A method and device for aligning pharmaceutical subject data based on multimodal feature fusion

Abstract

This application discloses a method and device for aligning pharmaceutical entity data based on multimodal feature fusion. The method includes: acquiring multi-source data records containing entity name, first identity identifier, second identity identifier, related person information, and address information; constructing a core identity feature tuple composed of the entity name and dual identity identifiers to perform a first-level analysis; when the first-level analysis result is a partial mismatch, performing a second-level analysis based on the consistency of related person information and semantic proximity of addresses to determine whether the records point to the same entity; and merging and updating data determined to be from the same entity based on the date of the valid certificate. This application effectively addresses the unique "separation of licenses and permits" and information inconsistency issues in the pharmaceutical industry, significantly improving the accuracy and automation level of entity data alignment.

Description

Technical Field

[0001] This invention relates to the field of big data processing technology, and more specifically, to a master data governance method and device applied to the pharmaceutical industry, particularly a technology for realizing the identification, data alignment and data quality improvement of pharmaceutical entities such as enterprises and institutions. Background Technology

[0002] In the pharmaceutical industry, the master data governance of entities such as enterprises, pharmacies, and hospitals faces unique challenges. First, the data sources are complex. Pharmaceutical entities are usually subject to dual supervision by market supervision and administration departments and drug supervision and administration departments, which issue business licenses and drug operation / production licenses respectively. This results in the same entity having different identity credentials under different regulatory systems.

[0003] Secondly, inconsistencies in information are widespread. Due to reasons such as registration standards, management requirements, or historical development, the name of the same pharmaceutical entity on its business license may differ from its name on its drug distribution license. For example, "XX Pharmacy" and "XX Pharmacy (Chain Store)" may refer to the same store. In addition, license numbers, legal person information, address information, etc., may also show inconsistencies in different data sources due to registration errors or delays in changes.

[0004] Existing data cleaning and alignment technologies mostly rely on single-dimensional matching rules, such as exact matching based solely on company names or precise matching based solely on unified social credit codes. These methods fall short when dealing with the unique challenges of the pharmaceutical industry, such as multiple licenses for a single company and inconsistent license information. Simple name matching cannot identify the same entity with subtle name differences, while matching a single license number may fail to link the entity's qualification information under another regulatory system. This technological limitation leads to data redundancy (the same entity is incorrectly identified as two or more independent records) and data silos (the inability to effectively link the same entity's business registration information and drug regulatory qualification information) in data governance systems. This not only increases data management costs but also introduces potential risks in compliance reviews and business decisions.

[0005] Furthermore, data records from upstream systems often suffer from missing key fields, such as administrative division codes and detailed enterprise attributes. Traditionally, manual querying and data entry are relied upon, which is not only inefficient and costly but also makes it difficult to guarantee the accuracy and timeliness of the data.

[0006] Therefore, how to design a data alignment method and system that can effectively integrate multi-source heterogeneous data, accurately identify the unique identity of pharmaceutical entities, and automatically complete key attributes has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] The main objective of this invention is to provide a method and system for aligning pharmaceutical main data based on multimodal feature fusion, in order to solve the technical problems mentioned in the background art, such as low data alignment accuracy and low degree of automation caused by problems such as "one enterprise with multiple certificates" and inconsistencies in information in the pharmaceutical industry.

[0008] To achieve the above objectives, this invention provides a method for aligning pharmaceutical subject data based on multimodal feature fusion, comprising:

[0009] Acquire data records from multiple data sources for a pharmaceutical entity, including the entity name, first identity identifier, second identity identifier, associated person information, address information, and the date of the validity certificate;

[0010] For the two data records to be matched, a core identity feature tuple consisting of the subject name, the first identity identifier, and the second identity identifier is constructed, and a first-level analysis is performed based on the core identity feature tuple;

[0011] When the result of the first-level analysis is a partial mismatch, the associated person information and the address information of the two data records to be matched are extracted, and based on the consistency of the associated person information and the address semantic proximity of the address information, a second-level analysis is performed to determine whether the two data records to be matched point to the same subject.

[0012] For data records determined to point to the same subject, the data record with the latest date is selected as the standard master data for merging and updating based on the date of its validity certificate.

[0013] Furthermore, the present invention also provides a pharmaceutical subject data governance system, comprising:

[0014] The data acquisition module is used to acquire data records from multiple data sources for pharmaceutical entities. The data records include the entity name, first identity identifier, second identity identifier, associated person information, address information, and the date of the validity certificate.

