Cross-system medical data interaction and multi-agent control method, system and device

Abstract

The application provides a cross-system medical data interaction and multi-agent control method, system and device. The scheme comprises the following steps: constructing a standardized data interaction interface of multiple types of medical data; performing mapping and unified packaging on multi-source heterogeneous medical data; establishing a multi-system data consistency verification mechanism; constructing a global resource state perception model to realize dynamic arrangement and optimal allocation of multi-agent resources; and forming a full-process track that can be played back and verified through a service tracing system. The scheme can realize high-reliability interaction and multi-agent collaborative control of traditional Chinese and western medical data and medical knowledge among multiple systems, and improve medical data sharing efficiency, intelligent reasoning stability, system scalability and compliance audit capability.

Description

Technical Field

[0001] This invention relates to the fields of medical artificial intelligence, medical data processing, multi-agent collaborative control and knowledge engineering, and in particular to a cross-system medical data interaction and multi-agent control method, system and device. Background Technology

[0002] In the context of integrated traditional Chinese and Western medicine data governance, there are simultaneously hospital information systems, electronic medical record systems, laboratory testing systems, TCM disease-specific systems, clinical knowledge bases, guideline databases, and various brain-like reasoning and intelligent agent service systems. These systems have long been heterogeneous in terms of data models, terminology, transmission methods, access boundaries, and processing logic. This makes it difficult to stably share data such as medical records, treatment orders, laboratory indicators, syndrome records, and medical books and guidelines between systems, thus limiting the implementation of intelligent collaborative applications between traditional Chinese and Western medicine.

[0003] Existing technologies typically build point-based solutions for single data sources or single models. On the one hand, there is a lack of unified interaction interfaces between different systems that are compatible with structured data, semi-structured documents, and unstructured medical knowledge texts, especially for knowledge-based data such as guidelines, books, ancient texts, and case documents, which lack consistent access and processing solutions. On the other hand, when multiple hospitals, multiple platforms, and multiple models participate in the processing, there is a lack of a comprehensive consistency verification mechanism between source data, transmitted data, processed data, and result data, which can easily lead to encoding mismatch, unit confusion, semantic drift, and result distortion.

[0004] Furthermore, with the increasing application of multi-agent technology in medical data management and decision support, different agents often undertake tasks such as data access, knowledge extraction, retrieval and verification, clinical reasoning, document generation and auditing. However, existing solutions often lack a unified control plane and global scheduling mechanism, making it difficult to dynamically orchestrate multiple agents based on task priority, resource consumption, skill capabilities and security policies. This results in resource contention, duplicate execution, increased response latency and insufficient abnormal recovery capabilities.

[0005] Furthermore, existing systems generally lack a complete traceability system that covers data sources, processing procedures, agent invocation, prompt template versions, tool invocation, and result publication. When system outputs errors, deviations, or compliance disputes occur, it is difficult to accurately pinpoint whether the problem lies with the source data, transmission, standardization, knowledge extraction, model illusion, or agent collaboration, severely impacting the credibility, auditability, and application value of data systems in medical settings. Summary of the Invention

[0006] The purpose of this invention is to provide a method, system, and device for cross-system medical data interaction and multi-agent control, particularly for medical data combining traditional Chinese and Western medicine, to solve at least one of the following problems in the prior art: lack of a unified interaction interface compatible with multiple data types between different medical systems and knowledge systems; lack of end-to-end consistency verification during the cross-system sharing of multi-source heterogeneous data, making it difficult to guarantee the accuracy of sharing; lack of a unified control plane and global resource awareness capability for multiple agents, resulting in low scheduling efficiency; and lack of complete traceability in the data flow and agent operation process.

