Cloud-native medical business platform construction method, device, equipment and storage medium

Abstract

The present application discloses a method, apparatus, equipment and storage medium for constructing a cloud-native medical business middle platform, the method comprising: obtaining multimodal medical data, the multimodal medical data comprising: structured data, unstructured data and graph data; inputting the multimodal medical data into a pre-trained intelligent classifier, the intelligent classifier being used to perform entity recognition on the unstructured data, mapping the recognized entities to the codes in the structured data, and generating dynamic medical information tags; constructing a microservice medical knowledge base based on the dynamic medical information tags, integrating the microservice medical knowledge base with a cloud-native business interface; connecting the cloud-native business interface and the distributed medical database through an interactive interface, thereby effectively solving the integration problem caused by the heterogeneity of medical data.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method, apparatus, equipment and storage medium for constructing a cloud-native medical business middle platform. Background Art

[0002] In the development of medical informatization, traditional medical systems face significant data silos. Business systems such as hospital information systems, testing systems, and imaging systems operate independently, using different technical architectures and data standards. This makes it difficult to interoperate with various types of medical data, including structured medical records, unstructured imaging reports, and diagnosis-related data. When doctors provide treatment across departments, they must switch between multiple systems to query patient information, preventing them from quickly obtaining complete diagnosis and treatment data. When patients seek treatment across different campuses, data from different hospitals cannot be synchronized in real time, leading to duplicate examinations and missed medical histories. This data isolation not only reduces diagnosis and treatment efficiency and increases the risk of medical errors, but also hinders the in-depth application of medical data in scenarios such as scientific research and analysis and regional medical collaboration, making it difficult to meet the data integration needs of modern healthcare.

[0003] Therefore, there is an urgent need for a cloud-native medical business middle platform construction method that can achieve data integration. Summary of the Invention

[0004] The main technical problem solved by this application is to provide a cloud-native medical business middle platform construction method, device, equipment and storage medium, which can achieve efficient integration of different data.

[0005] To solve the above technical problems, a technical solution adopted in this application is: to provide a method for constructing a cloud-native medical business middle platform, the method including: obtaining multimodal medical data, the multimodal medical data including: structured data, unstructured data and graph data; inputting the multimodal medical data into a pre-trained intelligent classifier, the intelligent classifier is used to perform entity recognition on the unstructured data, mapping the identified entities with the codes in the structured data, and generating dynamic medical information tags; constructing a microservice medical knowledge base based on the dynamic medical information tags, integrating the microservice medical knowledge base with the cloud-native business interface; connecting the cloud-native business interface and the distributed medical database through an interactive interface.

[0006] Among them, generating dynamic medical information labels further includes: converting the diagnostic codes in structured data into one-hot encoding vectors; performing word segmentation on unstructured text data in unstructured data and extracting semantic vectors; fusing the one-hot encoding vectors and semantic vectors at the splicing layer to generate a joint feature vector; mapping the semantic relationships in the medical ontology to the local terminology of the target medical institution; analyzing the time series of patient diagnosis and treatment events and dynamically adjusting the label weights.

[0007] Among them, the one-hot encoded vector and the semantic vector are fused at the splicing layer to generate a joint feature vector, including: calculating the correlation weights of structured fields in structured data and unstructured text features in unstructured data through a multi-head attention mechanism.

[0008] Among them, the multi-head attention mechanism is also used to: map the one-hot encoded vector of the structured field to the query vector through the fully connected layer; map the semantic vector of the unstructured text to the key vector and value vector respectively through two independent fully connected layers; and mask the structured fields or unstructured text features whose relevance weight is lower than the threshold.

[0009] Among them, obtaining multimodal medical data includes: obtaining multimodal medical data from medical institution information systems, image archiving systems and drug databases; and / or, structured data includes: diagnosis codes, drug codes and examination item codes in patient electronic files; and / or, unstructured data includes: medical imaging text reports and scanned copies of doctors' handwritten medical records; and / or, graph data includes: indication-contraindication relationships and drug interaction relationships in drug knowledge graphs.