[0015] The alignment analysis module is configured to perform a first-level analysis on two data records to be matched based on a core identity feature tuple composed of the subject name, the first identity identifier, and the second identity identifier. If the result of the first-level analysis is a partial mismatch, a second-level analysis is initiated. The second-level analysis is based on the consistency of the associated person information and the address semantic proximity of the address information to determine whether the two data records to be matched point to the same subject.

[0016] The data fusion module is used to select the latest data as the standard master data for merging and updating based on the date of the validity certificate after the alignment judgment module determines that the data records point to the same subject.

[0017] Furthermore, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0018] Furthermore, the present invention also provides an electronic device, including a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the electronic device enables the electronic device to perform the functions of the above-described system.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0020] 1. This invention constructs a core identity feature tuple consisting of "subject name + dual identity identifiers" and combines it with multimodal features such as legal person and address for two-level in-depth analysis. This effectively addresses the unique problems of "separation of licenses and permits" and information inconsistency in the pharmaceutical industry, significantly improves the accuracy of subject identification, and avoids data duplication and association errors.

[0021] 2. By introducing a knowledge-based attribute auto-completion mechanism, missing key fields can be automatically filled in according to certificate coding rules or associated drug information, which greatly reduces the workload of manual verification and data entry and improves the overall efficiency of data governance.

[0022] 3. This invention introduces the date of time-sensitive certificates as a key basis for data fusion, and can trace enterprise lifecycle events through time-series attribute graphs, ensuring the real-time nature and accuracy of master data, dynamically reflecting the enterprise's certificate change history, and providing reliable data support for compliance management and business decision-making. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the application environment of a pharmaceutical entity data governance system provided in an embodiment of the present invention.

[0025] Figure 2 This is a functional module structure diagram of a pharmaceutical entity data governance system provided in an embodiment of the present invention.

[0026] Figure 3 This is a flowchart of a pharmaceutical subject data alignment method based on multimodal feature fusion provided in an embodiment of the present invention.

[0027] Figure 4 This is a detailed flowchart of the alignment assessment steps in an embodiment of the present invention.

[0028] Figure 5 This is a detailed flowchart of calculating address semantic proximity in an embodiment of the present invention.

[0029] Figure 6 This is a schematic diagram of the timing attribute graph in an embodiment of the present invention.

[0030] Figure 7 This is a flowchart of a conflict resolution method based on a time-series attribute graph in an embodiment of the present invention. Detailed Implementation

[0031] 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.

[0032] It should be noted that the terms "first," "second," etc., in the specification, claims, and drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0033] Example 1

[0034] This embodiment provides a method and system for aligning pharmaceutical subject data based on multimodal feature fusion. For example... Figure 1 As shown, the system can be deployed on a server or distributed computing cluster 100 and interact with various data sources (such as industrial and commercial database 10, drug regulatory database 20, enterprise internal system 30, etc.) and downstream application systems 40 through the network.

[0035] like Figure 2 As shown, the pharmaceutical data governance system 200 provided in this embodiment includes, but is not limited to, a processor 210, a memory 220, and a network interface 230 at the hardware level. The processor 210, as the computing core, executes computer program instructions stored in the memory 220 to implement various system functions. The memory 220 can be used to store raw data, intermediate results, finally generated master data, and knowledge base data. The network interface 230 is used for communication with external data sources and application systems.

[0036] At the logical level, system 200 includes multiple functional modules, which are program units stored in memory 220 and executed by processor 210. These modules include: data acquisition module 201, data preprocessing module 202, alignment and judgment module 203, data fusion module 204, attribute completion module 205, knowledge base module 206, and visualization module 207.

[0037] Combination Figure 3 The specific process of the data alignment method provided in this embodiment will be described in detail.

[0038] Step S301: Data acquisition.

[0039] The data acquisition module 201 is responsible for acquiring data records related to the pharmaceutical entity from multiple heterogeneous data sources. These data sources may include publicly available data from the Administration for Industry and Commerce (providing business license information), databases from the Drug Administration (providing drug distribution or production license information), and internal business systems such as enterprise resource planning or warehouse management systems. In some embodiments, the data acquisition module 201 may also integrate optical character recognition technology to automatically extract key fields from scanned copies or images of certificates and licenses.