[0007] In view of the above problems, the present invention provides the following solution: In a first aspect, the present invention provides a method for cross-system medical data interaction and multi-agent control, comprising: S1. Receive raw data and task requests from multiple medical information systems, medical knowledge systems, and intelligent agent service systems, and construct standardized data interaction interfaces according to a unified data model to classify and encapsulate different types of data; the different types of data include structured data, semi-structured data, and unstructured data; different types of data include Liri electronic medical records, treatment orders, medical orders, test results, traditional Chinese medicine four diagnostic methods information, clinical guidelines, medical books, literature, and knowledge graph data, etc. S2. The data accessed through the standardized data interaction interface is sent to the AI ​​data platform, and the DataFlow pipeline is used to perform parsing, extraction, standardization, segmentation, recombination and unified representation processing on different types of data to form a standard data object that can be called by both brain-like systems and multi-agent systems. S3. Construct a multi-system data consistency verification system covering the data sending end, transmission link, processing node, inference node and result output end. Perform syntax consistency verification, pattern consistency verification, terminology encoding consistency verification, semantic evidence consistency verification and result traceability consistency verification on the standard data object, and output the data object that has passed the consistency verification or the data object to be reviewed based on the verification results. S4. Construct an intelligent agent collaboration layer to collect information on intelligent agent operation status, skill availability, tool permissions, session load, resource consumption, and task priority, and establish a global resource status awareness model. S5. Based on the global resource state awareness model, perform task decomposition, route scheduling, and resource allocation for each agent to generate cross-system data interaction results and multi-agent collaborative processing results; the agents include data access agents, knowledge extraction agents, terminology mapping agents, consistency verification agents, collaborative reasoning agents, and audit traceability agents. S6. Establish a service traceability record for the cross-system data interaction results and multi-agent collaborative processing results. The service traceability record includes at least the source system identifier, data version, input hash, output hash, calling agent identifier, calling tool identifier, prompt template version, timestamp, and evidence citation information, thereby forming a complete trajectory of data flow and agent operation.

[0008] Preferably, in step S1, the standardized data interaction interface adopts a layered adaptation structure, including an access adaptation layer, a standard encapsulation layer, a semantic mapping layer, and a security control layer; wherein, the standard encapsulation layer encapsulates data from different sources into a unified data envelope object, and the data envelope object is represented as... At least header metadata Effective information payload Basic metadata and security policies and access control information .

[0009] Preferably, in S2, the DataFlow pipeline is used to perform field standardization, unit unification, and timeline alignment on structured medical data; to perform title hierarchy, paragraph boundary, table structure, and key field extraction on semi-structured medical documents; and to perform document segmentation, entity recognition, relation extraction, terminology normalization, and evidence fragment index construction on clinical guidelines, medical books, ancient texts, and PDF documents, and output standard data objects, knowledge objects, and evidence indexes.

[0010] Preferably, in step S3, the multi-system data consistency verification mechanism includes source verification, transmission verification, processing verification, inference verification, and result verification, and is based on pattern consistency scores. Coding consistency score Unit consistency score Evidence alignment score and traceability integrity score Generate a comprehensive consistency score:

[0011] in, to This is the weighting coefficient. When the overall consistency score is lower than the preset threshold C0, at least one of the following actions is triggered: re-retrieval, regeneration, manual confirmation, or result downgrade output.

[0012] Preferably, in S4 and S5, the global resource state awareness model is constructed based on the gateway control plane and a multi-agent routing mechanism. It perceives the agent's online status, session activity, workspace isolation status, skill availability, tool permissions, CPU utilization, GPU memory usage, memory usage, task waiting time, historical success rate, output confidence, and the security domain to which the task belongs, and then applies this perception according to a preset objective function. It performs task decomposition, dynamic routing, parallel execution, load balancing, and failover. Indicates the benefits of task quality. Indicates the benefits of resource utilization. Indicates priority benefits. Indicates the cost of latency. Indicates the risk and cost. , , , , These are weight parameters; Preferably, in step S6, the service traceability record adopts an event chain structure to associate and record the processes of data entry, DataFlow processing, agent invocation, result generation, and result publication, and records any traceability event as: This includes at least the link identifier. Current event identifier Parent event identifier Source system identifier Execute intelligent agent identification Skill identifier invoked The identifier of the pipeline operator called Input summary Output summary timestamp Strategy Information Citation of evidence and execution status .

[0013] Secondly, the present invention provides a cross-system medical data interaction and multi-agent control system, the system being used to execute the method described in the first aspect, including a data access module, a standardization processing module, a consistency verification module, an agent collaboration module, a resource perception module, a service traceability module, and a result output module.

[0014] The data access module is used to receive multi-source medical data, medical documents, medical knowledge, and task requests from external systems. The standardization processing module is used to complete the parsing, extraction, normalization, and unified representation of multiple types of data through a standardized data interaction interface and the DataFlow pipeline; The consistency verification module is used to perform end-to-end multi-system data consistency verification and result reliability verification. The agent collaboration module is used to complete task decomposition, session orchestration, resource scheduling and fault switching based on the gateway control plane, multi-agent routing and multi-agent skill invocation, so as to achieve optimal allocation of multi-agent resources; The resource awareness module is used to build and update the global resource status awareness model; The service traceability module is used to record data flow, skill calls, tool calls, input / output summaries, and audit trails; The results output module is used to provide shared data objects, verification reports, collaborative processing results, traceability reports, and interface response results to external parties.

