A hospital information system data closed loop method and system based on edge intelligence and multi-agent

Abstract

The application discloses a hospital information system data closed loop method and system based on edge intelligence and multi-agent, which comprises a department end AI box, a central big data processing platform, a multi-agent collaboration layer, a hybrid arrangement engine, a large model reasoning layer, a medical heterogeneous system adapter and a unified interaction layer. The AI box is used for local lightweight reasoning, session management, caching and network outage degradation; the central platform performs layered storage, calculation management and interface service packaging on multi-source data such as HIS, EMR, LIS and PACS. The multi-agent system is based on task classification, planning and scheduling and parallel execution, and realizes the automatic conversion from natural language request to data query and business operation. The system adopts a dual-path data acquisition strategy of big data platform priority and real-time interface supplement, and combines technology blood relationship and business semantic blood relationship tracking mechanism to realize real-time acquisition, traceability and explainability of cross-system data, and is suitable for hospital intelligent interaction and data closed loop application.

Description

Technical Field

[0001] This invention belongs to the field of data governance and intelligent decision support technology, specifically relating to a data closed-loop method and system for hospital information systems based on edge intelligence and multi-agent systems. Background Technology

[0002] With the advancement of high-quality development, performance evaluation, accreditation, DRG / DIP payment reform, and smart hospital construction in public hospitals, hospital managers and clinical departments need to obtain real-time data on indicators such as outpatient volume, number of inpatients, average length of stay, bed occupancy rate, proportion of level IV surgeries, drug ratio, consumable ratio, medical service revenue ratio, CMI, RW, and DRG balance for operational management and clinical decision-making.

[0003] Hospitals typically run multiple business systems simultaneously, including HIS, EMR, LIS, PACS, anesthesia management, medical records, medical insurance, finance, human resources, and supplies. These systems are built by different vendors at different stages, resulting in significant differences in interface protocols, data standards, business coding, and data definitions. This leads to data fragmentation and disjointed business capabilities. While traditional point-to-point interfaces or enterprise service buses can solve some connectivity issues, they rely heavily on static routing and fixed processes. If the interface changes, the system times out, or the data format is abnormal, the process is prone to interruption and requires manual intervention.

[0004] Furthermore, existing medical AI or data integration systems are mostly deployed centrally on servers, with departments accessing them remotely via ordinary PCs. This model suffers from strong network dependence, high concurrent load on central services, high latency in departmental interactions, and a high risk of single points of failure. When medical staff need to query complete patient treatment loop information or comprehensive operational indicators, they may still need to repeatedly log into multiple systems, manually copy and compare data, making it difficult to achieve low-latency, traceable, and explainable hospital-wide data loop access capabilities. Therefore, a technical solution is needed that can unify and adapt departmental edge intelligent terminals, central big data platforms, multi-agent collaboration, and heterogeneous systems. Summary of the Invention

[0005] The purpose of this invention is to provide a data closed-loop method and system for hospital information systems based on edge intelligence and multi-agent systems, in order to solve the problems of data dispersion, rigid processes, high latency of department-level intelligent interaction, reliance on manual operation for cross-system data acquisition, and difficulty in tracing and interpreting data results in existing hospital information systems.

[0006] To achieve the above objectives, this invention provides a data closed-loop method for hospital information systems based on edge intelligence and multi-agent systems. This method is based on standardized data services provided by a central big data processing platform. Departmental edge intelligent terminals receive natural language requests and perform intent recognition, entity extraction, task complexity grading, and local cache lookup. For basic query tasks, the edge intelligent terminals respond directly based on local cache or lightweight end-side models. For cross-system data acquisition tasks, complex analysis tasks, or real-time business operation tasks, the edge intelligent terminals generate structured task descriptions and submit them to the multi-agent collaboration layer. Scheduling layer agents decompose tasks, route data sources, and schedule execution based on medical entity models. They prioritize calling the standardized data services of the central big data processing platform. When the platform does not cover these services or real-time business operations are required, they supplement real-time data acquisition or business operation execution through medical heterogeneous system adapters. Subsequently, the execution results are summarized, conflict resolved, and formatted. Technical and business semantic lineages are established for the fused results. Finally, the fused results carrying lineage information are returned to the edge intelligent terminals for presentation and written to the local cache.

