A medicine intelligent monitoring method and system

Abstract

This invention discloses a method and system for intelligent drug monitoring, belonging to the field of drug regulatory technology. It includes acquiring multi-source monitoring data, preprocessing the multi-source monitoring data and extracting multi-dimensional features to obtain multi-dimensional drug feature data; constructing a spatiotemporal graph neural network model to deeply fuse the multi-dimensional drug feature data to obtain a multi-dimensional fused feature representation; constructing a variant semantic knowledge base, matching variant keywords with prohibited drug prototype keywords through a variant semantic reasoning model, and combining the multi-dimensional fused feature representation to perform semantic mapping between the essential attributes of the drug and the prohibited category; integrating key intermediate results from the entire monitoring process, generating a structured evidence chain, and providing visual output. This invention, based on spatiotemporal graph neural networks and variant semantic reasoning, enables real-time monitoring of drugs in complex market environments.

Description

Technical Field

[0001] This invention relates to the field of drug regulatory technology, and in particular to a method and system for intelligent drug monitoring. Background Technology

[0002] Currently, the monitoring of prohibited drugs both domestically and internationally mainly employs the following technical methods, combined with commonly used intelligent monitoring algorithms, to form a preliminary monitoring system. However, the overall system is still in a relatively basic stage: The text rule and keyword comparison monitoring mode scans and marks text information such as drug titles, product detail page text, sales promotion content, and customer service chat logs by pre-setting keywords such as prohibited drug names and illegal expressions, using regular expressions. It focuses on identifying the original names of prohibited drugs (such as "compound mifepristone tablets" and "codeine tablets") and illegal terms such as "counterfeit drug" and "substandard drug." This type of method often uses basic algorithms such as Naive Bayes classifiers and decision trees to complete simple classification, achieving only preliminary screening of explicit violations.

[0003] The manual rule configuration and sampling verification mode involves regulatory personnel and platform reviewers combining their work experience to configure monitoring rules and conduct sampling inspections targeting specific drug categories, sales entities, and regions. Alternatively, manual review may be conducted after receiving public reports or discovering clues of violations. For suspected violations in complex scenarios, the core reliance is on manual judgment of the prohibited drug attributes, the facts of the violation, and the legal basis. In some scenarios, the k-means algorithm is used to assist in data clustering, narrowing the scope of manual review and improving verification efficiency.

[0004] In the single-modal automatic identification mode, a few monitoring systems attempt to introduce natural language processing technology or structured field verification technology to analyze single-modal data such as drug text descriptions and category tags, or combine support vector machines and nearest neighbor algorithms to complete simple classification and identification. However, such systems generally rely on a single data source, fail to achieve deep integration of multi-dimensional data, and do not optimize the technology for core characteristics such as prohibited drug variant keywords and dynamic timeliness, resulting in obvious monitoring limitations.

[0005] While the aforementioned technical solutions are relatively inexpensive to implement and easy to deploy, and can play a preliminary screening role in overt violations (such as selling drugs under the original names of prohibited drugs), their technical shortcomings have become increasingly prominent in the face of the increasingly covert sales of prohibited drugs and the increasingly complex monitoring scenarios, making it difficult to meet actual regulatory needs. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for intelligent drug monitoring, aiming to solve the problems of existing technologies.

[0007] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for intelligent drug monitoring, including acquiring multi-source monitoring data; preprocessing the multi-source monitoring data and extracting multi-dimensional features to obtain multi-dimensional drug feature data; constructing a spatiotemporal graph neural network model to deeply fuse the multi-dimensional drug feature data to obtain a multi-dimensional fused feature representation; constructing a variant semantic knowledge base, matching variant keywords with prohibited drug prototype keywords through a variant semantic reasoning model, and combining the multi-dimensional fused feature representation to perform semantic mapping between the essential attributes of the drug and the prohibited category; integrating key intermediate results of the entire monitoring process, generating a structured evidence chain, and visualizing the output.

[0008] Secondly, this application provides a drug intelligent monitoring system, including a data acquisition module for acquiring multi-source monitoring data, which includes drug-related data, sales-related data, spatiotemporal related data, legal and rule data, and violation clue data; a feature extraction module for preprocessing the multi-source monitoring data and extracting multi-dimensional features to obtain multi-dimensional drug feature data, which includes text features, image features, spatiotemporal features, and structured features; a spatiotemporal fusion module for deeply fusing the multi-dimensional drug feature data to obtain a multi-dimensional fused feature representation; a semantic reasoning and recognition module for constructing a variant semantic knowledge base, matching variant keywords with prohibited drug prototype keywords through a variant semantic reasoning model, and performing semantic mapping between the essential attributes of drugs and prohibited categories in conjunction with the multi-dimensional fused feature representation; and an evidence chain generation module for integrating key intermediate results throughout the monitoring process, generating a structured evidence chain including attribute basis, legal basis, and judgment explanation, and providing visual output.

[0009] Through the above technical solutions, the beneficial effects of the present invention are as follows: Based on spatiotemporal graph neural networks and variant semantic reasoning, it can realize systematic, continuous and dynamic monitoring of prohibited drugs in complex monitoring scenarios with multiple scenarios, multiple time periods and multiple variants, automatically identify explicit and implicit violations, and generate an interpretable, traceable and law enforcement-adaptable structured evidence chain, providing reliable technical support for regulatory authorities' law enforcement, platform compliance governance and public rights protection. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention, and the embodiments in the accompanying drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1A flowchart illustrating a smart drug monitoring method provided in this application embodiment; Figure 2 A flowchart illustrating yet another intelligent drug monitoring method provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of a spatiotemporal graph neural network model provided in an embodiment of this application; Figure 4 A schematic diagram illustrating a process for identifying prohibited drugs, provided as an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a drug intelligent monitoring system provided in an embodiment of this application; Figure 6 A schematic diagram of the structure of a computer device provided in an embodiment of this application; The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0012] It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the application. Rather, these embodiments are provided to make the disclosure more thorough and complete, and to fully convey the scope of the disclosure to those skilled in the art.

[0013] Drug safety is crucial to public health and safety. To ensure drug quality, protect public medication safety and legal rights, and safeguard public health, the state has clearly defined the scope of prohibited drugs and formulated relevant regulatory laws and regulations, including the "Drug Administration Law of the People's Republic of China," the "Administrative Measures for Internet Drug Information Services," and the "Administrative Measures for the Supervision and Management of Online Drug Sales (Draft for Comments)." According to these regulations, drugs subject to special state management, such as narcotic drugs, drugs for terminating pregnancy, toxic drugs, vaccines, blood products, and radioactive drugs, as well as counterfeit drugs, substandard drugs, ephedrine-containing compound preparations, and drugs prohibited from sale during specific periods (such as cold, fever-reducing, cough-suppressing, and antiviral drugs during the epidemic), are strictly prohibited from being sold illegally online or offline.

[0014] With the rapid development of internet technology, drug sales channels have become increasingly diversified. Online stores, third-party platforms, and social media have become important carriers for drug sales. At the same time, offline pharmacies and health product stores also pose risks of illegal drug sales. The sales methods of prohibited drugs are becoming more covert and diverse. Currently, my country has a large number of drug sales entities and complex product information. Prohibited drugs also have various time-sensitive types, such as long-term prohibitions, prohibitions in specific regions during specific periods, and prohibitions on specific batches from specific manufacturers. Relying solely on manual inspections, post-event complaint handling, and traditional keyword comparison methods is no longer sufficient to meet the needs of large-scale, routine, and precise monitoring. Building an intelligent and automated monitoring system for prohibited drugs has become a rigid requirement for regulatory authorities and drug sales platforms.