[0040] The collected data records should include at least the following fields: subject name, primary identifier, secondary identifier, associated person information, address information, and date of validity certificate.

[0041] It should be noted that although in this embodiment, the first identification identifier is preferably the unified social credit code of the business license, and the second identification identifier is preferably the drug business or production license number, this application is not limited to these. In other embodiments, these identifiers can also be other codes that can uniquely or quasi-uniquely identify the legal or business qualifications of the entity. Similarly, the related party information is preferably the name of the legal representative or person in charge, but it can also be extended to information on other strongly related persons such as major shareholders and senior executives. The date of the validity certificate usually refers to the approval date of the certificate or the most recent change date, which reflects the "freshness" of the data record.

[0042] Step S302: Data preprocessing.

[0043] After acquiring the raw data, the data preprocessing module 202 cleans and standardizes the data records to improve the accuracy of subsequent processing. Specific operations include:

[0044] Clean the main name, for example, by removing common suffixes such as "(chain)" or "branch" or the notes in parentheses, in order to extract the core name.

[0045] Normalize the address information. For example, unify "hao" to "-", convert Chinese numerals to Arabic numerals, and standardize the expression of address elements to prepare for subsequent semantic analysis.

[0046] Step S303: Alignment judgment.

[0047] The alignment judgment module 203 is the core of this method and is responsible for determining whether two data records to be matched point to the same pharmaceutical entity. This process uses a two-level judgment model, and the specific process is referred to Figure 4 .

[0048] S401: Construct the core identity feature tuple and perform the first-level judgment.

[0049] For two data records to be matched (such as record A and record B), the system first constructs a core identity feature tuple consisting of "entity name, first identity identifier, and second identity identifier". This tuple structurally reflects the dual supervision characteristics of the pharmaceutical entity and is the basis for high-precision identification.

[0050] The first-level judgment performs rule matching based on this tuple.

[0051] If the entity names, first identity identifiers, and second identity identifiers of record A and record B are all exactly the same, it is directly determined to point to the same entity (S402).

[0052] If the entity names are the same, but the first identity identifier and the second identity identifier are both not empty and are different, this usually means that they are two homonymous entities independently registered at the industrial and commercial and drug regulatory levels, so it is determined to point to different entities (S403).

[0053] If there are partial mismatches, it enters the second-level judgment (S404). Typical scenarios of partial mismatches include: at least one of the first identity identifier or the second identity identifier is the same, but the entity names are different. This may be due to enterprise name changes or slightly different names used on different licenses.

[0054] S405: Perform the second-level judgment.

[0055] When the first-level judgment cannot directly draw a conclusion, the system starts the second-level judgment and introduces related person information and address information as auxiliary judgment factors.

[0056] The system extracts the related person information (such as the name of the legal representative) of the two records for comparison and calculates the address semantic proximity between the address information of the two records.

[0057] S406: Make a judgment based on the results of the second-level judgment.

[0058] If the information of the related persons is consistent, and the semantic proximity of the addresses is higher than a preset first threshold. (For example, 0.9) In this case, the system determines that the two records point to the same entity (S407). This covers the common situation where a company changes its name but the legal representative and business address remain the same.

[0059] If the information of the related persons is inconsistent, but the semantic proximity of the addresses is higher than a more stringent second preset threshold, then... (For example, 0.95, where) This situation is highly suspicious. It could be two different companies operating at the same address at different times, or it could be that the change of legal representative and the change of company name occurred simultaneously. To ensure accuracy, the system marks this pair of records as "highly suspicious" and transfers them to the manual review process (S408).

[0060] If the above conditions are not met, it is determined that it refers to a different subject (S409).

[0061] Step S304: Data fusion.

[0062] For all data records that are determined by the alignment analysis module 203 to point to the same subject, the data fusion module 204 performs a merge update operation.

[0063] The core rule is to select the most recent data record as the standard master data based on the "date of validity certificate". Information from older records, such as former names, can be retained as historical tags.

[0064] In one specific implementation, when multiple records have the same validity period certificate date, the system will merge all non-conflicting attribute fields (i.e., fields with the same value) in all records, while for conflicting attribute fields (i.e., fields with different values), they will be marked as "pending manual review," and the data administrator will ultimately decide which value to use.

[0065] Step S305: Attribute auto-completion.

[0066] After the subject is aligned and a clean single subject record is generated, the attribute completion module 205 uses the rules and data stored in the knowledge base module 206 to intelligently fill in the missing fields in the record.