[0015] Thirdly, the present invention also provides a cross-system medical data interaction and multi-agent control device, including a processor, a memory and a bus, wherein the processor is used to call instructions in the memory to execute the method as described in the first aspect.

[0016] By adopting the above technical solution, the present invention has the following beneficial effects: First, by using a unified and standardized data interaction interface, covering both structured Chinese and Western medicine medical data and unstructured medical knowledge data, a seamless connection between the AI ​​data platform and the brain-like system can be achieved, which helps reduce the cost of cross-system access and improve the degree of data reuse.

[0017] Second, by using a multi-system data consistency verification mechanism, data consistency control is extended from a single processing node to the entire process of sending, transmitting, processing, inference, and publishing. This helps reduce the risk of error propagation caused by terminology mismatch, unit inconsistency, semantic drift, and model illusion, and improves the accuracy of shared data and the credibility of processing results.

[0018] Third, through the global resource state awareness model and multi-agent scheduling strategy, multiple agents can be optimally allocated and failover can be performed based on task priority, real-time load, skill availability and security constraints, thereby improving the system's response stability and resource utilization.

[0019] Fourth, the service traceability system records the entire data lineage, prompt template version, skill calls, tool calls, and output evidence. It supports anomaly playback, responsibility identification, version comparison, and audit export, meeting the compliance audit requirements in medical scenarios.

[0020] Fifth, this invention uses a multi-agent operating framework (such as OpenClaw) for multi-agent control and task orchestration, and DataFlow for data preparation and knowledge processing, forming a closed loop between interface unification, data verification, resource collaboration and trajectory tracing, which is suitable for scenarios such as integrated traditional Chinese and Western medicine data processing, medical knowledge management, decision support and multi-institutional data collaboration. Attached Figure Description

[0021] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the overall process of the method provided in this embodiment of the invention. Figure 2 This invention provides a standardized data interaction interface and a data platform processing flowchart for embodiments of the invention. Figure 3 A flowchart of a multi-system data consistency verification mechanism provided in an embodiment of the present invention; Figure 4 A flowchart of the multi-agent collaboration and global resource state awareness model provided in this embodiment of the invention; Figure 5 This is a schematic diagram of the event chain structure of the service traceability system provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the system architecture provided for an embodiment of the present invention. Detailed Implementation

[0023] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the accompanying drawings are merely schematic illustrations and are not drawn to scale, and some details may be simplified or omitted to highlight the core innovations. Those skilled in the art can understand the technical solution of the present invention after reading the following description and referring to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] Example 1 like Figure 1 As shown, the method provided in this embodiment is aimed at data flow and task collaboration between hospital electronic medical record systems, laboratory examination systems, traditional Chinese medicine diagnosis and treatment systems, clinical knowledge systems, brain-like reasoning systems, and intelligent agent service systems, and mainly includes the following steps: Step 1: Receive raw data and task requests from multiple medical information systems, medical knowledge systems, and intelligent agent service systems, and build a standardized data interaction interface through a unified data model. Classify and access different types of data, encapsulate them uniformly, and write them into the AI ​​data platform.

[0025] like Figure 2 As shown, Figure 2 This illustrates the standardized data access, unified encapsulation, semantic mapping, security control, and AI data platform processing procedures corresponding to steps 1 and 2. Specifically, step 1 includes: Step 11: Establish a source system adapter to interface with the Hospital Information System (HIS), Electronic Medical Record System (EMR), Laboratory Information System (LIS), Medical Image Storage and Transmission System (PACS), Traditional Chinese Medicine Four Diagnostic Methods System, Clinical Guideline Library, Medical Book Library, Document File System and Knowledge Graph System, and support API calls, message queue subscriptions, file subscriptions or batch import methods.

[0026] Step 12: After access adaptation, identify the data types and encapsulate patient data, examination and test data, TCM four diagnostic methods and syndrome data, medical document data, and knowledge-based text data into a unified data envelope object according to preset encapsulation rules. The unified data envelope object includes at least header metadata. Effective information payload Basic metadata and security policies and access control information The preset encapsulation rules include at least: determining the object type based on the data source, data structure, and business purpose; extracting common header fields and writing them into header metadata. The header metadata It should include at least the object type, source identifier, timestamp, data format identifier, and pre-assigned traceability fields; and include the medical record text, test values, four diagnostic records, prescription content, document paragraphs, or knowledge fragments in the effective information payload. The source institution, patient identifier, visitor identifier, knowledge subject identifier, terminology system version, unit system version, and data type are written into the basic metadata. Security levels, access subject scope, de-identification strategies, transmission encryption requirements, and sharing constraints are written into security policies and access control information. For structured data, it is generated directly according to the field mapping table. , , and For semi-structured and unstructured data, first perform page layout parsing, paragraph segmentation, and key field extraction, and then generate a unified data envelope object based on the same set of fields.

