Multi-agent collaborative reasoning method and system for complex trauma treatment

Abstract

The present application relates to the technical field of medical artificial intelligence, and in particular to a multi-agent collaborative reasoning method and system for complex trauma treatment. The method assesses the complexity of multi-modal medical data and decomposes tasks by coordinating agents, distributes sub-tasks to each professional agent for reasoning within the scope of decision-making suggestions. The coordinating agent collects suggestions and coordinates conflicts to form a comprehensive decision-making scheme, and dynamically adjusts the task decomposition and execution sequence according to the state update information after execution. The present application realizes the cooperation and dynamic adjustment of multi-professional knowledge, and improves the accuracy and timeliness of complex trauma treatment decision-making.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, and in particular to a multi-agent collaborative reasoning method and system for complex trauma treatment. Background Technology

[0002] In the field of complex trauma care, existing technologies typically rely on single-agent systems or fixed-process clinical decision support tools. These technologies aim to process patient data and provide diagnostic or treatment recommendations. Specifically, the conventional approach is to build a centralized, relatively single-function agent or decision model. This model receives multimodal medical data input from monitoring equipment, imaging examinations, laboratory reports, etc., and attempts to directly output a comprehensive treatment plan covering multiple professional fields through an integrated inference engine. This architecture attempts to simulate an experienced trauma care expert who oversees the overall situation, processes all information at once, and makes a judgment. The entire decision-making process is often linear—data input, model processing, and plan output—lacking explicit modeling of the complex structure and dynamic changes within the treatment task.

[0003] However, the aforementioned conventional practices have significant shortcomings. Complex trauma care involves multiple highly specialized medical fields, including surgery, emergency medicine, intensive care, anesthesia, and imaging, resulting in a vast knowledge base and diverse decision-making logics. Single-agent models struggle to deeply integrate such heterogeneous expertise, and their reasoning processes tend to become general and superficial. This leads to decision recommendations lacking depth and precision in specific professional dimensions, neglecting crucial professional details or cutting-edge treatment guidelines. Furthermore, trauma patients' conditions change rapidly, and treatment itself is a dynamic process involving multiple stages and iterative iterations, including assessment, resuscitation, definitive surgery, and monitoring. Existing single-agent systems or fixed-process tools lack effective task decomposition and dynamic coordination mechanisms. They cannot flexibly reorganize the priority and execution logic of treatment tasks based on real-time changes in the patient's condition (such as the emergence of new complications or poor response to initial treatment). This limits the adaptability and timeliness of their decisions, making it difficult to meet the clinical need for precise, personalized, and adaptive collaborative decision support. Summary of the Invention

[0004] The embodiments of the present invention provide a multi-agent collaborative reasoning method and system for complex trauma treatment, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a multi-agent cooperative reasoning method for complex trauma treatment, comprising:

[0006] Receive multimodal medical data from trauma patients;

[0007] The complexity of the multimodal medical data is assessed by a coordinating intelligent agent. Based on the complexity assessment results, the treatment task is decomposed into multiple sub-tasks, and the execution order and dependencies of the multiple sub-tasks are determined.

[0008] Based on the task description text of the multiple sub-tasks, each sub-task and its associated medical knowledge evidence are distributed to the corresponding professional intelligent agents. Each professional intelligent agent then performs reasoning on the assigned sub-task based on predefined medical role permissions to generate professional decision-making suggestions.

[0009] The coordinating agent collects the professional decision-making suggestions returned by each professional agent, determines whether there are conflicts between the professional decision-making suggestions, and when a conflict is detected, a coordinated comprehensive decision-making scheme is determined by the professional agents and the coordinating agent.

[0010] The comprehensive decision-making scheme is executed and new medical data is acquired as status update information. When the status update information meets the preset status change conditions, the task is decomposed and the execution order is adjusted according to the status update information.

[0011] The complexity of the multimodal medical data is assessed by a coordinating intelligent agent. Based on the complexity assessment results, the treatment task is decomposed into multiple sub-tasks, and the execution order and dependencies of the multiple sub-tasks are determined, including:

[0012] The coordinating agent is used to perform cross-modal consistency analysis and decision path conflict detection on the multimodal medical data. By calculating semantic contradiction indicators and treatment plan mutual exclusion indicators between different modal data, a complexity assessment result including data consistency score and decision conflict score is determined.

[0013] The number of sub-task levels is determined based on the data consistency score, and the number of sub-tasks within each level is determined based on the decision conflict score, thus generating a task decomposition structure containing multiple levels of sub-tasks.

[0014] Resource requirement analysis and temporal dependency analysis are performed on each subtask in the task decomposition structure to determine a subtask dependency graph containing node weights and edge weights. The node weights represent the resource consumption of the subtasks, and the edge weights represent the dependency strength between the subtasks.

[0015] The subtask dependency graph is constrained and solved to generate a full subtask execution sequence under the premise of satisfying the preceding execution relationship. The subtask execution sequence represents the execution order and dependency relationship of multiple subtasks.

[0016] Resource requirement analysis and temporal dependency analysis are performed on each subtask in the task decomposition structure to determine the subtask dependency graph, which includes node weights and edge weights, including:

[0017] For each subtask in the task decomposition structure, a mapping table between subtask identifiers and required resource type identifiers is determined.

[0018] Based on the number of resource types corresponding to each subtask in the mapping table and the scarcity coefficient of each resource type, the resource competition index of each subtask is calculated as the node weight in the subtask dependency graph;

[0019] Traverse each subtask in the task decomposition structure, determine the data field names of the input interface and the data field names of the output interface of each subtask, and establish directed edges in the subtask dependency graph based on the output interface and the data field names of the output interface of each subtask;

[0020] For each directed edge in the subtask dependency graph, determine the number of matches between the output data field of the starting subtask and the input data field of the ending subtask, and calculate the ratio of the number of matches to the total number of input data fields of the ending subtask as the data dependency.

[0021] The timing urgency is calculated based on the execution time of the starting subtask and the tolerable waiting time of the ending subtask. The edge weights of the directed edges are determined by the data dependency and the timing urgency to obtain the subtask dependency graph.

[0022] Based on the task description text of the multiple sub-tasks, each sub-task and its associated medical knowledge evidence are distributed to the corresponding professional intelligent agents. Each professional intelligent agent then performs reasoning on the assigned sub-task based on predefined medical role permissions, generating professional decision-making suggestions, including:

[0023] Based on the task description text of each of the multiple subtasks, medical professional field keywords and diagnosis and treatment operation keywords are obtained;

[0024] Based on a pre-constructed professional domain knowledge graph, the keywords in the medical professional domain are semantically expanded. The keywords in the medical professional domain and the keywords in the diagnosis and treatment operation are matched and calculated with the capability tags of the professional intelligent agents to obtain the matching degree matrix between each subtask and each professional intelligent agent.