[0007] This invention also provides a closed-loop data system for a hospital information system based on edge intelligence and multi-agent technology, including edge intelligent terminals deployed in departments, a central big data processing platform, a multi-agent collaboration layer, a hybrid orchestration engine, a medical heterogeneous system adapter, and a dual lineage tracing module. The edge intelligent terminal is preferably an AI box, serving as a department-level local intelligent gateway, responsible for natural language interaction, local lightweight inference, departmental agent operation, session context management, local caching, result presentation, and network outage degradation services. The central big data processing platform provides unified access, hierarchical storage, standardized governance, computational processing, and data service encapsulation for data from hospital business systems such as HIS, EMR, LIS, PACS, anesthesia management system, medical record system, medical insurance system, financial system, human resources system, and materials system.

[0008] The multi-agent collaboration layer comprises interaction layer agents, scheduling layer agents, and execution layer agents. Interaction layer agents receive natural language requests from healthcare professionals and perform intent recognition, entity extraction, and task structuring. Scheduling layer agents handle task planning, data source routing, execution scheduling, and exception handling. Execution layer agents invoke data service interfaces from the central big data processing platform or access real-time business systems not yet integrated into the central big data processing platform via the medical heterogeneous system adapter. The hybrid orchestration engine includes a visual orchestration sub-engine and a graph state orchestration sub-engine, used for orchestrating and controlling agent workflows of varying complexity. The medical heterogeneous system adapter connects different hospital business systems via API calls, direct database connections, or automated operation simulations. A dual lineage tracking module records both technical and business semantic lineage.

[0009] This invention employs a dual-path data acquisition strategy of "big data platform priority, supplemented by real-time interfaces." For data already processed at layers such as ODS, DWD, DWS, and ADS, acquisition is prioritized through the data service interfaces of the central big data processing platform. For data not yet covered or requiring real-time operation, API calls, direct database connections, or automated operation simulations are performed via medical heterogeneous system adapters. The system also establishes a dual tracking mechanism for technical lineage and business semantic lineage, recording the technical link of data from source system fields, platform processing levels, data service interfaces, intelligent agent calls, result aggregation to final presentation, as well as the corresponding medical entities, business events, organizational dimensions, and definition rules. Beneficial effects of the present invention

[0010] By establishing a unified intelligent entry point at the departmental level through edge intelligent terminals, it enables rapid response on the edge for simple tasks, in-depth processing at the center for complex tasks, and local service degradation in the event of network anomalies. Through collaboration between the central big data processing platform and medical heterogeneous system adapters, it takes into account both standardized data calls and real-time access to business systems. Through a multi-agent collaboration layer, it enables the decomposition, routing, and orchestration of complex medical data tasks. Through a dual lineage tracking module, it simultaneously records technical lineage and business semantic lineage, improving the traceability and interpretability of cross-system data fusion results. Attached Figure Description

[0011] Figure 1 This is a diagram of the overall technical architecture of the system. Figure 2 Hardware architecture diagram of the edge intelligent terminal AI box; Figure 3 Diagram illustrating the deployment and interaction of the AI-powered smart box in a department; Figure 4 This is a diagram of the architecture of a big data processing platform. Figure 5 Diagram of edge-cloud collaborative multi-agent architecture; Figure 6 System logic deployment diagram; Figure 7 This is a diagram of the system's physical deployment. Figure 8 This is a complete flowchart of the edge-cloud collaboration process.