[0015] Technical difficulties faced by existing technologies in complex scenarios: The wording related to prohibited drugs is often disguised and difficult to accurately identify. In actual illegal sales, some sellers deliberately avoid using the original keywords of prohibited drugs to evade supervision. Instead, they use methods such as sound changes, shape changes, splitting, combining, pinyin, homophones, and industry slang to describe prohibited drug information. For example, they might split "ibuprofen" into "bu". The use of terms like "fen" and "ibuprofen," abbreviating "mifepristone tablets" to "mifepristone tablets," using "painkillers" to imply abortion drugs, and using "special cough powder" to imply a codeine-containing compound preparation are all examples of variations that are difficult to identify effectively using traditional keyword comparison and simple algorithm-based technologies. This leads to a large number of hidden violations going undetected, which is one of the core challenges facing existing monitoring technologies.

[0016] The validity periods of banned drugs are dynamically changing, and existing technologies are ill-suited to them. Banned drugs have various validity periods, including long-term bans, bans in specific regions during specific periods, and bans on specific batches from specific manufacturers. Furthermore, the validity boundaries change dynamically with policy adjustments and risk assessment results (e.g., adjustments to the scope of banned drugs during pandemics, or emergency bans on certain batches of drugs from a specific manufacturer due to quality issues). Existing technologies often use fixed rule configurations and do not consider the dynamic correlation of time and space dimensions. This makes it impossible to automatically update and adapt the validity information of banned drugs, easily leading to problems such as "overdue monitoring" (drugs that have already been lifted from bans are still marked) or "missed monitoring" (newly added banned drugs and temporarily banned drugs are not included in the monitoring scope), failing to meet the needs of dynamic supervision.

[0017] The monitoring of prohibited drugs faces challenges in integrating multi-scenario and multi-dimensional data, resulting in insufficient comprehensiveness. Monitoring of prohibited drugs involves multiple scenarios (online third-party platforms, online stores, social media, offline pharmacies, health product stores, etc.) and multi-dimensional data (drug information, seller qualifications, sales records, spatiotemporal information, user feedback, legal clauses, etc.). Existing technologies primarily monitor single-scenario and single-modal data, failing to achieve collaborative integration of multi-scenario data. For example, online monitoring does not link to offline sales records, text information monitoring does not incorporate drug display information in images and videos, and it fails to effectively utilize spatiotemporal characteristics (such as the patterns of illegal sales in specific regions and time periods). This leads to blind spots in monitoring, making it impossible to comprehensively capture cross-scenario and multi-dimensional violations, such as illegal offline stockpiling and online-to-offline sales of prohibited drugs.

[0018] The attributes of drugs are difficult to match precisely with prohibited categories. The core of determining prohibited drugs depends on their essential attributes (such as ingredients, efficacy, uses, and production batches), rather than the superficial descriptions provided by the seller. For example, some sellers may sell non-pharmaceutical products (such as foot bath powder) as drugs, or disguise compound preparations containing ephedrine as ordinary health products. Current technology largely relies on category labels and superficial descriptions provided by sellers, failing to combine multi-dimensional data for precise judgment of the essential attributes of drugs. This makes it difficult to effectively distinguish between legal and prohibited drugs, and between non-pharmaceutical and prohibited drugs, easily leading to misjudgments and omissions.

[0019] The purpose of this invention is to provide an intelligent monitoring method and system for nationally prohibited drugs based on spatiotemporal graph neural networks and variant semantic reasoning. This method addresses the problems in existing technologies, such as difficulty in identifying variant expressions of prohibited drugs, inability to adapt to dynamic timeliness, insufficient multi-dimensional data fusion, difficulty in directly incorporating legal provisions into reasoning, and inefficient evidence chain generation. It enables automated, intelligent, and precise monitoring of the sale of nationally prohibited drugs, improves the efficiency and accuracy of supervision by regulatory authorities and drug sales platforms, standardizes the order of the drug sales market, and protects the public's drug safety and legitimate rights.

[0020] Through the technical solution of this invention, systematic, continuous and dynamic monitoring of prohibited drugs can be achieved in complex monitoring scenarios with multiple scenarios, multiple time periods and multiple variations. It can automatically identify explicit and implicit violations, generate an interpretable, traceable and law enforcement-adaptable structured evidence chain, and provide reliable technical support for regulatory authorities' law enforcement, platform compliance governance and protection of public rights.

[0021] The foregoing and other technical contents, features and effects of the present invention are described in conjunction with the appendix below. Figures 1-6 The detailed description of the embodiments will make this clear. All structural details mentioned in the following embodiments are based on the accompanying drawings. Exemplary embodiments of the present invention will now be described with reference to the accompanying drawings.

[0022] In one exemplary embodiment, such as Figure 1 As shown, a smart drug monitoring method is provided, the method including: S101, acquire multi-source monitoring data, preprocess the multi-source monitoring data and extract multi-dimensional features to obtain multi-dimensional drug feature data; S102, Construct a spatiotemporal graph neural network model to deeply fuse the multidimensional drug feature data and obtain a multidimensional fused feature representation; S103, Construct a variant semantic knowledge base, match variant keywords with prohibited drug prototype keywords through a variant semantic reasoning model, and combine the multi-dimensional fusion feature representation to perform semantic mapping between the essential attributes of drugs and prohibited categories; S104 integrates key intermediate results from the entire monitoring process, generates a structured chain of evidence, and provides a visual output.

[0023] like Figure 2 As shown, for step S101, multi-source, multi-scenario monitoring data of the area to be monitored and the subject to be monitored are collected through various means such as web crawlers, internal data interfaces, offline data collection equipment, and third-party data cooperation. The multi-source monitoring data includes at least: drug-related data, sales-related data, spatiotemporal related data, legal and rule data, and violation clue data.

[0024] The collected multi-source monitoring data is preprocessed. This data is categorized into text, image, and structured data. Noise is eliminated and missing data is filled in to create standardized input data suitable for subsequent modeling. Feature extraction is then performed on each of the preprocessed data types to support subsequent fusion modeling and semantic reasoning. This includes text feature extraction, image feature extraction, spatiotemporal feature extraction, and structured feature extraction, resulting in multidimensional drug feature data.

[0025] For step S102, a spatiotemporal graph neural network (ST-GNN) model is constructed. The extracted text feature vectors, image feature vectors, spatiotemporal feature vectors, and structured feature vectors are deeply fused to construct a multi-dimensional fused feature representation. Based on the spatiotemporal graph neural network, online and offline multi-scenario monitoring data (text, images, sales records, qualification information, etc.) are integrated, incorporating features from the time dimension (prohibition period, sales time) and spatial dimension (sales region, logistics information) to construct a spatiotemporal fusion model. This enables deep correlation and collaborative analysis of multi-dimensional data, eliminates monitoring blind spots, and improves the comprehensiveness of monitoring.

[0026] Further, a dynamic update and adaptation mechanism for the timeliness of prohibited drugs will be constructed. By combining the temporal feature learning ability of spatiotemporal graph neural networks, the timeliness adjustment information of prohibited drugs (such as policy adjustments, temporary prohibition notices, and batch prohibition information) will be automatically captured. This will enable dynamic optimization of monitoring rules and adapt to multiple timeliness scenarios such as long-term prohibition, prohibition during specific periods, and prohibition of specific batches, thereby avoiding expired monitoring and missed monitoring.