[0010] Among them, the cloud-native business interface and the distributed medical database are connected through the interactive interface, including: encapsulating the labeled data into an independent microservice module, generating the corresponding interactive interface description file, and injecting it into the gateway.

[0011] Among them, the relevance weights of structured fields in structured data and unstructured text features in unstructured data are calculated through a multi-head attention mechanism, including: generating a weight matrix after performing a dot product operation on the query vector and the key vector.

[0012] In order to solve the above technical problems, another technical solution adopted in this application is: to provide a cloud-native medical business middle-platform construction device, which includes: an acquisition module, a label generation module, an integration module and an interaction module; the acquisition module: acquires multimodal medical data, and the multimodal medical data includes: structured data, unstructured data and graph data; the label generation module: inputs the multimodal medical data into a pre-trained intelligent classifier, and the intelligent classifier is used to perform entity recognition on the unstructured data, and map the identified entities with the codes in the structured data to generate dynamic medical information labels; the integration module: builds a microservice medical knowledge base based on the dynamic medical information labels, and integrates the microservice medical knowledge base with the cloud-native business interface; the interaction module: connects the cloud-native business interface and the distributed medical database through an interactive interface.

[0013] To solve the above technical problems, another technical solution adopted in this application is: to provide a cloud-native medical business middle-office construction device, which includes: a memory and at least one processor, the memory stores instructions; at least one processor calls the instructions in the memory to enable the cloud-native medical business middle-office construction device to execute the steps of the cloud-native medical business middle-office construction method such as any one of the above-mentioned items.

[0014] In order to solve the above technical problems, another technical solution adopted in this application is: providing a computer-readable storage medium, on which instructions are stored, and when the instructions are processed and executed, the steps of the cloud-native medical business middle platform construction method as described in any of the above items are implemented.

[0015] Different from the existing technology, the beneficial effects of this application are: by integrating multimodal medical data and using pre-trained intelligent classifiers to achieve entity mapping of structured and unstructured data, dynamic medical information labels are generated, and the integration problems caused by the heterogeneity of medical data are effectively solved. The microservice knowledge base built based on labels combined with the cloud-native architecture realizes the efficient organization and elastic expansion of medical knowledge, while connecting the distributed database and business interface through an interactive interface to ensure data real-time and service response speed. The medical middle-end system finally constructed not only improves the semantic correlation accuracy of cross-modal data, but also enhances the maintainability and business scenario adaptability of the platform through modular design, providing medical institutions with an intelligent and highly available data support platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of an implementation method of the cloud-native medical business middle platform construction method of this application.

[0017] Figure 2 This is a structural framework diagram of an implementation method of a cloud-native medical business middle platform construction device of the present application.

[0018] Figure 3 This is a structural framework diagram of an implementation method of building a cloud-native medical business middle platform device in this application. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end that includes a series of steps or units is not limited to the listed steps or units, but optionally includes other steps or units that are not listed, or optionally includes other steps or units that are inherent to these processes, methods, products or device ends. The term "and / or" in this article is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the related objects before and after are in an "or" relationship. In addition, "many" in this article means two or more than two.

[0021] To facilitate understanding of this embodiment, we first introduce in detail a method for constructing a cloud-native medical service platform disclosed in an embodiment of the present invention. Figure 1 As shown, Figure 1 This is a flow chart of an implementation method of the cloud-native medical business middle platform construction method of this application, which includes the following steps.

[0022] Step S11: Acquire multimodal medical data, which includes structured data, unstructured data, and atlas data.

[0023] Specifically, it acquires structured, unstructured, and graph data from multiple source systems, covering multiple medical scenario data types.