[0025] Based on the matching degree matrix, the professional agent with the highest matching degree is selected as the executing agent for each subtask, and medical knowledge evidence related to the semantics of the task description text of the subtask is retrieved for each subtask.

[0026] Send the task description text of each subtask, the medical knowledge evidence corresponding to the subtask, and the medical role permission configuration information of the executing agent of the subtask to the executing agent;

[0027] The executing agent performs logical reasoning on the task description text and the medical knowledge evidence based on the medical role permission configuration information, obtains intermediate reasoning conclusions, and generates professional decision-making suggestions based on the intermediate reasoning conclusions and the medical knowledge evidence.

[0028] The executing intelligent agent performs logical reasoning on the task description text and the medical knowledge evidence based on the medical role permission configuration information, obtains intermediate reasoning conclusions, and generates professional decision-making suggestions based on the intermediate reasoning conclusions and the medical knowledge evidence, including:

[0029] The executing agent generates a set of permission constraints, which includes data access constraint rules and operation execution constraint rules, based on the data access permission list and operation execution permission list in the medical role permission configuration information.

[0030] The executing agent parses the task description text into an initial question statement and retrieves knowledge fragments from the medical knowledge evidence that are semantically related to the initial question statement.

[0031] Based on the set of permission constraints, data content in the knowledge fragment that exceeds the scope of accessible medical data is filtered to obtain valid knowledge fragments that conform to the permission constraints;

[0032] The executing agent determines the inference chain based on the valid knowledge fragments, verifies whether the required data is within the scope of the accessible medical data and whether the required operation is within the scope of the executable diagnostic and treatment operations in each inference step of the inference chain, and determines the intermediate inference conclusions generated by all verified inference steps;

[0033] The executing agent extracts evidence text fragments and evidence source identifiers that support the intermediate reasoning conclusion from the medical knowledge evidence, and combines the intermediate reasoning conclusion, the evidence text fragments, and the evidence source identifiers to generate professional decision-making suggestions.

[0034] The coordinating agent collects the professional decision-making suggestions returned by each specialized agent, determines whether there are conflicts among the professional decision-making suggestions, and when a conflict is detected, determines a coordinated comprehensive decision-making scheme through the specialized agents and the coordinating agent, including:

[0035] The coordinating agent performs semantic parsing on the professional decision-making suggestions returned by each professional agent, and extracts the decision-making elements from each professional decision-making suggestion;

[0036] The extracted decision elements are correlated with a pre-constructed medical constraint knowledge graph. When there are taboo or mutually exclusive relationships between the decision elements of different professional decision suggestions in the medical constraint knowledge graph, it is determined that there is a conflict between the professional decision suggestions, and the decision elements with taboo or mutually exclusive relationships are marked as conflicting elements.

[0037] The coordinating agent sends a consultation request, containing professional decision-making suggestions from the conflicting element and other professional agents, to the specialized agents corresponding to the conflicting elements.

[0038] The specialized intelligent agents corresponding to the conflicting elements generate decision adjustment schemes based on the strength of the role of the conflicting elements in their own decision-making logic.

[0039] The coordinating agent modifies the conflicting professional decision-making suggestions according to the decision adjustment scheme. After verifying that the modification results meet the constraints through the medical constraint knowledge graph, the modified professional decision-making suggestions are combined to generate a coordinated comprehensive decision-making scheme.

[0040] The comprehensive decision-making scheme is executed, and new medical data is acquired as status update information. When the status update information meets preset status change conditions, the task decomposition and execution order are readjusted based on the status update information, including:

[0041] The comprehensive decision-making scheme is executed, and the patient's physiological index data and treatment response data are collected, which are then used as new medical data.

[0042] Time-series analysis is performed on the new medical data to obtain key change features, and status update information is determined based on these key change features.

[0043] Based on the abnormal change characteristics in the status update information, when the abnormal change characteristics meet the conditions for determining disease deterioration or treatment ineffectiveness, it is determined that the status update information meets the preset status change conditions;

[0044] When it is determined that the status update information meets the preset status change conditions, the abnormal change features in the status update information are compared and analyzed with the target constraints of the original medical task to determine the degree of target deviation caused by the abnormal change features.

[0045] The original medical task is decomposed into an adjusted set of subtasks based on the degree of deviation from the target. The execution order is then recalculated based on the dependencies between the adjusted subtask sets and the resource availability information in the status update information.

[0046] A second aspect of the present invention provides a multi-agent collaborative reasoning system for complex trauma treatment, comprising:

[0047] A data receiving unit is used to receive multimodal medical data from trauma patients;

[0048] The task decomposition unit is used to assess the complexity of the multimodal medical data through a coordinating agent, decompose the treatment task into multiple sub-tasks based on the complexity assessment results, and determine the execution order and dependencies of the multiple sub-tasks.

[0049] The task distribution unit is used to distribute each subtask and its associated medical knowledge evidence to the corresponding professional intelligent agent based on the task description text of the multiple subtasks. Each professional intelligent agent then performs reasoning on the assigned subtask based on predefined medical role permissions to generate professional decision suggestions.

[0050] The conflict coordination unit is used to collect professional decision-making suggestions returned by each professional intelligent agent by the coordination intelligent agent, determine whether there is a conflict between the professional decision-making suggestions, and when a conflict is detected, determine a coordinated comprehensive decision-making scheme through each professional intelligent agent and the coordination intelligent agent.

[0051] The dynamic adjustment unit is used to execute the comprehensive decision-making scheme and acquire new medical data as status update information. When the status update information meets the preset status change conditions, the task decomposition and execution order are readjusted according to the status update information.

[0052] A third aspect of the present invention provides an electronic device, comprising:

[0053] processor;

[0054] Memory used to store processor-executable instructions;

[0055] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0056] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0057] This invention significantly improves decision-making efficiency and collaborative processing capabilities in complex trauma care. By coordinating an intelligent agent to assess the complexity of multimodal medical data, it achieves precise decomposition and structured orchestration of treatment tasks. Tasks are broken down into multiple sub-tasks with clear execution sequences and dependencies, avoiding confusion and resource conflicts in the treatment process and ensuring the orderliness and timeliness of treatment actions. This dynamic task planning mechanism can adapt to the rapidly changing nature of patients' conditions in trauma care, laying a solid foundation for subsequent parallelized professional processing.

[0058] In this invention, each subtask and its associated medical knowledge evidence are precisely distributed to the corresponding professional intelligent agents. Each agent independently reasones based on its predefined medical role permissions, generating highly specialized decision recommendations. This role-based task allocation ensures the professionalism and authority of the decisions, fully utilizes in-depth knowledge from various fields, and avoids the drawbacks of limited knowledge in a single decision model. Simultaneously, permission constraints guarantee the security and compliance of the decision-making process, preventing unauthorized operations.