[0027] For step S103, a variant semantic reasoning model is constructed. Combined with the list of prohibited drug prototype categories, this model achieves precise association and semantic reasoning between variant keywords and prototype keywords, as well as precise matching of drug attributes with prohibited categories. For variant forms of prohibited drug keywords, such as phonetic changes, shape changes, splitting, combination, pinyin, homophones, and industry slang, precise association and semantic reasoning between variant keywords and prohibited drug prototype keywords are achieved. This eliminates the impact of variant expressions on monitoring results, ensures accurate identification of hidden violations, and reduces the risk of missed detections.

[0028] By integrating multi-dimensional data (drug ingredients, efficacy, production information, sales descriptions, etc.) and leveraging the feature extraction capabilities of variant semantic reasoning and spatiotemporal graph neural networks, we can accurately determine the essential attributes of drugs, establish a mapping relationship between the essential attributes of drugs and prohibited drug categories, effectively distinguish the boundaries between legal drugs and prohibited drugs, and between non-drugs and prohibited drugs, and reduce the risk of misjudgment.

[0029] For step S104, record the key intermediate results of the entire process of monitoring, feature extraction, fusion modeling, semantic reasoning, and violation determination, and automatically generate a structured evidence chain.

[0030] Generating the structured evidence chain includes: generating a structured evidence chain with attribute basis, legal basis, and judgment explanation; the attribute basis is used to record OCR recognition results, variant matching records, and essential attribute features; the legal basis is used to associate specific clauses in drug management laws and regulations; the judgment explanation is used to clarify the violating entity, prohibited sales type, violation facts, conflict reasoning points, and disposal suggestions.

[0031] Compared with existing technologies, this application improves the accuracy and comprehensiveness of identifying prohibited drugs in complex scenarios. It achieves high-precision violation identification in complex scenarios such as variant descriptions of prohibited drugs, cross-scenario sales, dynamic timeliness, and non-pharmaceuticals masquerading as prohibited drugs. The accuracy rate of variant keyword identification is ≥95%, and the overall accuracy rate of prohibited drug identification is ≥98%. It significantly reduces the risk of misjudgment and omission, and comprehensively captures explicit and implicit violations.

[0032] To automate and dynamize the monitoring of prohibited drugs and reduce labor costs. By leveraging the synergy of spatiotemporal graph neural networks and variant semantic reasoning, the system enables automatic collection, analysis, and reasoning of data from multiple scenarios, as well as dynamic optimization of monitoring rules. This reduces reliance on manual rule configuration, manual review, and manual evidence preparation, thereby lowering labor costs for regulatory departments and platforms. Monitoring efficiency is improved by over 80%, enabling routine and dynamic monitoring.

[0033] It adapts to the monitoring needs of prohibited drugs across multiple timeframes and scenarios. It can flexibly adapt to various timeframe scenarios, such as long-term prohibitions, prohibitions in specific regions during specific periods, and prohibitions on specific batches from specific manufacturers, as well as multiple online and offline sales scenarios. This achieves full coverage and no blind spots in the monitoring of prohibited drugs, and improves the speed of responding to policy adjustments and risk assessments to the hour level, ensuring the timeliness and relevance of monitoring work.

[0034] In one specific embodiment, the method includes: the multi-source monitoring data includes drug-related data, sales-related data, spatiotemporal related data, legal and regulatory data, and violation clue data; The preprocessing of multi-source monitoring data includes word segmentation, stop word removal, standardization, and preliminary variant annotation of text data; size unification, format conversion, and OCR recognition of image data; format unification and coordinate calibration of spatiotemporal data; and structured parsing of legal rule data. The multidimensional drug feature data includes text features, image features, spatiotemporal features, and structured features.

[0035] Drug-related data: Text data such as drug title, product details page description, promotional copy, ingredient list, efficacy description, manufacturer, production batch, drug approval number, and category label; Image data such as drug main image, packaging image, demonstration video, and promotional images; Sales-related data: Sales entity information (name, qualifications, business scope, etc.), sales records (sales time, sales quantity, sales region, logistics information, etc.), pricing information, user reviews, customer service chat records, etc. Spatiotemporal related data: monitoring time, sales region (accurate to the district / county level), logistics trajectory, offline store location, etc.; Legal and regulatory data: List of prohibited drug categories, information on the duration of the ban, relevant legal and regulatory provisions, regulatory policy documents, etc. Data on violations: information from public reports, historical violation records, and information from regulatory authorities.

[0036] The collected multi-source monitoring data is preprocessed to eliminate data noise and fill in missing data. Specifically, this includes: word segmentation, stop word removal, standardization (unifying the expression format), and preliminary variant annotation of text data; size unification, format conversion, and OCR recognition (extracting text information from images) of image data; standardized encoding processing of structured data (sales records, qualification information, etc.); format unification and coordinate calibration of spatiotemporal data; and structured parsing of legal and prohibition rules data to form standardized input data that can be used for subsequent modeling.

[0037] Feature extraction is performed on different types of preprocessed data to support subsequent fusion modeling and semantic reasoning, specifically including: Text feature extraction: Input text data such as drug-related texts, sales promotion texts, and chat logs into a pre-trained language model (such as the BERT model) to extract text semantic feature vectors, focusing on capturing core semantic information such as drug names, ingredients, and efficacy, while also extracting variant keyword candidate features from the text; Image feature extraction: For image data such as drug images and packaging images, a deep visual feature extraction model (such as a CNN model) is used to obtain image visual feature vectors. Combined with text features obtained from OCR recognition, the semantic expression related to drug attributes and prohibition determination in the image modality is enhanced. Spatiotemporal feature extraction: Extract time features (sales period, correlation with the prohibition period, etc.) and spatial features (regional violation probability, regional correlation, etc.) from spatiotemporal data such as sales time, sales region, and logistics trajectory, and transform them into spatiotemporal feature vectors; Structured feature extraction: Structured data such as sales entity qualifications, drug approval numbers, and production batches are converted into structured feature vectors through predefined coding mapping rules, focusing on capturing core features such as qualification compliance and drug batch validity.

[0038] In one specific embodiment, constructing the spatiotemporal graph neural network model includes: A spatiotemporal graph is constructed with monitoring subject nodes, drug nodes, and spatiotemporal nodes as core nodes, and the edges between nodes in the spatiotemporal graph represent the association relationship. The monitoring entity node includes sales entity information; the drug node includes monitored drug information; and the spatiotemporal node includes time information and spatial information.

[0039] The spatiotemporal graph is constructed with “monitoring subject - drug - spatiotemporal node” as the core nodes. The monitoring subject node includes features such as the sales subject’s qualifications and historical violation records, the drug node includes information such as drug text and image features, and the spatiotemporal node includes time and space features. The edges between nodes represent the relationships (such as the sales relationship between the sales subject and the drug, the sales spatiotemporal relationship between the drug and the spatiotemporal node, and the regional relationship between the monitoring subject and the spatiotemporal node).

[0040] In one specific embodiment, the spatiotemporal graph neural network model adopts a fusion structure of graph convolutional neural network and temporal neural network: The spatial correlation features between different nodes are captured by a graph convolutional layer; the temporal correlation features in the time dimension are captured by a temporal attention layer; and the spatial correlation features and temporal correlation features are weighted and fused with the multidimensional drug feature data by a fusion layer to generate the multidimensional fused feature representation.

[0041] Spatiotemporal feature fusion uses the graph convolutional layer of a spatiotemporal graph neural network to propagate and update node features, capturing the correlation between features of different dimensions. At the same time, it incorporates a temporal attention mechanism to focus on the correlation between the prohibition period and the sales time, and the correlation between regional violation patterns and the current sales region, generating a unified fusion feature vector that can represent drug attributes, sales behavior, and spatiotemporal features.