[0024] Acquiring multimodal medical data includes obtaining multimodal medical data from medical institution information systems, image archiving systems, and drug databases. Structured data includes standardized data stored in coded form, supporting rapid retrieval and statistics. Examples include coded data such as ICD-10 (International Classification of Diseases, Tenth Revision) and ATC (Anatomical Therapeutic Chemical). In some application scenarios, structured data includes diagnosis codes, drug codes, and examination item codes in patient electronic records. In other application scenarios, structured data can also include laboratory test results coded using LOINC (Logical Observation Identifiers Names and Codes) and pathology test results coded using SNOMED CT (Systematized Nomenclature of Medicine - Clinical Terms).

[0025] Unstructured data includes non-standardized data such as text, images, and audio, requiring semantic analysis to extract key information. Examples include text reports of medical digital imaging and communication structured reporting formats and scans of doctors' handwritten medical records. In some application scenarios, unstructured data includes text reports of medical imaging and scans of doctors' handwritten medical records. In other application scenarios, unstructured data may also include dynamic ultrasound image sequences and voice recordings of remote medical consultations transcribed using automatic speech recognition.

[0026] Graph data includes inter-entity relationship networks, which are used to mine potential medical associations. Specifically, these include drug interactions, such as indication-contraindication triples within the drug knowledge graph. In some application scenarios, graph data includes indication-contraindication relationships and drug-drug interaction relationships within the drug knowledge graph. Furthermore, in other application scenarios, graph data can be expanded to include gene-disease association maps and microbiome-based drug resistance transmission networks.

[0027] The above solution breaks down the problem of traditional medical data silos by integrating data from heterogeneous systems, such as medical institution information systems, image archiving, and communication systems. In some specific embodiments, obtaining structured data from multimodal medical data can involve obtaining a patient's ICD-10 code "I25.1," which indicates a previous myocardial infarction, and the ATC code "C01CA05," which indicates aspirin, from medical institution information and communication systems. Obtaining unstructured data from multimodal medical data can involve obtaining a structured medical digital imaging and communication report of a coronary artery from an image archiving and communication system, such as a text description of "calcified plaque in the left anterior descending artery, 70% lumen stenosis," and a handwritten doctor's ward round note recognized using optical character recognition. Obtaining atlas data from multimodal medical data can involve obtaining the relationship triple "aspirin - bleeding risk - contraindication for combined use with warfarin" from a drug database, where the Anatomical Therapeutic Chemical Classification system code for warfarin is "B01AA02."

[0028] Furthermore, for emerging medical data, such as intestinal flora testing data and EEG monitoring data, access capabilities can be expanded through custom data adapters to support the parsing of structured data in formats such as JSON and XML, as well as unstructured data in binary formats. This maintains the openness of the system, adapts to the diversified development of medical data, and extends the life cycle of the middle platform.

[0029] Step S12: Input the multimodal medical data into a pre-trained intelligent classifier, which is used to perform entity recognition on unstructured data, map the recognized entities with the codes in the structured data, and generate dynamic medical information tags.

[0030] Specifically, intelligent classifiers are used to identify entities in unstructured data and map codes to structured data, generating medical tags with time-series dynamic weights. A medical entity is a uniquely identifiable, clinically significant discrete object or concept, the basic unit of medical data. These entities can be concrete, such as patients and medications, or abstract, such as diseases and symptoms. Accurately identifying and managing medical entities is fundamental to the structured and intelligent application of medical data. Entity recognition involves extracting medical entities, such as diseases, medications, and anatomical parts, from unstructured text using pre-trained natural language processing models. Code mapping involves establishing an association between medical entities and codes, such as mapping "myocardial infarction" to "ICD-10:I21."

[0031] In some embodiments, generating dynamic medical information labels further includes: converting diagnostic codes in structured data into one-hot encoded vectors; performing word segmentation on unstructured text data in unstructured data to extract semantic vectors; fusing the one-hot encoded vectors and semantic vectors at the splicing layer to generate a joint feature vector; mapping the semantic relationships in the medical ontology to the local terminology of the target medical institution; analyzing the time series of patient diagnosis and treatment events and dynamically adjusting label weights.