[0067] For example, missing province, city, and district / county fields in records can be automatically filled in based on the administrative division codes embedded in the unified social credit code.

[0068] For example, when a data record contains a drug registration approval number, the system can automatically link and complete or verify the entity's manufacturer name, drug dosage form, specifications, and other attributes by querying the drug knowledge base.

[0069] Step S306: Generate a unique master data code and provide services.

[0070] After completing all the above steps, the system generates a globally unique and stable master data ID for each cleaned, aligned, and completed pharmaceutical data entry. This ID can be used by various business systems within the enterprise, ensuring the consistency of the master data across the entire system.

[0071] In addition, the visualization module 207 can display the historical certificate change records, legal person change history, and relationships with other enterprises of a single entity in the form of a timeline or topology diagram based on the processed data, providing intuitive data support for business analysis.

[0072] Through the above embodiments, the method and system of the present invention can effectively solve the pain points in data governance in the pharmaceutical industry and achieve high-precision, automated alignment and quality improvement of pharmaceutical main data.

[0073] Example 2

[0074] Based on Example 1, this embodiment deepens and expands the functions of the alignment judgment module and the data fusion module, and introduces more advanced address semantic analysis technology and a conflict resolution engine based on graph database to handle more complex scenarios and further improve accuracy.

[0075] 1. Address Semantic Fingerprint Generation Based on Hierarchical Attention Mechanism

[0076] In the secondary analysis of Embodiment 1, calculating the address semantic proximity is a crucial step. This embodiment provides a high-precision calculation method, which is implemented by the address semantic analysis unit inside the alignment analysis module 203. The specific process is described in detail below. Figure 5 .

[0077] S501: Extraction of structured elements from address text.

[0078] This step uses a sequence labeling model to process the input unstructured address text and extract structured address elements from it.

[0079] In a preferred embodiment, the sequence labeling model is a bidirectional long short-term memory network combined with a conditional random field model (Bi-LSTM-CRF). This model can effectively capture contextual dependencies in text sequences and accurately identify entity boundaries in addresses. To further improve performance, a pre-trained model based on BERT-wwm-ext can be used as the underlying embedding layer.

[0080] Extracted structured address elements It can be represented as a set, for example: POI refers to a point of interest, such as "XX Building" or "XX Shopping Mall".

[0081] S502: Generate a weighted address feature vector.

[0082] After extracting the structured address elements, the system needs to convert them into a fixed-dimensional numerical vector, i.e., the address feature vector. This process employs a weighted method based on a hierarchical attention mechanism. The core idea is that elements at different levels in an address contribute differently to distinguishing geographical locations; for example, "house number" and "street name" have a much higher distinguishability than "province" and "city".

[0083] Specifically, this computational logic utilizes word vector models (such as the output of Word2Vec or BERT) to divide the address elements at each level ( Convert to a high-dimensional element vector Then, a hierarchical attention mechanism is applied to lower-level address elements (such as...) that can more accurately identify geographic locations. ) assigns a value to higher-level address elements (such as Higher weight Finally, a weighted aggregation of all element vectors is performed to generate the final address feature vector. .

[0084] In a preferred embodiment, the above calculation logic is implemented using the following formula:

[0085]

[0086] in, It is the generated address feature vector, for example, a 256-dimensional vector; It is the first The weighting coefficient of the level address element. Based on experience, it can be set... (Province, city, low differentiation) (district), (Roads, house numbers, high-discrimination core features) (POI).

[0087] S503: Calculate the semantic proximity of addresses.

[0088] For two addresses to be compared, generate their address feature vectors respectively. and Then, the semantic proximity of the address is obtained by calculating the cosine similarity between the two vectors. .

[0089] Specifically, this proximity is calculated using the following formula:

[0090]

[0091] in, Represents the dot product of vectors. and The similarity value represents the magnitude of the vector. The similarity value ranges from -1 to 1. The closer the value is to 1, the more semantically similar the two addresses are.

[0092] In practical applications, a condition can be set to determine if physical addresses overlap, for example, when and If the (address number) elements are completely identical, it is determined that the addresses are highly overlapping.

[0093] This generated high-dimensional address feature vector is resistant to interference from address text writing style. This can be called an "address semantic fingerprint". It can be indexed and stored in an address database for efficient address retrieval and comparison.