[0042] In one specific embodiment, the spatiotemporal graph neural network model further includes: Real-time collection and monitoring data; dynamic addition or marking of invalid nodes and edge relationships. Based on the real-time acquisition of information on the adjustment of the sales ban period, the timeliness characteristics and edge association weights of the spatiotemporal graph nodes are dynamically updated to adapt to dynamic timeliness scenarios.

[0043] The dynamic update mechanism combines real-time collected monitoring data and information on the adjustment of the sales ban period to dynamically update the spatiotemporal graph node features and edge relationships, ensuring that the fused features can adapt to the dynamic timeliness of the sales ban drugs and the real-time changes in sales behavior.

[0044] In one specific embodiment, constructing the variant semantic knowledge base includes: Based on the list of prohibited drug prototypes, we collect and organize various variant forms of prohibited drugs, construct a variant semantic knowledge base, and label the correlation and semantic similarity between variants and prototypes. The variant forms include sound variants, shape variants, split and combine variants, pinyin or English variants, and industry slang.

[0045] The variant semantic knowledge base is constructed based on the list of prohibited drug prototype categories. Common variant forms of various prohibited drugs (sound changes, shape changes, splitting, combination, pinyin, homophones, industry slang, etc.) are collected and organized to build a variant semantic knowledge base, and the correlation and semantic similarity between variants and prototypes are marked.

[0046] In one specific embodiment, the matching of variant keywords with the prototype keywords of prohibited drugs using a variant semantic reasoning model includes: The variant keyword candidate features obtained from text feature extraction are input into the variant semantic reasoning model. Through semantic similarity calculation and variant rule matching, the variant keywords are matched with the prototype keywords to identify statements related to prohibited drugs.

[0047] Variant Keyword Identification and Matching: Variant keyword candidate features extracted from text features are input into the variant semantic reasoning model. Through semantic similarity calculation (such as cosine similarity) and variant rule matching, the variant keywords are accurately matched with the prototype keywords, and the relevant statements of suspected prohibited drugs are identified.

[0048] In one specific embodiment, the semantic mapping between the essential attributes of drugs and prohibited categories includes: By combining the multi-dimensional fusion feature representation to extract the essential attributes of drugs, semantic mapping is performed with the prohibited drug categories to determine whether a drug falls within the prohibited sales scope and to clarify the prohibited sales type.

[0049] Drug attributes are matched with prohibited categories. By combining the unified fusion feature vector output by the spatiotemporal graph neural network, the essential attributes of the drug (ingredients, efficacy, production batch, etc.) are extracted and semantically mapped with prohibited drug categories to determine whether the drug falls within the prohibited scope and to clarify the type of prohibition (long-term prohibition, temporary prohibition, etc.).

[0050] In one specific embodiment, generating the structured evidence chain includes: generating a structured evidence chain with attribute basis, legal basis, and judgment explanation; Attributes are used to record OCR recognition results, variant matching records, and essential attribute features; The legal basis refers to specific clauses in the relevant laws and regulations governing drug administration; The judgment explanation is used to clarify the violating entity, the type of prohibited sale, the facts of the violation, the points of conflict reasoning, and the proposed handling.

[0051] The system records key intermediate results throughout the entire process of monitoring, feature extraction, fusion modeling, semantic reasoning, and violation determination, automatically generating a structured evidence chain. This chain includes: Structured evidence chain organization: integrating drug information (prototype keywords, variant keywords, essential attributes, etc.), sales entity information (qualifications, descriptions of violations, etc.), spatiotemporal information (sales time, region, etc.), violation reasoning process, applicable legal provisions, and relevant evidence (text, screenshots, etc.) to form a standardized evidence chain; Explainable output: visually outputting violation determination results, reasoning paths, legal basis, and evidence chains, clearly explaining the association between variant keywords and prototype keywords, the matching basis between drugs and prohibited categories, and the correspondence between violations and legal provisions, supporting manual review and regulatory enforcement retrieval; Evidence storage and traceability: storing the structured evidence chain in a graph database format, supporting evidence querying and traceability, ensuring the integrity, authenticity, and traceability of evidence, and adapting to enforcement archiving needs.

[0052] In one exemplary embodiment, such as Figure 3 As shown, the Spatiotemporal Graph Neural Network (ST-GNN) used in this invention is based on a fusion structure of Graph Convolutional Neural Network (GCN) and Temporal Neural Network (LSTM), taking into account both spatial feature correlation mining and dynamic learning of temporal features. The specific structure includes an input layer, a graph convolutional layer, a temporal attention layer, a fusion layer, and an output layer, with the functions of each layer as follows: Input layer: Receives preprocessed multi-dimensional feature vectors (text, image, spatiotemporal, and structured feature vectors) as initial features for nodes in the spatiotemporal graph. Node types include monitoring subject nodes, drug nodes, and spatiotemporal nodes. Node features are initialized differentially according to node type. Graph convolutional layer: Employs weighted graph convolution operations to propagate and update node features, capture the relationships between different nodes (such as the sales relationship between the sales entity and the drug, and the spatiotemporal relationship between the drug and spatiotemporal nodes), calculate the weighted sum of the neighbor features of the node, update the node features, and highlight the node relationships related to the prohibition of sale determination. Temporal Attention Layer: Integrates LSTM temporal modeling capabilities to learn the temporal features of spatiotemporal nodes, captures the changing patterns in the time dimension (such as adjustments to the sales ban period and changes in the probability of violations at different times), and introduces an attention mechanism to dynamically allocate the weights of nodes at different time nodes and in different regions, focusing on the node features within the sales ban period and in high-risk violation areas. Fusion layer: The spatial correlation features output by the graph convolutional layer and the temporal features output by the temporal attention layer are weighted and fused together. Combined with text, image and structured features, a unified multi-dimensional fused feature vector is generated, which eliminates redundancy between features of different modalities and strengthens core features. Output layer: The fused feature vector is output to the subsequent variant semantic reasoning model and violation judgment module, providing feature support for drug attribute recognition and prohibition judgment.

[0053] In an exemplary embodiment, the construction of nodes and edges in the spatiotemporal graph directly affects the effect of feature fusion. Based on the core requirement of monitoring prohibited drugs, this invention formulates the following construction rules: Node Construction: Monitoring Entity Nodes: Each sales entity (online store, offline pharmacy, third-party platform, etc.) is treated as an independent node. Node characteristics include the sales entity name, qualification number, business scope, historical violation count, violation type, and registered region. Drug Nodes: Each drug to be monitored (distinguished by name and production batch) is treated as an independent node. Node characteristics include drug text features, image features, manufacturer, production batch, drug approval number, and category label. Spatiotemporal Nodes: Spatiotemporal nodes are divided according to "time slice - region." Time slices can be set according to monitoring precision (e.g., hour, day), and regions are accurate to the district / county level. Node characteristics include time information, region information, historical violation records within that spatiotemporal range, and prohibition period information.

[0054] Edge Construction: Sales-Related Edges: Edges are established between monitoring subject nodes and drug nodes. Edge weights are determined based on sales volume, sales frequency, and sales duration. The higher the sales volume and frequency, the greater the edge weight, indicating a closer correlation between the two. Spatiotemporal-Related Edges: Edges are established between drug nodes and spatiotemporal nodes. Edge weights are determined based on the sales volume and sales proportion of the drug within that spatiotemporal range, while also considering the prohibition period. If the sales time falls within the prohibition period, the edge weight is appropriately increased. Regional-Related Edges: Edges are established between spatiotemporal nodes in adjacent regions to capture the patterns of illegal spread between regions (e.g., if illegal drug sales occur in one region, the risk of illegal spread in adjacent regions increases). Edge weights are determined based on regional distance and the probability of illegal spread. Qualification-Related Edges: Edges are established between monitoring subject nodes to associate with sales entities that have a relationship (e.g., chain pharmacies, sales entities under the same parent company). Edge weights are determined based on the degree of association, enabling cross-subject collaborative monitoring.