[0027] Step 13: After completing the unified encapsulation, perform semantic mapping and security control processing. Semantic mapping maps TCM terms, Western medicine terms, standard codes, and local extended fields to a unified semantic space. Security control performs identity authentication, permission verification, de-identification rule matching, and transmission encryption. The unified semantic space adopts a combined representation of "standard terminology codes + semantic type tags + relationship slots," including at least entity identifiers, entity types, standard codes, source text, time tags, and association fields. During semantic mapping, TCM terms, Western medicine terms, standard codes, and local extended fields are first normalized and merged according to the terminology dictionary, code mapping table, and local field lookup table. The mapping results are then written into the unified semantic space, converting different encapsulated data objects into a unified representation that is comparable, searchable, and inferable. The mapped data retains the original text fragments, standard codes, mapping confidence, and source location for subsequent evidence alignment and tracing.

[0028] Step 14: Assign a unique trace_id, source_id, and version_id to each data object, and record the creation time, source institution, patient or knowledge topic identifier, and security level. Then, write the data object into the unified storage layer of the AI ​​data platform. The trace_id is generated using a unique end-to-end identifier rule, combining the source institution identifier, object type identifier, timestamp, and random checksum, and is used throughout the entire process of access, processing, invocation, and release. The source_id uniquely identifies the original source system, source document, or source record, and must remain unique within the same source domain. The version_id is generated incrementally according to the object update sequence, identifying different processing versions under the same source_id and establishing a parent-child relationship with the previous version. All the above identifiers are written in a fixed field format. and And it is stored together with the object hash digest.

[0029] Step 2: Based on the AI ​​data platform and combined with the DataFlow pipeline, perform parsing, extraction, segmentation, standardization, recombination and unified representation processing on the data accessed through the standardized data interaction interface.

[0030] In a preferred embodiment, step 2 specifically includes: Step 21: Perform field standardization, null value handling, unit unification, encoding mapping, and timeline alignment on structured medical data to form standard data fragments that can be directly calculated and shared. Null value handling is performed according to field type: null value markers and missing reason codes are written for numeric fields; status codes such as "not collected," "not applicable," and "unknown" are written for enumerated fields; and empty strings are retained and missing identifiers are added for text fields, without defaulting to clinical values ​​to replace missing values. During encoding mapping, corresponding terminology mapping tables are called for diseases, symptoms, syndromes, laboratory tests, drugs, and procedures, mapping local codes, aliases, or free text to unified standard codes while preserving the original local values ​​and mapping relationships. During timeline alignment, based on the patient master index and unified timezone timestamps, outpatient, inpatient, laboratory, examination, medical orders, and four diagnostic methods information are merged into a unified event sequence. Data containing only dates is supplemented with daily time granularity; start and end times are recorded for interval-type data, and the data is sorted according to collection time, execution time, and reporting time.

[0031] Step 22: Extract title levels, paragraph boundaries, table structure, and key fields from semi-structured medical document data to form structured document fragments; for clinical guidelines, medical books, ancient texts, and PDF documents, execute document segmentation, entity recognition, relation extraction, fact filtering, and evidence fragment index construction sequentially based on DataFlow's knowledge extraction and normalization pipeline.

[0032] Step 23: Establish a unified representation template for the processed TCM and Western medicine knowledge items, representing "symptoms—signs—syndromes—diseases—treatment—sources of evidence," and output sequence representations, graph structure representations, or fusion representations according to different downstream tasks. The unified representation template is constructed using fixed fields and a directed relational structure, including at least node fields {entity ID, entity type, standard code, text value, time tag, source ID} and relation fields {predecessor entity, relation type, successor entity, evidence citation}. When establishing the template, various types of data, such as TCM symptoms, tongue and pulse characteristics, syndromes, Western medicine diagnoses, test indicators, treatment plans, and sources of evidence, are uniformly mapped to node-relational combinations. For sequential reasoning, report generation, and text retrieval tasks, output sequence representations; for syndrome-disease association reasoning, treatment path analysis, and evidence backtracking tasks, output graph structure representations; and for collaborative reasoning tasks that simultaneously utilize temporal and relational information, output fusion representations.