[0094] 2. Conflict resolution and fusion based on time-series attribute graphs

[0095] This embodiment upgrades the rule-based data fusion module in Embodiment 1 to a conflict resolution engine based on a time-series attribute graph, in order to more accurately handle complex lifecycle events such as enterprise renaming, splitting, and merging.

[0096] S601: Construction of the time series attribute graph.

[0097] like Figure 6 As shown, the system maintains a dynamic time-series attribute graph in a graph database (such as Neo4j). .

[0098] in, It is a collection of nodes, each node representing a pharmaceutical entity 601.

[0099] It is a set of edges, where each edge represents a relationship between entities. For example, the "holding" relationship connects the entity node 601 and its certificate node 602, the "holding" relationship connects the entity node 601 and its associated person node 603, and the "located" or "registered" relationship connects the entity node 601 and its address node 604.

[0100] The address node 604 is used to characterize the physical spatial attributes of a pharmaceutical entity within a specific time span, including but not limited to registered address, business address, or production address. By making address information an independent attribute node, the system can track the address evolution path of the entity due to relocation or changes in administrative divisions.

[0101] It is a time-sensitive attribute; the nodes and edges in the graph all have a valid time window. This indicates the time period during which the attribute or relationship is valid. For example, a drug business license node will have a start and end date for its validity period, or a residence / office start and end date corresponding to an address node 604.

[0102] This modeling approach constructs a time-series graph of the pharmaceutical entity and all its related information. During secondary analysis, the system not only compares the current address text but also checks the time-series attribute graph to see if two data records have ever pointed to the same historical address node 604, thereby identifying records with "apparent inconsistencies" caused by address changes.

[0103] S701: Merging decision based on fusion probability.

[0104] When it is necessary to determine two main nodes (e.g.) and When deciding whether to merge, the conflict resolution engine calculates the merging probability between them. The process is as follows: Figure 7 As shown.

[0105] Specifically, the fusion probability is a comprehensive score calculated by weighting the similarity across multiple dimensions. In a preferred embodiment, this calculation logic is implemented using the following formula:

[0106]

[0107] in:

[0108] These are the weight coefficients of each item, which can be obtained through training a machine learning model.

[0109] It is the similarity of certificate numbers, which is usually 0 (completely different) or 1 (at least one core certificate number is the same).

[0110] It is the address vector similarity, that is, the cosine similarity of the address semantic fingerprint calculated by the aforementioned method.

[0111] This is the legal entity correlation degree. This value can be determined based on whether the names of the related persons are the same, or by searching in the graph to find whether there is a preset strong correlation between the related persons of the two entities (such as the shortest path distance being less than or equal to 1).

[0112] The overlap of operating time windows is a key adjustment factor.

[0113] S702: Time overlap check.

[0114] The time overlap verification logic is a core innovation of this engine. It is used to accurately distinguish between two easily confused scenarios in the fusion decision: "two companies at the same address at different times" and "the same company changing its name".

[0115] Its specific mode of operation is as follows:

[0116] When the system identifies two entities and Their operating time windows highly overlap (e.g., the overlap period accounts for more than a certain threshold, such as 80%), but their core identity identifiers (such as the unified social credit code) are different. This strongly suggests that they are two independent entities operating simultaneously at the same or nearby addresses. The factor was set to a penalized value (e.g., 0.5) below the baseline value of 1.0, thereby significantly reducing the final fusion probability. .

[0117] Conversely, when a subject The end of the operating window and another entity If the starting points of a business's operating window are close in time or consecutive (e.g., the time interval is less than a preset threshold, such as 90 days), this could be a signal of a business name change or license transfer. The factor is set to a reward value (e.g., 1.2) that is higher than the baseline value of 1.0, thereby increasing the fusion probability and tending to merge the two.

[0118] S703: Execute fusion decisions.

[0119] Calculate the fusion probability Then, the system performs operations based on preset decision thresholds:

[0120] like The system automatically assigns nodes and Merge them and generate a unified master data ID for them.

[0121] like The system marks this pair of nodes as requiring manual review and highlights their difference fields on the interface.

[0122] like The system determines that they are two independent entities.

[0123] Furthermore, this time-series attribute graph can also be used to proactively discover potential relationships. For example, when a node's license is revoked, and a new node with a new license appears at a nearby time and geographical location, the system can automatically establish a candidate "inheritance" or "separation" relationship edge between the two nodes and calculate its confirmation probability for data analysts' reference.