[0055] In an exemplary embodiment, the present invention achieves deep fusion of spatiotemporal features and multi-dimensional features through three steps: "spatial feature propagation - temporal feature learning - multimodal feature fusion," as detailed below: Spatial feature propagation: Node features are propagated through graph convolutional layers. The calculation formula is: h_i = σ(Σ_j(w_ij) h_j) + b), where h_i is the updated feature of node i, w_ij is the edge weight between node i and node j, h_j is the initial feature of node j, b is the bias term, and σ is the activation function (using the ReLU function). This operation captures the spatial association features between different nodes. Temporal feature learning: The features of spatiotemporal nodes are input into the LSTM layer in chronological order to learn the dynamic changes in the time dimension, capture the temporal patterns of sales restriction period adjustment and sales behavior, and output a temporal feature vector; at the same time, a temporal attention mechanism is introduced to calculate the attention weights of different time slices, focusing on the temporal features within the sales restriction period and during periods of high incidence of violations. Multimodal feature fusion: Spatial correlation features output from graph convolutional layers and temporal features output from LSTM layers are weighted and fused with text features, image features, and structured features. The fusion formula is: h_fusion = α h_space + β h_time + γ h_text + δ h_image + ε h_struct, where α, β, γ, δ, and ε are fusion weights, which are automatically optimized through model training to ensure that the fusion features can comprehensively represent drug attributes, sales behavior, and spatiotemporal characteristics, providing reliable support for subsequent semantic reasoning and violation determination.

[0056] In an exemplary embodiment, to adapt to the dynamic timeliness of prohibited drugs and the real-time changes in sales behavior, this invention designs a spatiotemporal graph dynamic update mechanism, specifically including three aspects: node update, edge update, and feature update. Node Update: When a new sales entity, new drug, or new spatiotemporal node (such as a new time slice or a new monitoring region) is added, the corresponding node is automatically added and the node characteristics are initialized; when the sales entity's qualification expires, the drug is removed from the shelves, or the spatiotemporal node is outside the monitoring range, the node status is automatically marked and the feature propagation is stopped. Update edge: When the sales relationship between the sales entity and the drug changes (such as adding a new drug to be sold or stopping the sale of a certain drug), update the weight of the sales-related edge; when the sales data within the time and space range changes or the regional illegal spread pattern changes, update the weight of the time and space related edge and the regional related edge. Feature Update: Real-time collection of monitoring data and real-time updates of node features (such as historical violation records of sales entities, drug sales data, and violation records of spatiotemporal nodes). At the same time, combined with information on the adjustment of the ban period, the ban period features of spatiotemporal nodes are updated to ensure that the spatiotemporal map can reflect changes in the monitoring scenario in real time, thereby improving the timeliness and accuracy of monitoring.

[0057] In one exemplary embodiment, a variant semantic reasoning model is constructed to accurately associate and semantically reason with the original keywords of prohibited drugs for variant forms such as sound changes, shape changes, splitting, combination, pinyin, homophones, and industry slang. This eliminates the impact of variant expressions on monitoring results, ensures accurate identification of hidden violations, and reduces the risk of missed detections.

[0058] Construction and updating of the variant semantic knowledge base: The variant semantic knowledge base is the foundation for realizing variant keyword recognition and reasoning. This invention combines the list of prohibited drug categories, historical violation data, and industry code words to construct a dynamically updated variant semantic knowledge base. The specific process is as follows: Prototype Keyword Analysis: Based on the prohibited drug categories provided in the background (narcotic drugs, abortion drugs, toxic drugs, etc.), the prototype keywords of various prohibited drugs are analyzed, including the generic name, brand name, and ingredient name of the drug, such as "mifepristone tablets", "codeine tablets", and "mercury", as a benchmark for variant matching.

[0059] Variant Form Collection and Classification: This section collects common variant forms of the original keywords for various prohibited drugs, categorizing them by variant type, specifically including: Phonetic Changes: Homophones and near-homophone variants, such as "ibuprofen" → "buprofen" or "buprofen"; Graphic Changes: Abbreviations, abbreviations, and misspellings, such as "mifepristone tablets" → "mifepristone tablets," "mifepristone," or "mifepristone tablets"; Splitting and Combining: Keyword splitting and combination with other words, such as "compound licorice tablets" → "compound licorice tablets," "licorice compound tablets," or "compound licorice tablets"; Pinyin / English: Full Pinyin name, Pinyin abbreviation, English translation, and variants, such as "ibuprofen" → "Buluofen," "BLF," or "Ibuprofen"; Industry Code Words: Code words used in the pharmaceutical industry and in illegal sales, such as "painkillers" → drugs for terminating pregnancy, and "special cough powder" → compound preparations containing codeine.

[0060] Knowledge base annotation and initialization: For each variant keyword, annotate its corresponding prototype keyword, variant type, semantic similarity, and frequency of occurrence, and initialize the variant semantic knowledge base.

[0061] Dynamic updates: By combining real-time monitoring data, historical violation records, and public reports, new variant keywords are automatically captured. After manual review and confirmation, they are added to the knowledge base. At the same time, the semantic similarity and frequency of occurrence between variants and prototypes are updated to ensure that the knowledge base can cover the latest variant forms.

[0062] Structure and Working Principle of the Variant Semantic Reasoning Model: The variant semantic reasoning model constructed in this invention is based on the pre-trained language model (BERT) and improved attention mechanism to achieve accurate matching and semantic reasoning between variant keywords and prototype keywords. The model structure includes an input layer, a semantic encoding layer, an attention matching layer, an inference layer, and an output layer. The specific working principle is as follows: Input layer: Receives preprocessed text feature vectors, variant keyword candidate features, and prototype keyword features from the variant semantic knowledge base as model input; Semantic encoding layer: The BERT pre-trained language model is used to semantically encode variant keyword candidates and prototype keywords to generate high-dimensional semantic vectors, capture the deep semantic information of keywords, and eliminate the influence of surface expression differences; at the same time, the text context features are encoded and combined with contextual semantics to improve the accuracy of variant recognition (e.g., combined with the context of "used to terminate pregnancy", it is determined that "painkillers" is a code word for termination of pregnancy drugs). Attention Matching Layer: An attention mechanism is introduced to calculate the attention weight between the candidate semantic vector of variant keywords and the semantic vector of each prototype keyword. The focus is on prototype keywords with high semantic similarity. At the same time, the attention weight is adjusted by combining variant type and frequency of occurrence to improve matching accuracy. Inference layer: Based on the attention matching results, semantic reasoning is performed to calculate the semantic similarity between variant keyword candidates and prototype keywords (using cosine similarity calculation). Combined with variant rules (such as splitting and combining rules, sound change rules), the prototype keywords corresponding to variant keywords are determined to clarify whether they are related to prohibited drugs. At the same time, the fusion features output by the spatiotemporal graph neural network are combined to assist in judging drug attributes and improve the accuracy of reasoning. Output layer: Outputs the matching results, semantic similarity, variant type, and corresponding prohibited drug categories of variant keywords with prototype keywords, providing support for subsequent violation determination; For candidate keywords with semantic similarity below the preset threshold, they are marked as suspected variants and output to the manual review module.