[0032] One-hot encoding is a method for converting discrete features into vector representations. In medical scenarios, diagnostic codes are discrete, such as ICD-10 codes. Assuming there are 1,000 different ICD-10 codes, each code can be represented by a vector of length 1,000, where only the element corresponding to the code position is 1, and all other elements are 0. For example, a patient's diagnosis code is "ICD-10:E11.9," which refers to unspecified type 2 diabetes. In a system with 1,000 codes, "E11.9" corresponds to the 300th position. Therefore, its one-hot encoding vector is a vector of length 1,000, where the 300th element is 1 and the remaining 999 elements are 0. This approach has the advantage of converting diagnostic codes into a numerical form that computers can process, facilitating subsequent calculations by machine learning models.

[0033] Unstructured text data, such as medical records and examination reports, consists of natural language. To make these texts understandable to computers, they must first be segmented. Segmentation involves breaking down continuous text into meaningful words or phrases. For example, "The patient reported having cough and fever symptoms over the past week" can be segmented into "patient," "reported," "had," "cough," "fever," and "symptoms" (symptoms). After segmentation, these words need to be converted into vector representations, or semantic vectors. For example, the word "cough" would be converted into a vector of a specific dimension, which contains the semantic information of "cough" in the medical field.

[0034] Diagnostic codes in structured data and textual information in unstructured data both contain important patient information, but they are represented differently. Fusion of one-hot encoded vectors and semantic vectors combines this information to form a more comprehensive and richer feature representation, known as a joint feature vector. The concatenation layer concatenates the one-hot encoded vectors in a preset order. If the one-hot encoded vector has a length of 1000 and the semantic vector has a length of 768, the length of the concatenated joint feature vector is 1768.

[0035] Different medical institutions may use different terms to describe the same medical concepts, leading to data inconsistency. Mapping the semantic relationships in a medical ontology to the target medical institution's local terminology can eliminate this inconsistency and enable better integration and sharing of data across different medical institutions. For example, the concept of "diabetes" in a medical ontology can be mapped to "xiaoke disease" in a medical institution's local terminology, allowing them to be treated as the same concept during data analysis.

[0036] A patient's condition changes over time, and diagnosis and treatment information at different points in time has varying importance for determining the patient's current status and prognosis. Therefore, it's necessary to analyze the time series of a patient's diagnosis and treatment events to understand the progression of their condition. Based on the results of this time series analysis, the weight of a patient's medical label can be dynamically adjusted. For example, a patient is initially diagnosed with "Stage 1 Hypertension" with a label weight of 0.8. Over time, if the patient develops new symptoms, such as "Left Ventricular Hypertrophy," and analysis of the time series data reveals a worsening trend, the label can be adjusted to "Hypertension - Left Ventricular Hypertrophy - Extremely High Risk," and the label weight increased to 0.95. This allows the medical label to more accurately reflect the patient's current condition, providing a more reliable basis for clinical decision-making.

[0037] In some embodiments, the one-hot encoded vector and the semantic vector are fused at the concatenation layer to generate a joint feature vector, including: calculating the correlation weights of structured fields in structured data and unstructured text features in unstructured data through a multi-head attention mechanism.

[0038] Specifically, the one-hot encoded vector of the structured field is first mapped to a query vector through a fully connected layer, and the semantic vector of the unstructured text is mapped to a key vector and a value vector respectively through two independent fully connected layers. Then, a dot product operation is performed on the query vector and the key vector, and the result of the operation is divided by the square root of the key vector dimension to scale the numerical range. The result is then converted into a correlation weight in the form of a probability distribution through the softmax function. This weight reflects the degree of association between the structured field and the unstructured text features.

[0039] Among them, the multi-head attention mechanism divides the query vector, key vector and value vector into multiple independent subspaces, namely "heads", for parallel calculation. Each subspace focuses on capturing semantic relationships of different dimensions, such as the association between diseases and symptoms, the association between drugs and indications, etc., and then the calculation results of each subspace are spliced ​​into a complete joint feature vector.