[0124] Example 3

[0125] This embodiment provides alternative implementation methods or enhanced functions for some modules of the system.

[0126] In one embodiment, the system further includes a threshold adaptive adjustment module. This module's function is to dynamically optimize the first preset threshold in the secondary assessment based on feedback from manual review. Second preset threshold Specifically, each time a human reviewer makes a final judgment on a pair of records marked as "highly suspected" by the system—whether it is "the same entity" or "not the same entity"—this judgment result is fed back as a labeled training sample to the threshold adaptive adjustment module. This module can internally run a simple machine learning model (such as logistic regression) to periodically or in real-time recalculate the optimal threshold based on accumulated feedback data. This allows the system's judgment boundary to increasingly approximate the real-world business scenario, achieving self-optimization.

[0127] In another embodiment, the system can be deployed in a distributed computing cluster based on Spark or Hadoop. This is essential for processing massive (e.g., hundreds of millions) medical subject data records. Tasks such as data acquisition, preprocessing, feature extraction, and graph computation can all be designed as parallel computing jobs, executed efficiently on the cluster.

[0128] It should be noted that the address semantic fingerprint generation model mentioned in this embodiment is not limited to Bi-LSTM-CRF; other advanced natural language processing models, such as those based on the Transformer architecture, can also be used, as long as they can accurately extract structured elements from address text. Similarly, the graph database is not limited to Neo4j; any graph database that supports attribute graphs and time-series queries can be applied.

[0129] Those skilled in the art will understand that the pharmaceutical subject data alignment method based on multimodal feature fusion provided in the embodiments of the present invention can be implemented, in whole or in part, by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the foregoing method embodiments. Accordingly, the embodiments of the present invention also provide a computer-readable storage medium storing a computer program implementing the foregoing method embodiments.

[0130] Those skilled in the art will understand that the pharmaceutical data governance system provided in this embodiment of the invention can be implemented using corresponding hardware structures. Accordingly, this embodiment of the invention also provides an electronic device, which includes a processor and a memory. The memory stores a computer program, and when the computer program is executed by the processor, it can implement the functions of the aforementioned system embodiments.

[0131] In summary, this invention constructs a highly accurate and automated pharmaceutical entity data alignment and governance solution by combining multi-dimensional feature binding and a two-level deep analysis model with advanced address semantic fingerprinting technology and a temporal attribute graph conflict resolution engine. It not only effectively solves the challenges faced by existing technologies but also dynamically traces the enterprise lifecycle, providing strong technical support for the digital management and compliance monitoring of the pharmaceutical industry.

[0132] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A medical subject data alignment method based on multi-modal feature fusion, characterized in that, include: Acquire data records from multiple data sources for a pharmaceutical entity, including the entity name, first identity identifier, second identity identifier, associated person information, address information, and the date of the validity certificate; For the two data records to be matched, a core identity feature tuple consisting of the subject name, the first identity identifier, and the second identity identifier is constructed, and a first-level analysis is performed based on the core identity feature tuple; When the result of the first-level analysis is a partial mismatch, the associated person information and the address information of the two data records to be matched are extracted, and based on the consistency of the associated person information and the address semantic proximity of the address information, a second-level analysis is performed to determine whether the two data records to be matched point to the same subject. For data records determined to point to the same subject, the data record with the latest date is selected as the standard master data for merging and updating based on the date of its validity certificate.

2. The method of claim 1, wherein, The first identity identifier is the unified social credit code of the business license, and the second identity identifier is the drug business or production license number.

3. The method of claim 1, wherein, The conditions for partial mismatch in the first-level analysis include: at least one of the first identity identifier or the second identity identifier of the two data records to be matched is the same, but the subject names are different.

4. The method according to claim 1, characterized in that, In the secondary judgment, when the associated person information is consistent and the address semantic proximity is higher than a first preset threshold , it is determined that the two data records to be matched point to the same subject.

5. The method of claim 1, wherein, The first-level analysis also includes: if the subject names of the two data records to be matched are the same, but the first identity identifier and the second identity identifier are both not empty and are both different, then it is determined that the two data records to be matched point to different subjects.

6. The method according to claim 5, characterized in that, The secondary judgment further includes: when the associated person information is inconsistent, but the address semantic proximity is higher than a second preset threshold , marking the two data records to be matched as highly suspected and transferring into an artificial review process; wherein, .