[0012] The diagram numbers are as follows: 100 represents the hospital's multi-source heterogeneous business systems; 110 represents the HIS system; 120 represents the EMR system; 130 represents the LIS system; 140 represents the PACS system; 150 represents the systems for anesthesia, medical records, medical insurance, finance, human resources, and supplies; 200 represents the central big data processing platform; 210 represents the data access layer; 220 represents the data storage layer; 230 represents the data computing layer; 240 represents the data service layer; 300 represents the multi-agent collaboration layer; 310 represents the interaction layer agents; 320 represents the scheduling layer agents; and 330 represents the execution layer agents. 400 is the hybrid orchestration engine, 410 is the visual orchestration sub-engine, 420 is the graph state orchestration sub-engine, 500 is the medical heterogeneous system adapter, 600 is the large model inference layer, 700 is the edge intelligent terminal AI box, 710 is the main control processor, 720 is the AI ​​acceleration unit, 730 is the running memory, 740 is the local storage, 750 is the network interface, 760 is the display output interface, 770 is the peripheral interface, 780 is the security encryption module, 800 is the unified interaction layer, and 900 is the technical lineage and business semantic lineage tracking module. Detailed Implementation

[0013] The embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the following embodiments are used to explain the present invention and are not intended to unduly limit the scope of protection.

[0014] like Figure 1 As shown, the intelligent integration and data closed-loop system of the hospital information system in this embodiment includes a hospital multi-source heterogeneous business system 100, a central big data processing platform 200, a multi-agent collaboration layer 300, a hybrid orchestration engine 400, a medical heterogeneous system adapter 500, a large model inference layer 600, an edge intelligent terminal AI box 700, a unified interaction layer 800, and a technology lineage and business semantic lineage tracking module 900. The hospital multi-source heterogeneous business system 100 may include a HIS system 110, an EMR system 120, a LIS system 130, a PACS system 140, and systems for anesthesia, medical records, medical insurance, finance, human resources, and materials 150. The data and business capabilities of the above business systems are opened to the multi-agent collaboration layer 300 through the central big data processing platform 200 and the medical heterogeneous system adapter 500.

[0015] like Figure 2 and Figure 3As shown, the edge intelligent terminal AI box 700 is deployed near nurse stations, doctors' offices, conference rooms, or departmental display devices, serving as a departmental-level intelligent interaction entry point. The AI ​​box 700 includes a main control processor 710, an AI acceleration unit 720, RAM 730, local storage 740, a network interface 750, a display output interface 760, a peripheral interface 770, and a security encryption module 780. The main control processor 710 is used to run the terminal system, schedule tasks, and handle interface communication; the AI ​​acceleration unit 720 is used to execute edge-side model inference; the local storage 740 is used to store local models, departmental configurations, cached data, and operation logs; the display output interface 760 connects to a departmental monitor or touchscreen; and the peripheral interface 770 connects to a microphone, camera, barcode scanner, or card reader. The AI ​​box 700 deploys an edge-side software stack, including a local lightweight inference module, a departmental intelligent agent operation module, a session and context management module, a local caching module, an edge-cloud collaborative scheduling module, and a network disaster recovery and degradation module.

[0016] When medical staff submit natural language requests through the unified interaction layer 800, the AI ​​Box 700 first performs edge-side intent recognition and task classification. For example, for basic queries such as "How many inpatients are there in our department today?", the AI ​​Box 700 can call local cache or the edge-side intelligent agent to directly generate a response; for complex tasks such as "Generate a full-cycle postoperative rehabilitation report for a patient", the AI ​​Box 700 converts the natural language request into a structured task description and submits it to the central platform through a secure channel.

[0017] like Figure 4 As shown, the central big data processing platform 200 includes a data access layer 210, a data storage layer 220, a data computation layer 230, and a data service layer 240. The data access layer 210 collects business system data through at least one of the following methods: batch data synchronization, real-time change data capture, and real-time access via message queues. The data storage layer 220 stores raw data, standard detailed data, thematic summary data, and application data according to a hierarchical model of ODS, DWD, DWS, and ADS. The data computation layer 230 provides batch processing computation, streaming computation, interactive querying, and multidimensional pre-aggregation analysis. The data service layer 240 encapsulates the standardized data after governance into standard query interfaces, indicator calculation interfaces, batch data interfaces, and real-time push interfaces for use by the execution layer intelligent agent 330.