[0059] In this invention, the coordinating agent is responsible for collecting and integrating decision-making suggestions from various specialties and proactively detecting conflicts between different suggestions. When a conflict is detected, a coordination mechanism is activated, guiding relevant specialty agents to interact with the coordinating agent. Through negotiation or based on preset rules, a consensus is reached, ultimately forming a coordinated comprehensive decision-making plan. This process effectively solves the common problem of differing opinions in multidisciplinary consultations, enabling rapid transformation from multiple expert opinions to a unified and executable treatment plan, thus improving the efficiency of team collaboration and the consistency of decision-making.

[0060] This invention possesses dynamic feedback and adjustment capabilities. After executing a comprehensive decision-making plan, it acquires new medical data as status update information and continuously monitors whether preset status change conditions are met. Once the conditions are met, it can immediately trigger a reassessment of the current treatment task, decomposing tasks and adjusting the execution order in real time based on the latest patient status information. This allows the entire treatment process to dynamically evolve closely following the development of the patient's condition, forming a closed-loop optimization of "assessment-decision-execution-reassessment," greatly enhancing the adaptability and robustness of the treatment system to complex and uncertain conditions, and providing patients with a continuous, precise, and personalized treatment path. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating an embodiment of the multi-agent collaborative reasoning method for complex trauma treatment according to the present invention.

[0062] Figure 2 This is a flowchart illustrating the process of determining the subtask dependency graph according to the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0065] Figure 1 This is a flowchart illustrating an embodiment of the multi-agent collaborative reasoning method for complex trauma treatment according to the present invention. Figure 1 As shown, the method includes:

[0066] Receive multimodal medical data from trauma patients;

[0067] The complexity of the multimodal medical data is assessed by a coordinating intelligent agent. Based on the complexity assessment results, the treatment task is decomposed into multiple sub-tasks, and the execution order and dependencies of the multiple sub-tasks are determined.

[0068] Based on the task description text of the multiple sub-tasks, each sub-task and its associated medical knowledge evidence are distributed to the corresponding professional intelligent agents. Each professional intelligent agent then performs reasoning on the assigned sub-task based on predefined medical role permissions to generate professional decision-making suggestions.

[0069] The coordinating agent collects the professional decision-making suggestions returned by each professional agent, determines whether there are conflicts between the professional decision-making suggestions, and when a conflict is detected, a coordinated comprehensive decision-making scheme is determined by the professional agents and the coordinating agent.

[0070] The comprehensive decision-making scheme is executed and new medical data is acquired as status update information. When the status update information meets the preset status change conditions, the task is decomposed and the execution order is adjusted according to the status update information.

[0071] The complexity of the multimodal medical data is assessed by a coordinating intelligent agent. Based on the complexity assessment results, the treatment task is decomposed into multiple sub-tasks, and the execution order and dependencies of the multiple sub-tasks are determined, including:

[0072] The coordinating agent is used to perform cross-modal consistency analysis and decision path conflict detection on the multimodal medical data. By calculating semantic contradiction indicators and treatment plan mutual exclusion indicators between different modal data, a complexity assessment result including data consistency score and decision conflict score is determined.

[0073] The number of sub-task levels is determined based on the data consistency score, and the number of sub-tasks within each level is determined based on the decision conflict score, thus generating a task decomposition structure containing multiple levels of sub-tasks.

[0074] Resource requirement analysis and temporal dependency analysis are performed on each subtask in the task decomposition structure to determine a subtask dependency graph containing node weights and edge weights. The node weights represent the resource consumption of the subtasks, and the edge weights represent the dependency strength between the subtasks.

[0075] The subtask dependency graph is constrained and solved to generate a full subtask execution sequence under the premise of satisfying the preceding execution relationship. The subtask execution sequence represents the execution order and dependency relationship of multiple subtasks.

[0076] After receiving multimodal medical data including CT images, vital sign monitoring data, and laboratory test results, the coordinating agent initiates a cross-modal consistency analysis process. This process uses a semantic mapping network to convert different modalities of data into a unified medical concept space representation. For example, it correlates and matches the spleen rupture detected in CT images with the decreasing trend of hemoglobin in blood tests. The semantic contradiction index is obtained by calculating the semantic distance between the diagnostic conclusions supported by different modalities of data. When CT shows no obvious bleeding but blood pressure continues to drop, the contradiction index value increases. Decision path conflict detection is achieved by constructing a decision tree for treatment plans. For multiple treatment paths with the same symptom, mutual exclusivity is determined. For example, the degree of mutual exclusivity between immediate surgery and conservative observation is calculated by the overlap of their expected treatment effects. The data consistency score is obtained by weighted summation of the evidence strength of each modality of data supporting the same diagnostic conclusion, with a value ranging from 0 to 1. The decision conflict score is calculated by counting the number of mutually exclusive treatment plan pairs and combining their differences in clinical priority.

[0077] When the data consistency score is below 0.6, it indicates significant inconsistency in the multimodal data, requiring the task to be divided into a three-tiered structure: data verification, diagnostic reasoning, and treatment decision-making. When the decision conflict score is above 0.4, the diagnostic reasoning layer needs to be further refined into multiple parallel sub-tasks, such as organ system assessment, injury severity grading, and complication prediction. The task decomposition structure adopts a tree-like organization, with the root node representing the overall treatment task, the second layer consisting of sub-tasks divided according to physiological systems, and the third layer representing specific examination, diagnostic, or treatment procedures.

[0078] When performing resource requirement analysis on each subtask, the required types of medical equipment, professional staff configuration, and estimated execution time are statistically analyzed. For example, the abdominal CT scan subtask requires imaging equipment resources, and the resource usage is quantified into node weight values ​​based on equipment usage time. Temporal dependency analysis identifies the preconditional relationships between subtasks. For instance, pathological biopsy must be performed after image localization is completed. The degree of mandatory nature of this dependency is used as the edge weight, with a value of 1 indicating strict dependency and 0.5 indicating advisory dependency. The subtask dependency graph is stored using a directed acyclic graph structure. Nodes contain task identifiers, resource requirement vectors, and expected execution time attributes, while edges contain dependency type and waiting time tolerance attributes.

[0079] The constraint-solving process employs a topological sorting algorithm to handle strict dependencies. For cases with multiple feasible paths, a priority scheduling strategy selects the execution sequence with the least resource conflict. When a resource bottleneck is detected, the start time of parallel subtasks is adjusted to avoid excessive consumption of critical resources. The generated subtask execution sequence is output as a timestamp-annotated task list, including the earliest start time, latest completion time, and preceding task identifier for each subtask, ensuring efficient coordination of the treatment process while meeting medical safety constraints.