[0033] Step 24: Provide standard data objects to the brain-like system and multi-agent system for invocation, and provide knowledge objects and evidence indexes to the retrieval and reasoning modules for invocation, and generate interface response results and sharing results based on external requests. The evidence index is generated from the document segmentation, entity recognition, relation extraction, and fact filtering results in Step 22, specifically by establishing an inverted index mapping of fragment identifiers, entity identifiers, keywords, time tags, and source locations for guideline clauses, book excerpts, case document excerpts, and knowledge triples.

[0034] Step 3: Construct a multi-system data consistency verification mechanism covering the sending end, transmission link, processing node, inference node and result output end, and perform multi-level consistency verification on standard data objects.

[0035] like Figure 3 As shown, step 3 specifically includes: Step 31: Perform source verification, check the field integrity of the data before sending, and generate a signature or source hash digest to confirm the integrity of the original data before it leaves the sending end.

[0036] Step 32: Perform transmission verification. Verify and compare data packets during transmission, control retransmission, and check version consistency to confirm that data has not been lost, corrupted, or versioned in the link.

[0037] Step 33: Perform processing verification. The standardized data objects undergo pattern constraint checks, unit conversion verification, terminology coding consistency comparison, and cross-source conflict detection to confirm that the standardization results meet preset requirements at both the structural and semantic levels. Pattern constraint checks include at least checking field existence, field type, value range, and mandatory relationships. Unit conversion verification verifies the legality of test indicators, vital signs, traditional Chinese medicine dosages, and administration frequencies according to a preset unit conversion table. Terminology coding consistency comparison checks diseases, symptoms, syndromes, drugs, and examination items based on a mapping dictionary from local to standard codes. Cross-source conflict detection identifies conflicts in diagnosis, test results, prescriptions, and key treatment records for the same patient within adjacent time windows. For traditional Chinese medicine prescription data, further verification checks the matching of herb names, dosage units, dosage ranges, decoction methods, and frequency fields. For key test items, further verification checks the correspondence between item names, values, units, and reference ranges.

[0038] Step 34: Perform inference verification. Align and compare the agent's or model output with the original evidence fragments, clinical rules, and knowledge graph constraints, and check temporal and logical consistency to suppress model illusions and evidence-free expansion.

[0039] Step 35: Perform result verification, verify whether the output result retains the source link, version information, evidence reference and trace identifier, and send the verification results of all levels to the consistency scoring model.

[0040] In a preferred embodiment, the comprehensive consistency test result C is calculated by the following formula:

[0041] in which, represents the schema consistency score, which is calculated according to the field completeness rate, field type matching rate and schema constraint passing rate; represents the coding consistency score, which is calculated according to the consistency rate of term text and standard code, local code mapping success rate and cross-source code conflict rate; represents the unit consistency score, which is calculated according to the unit legality rate, unit conversion accuracy rate and value-unit matching rate; represents the evidence alignment score, which is calculated according to the coverage rate of evidence fragments corresponding to the output conclusion, time consistency rate and logical constraint satisfaction rate; represents the trace completeness score, which is calculated according to the completeness rate of trace_id, source_id, version_id, evidence reference and parent-child event chain. Each score takes a normalized value between 0 and 1. to are weight coefficients. The system compares the comprehensive consistency score with a preset threshold C0; when C≥C0, the current data object is marked as a sharable or publishable state, and a consistency verification report is output; when C<C0, a supplementary retrieval, re-extraction, rollback to the previous stable version or manual confirmation process is triggered.

[0042] Step 4: Construct an agent collaboration layer based on the multi-agent operation framework, and establish a global resource state awareness model.

[0043] as Figure 4 shown, Figure 4 illustrates the multi-agent collaboration and resource scheduling process corresponding to step 4 and step 5, and the operation framework may adopt, for example, OpenClaw. Wherein, step 4 specifically includes: Step 41: Using the gateway control plane as the task entry point and control plane, complete task access, routing decisions, session management, and permission policy distribution. Maintain the shared workspace, session context, and skill package index within the control plane. The skill package index is generated from the capability descriptions reported by each agent during the registration phase. These capability descriptions include at least the agent identifier, skill identifier, input type, output type, required tools, resource requirements, security domain requirements, and callable DataFlow pipeline information. After receiving the capability descriptions, the gateway control plane standardizes and organizes them to establish a searchable skill package index.