[0018] In one implementation, the central big data processing platform 200 also establishes a medical entity model, abstracting business objects such as patients, visits, departments, wards, treatment groups, doctors, diagnoses, surgeries, examinations, tests, drugs, consumables, costs, medical record homepages, and DRG settlements into medical entities. It also defines entity attributes, entity relationships, and mapping relationships with data models at each layer: ODS, DWD, DWS, and ADS. Through this entity model, the multi-agent collaboration layer 300 can understand data based on medical business semantics, rather than simply querying the underlying database tables.

[0019] like Figure 5 As shown, the multi-agent collaboration layer 300 includes an interaction layer agent 310, a scheduling layer agent 320, and an execution layer agent 330. The interaction layer agent 310 performs intent recognition, entity extraction, short-term memory management, and multi-turn dialogue dereference on natural language input. After receiving a structured task description, the scheduling layer agent 320 breaks down the complex task into a sequence of sub-tasks with dependencies and determines whether the data required for the task has been standardized and stored in the big data platform. If the data already exists in the DWD, DWS, or ADS layers, the data query agent is scheduled to obtain standardized data from the central big data processing platform 200; if the data is not yet covered or involves real-time business operations, the corresponding execution layer agent is scheduled to obtain real-time data or perform business operations through the medical heterogeneous system adapter 500.

[0020] The execution layer agent 330 may include a data query agent, an HIS operation agent, a LIS query agent, a PACS query agent, an EMR query agent, and a manual anesthesia query agent, etc. Each execution layer agent receives tasks and returns results according to a unified inter-agent communication protocol. This communication protocol includes a task description format, a result return format, an asynchronous callback mechanism, and a result aggregation and conflict handling mechanism. The task description format includes task type, parameters, priority, and timeout fields; the result return format includes execution status, data content, lineage information, and exception information. When results returned from multiple data sources are inconsistent, the scheduling layer agent 320 can merge them according to preset rules such as the big data platform's master data priority.

[0021] like Figure 6As shown, the hybrid orchestration engine 400 includes a visual orchestration sub-engine 410 and a graph state orchestration sub-engine 420. The visual orchestration sub-engine 410 is used to quickly configure large model nodes, tool nodes, code nodes, branch nodes, and knowledge base retrieval nodes through canvas-style nodes; the graph state orchestration sub-engine 420 is used to implement loops, conditional branches, parallel execution, transactional operations, error handling, and degradation strategies. For simple workflows, the visual orchestration sub-engine 410 can quickly configure them; for complex tasks requiring fine-grained control, the visual orchestration sub-engine 410 can send tasks to the graph state orchestration sub-engine 420 for processing via a complex task forwarding node, and the processing results are then returned to the visual orchestration sub-engine 410 for further formatted output.

[0022] The Medical Heterogeneous System Adapter 500 is used to connect business systems or real-time operation scenarios that are not fully integrated into a big data platform. Adapter 500 supports API interface mode, direct database connection mode, and automated operation simulation mode. In API interface mode, the adapter calls the RESTful API or WebService interface provided by the business system and parses JSON or XML data; in direct database connection mode, the adapter executes standardized queries through a database connection; in automated operation simulation mode, the adapter obtains data or performs business operations by simulating human interface operations. Adapter 500 can also perform interface retries, timeout control, circuit breaking and degradation, and data format conversion.

[0023] like Figure 7 As shown, the system can be deployed within a hospital's intranet environment, and can be partitioned into access areas, application service areas, AI computing areas, big data processing areas, and data storage areas. Multiple departments deploy AI Box 700 units, each operating independently and connected to the central platform via the hospital's intranet. This physical deployment method helps reduce access latency at the departmental level and avoids impacting other departments' use when a device malfunctions in one department.

[0024] like Figure 8 As shown, the complete edge-cloud collaborative processing flow of the system includes: S1, the department end inputs a natural language request through the unified interaction layer 800; S2, the AI ​​box 700 performs edge-side intent recognition and task classification; S3, for complex tasks, the AI ​​box 700 submits a structured task description to the central platform; S4, the scheduling layer agent 320 performs task planning and determines the data source route; S5, each execution layer agent 330 executes data queries or business system calls in parallel and returns the results; S6, the scheduling layer agent 320 summarizes, merges, and handles conflicts in the results, and returns the results to the AI ​​box 700; S7, the AI ​​box 700 updates its local cache; S8, when the network is interrupted, the AI ​​box 700 switches to local degradation mode, using the local cache and edge-side model to continue providing basic services.