[0080] Figure 2 This is a schematic diagram illustrating the process of determining the subtask dependency graph according to the present invention. Figure 2 As shown, resource requirement analysis and temporal dependency analysis are performed on each subtask in the task decomposition structure to determine the subtask dependency graph containing node weights and edge weights, including:

[0081] For each subtask in the task decomposition structure, a mapping table between subtask identifiers and required resource type identifiers is determined.

[0082] Based on the number of resource types corresponding to each subtask in the mapping table and the scarcity coefficient of each resource type, the resource competition index of each subtask is calculated as the node weight in the subtask dependency graph;

[0083] Traverse each subtask in the task decomposition structure, determine the data field names of the input interface and the data field names of the output interface of each subtask, and establish directed edges in the subtask dependency graph based on the output interface and the data field names of the output interface of each subtask;

[0084] For each directed edge in the subtask dependency graph, determine the number of matches between the output data field of the starting subtask and the input data field of the ending subtask, and calculate the ratio of the number of matches to the total number of input data fields of the ending subtask as the data dependency.

[0085] The timing urgency is calculated based on the execution time of the starting subtask and the tolerable waiting time of the ending subtask. The edge weights of the directed edges are determined by the data dependency and the timing urgency to obtain the subtask dependency graph.

[0086] After the task decomposition structure is constructed, resource requirements analysis and temporal dependency analysis are performed on each subtask. The resource requirements analysis establishes a mapping table between the identifier of each subtask and the required resource type identifiers. Resource types include equipment resources, human resources, and information resources. For example, the resource type identifiers corresponding to the subtask "Image Diagnosis" include CT equipment, radiologists, and PACS systems. The mapping table is stored in key-value pair format, where the key is the unique identifier of the subtask, and the value is the set of all resource type identifiers required by that subtask.

[0087] The resource competition index is calculated based on the number of resource types corresponding to each subtask in the mapping table and the scarcity coefficient of each resource type. The scarcity coefficient is dynamically determined according to the availability of each resource in the current medical environment, and its value ranges from 0 to 1, with a larger value indicating a scarcer resource. The calculation formula is to sum the scarcity coefficients of all resource types required by the subtask and divide by the total number of resource types to obtain the resource competition index of the subtask. This index serves as the node weight of the corresponding node in the subtask dependency graph, reflecting the strength of the resource constraints on task execution.

[0088] Establishing directed edges between subtasks requires traversing each subtask in the task decomposition structure to obtain the input and output interface definitions for each subtask. The input interface contains the names of all data fields required for the subtask's execution, and the output interface contains the names of all data fields generated upon completion of the subtask. By comparing the output interface data fields of each subtask with the input interface data fields of other subtasks, when a certain output data field of subtask A matches a certain input data field of subtask B, a directed edge from A to B is established in the subtask dependency graph, indicating that B depends on the output data of A.

[0089] For each established directed edge, edge weights are calculated to quantify the strength of the dependency relationship. Data dependency is calculated by counting the number of matches between the output data fields of the starting subtask and the input data fields of the ending subtask, and then dividing this number by the total number of input data fields in the ending subtask to obtain the data dependency value. Temporal urgency is calculated based on the estimated execution time of the starting subtask and the tolerable waiting time of the ending subtask, by dividing the execution time by the tolerable waiting time; a larger ratio indicates a tighter temporal constraint. Multiplying the data dependency by the temporal urgency yields the edge weight of the directed edge, comprehensively characterizing the dependency strength and temporal sensitivity between subtasks, ultimately forming a subtask dependency graph containing node weights and edge weights.

[0090] Based on the task description text of the multiple sub-tasks, each sub-task and its associated medical knowledge evidence are distributed to the corresponding professional intelligent agents. Each professional intelligent agent then performs reasoning on the assigned sub-task based on predefined medical role permissions, generating professional decision-making suggestions, including:

[0091] Based on the task description text of each of the multiple subtasks, medical professional field keywords and diagnosis and treatment operation keywords are obtained;

[0092] Based on a pre-constructed professional domain knowledge graph, the keywords in the medical professional domain are semantically expanded. The keywords in the medical professional domain and the keywords in the diagnosis and treatment operation are matched and calculated with the capability tags of the professional intelligent agents to obtain the matching degree matrix between each subtask and each professional intelligent agent.

[0093] Based on the matching degree matrix, the professional agent with the highest matching degree is selected as the executing agent for each subtask, and medical knowledge evidence related to the semantics of the task description text of the subtask is retrieved for each subtask.

[0094] Send the task description text of each subtask, the medical knowledge evidence corresponding to the subtask, and the medical role permission configuration information of the executing agent of the subtask to the executing agent;

[0095] The executing agent performs logical reasoning on the task description text and the medical knowledge evidence based on the medical role permission configuration information, obtains intermediate reasoning conclusions, and generates professional decision-making suggestions based on the intermediate reasoning conclusions and the medical knowledge evidence.

[0096] When distributing and reasoning about subtasks, keywords are first extracted from the task description text. For this text, named entity recognition technology based on a medical domain dictionary is used to extract medical professional domain keywords and diagnostic / treatment operation keywords. Medical professional domain keywords include, but are not limited to, organ system names such as "heart," "respiratory system," and "nervous system," as well as pathological state terms such as "hemorrhagic shock" and "traumatic brain injury." Diagnostic / treatment operation keywords include specific operation names such as "endotracheal intubation," "debridement and suturing," and "infusion resuscitation."

[0097] After extracting keywords, a pre-constructed professional domain knowledge graph is used to semantically expand the medical professional domain keywords. This knowledge graph uses a triplet structure to store medical concepts and their relationships, such as "hemorrhagic shock - belongs to - circulatory system diseases" and "tracheal intubation - applicable to - respiratory failure". Through semantic links such as "hyper-paragraph relationships", "synonyms", and "related relationships" in the graph, the initial keywords are expanded into an expanded vocabulary containing their parent concepts, child concepts, and related concepts. For example, "traumatic brain injury" can be expanded into related terms such as "neurosurgery", "cerebral contusion and laceration", and "intracranial hemorrhage".

[0098] After semantic expansion, the expanded medical domain keywords and diagnostic operation keywords are matched with the capability tags of each professional agent. Each professional agent is assigned a set of capability tags during system initialization; for example, the capability tags of the neurosurgical agent include "traumatic brain injury," "intracranial pressure monitoring," and "craniotomy." Vectorization methods are used to convert keywords and capability tags into word embedding vectors, and the matching score is obtained by calculating cosine similarity. The matching scores of all subtasks and all professional agents are organized into a matrix, where rows represent subtasks, columns represent professional agents, and element values ​​are the corresponding matching scores.