[0040] In some embodiments, the multi-head attention mechanism is also used to: map the one-hot encoded vector of the structured field to a query vector through a fully connected layer; map the semantic vector of the unstructured text to a key vector and a value vector through two independent fully connected layers respectively; and mask the structured fields or unstructured text features whose relevance weights are lower than a threshold.

[0041] Specifically, a linear transformation in a fully connected layer converts the sparse representation of the one-hot encoded vector into a query vector suitable for semantic matching. The semantic vector is then mapped into a key vector for calculating relevance and a value vector for extracting feature information. After the mapping is complete, the calculated relevance weights are evaluated. When the weights of certain structured fields and unstructured text features fall below a preset threshold, the corresponding feature pairs are marked as invalid and ignored in subsequent calculations.

[0042] In some embodiments, the relevance weights of structured fields in structured data and unstructured text features in unstructured data are calculated through a multi-head attention mechanism, including: generating a weight matrix after performing a dot product operation on the query vector and the key vector.

[0043] Specifically, a dot product operation is performed on the query vector and the key vector to obtain a raw score matrix reflecting the semantic similarity between the two. The score is then scaled by dividing it by the square root of the key vector dimension to prevent gradient instability caused by excessively large values. The score matrix is ​​then converted into a weight matrix in the form of a probability distribution using the softmax function, where each element represents the correlation weight between the corresponding structured field and the unstructured text feature. Each row vector of the weight matrix represents the degree of attention a structured field pays to all unstructured text features, while the column vector represents the strength of the association between an unstructured text feature and all structured fields. This matrix enables dynamic weighted fusion of cross-modal features.

[0044] Step S13: Build a microservice medical knowledge base based on dynamic medical information tags, and integrate the microservice medical knowledge base with the cloud-native business interface.

[0045] Specifically, dynamic medical information tags are structured and stored and service-encapsulated to form independently deployable and elastically scalable microservice modules, which are then integrated with cloud-native business interfaces through standardized interfaces to enable real-time calling and display of tag data.

[0046] In some specific embodiments, the Spring Boot framework is used to split tagged data into independent microservice units based on business scenarios, such as patient tag services, drug indication services, and examination item association services. Each microservice module encapsulates tagged data for a specific domain and exposes data access capabilities through declarative APIs, such as RESTful interfaces. Swagger UI is used to generate visual interface documentation for easy access by front-end developers.

[0047] In some embodiments, the integration of a microservices-based medical knowledge base and a cloud-native business interface uses an Nginx gateway for load balancing and traffic control. The cloud-native interface, developed with front-end frameworks like Vue.js, uses tools like Axios to send HTTP requests to the microservice interface. For example, the electronic medical record interface calls the "Patient Diagnosis Tags" microservice to retrieve the patient's latest condition tags in real time and render them on the interface.

[0048] Step S14: Connect the cloud native business interface and the distributed medical database through the interactive interface.

[0049] Among them, the interactive interface undertakes the functions of data transmission and protocol conversion, supports query, write and other operations between the cloud native interface and the distributed database, and ensures the efficiency and security of data interaction.

[0050] In some specific embodiments, the GraphQL protocol (an API query language) is used to implement dynamic data queries, dynamically generating query statements based on the needs of the front-end interface and reducing redundant data transmission. For example, when a doctor queries a patient's blood sugar test results for the past three months on the outpatient interface, the GraphQL interface only returns data related to the query fields (such as test values ​​and dates), avoiding the transmission of irrelevant fields.

[0051] Specifically, the workflow of the interactive interface includes three stages: query parsing, execution plan generation, and result return. First, the front-end request is parsed into an abstract syntax tree and the required fields are extracted. Then, the execution strategy is selected based on the query complexity. Simple queries directly access the remote dictionary service cache database, while complex queries generate parallel query plans through the query optimizer and access the distributed database. Finally, the query results are serialized and returned to the front-end interface.