[0044] Step 42: Register multiple functional agents in the shared workspace. These functional agents include at least a data access agent, a knowledge extraction agent, a terminology mapping agent, a consistency verification agent, a collaborative reasoning agent, and an audit traceability agent. The collaboration of multiple agents is achieved through a "task message + object reference / index reference + callback confirmation" approach: After completing processing, the upstream agent writes the output object and its trace_id into the shared workspace and submits the object reference or evidence index reference for the next subtask to the gateway control plane. The gateway control plane selects the downstream agent based on the subtask dependency, skill package index, and security domain constraints. The downstream agent reads the object reference and continues processing, and writes back the processing status and result reference upon completion, thereby forming a serial, parallel, or converged collaborative link.

[0045] Step 43: Continuously collect state variables such as agent online status, session activity, workspace isolation status, skill availability, tool permissions, CPU utilization, GPU memory usage, memory usage, task waiting time, historical success rate, output confidence, and the security domain to which the task belongs, and establish a global resource status awareness model.

[0046] Step 5: Based on the global resource state awareness model, perform task decomposition, route scheduling, and resource allocation for multiple agents. The global resource state awareness model is not a single black-box model, but rather consists of a task decomposition rule set, a state vector fusion module, and a scheduling decision module. Specifically, the task decomposition rule set generates a sequence of sub-tasks based on task type and input object type; the state vector fusion module calculates the suitability of each candidate agent for the current sub-task; and the scheduling decision module selects the final execution path based on the objective function.

[0047] In a preferred embodiment, step 5 specifically includes: Step 51: Based on the current task type and data object characteristics, break down the task into sub-tasks such as data access, knowledge extraction, terminology mapping, consistency verification, collaborative reasoning, and audit traceability, and perform capability matching and security domain matching for each sub-task.

[0048] Step 52: Based on the objective function Perform resource allocation, among which, Indicates the benefits of task quality. Indicates the benefits of resource utilization. Indicates priority benefits. Indicates the cost of latency. Indicates the risk and cost. , , , , The weight parameter is used; the system selects the scheduling scheme with the greatest overall benefit among multiple candidate execution paths based on the objective function.

[0049] Step 53: Implement dynamic routing, parallel execution, load balancing, and failover based on the scheduling results; when the target agent is overloaded, offline, or does not meet the permission conditions, switch the task to the alternative agent, and if necessary, degrade the execution result output or manually take over.

[0050] Step 6: Establish service traceability records for cross-system data interaction results and multi-agent collaborative processing results.

[0051] like Figure 5 As shown, step 6 specifically includes: Step 61: Record the data access, DataFlow processing, agent invocation, result generation, and result publication processes as independent traceable events, and assign a unique current event identifier (event_id) to each traceable event; assign the same trace_id to multiple traceable events belonging to the same business processing flow; for any traceable event other than the initial event, record the direct upstream event identifier (parent_id) that triggered the event. Specifically, the source system access event corresponds to the data entry process, the DataFlow processing event corresponds to the standardization and knowledge processing process, the agent invocation event corresponds to the collaborative execution process, and the result generation event corresponds to the output formation process.

[0052] Step 62: Record the traceability events according to a unified structure. Any traceability event is denoted as:

[0053] in, For link identification, This serves as the current event identifier. As the parent event identifier, For source system identification, To implement agent identification, The identifier for the skill being invoked. Identifier for the pipeline operator being invoked. and These are the input and output summaries, respectively. For timestamps, For strategy information, For the purpose of citing evidence, In execution status.

[0054] Step 63, through Multiple traceable events are chained together to form an event chain, enabling end-to-end association from source system events to result release events, and supporting trajectory queries based on patients, institutions, tasks, agents, or time windows. Specifically, when a downstream event is generated, the event_id of the upstream event that triggered the downstream event is written into the parent_id field of the downstream event. At the same time, the same trace_id is maintained within the same business processing chain, and the source_id and version_id are inherited. For parallel branch events, a common parent event identifier is recorded. For converged events, the main parent event identifier is recorded, and other related event identifiers are retained in the strategy information, thus forming a traversable parent-child event chain.

[0055] Step 64: Generate replay results, responsibility location results, version comparison results, and audit export results based on the event chain, which are used for anomaly replay, problem location, and compliance audit.