7. The method of claim 1, wherein, Following the acquisition step, the data record is further preprocessed. The preprocessing includes cleaning the subject name to remove common suffixes and remarks; and normalizing the address information to unify the expression of address elements.

8. The method of claim 1, wherein, The method further includes an attribute auto-completion step based on a knowledge base; after the data record is determined to point to the same subject, the missing administrative division field in the data record is automatically filled in according to the administrative division code contained in the first identity identifier or the second identity identifier.

9. The method of claim 8, wherein, The automatic attribute completion step further includes: when the data record is associated with a drug registration approval number, automatically completing or verifying the production enterprise attribute of the subject by querying the drug knowledge base.

10. The method of claim 1, wherein, The step of calculating the semantic proximity of the address in the secondary judgment includes: using a sequence labeling model to extract structured address elements from the address text of the address information.

11. The method of claim 10, wherein, The sequence labeling model is a bidirectional long short-term memory network combined with a conditional random field model. The structured address elements include administrative division features such as province, city, district, street, and house number, as well as point of interest features.

12. The method of claim 11, wherein, The step of calculating the address semantic proximity further includes: using a word vector model to convert the address elements at all levels in the structured address elements into high-dimensional element vectors; applying a hierarchical attention mechanism to assign higher weights to low-level address elements that can more accurately identify geographical locations, the low-level address elements including streets, house numbers, and points of interest, compared to high-level address elements, and performing weighted aggregation on the high-dimensional element vectors to generate a fixed-dimensional address feature vector; and obtaining the address semantic proximity by calculating the cosine similarity between the address feature vectors of two data records.

13. The method according to claim 12, characterized in that, The generated address feature vector serves as an address semantic fingerprint, which is used for efficient retrieval and comparison in the address database.

14. The method according to claim 1, characterized in that, The data merging and updating step is replaced by a conflict resolution method based on a time-series attribute graph. This method includes: modeling the pharmaceutical entities as nodes in a graph database, and modeling the entity names, identifiers, and other attributes, as well as the relationships between them, as edges with time-series attribute attributes to form a time-series attribute graph.

15. The method according to claim 14, characterized in that, The conflict resolution method further includes: making a merging decision by calculating the merging probability between two nodes in the time-series attribute graph; if the merging probability is greater than a preset merging threshold, then merging the two nodes.

16. The method according to claim 15, characterized in that, The fusion probability is a comprehensive score calculated by weighting the similarity of licenses and permits, address vector similarity, legal entity relevance, and overlap of operating time windows.

17. The method according to claim 16, characterized in that, The overlap of the operating time windows is used as an adjustment factor. Specifically, when it is determined that the operating time windows of two entities have a high degree of overlap, that is, the proportion of the overlap time to the shorter time window is greater than the first time proportion threshold, but their core identity identifiers are different, the overlap of the operating time windows is set to a punitive value lower than the baseline value of 1.

0. When the end point of the operating time window of one entity is close to or continuous with the start point of the operating time window of another entity in time, that is, the time interval is less than the preset time difference threshold, the overlap of the operating time windows is set to a reward value higher than the baseline value of 1.0, so as to distinguish between the scenarios of "operating at the same address in succession" and "changing the name or relocating the enterprise".

18. The method according to claim 14, characterized in that, The time-series attribute graph is used to record and trace lifecycle events of pharmaceutical entities, such as name changes, legal entity changes, address relocations, mergers, or spin-offs.

19. The method according to claim 14, characterized in that, The method further includes: when the license of a node in the time-series attribute graph is revoked, and a new node obtains a new license at a nearby time point and geographical location, establishing a candidate relationship edge of "inheritance" or "separation" between the two nodes, and calculating its confirmation probability.

20. The method according to claim 14, characterized in that, The legal entity correlation degree is determined based on whether the names of the related persons are the same, or whether there is a pre-set strong correlation in the known corporate executive correlation network.

21. The method according to claim 1, characterized in that, In the data merging and updating step, when multiple data records that are determined to point to the same subject have conflicting fields, priority overwriting is performed based on the authority weight of the data source.

22. A computer program product comprising computer instructions that, when executed by a processor, implement the method of any one of claims 1-21.

23. An electronic device comprising a processor and a memory, wherein the memory stores a computer program, which, when executed by the processor, causes the electronic device to perform the method of any one of claims 1-21.

24. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method of any one of claims 1-21.

Images