[0099] Based on the matching degree matrix, the professional agent with the highest matching degree is selected as the executing agent for each subtask. When the matching degree of a subtask is close with multiple professional agents, the agent with the lowest current load is selected first to achieve load balancing. For each subtask, medical knowledge evidence related to the semantics of the task description text is retrieved from the medical knowledge base. The retrieval process uses dense vector retrieval technology, encoding the task description text into vectors, and searching the knowledge base for the several medical documents, clinical guidelines, or case records with the highest cosine similarity as knowledge evidence.

[0100] When distributing tasks, a message data packet is constructed containing task description text, medical knowledge evidence, and medical role permission configuration information. The medical role permission configuration information defines the data scope that the executing agent can access, the types of operations that can be performed, and the permission boundaries for decision-making. For example, the permission configuration of an orthopedic agent allows it to access imaging examination data and formulate fracture fixation plans, but it does not have the right to assess cardiac function.

[0101] After receiving the message data packet, the executing agent filters available medical knowledge evidence based on the medical role's permission configuration information, ensuring that only information within its authorized scope is processed. The filtered task description text and medical knowledge evidence are input into the inference module. A semantic understanding model based on an attention mechanism is used to extract key information, and inference is performed using a logical rule base. The inference process generates intermediate inference conclusions, such as "The patient has active bleeding and a persistent drop in blood pressure, requiring immediate hemostasis." Based on these intermediate inference conclusions and the treatment plan from the medical knowledge evidence, a professional decision-making recommendation containing diagnostic basis, treatment suggestions, and expected outcomes is generated and returned to the coordinating agent.

[0102] The executing intelligent agent performs logical reasoning on the task description text and the medical knowledge evidence based on the medical role permission configuration information, obtains intermediate reasoning conclusions, and generates professional decision-making suggestions based on the intermediate reasoning conclusions and the medical knowledge evidence, including:

[0103] The executing agent generates a set of permission constraints, which includes data access constraint rules and operation execution constraint rules, based on the data access permission list and operation execution permission list in the medical role permission configuration information.

[0104] The executing agent parses the task description text into an initial question statement and retrieves knowledge fragments from the medical knowledge evidence that are semantically related to the initial question statement.

[0105] Based on the set of permission constraints, data content in the knowledge fragment that exceeds the scope of accessible medical data is filtered to obtain valid knowledge fragments that conform to the permission constraints;

[0106] The executing agent determines the inference chain based on the valid knowledge fragments, verifies whether the required data is within the scope of the accessible medical data and whether the required operation is within the scope of the executable diagnostic and treatment operations in each inference step of the inference chain, and determines the intermediate inference conclusions generated by all verified inference steps;

[0107] The executing agent extracts evidence text fragments and evidence source identifiers that support the intermediate reasoning conclusion from the medical knowledge evidence, and combines the intermediate reasoning conclusion, the evidence text fragments, and the evidence source identifiers to generate professional decision-making suggestions.

[0108] After receiving the task description text and medical knowledge evidence distributed by the coordinating agent, the executing agent first parses the medical role permission configuration information. This configuration information includes a data access permission list and an operation execution permission list. The data access permission list defines the range of medical data types that the agent can read. For example, a radiologist agent can only access medical image data, image diagnosis reports, and related examination records, but cannot access surgical records or medication information. The operation execution permission list defines the types of diagnostic and treatment operations that the agent can suggest. For example, a surgeon agent can suggest surgical plans, but cannot directly recommend drug prescriptions. The executing agent constructs a set of permission constraints based on these two lists. The data access constraint rules are represented by a whitelist of accessible data fields, and the operation execution constraint rules are represented by a set of allowed operation type codes.

[0109] The executing agent parses the task description text into a structured initial problem statement, extracting key medical entities and treatment objectives. For example, the task description "assess the surgical indications for a patient with a left femoral fracture" is parsed into a structured representation containing the anatomical location "left femur," the injury type "fracture," and the treatment intention "surgical indication assessment." Subsequently, semantically relevant knowledge fragments are retrieved from a medical knowledge evidence base. The retrieval process uses vectorized representations of medical terms to calculate semantic similarity; fragments with a similarity exceeding a threshold of 0.75 are considered relevant. The retrieved knowledge fragments include medical knowledge items such as fracture classification standards, rules for determining surgical indications, and lists of surgical contraindications.

[0110] The executing agent performs permission checks on each retrieved knowledge fragment, iterating through the medical data fields referenced in the fragment and matching them against accessible fields in the data access permission list. If a knowledge fragment contains data types that the executing agent is not authorized to access, that portion is marked as restricted and removed from the knowledge fragment. For example, when an orthopedic surgeon agent is processing a surgical indication assessment task, if a knowledge rule relies on a patient's psychiatric medication history, and this data is outside its access permissions, that rule is filtered out. The knowledge fragments retained after permission filtering constitute the set of valid knowledge fragments.

[0111] The executing agent constructs inference chains based on valid knowledge fragments. Each inference chain consists of multiple inference steps, representing a logical process of deriving a new conclusion from currently known information. During the construction of the inference chain, the executing agent performs dual verification on each inference step: first, it checks whether the input data required for the step falls within the scope of accessible medical data by comparing data field identifiers with data access permission lists; second, it checks whether the diagnostic and treatment operation suggestions involved in the step fall within the scope of executable diagnostic and treatment operations by comparing operation type codes with operation execution permission lists. Only when a inference step passes both verifications is it included in the valid inference chain, and the resulting inference is stored as an intermediate inference conclusion. For example, when assessing surgical indications, intermediate inference conclusions include progressively derived judgments such as "the patient's fracture type is AO classification C3" and "there is an absolute indication for surgery."

[0112] The executing agent locates the source of evidence supporting each intermediate inference conclusion from medical knowledge evidence. For each intermediate inference conclusion, it traces the knowledge fragment upon which the reasoning is based, extracts the original text content of that knowledge fragment as the evidence text fragment, and records the source identifier of that knowledge. The source identifier includes tracing information such as the knowledge base name, document number, and chapter position. For example, the evidence text fragment is "AO classification C3 type femoral fracture with articular surface comminutedness requires emergency surgical reduction," and the evidence source identifier is "《Orthopedic Diagnosis and Treatment Guidelines》3rd Edition - Page 127 - Section 4.3". The executing agent assembles the intermediate inference conclusion, the evidence text fragment supporting the conclusion, and the evidence source identifier according to a predefined structured format to form a complete professional decision recommendation and returns it to the coordinating agent.