[0025] The technical lineage and business semantic lineage tracking module 900 runs through the entire process described above. Technical lineage records the data's processing path, from source system table fields, through various layers of processing on the central big data processing platform 200, via data service interfaces and intelligent agent calls, to result aggregation and final output. Business semantic lineage records the corresponding departments, medical entities, business events, organizational dimensions, and definition rules. Through dual lineage tracking, the natural language responses or visual dashboards ultimately displayed to medical staff not only include the results but also provide explanations of the result's source, processing path, and business meaning, thereby improving the traceability and interpretability of hospital data closed-loop access.

Claims

1. A data closed-loop method for hospital information systems based on edge intelligence and multi-agent systems, used to form a traceable data access closed loop between the hospital's multi-source heterogeneous business systems and the departmental interaction portal, wherein the multi-source heterogeneous business systems include at least two of HIS, EMR, LIS, PACS, anesthesia management system, medical record system, medical insurance system, financial system, human resources system, and material supply system, characterized in that, The method includes the following steps: S1. The data from the hospital's multi-source heterogeneous business systems are accessed, stored in layers, standardized and managed, and packaged into data services through the central big data processing platform to form standardized data services that can be called by intelligent agents. S2. Receive natural language requests from medical staff through edge intelligent terminals deployed in the department, and perform intent recognition, entity extraction, and task complexity classification on the edge intelligent terminal side of the natural language requests. S3. When the natural language request is determined to be a basic departmental query task, the edge intelligent terminal calls the local cache and / or the terminal-side lightweight model to generate a response; when the natural language request is determined to be a cross-system data acquisition task or a complex analysis task, the edge intelligent terminal generates a structured task description and submits it to the central-side scheduling layer intelligent agent. S4. The central-side scheduling layer intelligent agent performs task decomposition and data source routing on the structured task description based on the medical entity model, and prioritizes determining whether the data required by the task has been standardized and stored by the central big data processing platform. S5. When the required data has been standardized and stored, the scheduled data query agent calls the standardized data service to obtain the data; when the required data is not covered by the central big data processing platform or requires real-time business system operation, the corresponding execution layer agent is scheduled to obtain real-time data or perform business operation through the medical heterogeneous system adapter. S6. The central-side scheduling layer agent summarizes, handles conflicts, and formats the execution results from the standardized data service and / or the medical heterogeneous system adapter to generate a fusion result; S7. Establish technical lineage and business semantic lineage for the fusion result, wherein the technical lineage records at least the source system, source table fields, data platform layering, data service interface, intelligent agent call and result output path, and the business semantic lineage records at least the department, medical entity, business event, organizational dimension and caliber rules. S8. The fusion result carrying the technical lineage and the business semantic lineage is returned to the edge intelligent terminal, which presents it through the unified interaction layer and writes the request, fusion result and lineage summary into the local cache.

2. The method according to claim 1, characterized in that, In step S1, the central big data processing platform includes a data access layer, a data storage layer, a data calculation layer, and a data service layer. The data access layer accesses data using at least one of the following methods: batch data synchronization, real-time change data capture, and real-time access via message queues. The data storage layer stores raw data, standard detailed data, thematic summary data, and application data according to the ODS layer, DWD layer, DWS layer, and ADS layer. The data service layer encapsulates the processed data into standard query interfaces, indicator calculation interfaces, batch data interfaces, and real-time push interfaces.

3. The method according to claim 1, characterized in that, In step S2, the edge intelligent terminal is an AI box deployed in the department. The AI ​​box includes at least a main control processor, an AI acceleration unit, local storage, a network interface, and a display output interface. The AI ​​box runs the department-side interactive intelligent agent and maintains the session context, department configuration, and high-frequency query cache related to the department's business scenarios.

4. The method according to claim 1, characterized in that, In step S3, the structured task description includes at least the task type, medical entity parameters, time range, department parameters, task priority, and timeout field; the task complexity classification is determined at least based on whether the request involves cross-system data, whether historical batch data is required, whether real-time business system operations are required, and whether the local cache is hit.