[0052] For example, when querying "Glycated hemoglobin trends in diabetic patients", the interface will query the distributed stored patient tag data in parallel, use the stream processing framework to calculate the group aggregation results in real time, and return structured data within 100ms.

[0053] In some embodiments, the cloud-native business interface and the distributed medical database are connected through an interactive interface, including: encapsulating the labeled data into an independent microservice module, generating a corresponding interactive interface description file, and injecting it into the gateway.

[0054] Specifically, each microservice module generates an interface description file, such as a YAML file, according to the OpenAPI 3.0 specification. This file defines the interface path, request method, parameter type, and response format. These description files are injected into the Nginx gateway through automated tools, which then centrally manages interface routing, permission verification, and traffic restrictions.

[0055] In some specific embodiments, the interactive interface supports multi-tenant isolation, allocating independent database connection pools and API call quotas to different medical institutions through K8s (Kubernetes, a container orchestration platform). For example, high-level medical institution tenants can access full diagnostic code tags, while low-level medical institution tenants can only obtain simplified local terminology tags, meeting data tiered authorization requirements.

[0056] In the above scheme, this application solves the problem of integrating multi-source heterogeneous data by integrating multimodal medical data from medical institution information systems, image archiving systems and drug databases, including structured diagnostic codes, drug codes, examination item codes, unstructured medical imaging text reports, scanned copies of doctors' handwritten medical records, and indication-contraindication relationships and drug interaction relationships in drug knowledge graphs. Pre-trained intelligent classifiers are used to perform entity recognition on unstructured texts, and structured diagnostic codes are converted into one-hot encoding vectors, which are then spliced ​​and fused with semantic vectors generated by unstructured text segmentation. The multi-head attention mechanism is used to calculate the correlation weights between structured fields and unstructured text features, and masking is performed on features below the threshold to achieve accurate semantic matching of cross-modal data and generate dynamic medical information labels. A standardized medical concept system is established by mapping the semantic relationships of medical ontology with the local terminology table, and the label weights are dynamically adjusted in combination with the analysis of patient diagnosis and treatment time series to enhance the time sensitivity of clinical data. The labeled data is encapsulated into independent microservice modules and a standardized interface description file is generated. Through the gateway, the microservice medical knowledge base is interconnected with the cloud-native business interface and the distributed medical database to improve the system scalability and service response efficiency.

[0057] See also Figure 2 , Figure 2 This is a schematic diagram of the structural framework of an embodiment of the cloud native medical business platform construction device of this application. Figure 2As shown, the cloud-native medical business middle platform construction device 20 includes: an acquisition module 21, a label generation module 22, an integration module 23 and an interaction module 24; the acquisition module 21: acquires multimodal medical data, and the multimodal medical data includes: structured data, unstructured data and graph data; the label generation module 22: inputs the multimodal medical data into a pre-trained intelligent classifier, and the intelligent classifier is used to perform entity recognition on the unstructured data, and map the recognized entities with the codes in the structured data to generate dynamic medical information labels; the integration module 23: constructs a microservice medical knowledge base based on the dynamic medical information labels, and integrates the microservice medical knowledge base with the cloud-native business interface; the interaction module 24: connects the cloud-native business interface and the distributed medical database through an interactive interface.

[0058] In some embodiments, the label generation module 22 generates dynamic medical information labels further including: converting the diagnosis codes in the structured data into one-hot encoding vectors; performing word segmentation on the unstructured text data in the unstructured data and extracting semantic vectors; fusing the one-hot encoding vectors and the semantic vectors at the splicing layer to generate a joint feature vector; mapping the semantic relationships in the medical ontology to the local terminology of the target medical institution; analyzing the time series of patient diagnosis and treatment events and dynamically adjusting the label weights.

[0059] In some embodiments, the label generation module 22 fuses the one-hot encoding vector and the semantic vector at the splicing layer to generate a joint feature vector, including: calculating the correlation weights of structured fields in structured data and unstructured text features in unstructured data through a multi-head attention mechanism.