[0113] The coordinating agent collects the professional decision-making suggestions returned by each specialized agent, determines whether there are conflicts among the professional decision-making suggestions, and when a conflict is detected, determines a coordinated comprehensive decision-making scheme through the specialized agents and the coordinating agent, including:

[0114] The coordinating agent performs semantic parsing on the professional decision-making suggestions returned by each professional agent, and extracts the decision-making elements from each professional decision-making suggestion;

[0115] The extracted decision elements are correlated with a pre-constructed medical constraint knowledge graph. When there are taboo or mutually exclusive relationships between the decision elements of different professional decision suggestions in the medical constraint knowledge graph, it is determined that there is a conflict between the professional decision suggestions, and the decision elements with taboo or mutually exclusive relationships are marked as conflicting elements.

[0116] The coordinating agent sends a consultation request, containing professional decision-making suggestions from the conflicting element and other professional agents, to the specialized agents corresponding to the conflicting elements.

[0117] The specialized intelligent agents corresponding to the conflicting elements generate decision adjustment schemes based on the strength of the role of the conflicting elements in their own decision-making logic.

[0118] The coordinating agent modifies the conflicting professional decision-making suggestions according to the decision adjustment scheme. After verifying that the modification results meet the constraints through the medical constraint knowledge graph, the modified professional decision-making suggestions are combined to generate a coordinated comprehensive decision-making scheme.

[0119] After receiving professional decision-making suggestions from various specialized agents, the coordinating agent initiates a conflict detection and coordination process. It performs semantic analysis on each professional decision-making suggestion, transforming the unstructured decision text into structured decision elements using natural language processing technology. Specifically, it extracts key elements such as "emergency laparotomy" and "preoperative blood transfusion of 2 units" from the trauma surgery agent's suggestions; elements such as "general anesthesia" and "propofol (an anesthetic induction drug)" from the anesthesia agent's suggestions; and elements such as "enhanced CT scan" and "iodine contrast agent injection" from the imaging diagnosis agent's suggestions. Each decision element includes attribute fields such as operation type, time constraint, drug name, and dosage parameters.

[0120] The extracted decision elements are input into a medical constraint knowledge graph for correlation queries. This knowledge graph is pre-constructed with medical rules such as drug interactions, operational timing constraints, and physiological indicator thresholds. For example, when querying the relationship between "iodine contrast agent injection" and "acute renal failure," the knowledge graph returns a contraindication flag because the patient's current serum creatinine level is 280 μmol / L, and the use of iodine contrast agents carries the risk of contrast-induced nephropathy. Similarly, when querying the relationship between "preoperative blood transfusion" and "enhanced CT scan," the knowledge graph returns a temporal mutual exclusion relationship because blood transfusion can cause hemodynamic fluctuations, affecting image quality. These decision elements with contraindications or mutual exclusions are marked as conflicting elements; in this example, "iodine contrast agent injection" and "preoperative blood transfusion-CT scan timing" are marked as conflicting elements.

[0121] The coordinating agent generates a negotiation request containing detailed information on the conflicting elements, specifying the conflict type, the decision-making factors involved, and relevant suggestions from other specialized agents. A negotiation request is sent to the imaging diagnostic agent, informing it of the conflict between iodine contrast agent use and the patient's renal function status; negotiation requests are simultaneously sent to the trauma surgery and imaging diagnostic agents, explaining the timing conflict between blood transfusion and imaging examinations.

[0122] After receiving the negotiation request, each specialized agent assesses the strength of conflicting factors in its decision-making based on its domain knowledge. The imaging diagnostic agent calculates that although enhanced CT can more clearly show intra-abdominal bleeding, given the patient's relatively stable hemodynamics and the preliminary ultrasound location of the bleeding source, it can be downgraded to plain CT. The strength of the iodine contrast agent is assessed as moderate, generating an adjustment plan of "replacing enhanced CT with plain CT." The trauma surgery agent, after evaluation, believes that identifying the bleeding site preoperatively is crucial for surgical planning. The strength of imaging examinations is assessed as high priority, but it is acceptable to complete a rapid CT scan first and then prepare for blood transfusion, generating an adjustment plan of "adjusting the sequence to CT scan first, with blood transfusion performed en route to the operating room."

[0123] The coordinating agent integrates various decision adjustment schemes, revising the original professional decision recommendations. The imaging examination plan is revised to "plain CT scan to assess abdominal injury," and the timing of blood transfusion and imaging examination is revised to "completing the CT scan within 15 minutes, and immediately establishing a second intravenous access to initiate blood transfusion after the scan." After revision, each decision element is resubmitted to the medical constraint knowledge graph for verification, confirming that the revised scheme no longer triggers contraindications or mutually exclusive constraints. Once verification is successful, the revised professional decision recommendations are integrated according to time priority and logical dependencies to generate a comprehensive decision scheme containing the complete process of "immediate plain CT scan - simultaneous preparation of the operating room - blood transfusion during transport - initiation of surgery within 30 minutes," ensuring that the professional decisions are conflict-free on the timeline and comply with medical safety regulations.

[0124] The comprehensive decision-making scheme is executed, and new medical data is acquired as status update information. When the status update information meets preset status change conditions, the task decomposition and execution order are readjusted based on the status update information, including:

[0125] The comprehensive decision-making scheme is executed, and the patient's physiological index data and treatment response data are collected, which are then used as new medical data.

[0126] Time-series analysis is performed on the new medical data to obtain key change features, and status update information is determined based on these key change features.

[0127] Based on the abnormal change characteristics in the status update information, when the abnormal change characteristics meet the conditions for determining disease deterioration or treatment ineffectiveness, it is determined that the status update information meets the preset status change conditions;

[0128] When it is determined that the status update information meets the preset status change conditions, the abnormal change features in the status update information are compared and analyzed with the target constraints of the original medical task to determine the degree of target deviation caused by the abnormal change features.

[0129] The original medical task is decomposed into an adjusted set of subtasks based on the degree of deviation from the target. The execution order is then recalculated based on the dependencies between the adjusted subtask sets and the resource availability information in the status update information.

[0130] During the implementation of the comprehensive decision-making plan, it is necessary to collect patients' physiological indicators and treatment response data in real time. Physiological indicators include continuously monitored vital signs such as heart rate, blood pressure, blood oxygen saturation, respiratory rate, and body temperature. These are acquired through a data interface with monitoring equipment at a preset sampling frequency, which is dynamically adjusted according to the severity of the patient's condition; for critically ill patients, it can be set to collect data once per second. Treatment response data includes feedback information after treatment measures are implemented, such as changes in symptoms following drug intervention, tissue reactions after surgical procedures, and changes in blood volume after intravenous infusion. This data is obtained through the medical record system and the observation records of medical staff. The collected physiological indicator data and treatment response data are then input as new medical data into the analysis module.