[0063] A semantic mapping method for drug attributes and prohibited categories is proposed. This method combines the results of variant semantic reasoning with the fusion features of spatiotemporal graph neural network output to achieve accurate semantic mapping between the essential attributes of drugs and prohibited categories. The specific method is as follows: Drug essential attribute extraction: Extract essential attribute features of drugs from spatiotemporal graph fusion feature vectors, including component features, efficacy features, production batch features, and usage features, and combine them with image text information obtained by OCR recognition and drug approval numbers in structured data to supplement drug attribute information; Semantic modeling of prohibited drug categories: Semantic modeling is performed on prohibited drug categories. Each prohibited category corresponds to a semantic vector, which includes the core attributes of the category (such as the composition characteristics of narcotic drugs and the efficacy characteristics of drugs for terminating pregnancy) and the time limit of the prohibition. Semantic mapping and matching: Calculate the similarity between the semantic vector of the essential attributes of the drug and the semantic vector of each prohibited category. Combine the keyword matching results of prohibited drugs obtained by variant semantic reasoning to determine whether the drug belongs to a certain prohibited category. If the similarity exceeds the preset threshold and the keyword matching results are consistent, the drug is determined to belong to the corresponding prohibited category. If there are multiple suspected prohibited categories, further reasoning is used to confirm the status by combining contextual semantics and spatiotemporal features. Determining if a product is not matched with a drug approval number and has significant differences in its essential attributes from those of a drug, the product description and efficacy claims are used to determine whether there is any violation of "non-drug masquerading as drug". If the product is advertised as having drug efficacy and matches keywords of prohibited drug variants, it is determined to be a suspected prohibited drug sold illegally.

[0064] In one exemplary embodiment, the dynamic timeliness adaptation scheme for prohibited drugs includes structured modeling of the timeliness information and a dynamic update and adaptation mechanism for the timeliness. For the multiple timeliness types of prohibited drugs (long-term prohibition, prohibition in certain regions during a specific period, and prohibition of specific batches from specific manufacturers), a structured model of the timeliness of the prohibition is constructed, transforming the timeliness information into computable and associative features, specifically including: Time Limit Type Labeling: For each prohibited drug and prohibited category, label its time limit type and clearly define the time limit boundaries, as follows: Long-term prohibition: Indicate the start time of the prohibition, with no end time (or label it as "long-term"), such as narcotic drugs, toxic drugs, etc.; Prohibition in certain areas during specific periods: Indicate the start time, end time, and prohibited areas (accurate to the district / county level), such as prohibiting the sale of cold and fever-reducing drugs in a certain area during the epidemic; Prohibition of specific batches from specific manufacturers: Indicate the start time, end time, manufacturer, production batch, and prohibited areas (optional), such as prohibiting the sale of a certain batch of ibuprofen from a certain manufacturer due to quality issues.

[0065] Timeliness feature extraction: The prohibition timeliness information is transformed into a timeliness feature vector, including timeliness type code, start time, end time, prohibition area code, remaining time, etc., and integrated into the spatiotemporal graph node features and prohibition category semantic vector.

[0066] Structuring of time limit rules: Regulatory policies and notices related to the time limit for prohibition of sales are broken down into structured rules, clarifying the conditions for time limit adjustment (such as policy adjustments and risk investigation results) and the time limit change process, providing support for dynamic updates of time limits.

[0067] A dynamic update and adaptation mechanism for the time limit is constructed by combining the temporal learning capability of spatiotemporal graph neural networks with real-time monitoring data to ensure that the monitoring rules can adapt to changes in the time limit in real time, as detailed below: Real-time collection of timeliness information: Information on adjustments to the timeliness of sales bans (such as adding temporarily banned drugs, lifting the sales ban on a certain drug, adjusting the scope of the sales ban area, etc.) is collected in real time through channels such as regulatory department interfaces, official notices, and news announcements. After manual review and confirmation, the information is entered into the system.

[0068] Dynamic updates of timeliness features: Based on the collected timeliness adjustment information, the timeliness features of the prohibited sales timeliness structured model, spatiotemporal graph nodes, and prohibited sales category semantic vectors are automatically updated. At the same time, the edge association weights of the spatiotemporal graph are updated (e.g., if a drug exceeds the prohibited sales timeliness, its edge weight with the spatiotemporal node is reduced).

[0069] Time-based reasoning: In the process of determining violations, the sales time and sales region are accurately correlated with the prohibition period information to determine whether the drug was within the prohibition period and whether the sales region was a prohibited region when the sales occurred; for specific batches of prohibited drugs, the production batch of the drug is considered to determine whether it belongs to the prohibited batch.

[0070] Timeliness warning mechanism: Set timeliness warning thresholds. For drugs that are close to the end of the sales ban or are about to be added to the sales ban, the timeliness warning will be automatically generated to remind regulatory personnel and the platform to pay attention to the timeliness changes and adjust the monitoring focus in a timely manner. For drugs that have exceeded the sales ban period, the monitoring mark will be automatically removed to avoid expired monitoring.

[0071] In an exemplary embodiment, a specific example flow of a drug intelligent monitoring method of this application is as follows: Figure 6 As shown, the specific implementation process is as follows: System runtime environment preparation: The system runs on a cloud computing platform, deploying core services such as data acquisition, preprocessing, spatiotemporal graph neural networks, variant semantic reasoning, prohibited sales determination, and evidence chain generation through containerization. The system pre-loads a trained spatiotemporal graph neural network (integrating BP neural networks and graph attention mechanisms), a variant semantic recognition model (optimized based on a pre-trained language model), and databases of prohibited drugs, drug approval numbers, and violation monitoring keywords stored in a MySQL database, as well as a structured graph of laws and regulations stored in the Neo4j graph database, ensuring that relevant data and models can be quickly accessed during the monitoring process.

[0072] S201. Full Collection of Multi-Source Spatiotemporal Data: The system collects multi-dimensional spatiotemporal data for three major scenarios: online e-commerce platforms, offline pharmacies, and batch-prohibited drugs. Taking online e-commerce platform monitoring (prohibited sale of the anesthetic drug "Compound Glycyrrhiza Oral Solution"), offline pharmacy monitoring (prohibited sale of the abortion drug "Mifepristone Tablets"), and batch-prohibited drug monitoring (a certain batch of "Tromadol Tablets") as examples, the following multimodal and spatiotemporal information is collected: Online e-commerce platform (Compound Glycyrrhiza Oral Solution): Text data: The product title is "Compound Glycyrrhiza Oral Solution: Powerful Cough Relief, Essential for Home Use," the details page states "No prescription required, bulk shipping," customer service chat records mention "concealed packaging to avoid inspection," and user reviews state "Significant cough relief, no need for medical treatment." Image data: The main product image shows the complete packaging of the medicine (labeled "Compound Glycyrrhiza Oral Solution"), attached images include photos of the actual medicine and shipping packaging (no obvious medicine markings), and no clear approval number is marked. Structured data: The category is labeled "Daily Necessities (subtly labeled)," with no clear specifications, priced at 58 yuan / bottle, and the logistics method is "Ordinary Express, contents not labeled." The merchant does not have a drug business license number. Spatiotemporal data: The merchant's shipping address is an unlicensed small workshop (39.9°N, 116.4°E), transaction times are concentrated between 22:00 and 2:00, delivery addresses are scattered across multiple provinces nationwide, and the cumulative sales volume in 3 days is 1020 orders.

