagation matrix. Based on this, the location information, node attributes, and propagation weights of each component node are integrated to construct a risk propagation topology graph. Nodes in the risk propagation topology graph correspond to component nodes in the target allergen set, and the edge weights are determined by... The corresponding elements determine the weighted directed graph structure, which fully depicts the risk propagation network between the various ingredients in health products.

[0128] After labeling allergy risk levels onto each node of the risk propagation topology, a set of high-risk level nodes is selected based on a preset high-risk level threshold. For each high-risk level node, directly adjacent and reachable nodes within a set hop count are extracted from the risk propagation topology, forming a set of associated nodes. Based on the topological distance between high-risk level nodes and associated nodes, the risk diffusion range is calculated. Specifically, for high-risk nodes... Its risk diffusion radius The value of the allergy risk level at this node is compared with the preset diffusion ratio coefficient. The product is determined, that is ,in This represents the allergy risk level value corresponding to this node. The topological distance should not exceed... All nodes are included in the diffusion region of the high-risk node, and the diffusion regions of all high-risk nodes are merged to obtain the diffusion region set. The diffusion region set covers the range of all component nodes in the product that are directly or indirectly affected by high-risk allergens, providing spatial boundaries for subsequent regional risk density calculations.

[0129] When extracting the region boundary nodes of the diffusion region set, the topological distance of each diffusion region is exactly equal to the diffusion radius. The nodes are marked as boundary nodes, and the set of boundary nodes reflects the outer edge of risk diffusion. The regional risk density is calculated for each diffusion region. The calculation method is to sum the allergy risk level values ​​of all nodes in the region and divide by the total number of nodes in the region, i.e. ,in This represents the total number of nodes in the region. For the set of nodes in the diffusion region, For nodes Allergy risk level numerical value. Regional risk density. This reflects the concentration of allergy risk within the diffusion area; a higher value indicates a denser distribution of high-risk components within that area. Based on regional risk density... Based on the size of the risk level, each diffusion region is divided into different risk levels according to a preset stratification threshold sequence. The stratification results are then labeled to the corresponding areas of the risk propagation topology map, resulting in a stratified labeled topology map. The stratified labeled topology map visually presents the risk density distribution of each region through color depth or labels, providing a visualized spatial distribution reference for compliance verification.

[0130] When extracting the compliance rule set corresponding to health products from the pre-set standard product library, the product category code, applicable population label, and sales region identifier are used as search keys to match the corresponding compliance rule entries in the standard product library. The compliance rule set covers various rules, including the upper limit of allergen ingredient content, mandatory warning label requirements, and a list of prohibited ingredients. The component content values ​​of allergen nodes in the hierarchical labeling topology map are extracted, and the actual content of each allergen node is compared with the upper limit specified in the compliance rule set. If the component content of an allergen node exceeds the upper limit of the corresponding rule, the node is marked as a content exceeding the limit violation. Allergen warning labels in the product information map are extracted and matched with the list of mandatory allergen warning labels in the compliance rule set. If there are cases in the target allergen set where the rule requires labeling but the corresponding warning label does not appear on the product label, the missing item is recorded as a labeling violation. Content exceeding the limit violations and labeling violations are summarized, and the severity of the violations is weighted and sorted according to the allergy risk level to generate a compliance verification result. The compliance verification results include a detailed description of each violation, the corresponding rule entry number, the severity rating of the violation, and rectification suggestions, providing a complete structured output for the compliance review of health products and supporting subsequent regulatory decisions and product formulation optimization.

[0131] Figure 3 This is a risk propagation path attenuation curve of the health product allergy risk compliance verification method based on multimodal analysis, as described in this embodiment of the invention. It shows the trend of allergy risk gradually decreasing as the number of hops increases in the propagation topology. The attenuation curve is calculated based on the power iteration of the propagation adjacency matrix. The horizontal axis represents the number of propagation hops, and the vertical axis represents the attenuation coefficient. A lower attenuation coefficient indicates greater risk loss during transmission and weaker cumulative impact at distant nodes.