[0056] Example 2 like Figure 6 As shown, this embodiment provides a cross-system medical data interaction and multi-agent control system. The system as a whole is located between an external data and knowledge system and an external calling side. The external data and knowledge system on the left includes at least HIS, EMR, LIS, PACS, TCM diagnosis and treatment system, guideline library, book library, file system and atlas library. The external calling side on the right includes at least a brain-like reasoning system, clinical applications, sharing interface, reporting system and auditing terminal.

[0057] The system internally includes a data access module, a standardization processing module, a consistency verification module, an intelligent agent collaboration module, a resource awareness module, a service traceability module, and a result output module. Specifically, the data access module performs multi-source system access and raw task reception; the standardization processing module performs interface adaptation, DataFlow processing, and unified representation; the consistency verification module performs mode, encoding, unit, evidence, and traceability verification; the intelligent agent collaboration module performs OpenClaw routing, session isolation, and skill loading; the resource awareness module performs global state awareness and scheduling decisions; the service traceability module performs event chain recording, playback, and audit export; and the result output module outputs shared data objects, verification reports, inference results, and traceability reports.

[0058] In a preferred embodiment, the system can be deployed on a local hospital server, a cloud-edge collaborative node, or a dedicated medical intelligent computing device, and supports isolated deployment across multiple institutions. Furthermore, the system provided in this embodiment can be used to perform at least some of the steps of the method provided in Embodiment 1.

[0059] Example 3 This embodiment provides an apparatus including a processor, a memory, and a bus. The memory stores program instructions and related data for executing the above-described method steps. After the processor invokes the program instructions, it performs steps such as standardized data interaction, multi-system consistency verification, global resource status awareness, multi-agent collaborative scheduling, and service traceability recording.

[0060] Preferably, the device can be deployed on a local hospital server, a cloud-edge collaborative node, or a dedicated medical intelligent computing device, and supports extended deployment through containerization.

[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for cross-system medical data interaction and multi-agent control, characterized in that, include: S1. Receive raw data and raw task requests from multiple medical information systems, medical knowledge systems and intelligent agent service systems; A standardized data interaction interface is built based on a unified data model to classify and encapsulate different types of data; the different types of data include structured data, semi-structured data, and unstructured data. S2. The data accessed through the standardized data interaction interface is sent to the data platform, and the DataFlow pipeline is used to perform parsing, extraction, standardization, segmentation, recombination and unified representation processing on different types of data to form a standard data object that can be called by both the brain-like reasoning system and the multi-agent system. S3. Perform multi-system data consistency verification covering the data sending end, transmission link, processing node, inference node and result output end. Perform syntax consistency verification, pattern consistency verification, terminology encoding consistency verification, semantic evidence consistency verification and result traceability consistency verification on the standard data object to obtain a data object that passes the consistency test. S4. Construct an intelligent agent collaboration layer, collect the running status, tool capabilities, skill loading status, conversation queue length, computing resource consumption and task priority of each intelligent agent, and establish a global resource status awareness model. S5. Based on the global resource state awareness model, perform task decomposition, route scheduling, and resource allocation for each agent to generate cross-system data interaction results and multi-agent collaborative processing results; the agents include data access agents, knowledge extraction agents, terminology mapping agents, consistency verification agents, collaborative reasoning agents, and audit traceability agents. S6. Establish a service traceability record for the cross-system data interaction results and multi-agent collaborative processing results. The service traceability record includes at least the source system identifier, data version, input hash, output hash, calling agent identifier, calling tool identifier, prompt template version, timestamp, and evidence citation information, thereby forming a complete trajectory of data flow and agent operation.

2. The method according to claim 1, characterized in that, In step S1, the standardized data interaction interface adopts a layered adaptation structure, including an access adaptation layer, a standard encapsulation layer, a semantic mapping layer, and a security control layer. The access adaptation layer is used to connect to multiple medical information systems and medical knowledge systems, and supports API, message queue, file subscription or batch import methods; The standard encapsulation layer is used to encapsulate data from different sources into a unified data envelope object. Where H represents header metadata, P represents the original or parsed valid information payload, M represents metadata, which includes at least source, version, time, patient identifier, institution identifier and data type, and S represents security policy and access control information; The semantic mapping layer is used to map TCM terms, Western medicine terms, standard codes, and local extended fields to a unified semantic space; The security control layer is used to perform identity authentication, permission verification, de-identification rule matching, and transmission encryption.