[0073] Offline pharmacy (Mifepristone tablets): Text data: The drug label indicates "Mifepristone tablets," with no notice stating "Drugs for terminating pregnancy are prohibited from retail sale." The salesperson mentions to the customer that "no prescription is required; you can purchase directly." The receipt indicates "1 box of medicine costs 50 yuan." Image data: The pharmacy shelf image shows the drug is placed in the over-the-counter drug area. The drug packaging image indicates the approval number "National Drug Approval Number H43021942." The transaction record screenshot at the cashier clearly shows the transaction information. Structured data: The pharmacy does not have the qualifications to sell drugs for terminating pregnancy. The drug's inventory records do not show legal registration information for this drug. The transaction was a cash payment, and the customer's identity information was not registered. Spatiotemporal data: The pharmacy's latitude and longitude are 30.6°N, 114.3°E. The transaction time was 2024-05-21 15:30:00. The average weekly sales recently are 6 boxes, with no obvious pattern in transaction time.

[0074] Batch Banned (A Certain Batch of Acetaminophen Tramadol Tablets): Text Data: The product title is "Acetonaminophen Tramadol Tablets, Analgesic and Anti-inflammatory, Genuine Product Guaranteed," but the product details page does not indicate batch information. The distributor's supply record indicates batch number "20240101" (banned). Structured Data: The manufacturer is a pharmaceutical company; batch number 20240101 is banned due to quality issues. The distributor lacks the necessary qualifications to sell narcotic drugs, and pharmacy purchase records show 100 boxes of this batch. Spatiotemporal Data: The production address is an industrial park; the distributor's shipping addresses are multiple unlicensed pharmacies; transactions occurred within one week of the ban notice being issued.

[0075] S202. Multi-source data preprocessing includes: text data: all text is segmented and stop words are removed. Through the variant semantic preprocessing module, variant keywords such as "compound licorice oral solution" and "MFSTP" are uniformly normalized into regular keywords. Obscure expressions are marked, such as "cough-relieving miracle drug" corresponding to "compound codeine phosphate solution". Offline dialogue recordings are transcribed and core semantics are extracted.

[0076] Image data: The image size is uniformly adjusted to 800×600 pixels, and grayscale and noise reduction are performed. Text information such as drug name and approval number are extracted from the image through the Tesseract-OCR engine. Blurry images are enhanced to ensure that the OCR recognition accuracy is ≥98%.

[0077] Structured data: Verify the qualifications of merchants, distributors, and pharmacies, and mark entities without qualifications or with expired qualifications; detect abnormal sales and pricing values, and mark bulk transactions and abnormally priced products; verify drug approval numbers and batch information, and confirm the validity of prohibited batches.

[0078] S203. Spatiotemporal Alignment: The preprocessed text, images, and structured data are associated and aligned with the corresponding spatiotemporal data (latitude and longitude, transaction timestamps) to construct "spatiotemporal-data" association pairs, such as "2024-05-21 15:30:00 + 30.6° North Latitude, 114.3° East Longitude + Mifepristone Tablets + Cash Transaction". This ensures that each data point contains complete spatiotemporal dimension information, providing support for subsequent modeling.

[0079] S204. Spatiotemporal graph neural network modeling includes: Spatiotemporal graph construction: with "transaction nodes" as the core, a spatiotemporal graph structure is constructed. The nodes include product nodes (associated with product ID, drug name, and prohibited sales attributes), merchant / pharmacies / distributors nodes (associated with qualification information and address), transaction nodes (associated with transaction time and amount), and region nodes (associated with latitude and longitude). The edge structure associates the correspondence between each node and the spatiotemporal association strength.

[0080] Feature embedding: Preprocessed text features, image features, and structured features are embedded into the corresponding nodes of the spatiotemporal graph. A backpropagation neural network is used to optimize the feature embedding effect, and the network weights and biases are adjusted to reduce feature errors. Spatiotemporal features are embedded into the edge structure to characterize the spatiotemporal correlation strength between nodes.

[0081] Spatiotemporal correlation reasoning: Through graph attention mechanism, the correlation weight between each node is calculated to discover abnormal spatiotemporal patterns, such as the nighttime bulk transaction of compound licorice oral solution online, the unregistered cash transaction of mifepristone tablets offline, and the illegal circulation of batches of prohibited drugs after the issuance of the prohibition notice. The correlation weight is marked as spatiotemporal abnormal if it is higher than the preset threshold (0.8).

[0082] S205. Preliminary Screening Output a list of spatiotemporal anomalies and a list of anomalies. Initially filter out transactions, merchants, pharmacies, and distributors suspected of involving prohibited drugs. Use the Adaboost classifier to optimize the filtering results and improve the accuracy of anomaly identification.

[0083] S206. Variant semantic reasoning and qualification verification include: Variant semantic matching: The selected abnormal transaction-related text and OCR-recognized text are semantically matched with the prohibited drug keyword database. A variant semantic reasoning model combined with a Bayesian classifier is used to calculate semantic similarity. For example, the semantic similarity between compound licorice oral solution and prohibited anesthetic drugs is 0.96, and the semantic similarity between mifepristone tablets and prohibited abortion drugs is 0.97, both of which are higher than the threshold of 0.85, and are judged as suspected prohibited drugs; and prohibited drugs corresponding to cryptic expressions are identified.

[0084] Qualification and batch verification: Verify the approval number of suspected prohibited drugs, such as the approval number of Compound Licorice Oral Solution being a fake number, and the approval number of Mifepristone Tablets being valid but the pharmacy not having the corresponding business qualifications; verify batch information, confirm that a certain batch of Paracetamol Tramadol Tablets (20240101) is a prohibited batch, and the distributor and pharmacy do not have legal business qualifications.

[0085] S207. Prohibition determination and conflict reasoning include: Prohibition determination: Decision tree algorithms are used to construct determination rules. A violation is determined if any of the following conditions are met: spatiotemporal anomaly + semantic matching of prohibited drugs + unqualified qualifications; semantic matching of prohibited drugs + lack of valid approval number / fake approval number; unqualified entities selling suspected prohibited drugs; illegal circulation of prohibited batches of drugs. Final determination: Online merchants illegally selling prohibited anesthetic drugs (compound licorice oral solution) + counterfeit drugs; offline pharmacies illegally selling prohibited abortion drugs (mifepristone tablets); distributors and pharmacies illegally selling prohibited batches of acetaminophen tramadol tablets.

[0086] Conflict reasoning: Identify various conflicts, such as online merchants claiming "the product is a daily necessity, not a medicine," which conflicts with the OCR-recognized medicine packaging and semantic matching results; pharmacies claiming "no batch sales ban notice received," which conflicts with the sales ban notice pre-loaded by the system; and merchants implicitly labeling the product category, which conflicts with the actual drug attributes.

[0087] S208. The structured evidence chain output includes: The system automatically generates a test report containing the following structured evidence, which is simultaneously pushed to regulatory terminals, relevant platforms, and entities: Attribute basis: OCR-recognized drug packaging images, variant keyword matching records, spatiotemporal abnormal transaction data, and batch information screenshots, proving that the product is a prohibited drug and exhibits characteristics of illegal transactions. Legal basis: Related to Articles 61 and 98 of the *Drug Administration Law*, Article 9 of the *Administrative Measures for Internet Drug Information Services*, and relevant clauses of the *Regulations on Prohibiting Non-Medical Sex Determination of Fetuses and Sex-Selective Termination of Pregnancy*. Judgment explanation: Clearly identify the violating entity, the violating product, the type of prohibition (long-term / temporary / batch), the facts of the violation, and the points of conflict. The judgment result is "illegal sale of prohibited drugs," with the violation level and proposed handling simultaneously indicated.