[0131] When performing time-series analysis on new medical data, a time-series data window is first established. The window length is set according to the clinical significance of different indicators; for example, the analysis window for blood pressure data is set to 5 minutes, and the analysis window for heart rate data is set to 3 minutes. During the time-series analysis, the rate of change, trend, and amplitude of fluctuation of each physiological indicator are calculated to identify abnormal change patterns that exceed the normal fluctuation range. Key change features include abrupt change points of indicators, persistent abnormal intervals, and abnormal patterns of association among multiple indicators. For example, when blood pressure drops by more than 20 mmHg within 3 minutes and heart rate rises simultaneously by more than 30 beats per minute, this is identified as a key change feature indicating a tendency towards shock. Based on these key change features, status update information is generated and stored in structured data format, including the type, amplitude, timestamp, and clinical significance label of the change feature.

[0132] When determining whether the status update information meets the preset status change conditions, abnormal change features in the status update information are extracted and matched against predefined conditions for disease deterioration and treatment ineffectiveness. Conditions for disease deterioration include specific threshold rules such as a sharp deterioration in vital organ function indicators, a sustained increase in infection indicators, and progressive abnormalities in coagulation function. Conditions for treatment ineffectiveness include situations where expected improvement indicators fail to appear after the implementation of standard treatment plans, symptoms continue to worsen, or new complications occur. When any abnormal change feature meets any of the conditions, the status update information is determined to meet the preset status change conditions, triggering a task replanning process.

[0133] When a replanning is deemed necessary, the abnormal changes in the status update information are compared and analyzed with the target constraints of the original medical task. These target constraints include expected physiological indicator recovery goals, treatment time window requirements, and complication prevention goals. The degree of deviation from the target state is determined by calculating the deviation metric between the current state and the target state. A weighted distance calculation method is used to calculate the deviation metric, assigning different weights to different target constraints, with higher weights assigned to life-threatening indicators.

[0134] The original medical task is re-decomposed based on the degree of deviation from the target, generating an adjusted set of subtasks. This decomposition process considers the urgency of the current abnormal state and resource constraints. For example, when signs of hemorrhagic shock are detected in a patient, the generated adjusted set of subtasks includes emergency hemostasis, rapid fluid resuscitation, and administration of vasoactive drugs. These subtasks have higher priority than non-urgent treatment items in the original task. When determining the execution order, the dependencies between the adjusted subtask sets are analyzed, identifying task pairs that must be executed sequentially and task groups that can be executed in parallel. Simultaneously, resource availability information is extracted from the state update information, including the number of currently available medical personnel, equipment occupancy status, and drug inventory. Resource constraints are used as limitations for calculating the execution order, ultimately generating an execution order scheme that satisfies dependencies and is resource-feasible.

[0135] The multi-agent collaborative reasoning system for complex trauma treatment according to embodiments of the present invention includes:

[0136] A data receiving unit is used to receive multimodal medical data from trauma patients;

[0137] The task decomposition unit is used to assess the complexity of the multimodal medical data through a coordinating agent, decompose the treatment task into multiple sub-tasks based on the complexity assessment results, and determine the execution order and dependencies of the multiple sub-tasks.

[0138] The task distribution unit is used to distribute each subtask and its associated medical knowledge evidence to the corresponding professional intelligent agent based on the task description text of the multiple subtasks. Each professional intelligent agent then performs reasoning on the assigned subtask based on predefined medical role permissions to generate professional decision suggestions.

[0139] The conflict coordination unit is used to collect professional decision-making suggestions returned by each professional intelligent agent by the coordination intelligent agent, determine whether there is a conflict between the professional decision-making suggestions, and when a conflict is detected, determine a coordinated comprehensive decision-making scheme through each professional intelligent agent and the coordination intelligent agent.

[0140] The dynamic adjustment unit is used to execute the comprehensive decision-making scheme and acquire new medical data as status update information. When the status update information meets the preset status change conditions, the task decomposition and execution order are readjusted according to the status update information.

[0141] A third aspect of the present invention provides an electronic device, comprising:

[0142] processor;

[0143] Memory used to store processor-executable instructions;

[0144] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0145] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0146] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0147] 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 multi-agent collaborative reasoning method for complex trauma treatment, characterized in that, include: Receive multimodal medical data from trauma patients; The complexity of the multimodal medical data is assessed by a coordinating intelligent agent. Based on the complexity assessment results, the treatment task is decomposed into multiple sub-tasks, and the execution order and dependencies of the multiple sub-tasks are determined. Based on the task description text of the multiple sub-tasks, each sub-task and its associated medical knowledge evidence are distributed to the corresponding professional intelligent agents. Each professional intelligent agent then performs reasoning on the assigned sub-task based on predefined medical role permissions to generate professional decision-making suggestions. The coordinating agent collects the professional decision-making suggestions returned by each professional agent, determines whether there are conflicts between the professional decision-making suggestions, and when a conflict is detected, a coordinated comprehensive decision-making scheme is determined by the professional agents and the coordinating agent. The comprehensive decision-making scheme is executed and new medical data is acquired as status update information. When the status update information meets the preset status change conditions, the task decomposition and execution order are readjusted according to the status update information. The process of coordinating an intelligent agent to assess the complexity of the multimodal medical data, decomposing the treatment task into multiple sub-tasks based on the complexity assessment results, and determining the execution order and dependencies of the multiple sub-tasks includes: The coordinating agent is used to perform cross-modal consistency analysis and decision path conflict detection on the multimodal medical data. By calculating semantic contradiction indicators and treatment plan mutual exclusion indicators between different modal data, a complexity assessment result including data consistency score and decision conflict score is determined. The number of sub-task levels is determined based on the data consistency score, and the number of sub-tasks within each level is determined based on the decision conflict score, thus generating a task decomposition structure containing multiple levels of sub-tasks. Resource requirement analysis and temporal dependency analysis are performed on each subtask in the task decomposition structure to determine a subtask dependency graph containing node weights and edge weights. The node weights represent the resource consumption of the subtasks, and the edge weights represent the dependency strength between the subtasks. The subtask dependency graph is constrained and solved to generate a full subtask execution sequence under the premise of satisfying the preceding execution relationship. The subtask execution sequence represents the execution order and dependency relationship of multiple subtasks. The step of performing resource requirement analysis and temporal dependency analysis on each subtask in the task decomposition structure to determine the subtask dependency graph including node weights and edge weights includes: For each subtask in the task decomposition structure, a mapping table between subtask identifiers and required resource type identifiers is determined. Based on the number of resource types corresponding to each subtask in the mapping table and the scarcity coefficient of each resource type, the resource competition index of each subtask is calculated as the node weight in the subtask dependency graph; Traverse each subtask in the task decomposition structure, determine the data field names of the input interface and the data field names of the output interface of each subtask, and establish directed edges in the subtask dependency graph based on the output interface and the data field names of the output interface of each subtask; For each directed edge in the subtask dependency graph, determine the number of matches between the output data field of the starting subtask and the input data field of the ending subtask, and calculate the ratio of the number of matches to the total number of input data fields of the ending subtask as the data dependency. The timing urgency is calculated based on the execution time of the starting subtask and the tolerable waiting time of the ending subtask. The edge weights of the directed edges are determined by the data dependency and the timing urgency to obtain the subtask dependency graph.