[0132] from Figure 3It can be observed that the attenuation coefficient of the first-hop propagation is the highest, approximately 0.82, indicating that the loss of allergy risk is relatively small when it propagates within the first-level neighborhood, and the neighboring component nodes are most significantly affected. As the number of hops increases to two hops, the attenuation coefficient decreases to 0.49, a decrease of about 40%, indicating that the risk has been significantly reduced after two hops. The attenuation coefficients of the three-hop and four-hop propagation paths are 0.30 and 0.18, respectively, indicating that the effectiveness of risk transmission is further reduced, and distant nodes are almost no longer substantially affected. The attenuation trend is consistent with the design of the attenuation factor being compressed according to the exponential rule in multi-hop propagation paths, that is, the higher the number of hops and the farther the propagation distance, the lower the cumulative contribution of the target allergen set. The attenuation curve provides a quantitative basis for defining the diffusion range in the risk propagation topology map, and also lays a data foundation for the identification of high-risk nodes and the calculation of regional risk density.

[0133] Figure 4 This is a flowchart illustrating the risk propagation and compliance verification process of the health product allergy risk compliance verification method based on multimodal analysis, as described in an embodiment of the present invention.

[0134] A second aspect of this invention provides a compliance verification system for allergy risk of health products based on multimodal analysis, comprising:

[0135] The data acquisition unit is used to acquire product formula data and label identification data of health products, extract the product formula data to obtain a set of ingredient entities, perform multimodal recognition on the label identification data to obtain annotation information, and associate and map it with the set of ingredient entities to obtain a product information map.

[0136] The graph construction unit is used to link the component entities in the component entity set with the nodes in the pre-constructed domain knowledge graph and expand them along the corresponding knowledge association path to obtain an allergen association subgraph. The allergen association subgraph is then graph-embedded to obtain a component risk representation vector. Based on the component risk representation vector, the allergen database is queried to obtain a target allergen set.

[0137] The risk calculation unit is used to extract the component content parameters based on the product information map and couple them with the component risk characterization vector to obtain the component risk intensity, obtain the cross-reactivity relationship between different allergens in the target allergen set, solve the comprehensive risk score based on the cross-reactivity relationship and the component risk intensity, and perform segmented mapping to obtain the allergy risk level.

[0138] The verification output unit is used to construct a risk propagation topology map based on the distribution location of the target allergen set in the product information map and label the allergy risk level. It also extracts the compliance rules corresponding to the health product from a preset standard product library and performs matching verification in combination with the allergy risk level to obtain the compliance verification result.

[0139] A third aspect of the present invention provides an electronic device, comprising:

[0140] A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0141] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0142] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0143] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A compliance verification method for allergy risk of health products based on multimodal analysis, characterized in that, include: The product formula data and label identification data of health products are obtained. Entity extraction is performed on the product formula data to obtain a set of ingredient entities. Multimodal recognition is performed on the label identification data to obtain annotation information and associate it with the set of ingredient entities to obtain a product information map. The component entities in the component entity set are linked with the nodes in the pre-constructed domain knowledge graph and expanded along the corresponding knowledge association path to obtain an allergen association subgraph. The allergen association subgraph is then graph-embedded to obtain a component risk representation vector. Based on the component risk representation vector, the allergen database is queried to obtain the target allergen set. Based on the product information map, extract the component content parameters and couple them with the component risk characterization vector to obtain the component risk intensity. Obtain the cross-reactivity relationship between different allergens in the target allergen set. Based on the cross-reactivity relationship and the component risk intensity, solve the comprehensive risk score and perform segmented mapping to obtain the allergy risk level. Based on the distribution location of the target allergen set in the product information map, a risk propagation topology map is constructed and the allergy risk level is marked. The compliance rules corresponding to the health products are extracted from the preset standard product library and matched with the allergy risk level to obtain the compliance verification result.

2. The method according to claim 1, characterized in that, The process involves acquiring product formula data and label identification data for health products, performing entity extraction on the product formula data to obtain a set of ingredient entities, and performing multimodal recognition on the label identification data to obtain annotation information. This information is then associated and mapped with the set of ingredient entities to obtain a product information map, including: The product formula data is obtained by querying the product database based on the product identifier, the label identification data is obtained by acquiring the label image data of the product packaging through the image acquisition device, the text description field and the structured formula table are extracted from the product formula data, the text description field is sequence labeled to obtain the text ingredient entity, the ingredient name field and the content field are extracted from the structured formula table to obtain the table ingredient entity, and the text ingredient entity and the table ingredient entity are merged to obtain the ingredient entity set. Visual and semantic features are extracted from the label data and fused to obtain multimodal annotation features. Allergen warning labels and ingredient labeling labels are identified based on the multimodal annotation features to obtain the annotation information. Using the set of ingredient entities as nodes, ingredient reference identifiers are extracted from the annotation information and associated edges are established between different nodes. Supplementary associated edges are established based on the positional distance between the annotation information and the ingredient entities in the structured formula table. The ingredient names and contents in the set of ingredient entities are used as node attributes, and the allergen warning identifiers and ingredient label identifiers in the annotation information are used as edge attributes. The product information graph is constructed by combining the associated edges and the supplementary associated edges.