[0060] In some embodiments, the multi-head attention mechanism of the label generation module 22 is also used to: map the one-hot encoded vector of the structured field to a query vector through a fully connected layer; map the semantic vector of the unstructured text to a key vector and a value vector through two independent fully connected layers respectively; and mask the structured fields or unstructured text features whose relevance weights are lower than a threshold.

[0061] In some embodiments, the acquisition module 21 acquires multimodal medical data including: acquiring multimodal medical data from medical institution information systems, image archiving systems and drug databases; and / or, structured data including: diagnosis codes, drug codes and examination item codes in patient electronic files; and / or, unstructured data including: medical imaging text reports and scanned copies of doctors' handwritten medical records; and / or, graph data including: indication-contraindication relationships and drug interaction relationships in the drug knowledge graph.

[0062] In some embodiments, the interaction module 24 connects the cloud-native business interface and the distributed medical database through an interaction interface, including: encapsulating the labeled data into an independent microservice module, generating a corresponding interaction interface description file, and injecting it into the gateway.

[0063] In some embodiments, the label generation module 22 calculates the relevance weights of structured fields in structured data and unstructured text features in unstructured data through a multi-head attention mechanism, including: generating a weight matrix after performing a dot product operation on the query vector and the key vector.

[0064] The above solution solves the problem of integrating multi-source heterogeneous data by integrating multimodal medical data from medical institution information systems, image archiving systems, and drug databases, including structured diagnostic codes, drug codes, examination item codes, unstructured medical imaging text reports, scanned copies of doctors' handwritten medical records, and indication-contraindication relationships and drug interaction relationships in the drug knowledge graph. Pre-trained intelligent classifiers are used to perform entity recognition on unstructured text. By converting structured diagnostic codes into one-hot encoding vectors, they are concatenated and fused with the semantic vectors generated by unstructured text word segmentation. A multi-head attention mechanism is used to calculate the correlation weights between structured fields and unstructured text features. Features below the threshold are masked to achieve accurate semantic matching of cross-modal data and generate dynamic medical information labels. A standardized medical concept system is established by mapping medical ontology semantic relationships with local terminology. Label weights are dynamically adjusted in combination with patient diagnosis and treatment time series analysis to enhance the time sensitivity of clinical data. The labeled data is encapsulated into independent microservice modules and a standardized interface description file is generated. Through the gateway, the microservice medical knowledge base is interconnected with the cloud-native business interface and the distributed medical database to improve the system scalability and service response efficiency.

[0065] Figure 3 This is a schematic diagram of the structural framework of an embodiment of the cloud-native medical service middle platform construction device of the present application. The cloud-native medical service middle platform construction device 30 may have relatively large differences due to different configurations or performances, and may include one or more processors 31 and memory 32. The processor 31 may be configured to communicate with the memory 32, and execute a series of instruction operations in the memory on the cloud-native medical service middle platform construction device to implement the steps of the above-mentioned cloud-native medical service middle platform construction method. Those skilled in the art will understand that Figure 3 The cloud-native medical business middle platform construction device structure shown does not constitute a limitation on the cloud-native medical business middle platform construction device provided by the present invention. It may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0066] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions, and when the instructions are run on a computer, the computer executes the steps of the cloud-native medical business middle platform construction method.

[0067] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, devices, and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0068] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0069] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a cloud-native medical service middle platform, characterized in that: The method includes: acquiring multimodal medical data, the multimodal medical data including structured data, unstructured data, and graph data; inputting the multimodal medical data into a pre-trained intelligent classifier, the intelligent classifier being configured to perform entity recognition on the unstructured data, mapping the recognized entities to codes in the structured data, and generating dynamic medical information tags; constructing a microservice-based medical knowledge base based on the dynamic medical information tags, integrating the microservice-based medical knowledge base with a cloud-native business interface; and connecting the cloud-native business interface and a distributed medical database via an interactive interface. The acquisition of multimodal medical data includes: acquiring multimodal medical data from medical institution information systems, image archiving systems, and drug databases; the structured data includes: diagnosis codes, drug codes, and examination item codes in patient electronic files; the unstructured data includes: medical imaging text reports and scanned copies of doctors' handwritten medical records; the graph data includes: indication-contraindication relationships and drug interaction relationships in the drug knowledge graph; The generation of dynamic medical information labels includes: converting the diagnosis codes in the structured data into one-hot encoding vectors; performing word segmentation processing on the unstructured text data in the unstructured data to extract semantic vectors; fusing the one-hot encoding vectors and the semantic vectors at the splicing layer to generate a joint feature vector; mapping the semantic relationships in the medical ontology to the local terminology of the target medical institution; analyzing the time series of patient diagnosis and treatment events and dynamically adjusting the label weights.