3. The method according to claim 1, characterized in that, In S2, the unified representation processing specifically includes: For structured data, field standardization, null value handling, unit unification, and time axis alignment are completed according to preset field mapping rules; For semi-structured data, extract heading levels, paragraph boundaries, table structure, and key fields to form structured fragments; For unstructured data, the DataFlow-based knowledge extraction and normalization pipeline performs document segmentation, entity recognition, relation extraction, terminology normalization, and evidence fragment index construction. For knowledge items in Traditional Chinese Medicine and Western Medicine, a unified representation template of "symptoms-signs-syndromes-diseases-treatment-sources of evidence" is established, and the output is a sequence representation, graph structure representation, or a fusion of the two representations that can be processed by a brain-like system.

4. The method according to claim 1, characterized in that, The multi-system data consistency verification in S3 further includes: Data sender verification is used to perform field integrity checks on data before it is sent, as well as to generate signatures or hash digests. Transmission verification is used to perform verification and comparison, retransmission control, and version consistency checks on data packets during transmission. The processing and verification are used to perform pattern constraint checks, unit conversion verification, terminology coding consistency comparison, and cross-source conflict detection on standardized data. Reasoning verification is used to align and compare the output of an agent or model with original evidence fragments, clinical rules, and knowledge graph constraints. Result verification is used to verify whether the output results retain the source link, evidence citation, and version information.

5. The method according to claim 4, characterized in that, The inference verification includes result credibility control for suppressing model hallucinations: S301, The consistency verification agent extracts evidence requirement items from the key conclusions involved in the output data to be verified. S302. Retrieve original records, guideline provisions, book excerpts, and / or structured knowledge triples corresponding to the evidence requirement from standard data objects and knowledge indexes; S303. Based on evidence coverage, terminology matching degree, time consistency, and logical constraints, perform consistency calculations on each item of the output data to be verified. The consistency calculation method is as follows: in Representation of schema consistency, Indicates coding consistency. Indicates consistency of units. Indicates the score for evidence alignment. Indicates traceability completeness. to These are the weighting coefficients; S304. When the consistency calculation result is lower than the preset threshold, trigger at least one of the following processes: re-retrieval, regeneration, manual confirmation, or result downgrade output. S305. Only after attaching evidence citation marks to the verified output results should they be released to the public.

6. The method according to claim 1, characterized in that, In S4, the global resource state awareness model is built based on the OpenClaw gateway control plane and multi-agent routing method, and it is aware of at least the following state variables: agent online status, session activity, workspace isolation status, skill availability, tool permissions, CPU utilization, GPU memory usage, memory usage, task waiting time, historical success rate, output confidence and the security domain to which the task belongs. The objective function used to perform resource allocation is: in Indicates the benefits of task quality. Indicates the benefits of resource utilization. Indicates priority benefits. Indicates the cost of latency. Indicates the risk and cost. , , , , These are weight parameters; The objective function described above enables dynamic routing, parallel execution, load balancing, and failover for multi-agent tasks.

7. The method according to claim 1, characterized in that, In S6, the service traceability record adopts an event chain record structure, and any traceability event is recorded as: in, For link identification, This serves as the current event identifier. As the parent event identifier, For source system identification, To implement agent identification, The identifier for the skill being invoked. Identifier for the pipeline operator being invoked. and These are the input and output summaries, respectively. For timestamps, For strategy information, For the purpose of citing evidence, In execution status; The event chain record structure supports trajectory query, anomaly replay, responsibility location, version rollback, and audit export.

8. A cross-system medical data interaction and multi-agent control system, characterized in that, For performing the method as described in any one of claims 1-7, comprising: The data access module is used to receive multi-source medical data, medical documents, medical knowledge, and task requests from external systems. The standardization processing module is used to complete the parsing, extraction, normalization, and unified representation of multiple types of data through a standardized data interaction interface and the DataFlow pipeline; The consistency verification module is used to perform end-to-end multi-system data consistency verification and result reliability verification. The agent collaboration module is used to complete task decomposition, session orchestration, resource scheduling and fault switching based on the gateway control plane, multi-agent routing and multi-agent skill invocation, so as to achieve optimal allocation of multi-agent resources; The resource awareness module is used to build and update the global resource status awareness model; The service traceability module is used to record data flow, skill calls, tool calls, input / output summaries, and audit trails; The results output module is used to provide shared data objects, verification reports, collaborative processing results, traceability reports, and interface response results to external parties.

9. A cross-system medical data interaction and multi-agent control device, characterized in that, The system includes a processor, a memory, and a bus. The memory stores instructions and data read by the processor. The processor is used to invoke the instructions and data in the memory to execute the method as described in any one of claims 1-7. The bus is used to connect the functional components and transmit information.

Images