3. The method according to claim 1, characterized in that, The component entities in the component entity set are linked with nodes in the pre-constructed domain knowledge graph, and then expanded along the corresponding knowledge association paths to obtain the allergen association subgraph, including: Extract the standardized names of each component entity in the component entity set; extract the standardized names of the nodes corresponding to the candidate nodes from the domain knowledge graph and calculate the semantic similarity between them and the standardized names of the components; extract the contextual neighborhood information of each component entity in the product information graph and perform structural matching with the associated node information of the candidate nodes to obtain the structural matching degree; calculate the comprehensive link score based on the semantic similarity and the structural matching degree; and determine the optimal matching node corresponding to each component entity based on the comprehensive link score to obtain the complete entity link mapping. Extract the edge relationships associated with the linked nodes in the complete entity link mapping from the domain knowledge graph and identify the paths pointing to the allergen nodes. Extract the attribute labels of the intermediate nodes of the paths. Based on the attribute labels, determine the allergy risk transmission nature of the intermediate nodes and filter the paths with transmission nature to obtain effective knowledge association paths. The process extends outward from the linked nodes along the effective knowledge association path and collects the relationships between nodes and edges. The linked nodes are used as starting nodes, and a subgraph structure is constructed by combining the collected node and edge relationships. The importance of nodes is scored based on the path length in the subgraph structure, and nodes and associated edges are selected and retained based on the importance scores to obtain the allergen association subgraph.

4. The method according to claim 1, characterized in that, Graph embedding encoding is performed on the allergen association subgraph to obtain a component risk characterization vector. Based on the component risk characterization vector, the target allergen set is obtained by querying the allergen database, including: Node attributes and edge relationship types are extracted from the allergen association subgraph, and node feature matrix and edge type matrix are constructed. Based on the topological structure of the allergen association subgraph, a neighborhood aggregation matrix is ​​calculated and convolved with the node feature matrix to obtain neighborhood features. Allergy risk transmission edges are extracted from the edge type matrix and a risk transmission weight matrix is ​​constructed. The risk transmission weight matrix and the neighborhood features are weighted and aggregated to obtain risk-oriented features and mapped to obtain node embedding vectors. Extract the node embedding vector corresponding to the starting node in the allergen association subgraph as the component representation. Calculate the path weight based on the length of each path and the edge relationship type in the allergen association subgraph, and obtain the path embedding vector by weighted summation of the node embedding vectors corresponding to the nodes on the path. Solve for the component risk representation vector based on the component representation and the path embedding vector. Allergen feature vectors are extracted from a pre-defined allergen database and an allergen feature matrix is ​​constructed. The similarity score between the component risk characterization vector and each allergen feature vector in the allergen feature matrix is ​​calculated. Based on the similarity score, a candidate allergen set is extracted and merged with the allergen identifier corresponding to the termination node in the allergen association subgraph to obtain the target allergen set.

5. The method according to claim 1, characterized in that, Based on the product infographic, ingredient content parameters are extracted and coupled with the ingredient risk characterization vector to obtain ingredient risk intensity. The cross-reactivity relationships between different allergens in the target allergen set are obtained, including: The content values ​​corresponding to each component entity are extracted from the product information map and normalized to obtain a standardized content vector. Risk-oriented dimension features are extracted from the component risk characterization vector and the correlation between the risk-oriented dimension features and the standardized content vector is calculated to obtain a content sensitivity coefficient. The standardized content vector is weighted and modulated based on the content sensitivity coefficient and concatenated with the component risk characterization vector to obtain a coupled feature vector. The component risk intensity is obtained by performing a nonlinear transformation on the coupled feature vector. The system queries a pre-defined allergen cross-reaction database to retrieve cross-reaction records between any two allergens in the target allergen set. Based on the cross-reaction records, it extracts cross-reaction types and assigns differential weight coefficients based on the differences in sensitization mechanisms of the cross-reaction types. It then extracts reaction intensity identifiers from the cross-reaction records, calculates weighted reaction intensity based on the reaction intensity identifiers and the differential weight coefficients, and constructs a cross-reaction relationship graph. In the cross-reaction relationship graph, it establishes directed edges between vertices with allergens as vertices based on the weighted reaction intensity to obtain the cross-reaction relationship.