5. The method according to claim 1, characterized in that, In step S4, the medical entity model abstracts at least one of the following as medical entities: patient, consultation, department, ward, treatment group, doctor, diagnosis, surgery, examination, test, medicine, consumables, cost, medical record cover page, and DRG settlement. It also establishes a mapping relationship between the medical entities and the data models of the ODS layer, DWD layer, DWS layer, and ADS layer, so that the central side scheduling layer agent can route data sources according to medical business semantics.

6. The method according to claim 1, characterized in that, In step S5, the medical heterogeneous system adapter accesses the business system using at least one of the following modes: API interface mode, database direct connection mode, and automated operation simulation mode. It performs interface retry, timeout control, circuit breaker degradation, and data format conversion during the access process. When multiple data sources return conflicting results, step S6 merges them according to a preset priority. The preset priority includes at least the master data priority rule of the central big data processing platform.

7. The method according to claim 1, characterized in that, After step S8, when the network connection between the edge intelligent terminal and the central platform is interrupted, the edge intelligent terminal switches to local degradation mode, continues to respond to basic query tasks based on local caching, department configuration and end-side lightweight model, and synchronizes the request logs and cache update records during the degradation period to the central platform after the network is restored.

8. A data closed-loop system for hospital information systems based on edge intelligence and multi-agent systems, characterized in that, include: The central big data processing platform is used to access data from the hospital's multi-source heterogeneous business systems, and to perform hierarchical storage, standardized governance, indicator processing, and data service encapsulation on the accessed data to form standardized data services. Edge intelligent terminals deployed in departments are used to receive natural language requests input by users and perform intent recognition, entity extraction, task complexity classification, local cache query, result presentation and cache writing locally; The multi-agent collaboration layer includes an interaction layer agent, a scheduling layer agent, and an execution layer agent; wherein, the interaction layer agent is used to convert natural language requests into structured task descriptions, the scheduling layer agent is used to decompose tasks, route data sources, and determine the execution order based on the medical entity model, and the execution layer agent is used to call the standardized data services of the central big data processing platform, or access the corresponding hospital business system through the medical heterogeneous system adapter; Medical heterogeneous system adapters are used to connect different hospital business systems through interface calls, database access, or automated operations to obtain real-time data or perform business operations. A hybrid orchestration engine is used to orchestrate and manage the call relationships between the interaction layer intelligent agents, scheduling layer intelligent agents, execution layer intelligent agents, central big data processing platform, and medical heterogeneous system adapter; The dual lineage tracking module is used to generate technical lineage information and business semantic lineage information based on the fusion results. The technical lineage information includes data source, interface, fields, processing node and processing time, and the business semantic lineage information includes medical entity, business meaning, indicator definition and association rules. The system, through the collaboration of the edge intelligent terminal, multi-agent collaboration layer, central big data processing platform, medical heterogeneous system adapter, hybrid orchestration engine and dual lineage tracking module, enables user requests, data acquisition, data fusion, lineage generation, result presentation and local cache writing to form a closed loop.

9. The system according to claim 8, characterized in that, The edge intelligent terminal is an AI box, which includes a main control processor, an AI acceleration unit, running memory, local storage, a network interface, a display output interface, a peripheral interface, and a security encryption module. The AI ​​box is connected to a department monitor or touch screen through the display output interface, and is connected to at least one of a microphone, camera, barcode scanner, or card reader through the peripheral interface.

10. The system according to claim 8, characterized in that, The hybrid orchestration engine includes a visual orchestration sub-engine and a graph state orchestration sub-engine; The visual orchestration sub-engine is used to configure large model nodes, tool nodes, code nodes, branch nodes, and knowledge base retrieval nodes. The graph state orchestration sub-engine is used to handle loops, conditional branches, parallel execution, transactional operations, error handling, and degradation strategies. The visual orchestration sub-engine sends tasks requiring fine-grained control to the graph state orchestration sub-engine through the complex task forwarding node, and receives the processing results returned by the graph state orchestration sub-engine.

Images