2. The cloud-native medical service platform construction method according to claim 1 is characterized in that: The one-hot encoding vector and the semantic vector are fused at a concatenation layer to generate a joint feature vector, including: calculating the correlation weights of structured fields in structured data and unstructured text features in unstructured data through a multi-head attention mechanism.

3. The cloud-native medical service platform construction method according to claim 2 is characterized in that: The multi-head attention mechanism is also used to: map the one-hot encoded vector of the structured field into a query vector through a fully connected layer; map the semantic vector of the unstructured text into a key vector and a value vector through two independent fully connected layers respectively; and mask the structured fields or unstructured text features whose relevance weights are lower than a threshold.

4. The method for constructing a cloud-native medical service platform according to any one of claim 3, characterized in that: The connecting of the cloud-native business interface and the distributed medical database through the interactive interface includes: encapsulating the labeled data into an independent microservice module, generating a corresponding interactive interface description file, and injecting it into the gateway.

5. The cloud-native medical service platform construction method according to claim 3 is characterized in that: The method of calculating the relevance weights of structured fields in structured data and unstructured text features in unstructured data through a multi-head attention mechanism includes: performing a dot product operation on the query vector and the key vector to generate a weight matrix.

6. A cloud-native medical service middle platform construction device, characterized in that: The cloud-native medical business middle platform construction device includes: an acquisition module: acquiring multimodal medical data, the multimodal medical data including structured data, unstructured data and graph data; a label generation module: inputting the multimodal medical data into a pre-trained intelligent classifier, the intelligent classifier is used to perform entity recognition on the unstructured data, mapping the recognized entities with the codes in the structured data, and generating dynamic medical information labels; an integration module: constructing a microservice medical knowledge base based on the dynamic medical information labels, and integrating the microservice medical knowledge base with the cloud-native business interface; an interaction module: connecting the cloud-native business interface and the distributed medical database through an interactive interface; The acquisition module acquires multimodal medical data including: acquiring multimodal medical data from medical institution information systems, image archiving systems, and drug databases; the structured data includes: diagnosis codes, drug codes, and examination item codes in patient electronic files; the unstructured data includes: medical imaging text reports and scanned copies of doctors' handwritten medical records; the graph data includes: indication-contraindication relationships and drug interaction relationships in the drug knowledge graph; The intelligent classifier generates dynamic medical information labels further including: converting diagnostic codes in structured data into one-hot encoding vectors; performing word segmentation on unstructured text data in unstructured data to extract semantic vectors; fusing the one-hot encoding vectors and the semantic vectors at the splicing layer to generate a joint feature vector; mapping the semantic relationships in the medical ontology to the local terminology of the target medical institution; analyzing the time series of patient diagnosis and treatment events and dynamically adjusting label weights.

7. A cloud-native medical service platform construction device, characterized in that: The cloud-native medical business middle platform construction device includes: a memory and at least one processor, the memory stores instructions; the at least one processor calls the instructions in the memory to enable the cloud-native medical business middle platform construction device to execute the steps of the cloud-native medical business middle platform construction method as described in any one of claims 1-5.

8. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instruction is processed and executed, the steps of the cloud-native medical business middle platform construction method as described in any one of claims 1-5 are implemented.

Images