6. The method according to claim 1, characterized in that, Based on the cross-reactivity relationship and the risk intensity of the components, a comprehensive risk score is calculated and segmented to obtain the allergy risk level, including: Directed edges are extracted from the cross-reaction graph and a cross-reaction adjacency matrix is ​​constructed. Based on the cross-reaction adjacency matrix, the in-degree and out-degree of each allergen are calculated and strong co-existing allergen nodes are identified. Risk values ​​are extracted from the component risk intensity and a component co-existence matrix is ​​constructed by combining the mapping relationship between the strong co-existing allergen nodes and component entities. The cross-reaction adjacency matrix is ​​iterated exponentially and associated with the component co-existence matrix to obtain a multi-level co-propagation matrix. Based on the multi-level co-propagation matrix, the cumulative contribution of each component in the cross-reaction propagation path is calculated and the co-enhanced risk value is obtained by combining the component risk intensity. The basic sensitization intensity of each allergen in the target allergen set is obtained and a sensitization intensity vector is constructed. Bidirectional cross-reaction pairs are extracted from the cross-reaction relationship and the corresponding elements in the sensitization intensity vector are amplified and modulated to obtain a modulated sensitization intensity vector. The vector inner product between the modulated sensitization intensity vector and the co-enhanced risk value is calculated and a hyperbolic tangent transform is performed to obtain a comprehensive risk score. Segmentation thresholds are extracted from a preset risk grading standard library and a set of segmented intervals is constructed. The interval to which the individual belongs is determined based on the comprehensive risk score and the interval boundaries in the set of segmented intervals, and the allergy risk level is determined.

7. The method according to claim 1, characterized in that, A risk propagation topology map is constructed based on the distribution of the target allergen set in the product information map, and the allergy risk level is labeled. Compliance rules corresponding to the health products are extracted from a pre-set standard product library and matched with the allergy risk level to obtain compliance verification results, including: The component node positions corresponding to the target allergen set are extracted from the product information graph, and the topological distance between the nodes is calculated to construct a propagation adjacency matrix. Allergen reaction paths are extracted from the cross-reaction relationship graph and mapped to the component nodes of the propagation adjacency matrix to obtain a reaction path mapping matrix. Multi-hop propagation paths are identified based on the reaction path mapping matrix, and the path propagation attenuation coefficient is calculated. The propagation adjacency matrix is ​​iterated exponentially, and the attenuation propagation matrix is ​​solved by combining the path propagation attenuation coefficient. A risk propagation topology graph is constructed based on the attenuation propagation matrix. The allergy risk level is labeled to the risk propagation topology map and high-risk level nodes and their corresponding associated nodes are extracted. The risk diffusion range is calculated based on the topological distance between the high-risk level nodes and the associated nodes, and a diffusion region set is constructed. The region boundary nodes are extracted based on the diffusion region set and the region risk density is calculated. The region risk density is then layered and labeled to obtain a layered labeled topology map. The set of compliance rules corresponding to the health products is extracted from the preset standard product library. The component content values ​​of the allergen nodes in the hierarchical labeling topology map and the allergen warning labels in the product information map are extracted. Compliance judgments are then performed based on the set of compliance rules to obtain compliance verification results.

8. A health product allergy risk compliance verification system based on multimodal analysis, used to implement the method of any one of claims 1-7, characterized in that, include: The data acquisition unit is used to acquire product formula data and label identification data of health products, extract the product formula data to obtain a set of ingredient entities, perform multimodal recognition on the label identification data to obtain annotation information, and associate and map it with the set of ingredient entities to obtain a product information map. The graph construction unit is used to link the component entities in the component entity set with the nodes in the pre-constructed domain knowledge graph and expand them along the corresponding knowledge association path to obtain an allergen association subgraph. The allergen association subgraph is then graph-embedded to obtain a component risk representation vector. Based on the component risk representation vector, the allergen database is queried to obtain a target allergen set. The risk calculation unit is used to extract the component content parameters based on the product information map and couple them with the component risk characterization vector to obtain the component risk intensity, obtain the cross-reactivity relationship between different allergens in the target allergen set, solve the comprehensive risk score based on the cross-reactivity relationship and the component risk intensity, and perform segmented mapping to obtain the allergy risk level. The verification output unit is used to construct a risk propagation topology map based on the distribution location of the target allergen set in the product information map and label the allergy risk level. It also extracts the compliance rules corresponding to the health product from a preset standard product library and performs matching verification in combination with the allergy risk level to obtain the compliance verification result.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images