[0088] Compared with existing technologies, the identification method of this application has been changed from "fixed keyword surface comparison" to "variant semantic recognition"; the data processing and fusion has been changed from isolated monitoring of a single scene and a single modality to the fusion of multimodal and spatiotemporal data based on spatiotemporal graph neural networks to achieve cross-scene and multi-dimensional collaborative monitoring; the matching method of drug attributes and prohibited categories has been changed from surface label matching to combining semantic mapping to achieve accurate matching and identify hidden violations.

[0089] The model boasts high accuracy in variant identification and strong ability to capture hidden violations; it exhibits excellent dynamic adaptability, meeting real-time monitoring needs across multiple timeframes and scenarios. It possesses powerful multi-dimensional data fusion capabilities, achieving comprehensive, blind-spot-free monitoring. It can significantly improve the accuracy of semantic recognition of prohibited drug variants, spatiotemporal fusion of multimodal data, and dynamic timeframe adaptation, enhancing the interpretability and full traceability of violation reasoning, while simultaneously improving the efficiency and compliance level of structured evidence chain generation.

[0090] Based on the same inventive concept, this application also provides a drug intelligent monitoring system. The solution provided by this system is similar to the solution described in the above method; therefore, the specific limitations in one or more system embodiments provided below can be found in the limitations described above, and will not be repeated here.

[0091] In one exemplary embodiment, such as Figure 5 As shown, a drug intelligent monitoring system is provided, the system comprising: The data acquisition module is used to acquire multi-source monitoring data, which includes drug-related data, sales-related data, spatiotemporal related data, legal and regulatory data, and violation clue data. The feature extraction module is used to preprocess the multi-source monitoring data and extract multi-dimensional features to obtain multi-dimensional drug feature data; the multi-dimensional drug feature data includes text features, image features, spatiotemporal features, and structured features; The spatiotemporal fusion module is used to deeply fuse the multidimensional drug feature data to obtain a multidimensional fused feature representation; The semantic reasoning and recognition module is used to construct a variant semantic knowledge base, match variant keywords with the prototype keywords of prohibited drugs through a variant semantic reasoning model, and perform semantic mapping between the essential attributes of drugs and prohibited categories by combining the multi-dimensional fusion feature representation. The evidence chain generation module is used to integrate key intermediate results throughout the monitoring process, generate a structured evidence chain including attribute basis, legal basis and judgment explanation, and output it visually.

[0092] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 6 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and databases. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media to run. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection.

[0093] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0094] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0095] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0096] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0097] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (Read-Only Memory). Memory includes ROM, magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0098] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0099] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0100] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for intelligent drug monitoring, characterized in that, The methods include: Acquire multi-source monitoring data, preprocess the multi-source monitoring data and extract multi-dimensional features to obtain multi-dimensional drug feature data; A spatiotemporal graph neural network model is constructed to deeply fuse the multidimensional drug feature data, thereby obtaining a multidimensional fused feature representation; A variant semantic knowledge base is constructed, and a variant semantic reasoning model is used to match variant keywords with the prototype keywords of prohibited drugs. The semantic mapping between the essential attributes of drugs and prohibited categories is performed by combining the multi-dimensional fusion feature representation. Integrate key intermediate results from the entire monitoring process, generate a structured chain of evidence, and provide a visual output.

2. The method according to claim 1, characterized in that, The method includes: The multi-source monitoring data includes drug-related data, sales-related data, spatiotemporal related data, legal and regulatory data, and data on clues of violations; The preprocessing of multi-source monitoring data includes word segmentation, stop word removal, standardization, and preliminary variant annotation of text data; size unification, format conversion, and OCR recognition of image data; format unification and coordinate calibration of spatiotemporal data; and structured parsing of legal rule data. The multidimensional drug feature data includes text features, image features, spatiotemporal features, and structured features.

3. The method according to claim 1, characterized in that, The construction of the spatiotemporal graph neural network model includes: A spatiotemporal graph is constructed with monitoring subject nodes, drug nodes, and spatiotemporal nodes as core nodes, and the edges between nodes in the spatiotemporal graph represent the association relationship. The monitoring entity node includes sales entity information; the drug node includes monitored drug information; and the spatiotemporal node includes time information and spatial information.

4. The method according to claim 3, characterized in that, The spatiotemporal graph neural network model adopts a fusion structure of graph convolutional neural network and temporal neural network: Capture spatial correlation features between different nodes using graph convolutional layers; Capture temporal correlation features in the time dimension through a temporal attention layer; The spatial correlation features and temporal correlation features are weighted and fused with the multidimensional drug feature data through the fusion layer to generate the multidimensional fused feature representation.

5. The method according to claim 4, characterized in that, The spatiotemporal graph neural network model also includes: Real-time collection and monitoring data; dynamic addition or marking of invalid nodes and edge relationships. Based on the real-time acquisition of information on the adjustment of the sales ban period, the timeliness characteristics and edge association weights of the spatiotemporal graph nodes are dynamically updated to adapt to dynamic timeliness scenarios.

6. The method according to claim 1, characterized in that, The construction of the variant semantic knowledge base includes: Based on the list of prohibited drug prototypes, we collect and organize various variant forms of prohibited drugs, construct a variant semantic knowledge base, and label the correlation and semantic similarity between variants and prototypes. The variant forms include sound variants, shape variants, split and combine variants, pinyin or English variants, and industry slang.

7. The method according to claim 6, characterized in that, The matching of variant keywords with the prototype keywords of prohibited drugs using a variant semantic reasoning model includes: The variant keyword candidate features obtained from text feature extraction are input into the variant semantic reasoning model. Through semantic similarity calculation and variant rule matching, the variant keywords are matched with the prototype keywords to identify statements related to prohibited drugs.

8. The method according to claim 7, characterized in that, The semantic mapping between the essential attributes of drugs and prohibited categories includes: By combining the multi-dimensional fusion feature representation to extract the essential attributes of drugs, semantic mapping is performed with the prohibited drug categories to determine whether a drug falls within the prohibited sales scope and to clarify the prohibited sales type.

9. The method according to any one of claims 1 to 8, characterized in that, Generating the structured chain of evidence includes: generating a structured chain of evidence containing attribute basis, legal basis, and judgment explanation; Attributes are used to record OCR recognition results, variant matching records, and essential attribute features; The legal basis refers to specific clauses in the relevant laws and regulations governing drug management; The judgment explanation is used to clarify the violating entity, the type of prohibited sale, the facts of the violation, the points of conflict reasoning, and the proposed handling.

10. A drug intelligent monitoring system, characterized in that, The system includes: The data acquisition module is used to acquire multi-source monitoring data, which includes drug-related data, sales-related data, spatiotemporal related data, legal and regulatory data, and violation clue data. The feature extraction module is used to preprocess the multi-source monitoring data and extract multi-dimensional features to obtain multi-dimensional drug feature data; the multi-dimensional drug feature data includes text features, image features, spatiotemporal features, and structured features; The spatiotemporal fusion module is used to deeply fuse the multidimensional drug feature data to obtain a multidimensional fused feature representation; The semantic reasoning and recognition module is used to construct a variant semantic knowledge base, match variant keywords with the prototype keywords of prohibited drugs through a variant semantic reasoning model, and perform semantic mapping between the essential attributes of drugs and prohibited categories by combining the multi-dimensional fusion feature representation. The evidence chain generation module is used to integrate key intermediate results throughout the monitoring process, generate a structured evidence chain including attribute basis, legal basis and judgment explanation, and output it visually.

Images