2. The method according to claim 1, characterized in that, Based on the task description text of the multiple sub-tasks, each sub-task and its associated medical knowledge evidence are distributed to the corresponding professional intelligent agents. Each professional intelligent agent then performs reasoning on the assigned sub-task based on predefined medical role permissions, generating professional decision-making suggestions, including: Based on the task description text of each of the multiple subtasks, medical professional field keywords and diagnosis and treatment operation keywords are obtained; Based on a pre-constructed professional domain knowledge graph, the keywords in the medical professional domain are semantically expanded. The keywords in the medical professional domain and the keywords in the diagnosis and treatment operation are matched and calculated with the capability tags of the professional intelligent agents to obtain the matching degree matrix between each subtask and each professional intelligent agent. Based on the matching degree matrix, the professional agent with the highest matching degree is selected as the executing agent for each subtask, and medical knowledge evidence related to the semantics of the task description text of the subtask is retrieved for each subtask. Send the task description text of each subtask, the medical knowledge evidence corresponding to the subtask, and the medical role permission configuration information of the executing agent of the subtask to the executing agent; The executing agent performs logical reasoning on the task description text and the medical knowledge evidence based on the medical role permission configuration information, obtains intermediate reasoning conclusions, and generates professional decision-making suggestions based on the intermediate reasoning conclusions and the medical knowledge evidence.

3. The method according to claim 2, characterized in that, The executing intelligent agent performs logical reasoning on the task description text and the medical knowledge evidence based on the medical role permission configuration information, obtains intermediate reasoning conclusions, and generates professional decision-making suggestions based on the intermediate reasoning conclusions and the medical knowledge evidence, including: The executing agent generates a set of permission constraints, which includes data access constraint rules and operation execution constraint rules, based on the data access permission list and operation execution permission list in the medical role permission configuration information. The executing agent parses the task description text into an initial question statement and retrieves knowledge fragments from the medical knowledge evidence that are semantically related to the initial question statement. Based on the set of permission constraints, data content in the knowledge fragment that exceeds the scope of accessible medical data is filtered to obtain valid knowledge fragments that conform to the permission constraints; The executing agent determines the inference chain based on the valid knowledge fragments, verifies whether the required data is within the scope of the accessible medical data and whether the required operation is within the scope of the executable diagnostic and treatment operations in each inference step of the inference chain, and determines the intermediate inference conclusions generated by all verified inference steps; The executing agent extracts evidence text fragments and evidence source identifiers that support the intermediate reasoning conclusion from the medical knowledge evidence, and combines the intermediate reasoning conclusion, the evidence text fragments, and the evidence source identifiers to generate professional decision-making suggestions.

4. The method according to claim 1, characterized in that, The coordinating agent collects the professional decision-making suggestions returned by each specialized agent, determines whether there are conflicts among the professional decision-making suggestions, and when a conflict is detected, determines a coordinated comprehensive decision-making scheme through the specialized agents and the coordinating agent, including: The coordinating agent performs semantic parsing on the professional decision-making suggestions returned by each professional agent, and extracts the decision-making elements from each professional decision-making suggestion; The extracted decision elements are correlated with a pre-constructed medical constraint knowledge graph. When there are taboo or mutually exclusive relationships between the decision elements of different professional decision suggestions in the medical constraint knowledge graph, it is determined that there is a conflict between the professional decision suggestions, and the decision elements with taboo or mutually exclusive relationships are marked as conflicting elements. The coordinating agent sends a consultation request, containing professional decision-making suggestions from the conflicting element and other professional agents, to the specialized agents corresponding to the conflicting elements. The specialized intelligent agents corresponding to the conflicting elements generate decision adjustment schemes based on the strength of the role of the conflicting elements in their own decision-making logic. The coordinating agent modifies the conflicting professional decision-making suggestions according to the decision adjustment scheme. After verifying that the modification results meet the constraints through the medical constraint knowledge graph, the modified professional decision-making suggestions are combined to generate a coordinated comprehensive decision-making scheme.

5. The method according to claim 1, characterized in that, The comprehensive decision-making scheme is executed, and new medical data is acquired as status update information. When the status update information meets preset status change conditions, the task decomposition and execution order are readjusted based on the status update information, including: The comprehensive decision-making scheme is executed, and the patient's physiological index data and treatment response data are collected, which are then used as new medical data. Time-series analysis is performed on the new medical data to obtain key change features, and status update information is determined based on these key change features. Based on the abnormal change characteristics in the status update information, when the abnormal change characteristics meet the conditions for determining disease deterioration or treatment ineffectiveness, it is determined that the status update information meets the preset status change conditions; When it is determined that the status update information meets the preset status change conditions, the abnormal change features in the status update information are compared and analyzed with the target constraints of the original medical task to determine the degree of target deviation caused by the abnormal change features. The original medical task is decomposed into an adjusted set of subtasks based on the degree of deviation from the target. The execution order is then recalculated based on the dependencies between the adjusted subtask sets and the resource availability information in the status update information.

6. A multi-agent collaborative reasoning system for complex trauma treatment, used to implement the method as described in any one of claims 1-5, characterized in that, include: A data receiving unit is used to receive multimodal medical data from trauma patients; The task decomposition unit is used to assess the complexity of the multimodal medical data through a coordinating agent, decompose the treatment task into multiple sub-tasks based on the complexity assessment results, and determine the execution order and dependencies of the multiple sub-tasks. The task distribution unit is used to distribute each subtask and its associated medical knowledge evidence to the corresponding professional intelligent agent based on the task description text of the multiple subtasks. Each professional intelligent agent then performs reasoning on the assigned subtask based on predefined medical role permissions to generate professional decision suggestions. The conflict coordination unit is used to collect professional decision-making suggestions returned by each professional intelligent agent by the coordination intelligent agent, determine whether there is a conflict between the professional decision-making suggestions, and when a conflict is detected, determine a coordinated comprehensive decision-making scheme through each professional intelligent agent and the coordination intelligent agent. The dynamic adjustment unit is used to execute the comprehensive decision-making scheme and acquire new medical data as status update information. When the status update information meets the preset status change conditions, the task decomposition and execution order are readjusted according to the status update information.

7. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 5.

Images