Agent collaborative scheduling method and system for multi-modal medical intent recognition

Abstract

The application relates to the technical field of data processing, and discloses an intelligent agent collaborative scheduling method and system for multi-modal medical intent recognition, which comprises the following steps: accessing multi-modal health consultation data to output an intent feature set, fusing the intent feature set with a portrait data structure matrix, outputting a task scheduling sequence, and constructing a directed acyclic graph topology structure; extracting an initial feedback node to perform output mode mapping and state marker bit coverage, obtaining a feedback ready topology graph, and converting the feedback ready topology graph into an asynchronous event message; listening to a state data output state change queue, performing local correction of node weights, outputting a local correction path, and outputting a closed loop adjustment log. The application eliminates scheduling deadlocks and time sequence competitions, improves intent recognition accuracy, completes adaptive scheduling without triggering global reasoning of a large model again, and significantly reduces computing power consumption and network bandwidth load.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically to an intelligent agent collaborative scheduling method and system for multimodal medical intent recognition. Background Technology

[0002] With the application of artificial intelligence technology in related business fields, existing multimodal consulting systems typically employ a central dispatch system combined with a domain-specific large model to perform semantic parsing and intent recognition on users' multimodal input in order to output corresponding consulting suggestions. During system operation, the central dispatch system receives users' voice, text, or image input and transmits it to the cloud-based large model. The large model identifies the user's core needs and, combined with the user's current static user profile or profiling data, performs reasoning to generate decision results. Subsequently, the dispatch system distributes instructions based on these results, triggering multiple independent downstream business components to execute corresponding long-cycle tasks such as data monitoring, message intervention, or business recommendations.

[0003] However, in long-cycle collaborative scheduling scenarios, because downstream business components operate in heterogeneous and asynchronous states when executing tasks, the one-time task distribution by the central scheduling system based on static inference results at a specific moment cannot dynamically adapt to the real-time state changes and data feedback generated by downstream components during multi-module interactions and long-term operation. This leads to timing competition for state synchronization and interface call blind spots between the main scheduling node and downstream asynchronous components, causing state deadlock in the collaborative network, task flow execution interruption, and information loss due to dimensionality reduction during data transmission. Furthermore, if high-computing-power large models are frequently invoked for global secondary inference to cope with the above state changes, it will cause huge computational overhead and network bandwidth load. Therefore, how to avoid asynchronous state conflicts, reduce model computational overhead, and achieve closed-loop adaptive scheduling in the long-cycle collaborative scheduling of multi-agent systems for multimodal intent recognition is an urgent problem to be solved in this field. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention aims to provide an intelligent agent collaborative scheduling method for multimodal medical intent recognition. This method involves matrix-fusing the intent feature set output from multimodal health consultation data with a profile data structure to output a task scheduling sequence and construct a directed acyclic graph (DAG) topology for multi-stage task execution. Initial feedback nodes are extracted, and based on predefined output channel type identifiers within these nodes, feedback content is mapped to voice, text / image, or video output channels for real-time presentation, resulting in a feedback-ready topology. The task nodes awaiting dispatch are encapsulated into asynchronous event messages. Furthermore, a global state manager monitors the returned state data in real-time, outputting a state change queue and performing local adjustments to the node weights within the DAG topology. This eliminates scheduling deadlocks and task interruptions, improves intent recognition accuracy, and achieves closed-loop adaptive scheduling without triggering global inference in a large vertical model, thereby significantly reducing the computational overhead and network bandwidth load of large models.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] In a first aspect, the present invention provides an agent cooperative scheduling method for multimodal medical intent recognition, comprising:

[0007] Acquire multimodal health consultation data, access the multimodal health consultation data through a centralized intelligent agent, output intent feature set, obtain profile data structure, and perform matrix fusion of profile data structure and intent feature set through a centralized intelligent agent to output task scheduling sequence, and construct a directed acyclic graph topology structure for multi-stage task execution based on task scheduling sequence.

[0008] The initial feedback node is extracted from the directed acyclic graph topology. Output mode mapping and state flag bit overwriting are performed on the initial feedback node to obtain the feedback ready topology. The task nodes to be dispatched are extracted from the feedback ready topology. The task nodes to be dispatched are aggregated into a task set to be dispatched and encapsulated into asynchronous event messages.

[0009] Real-time monitoring of the status data returned by asynchronous event messages, outputting a status change queue, performing local correction of node weights on the directed acyclic graph topology based on the status change queue, and outputting the local correction path and closed-loop adjustment log.

[0010] Furthermore, the method for obtaining the intent feature set includes:

[0011] Multimodal health consultation data is preprocessed to generate standardized multimodal health consultation data.

[0012] Construct a request payload containing standardized multimodal health consultation data and output format constraints, and send the request payload to the preset vertical large model, outputting a structured JSON string;

[0013] Receive the structured JSON string returned by the vertical large model and parse it into a structured data object using a standard JSON deserializer;

[0014] Extract medical intent categories, feature entity lists, severity levels, and specific needs from structured data objects, and combine them to define an intent feature set.

[0015] Furthermore, the method for obtaining the task scheduling sequence includes:

[0016] Obtain the profile data structure, including physiological baseline parameters, past medical history labels, and time series of follow-up indicators;

[0017] Vectorization encoding is performed on the intent feature set, physiological baseline vector, historical business status label sequence and follow-up indicator time series sequence to generate the intent embedding vector, physiological baseline feature vector, historical status embedding vector and follow-up indicator time series feature vector respectively.

[0018] The intent embedding vector, physiological baseline feature vector, historical state embedding vector and follow-up index time series feature vector are used to construct a fusion input matrix and perform element-wise weighted summation to reduce the dimension and compress it into an intent profile fusion vector.

[0019] A secondary inference request containing an intent profile fusion vector and an intent feature set is sent to the vertical large model, and a structured data object is output from which a task scheduling sequence is extracted. The task scheduling sequence contains multiple scheduling task nodes.

[0020] Furthermore, the method for constructing the directed acyclic graph topology includes:

[0021] Traverse the scheduled task nodes in the task scheduling sequence and register them as graph nodes to form a set of scheduled task nodes.

[0022] Extract the predecessor dependency task numbers of the scheduled task nodes, and establish directed edges between the graph nodes corresponding to the predecessor dependency task numbers and the current graph node to form a set of task dependency directed edges.

[0023] The number of preceding dependent task IDs is counted as the in-degree value, and the node weight value is obtained using the in-degree value.

[0024] An initial topology graph structure is generated from the set of scheduled task nodes and the set of directed edges that depend on the tasks. If a dependency cycle exists, the directed edge that causes the cycle is deleted. The directed edges pointing to the destination graph nodes are reconstructed from the created global starting node, generating a directed acyclic graph topology structure.

[0025] Furthermore, the method for constructing the directed acyclic graph topology also includes:

[0026] Perform loop detection on the initial topology graph structure, define the node sequence formed by the graph nodes traversed in sequence as the current recursive traversal path, and define the graph nodes that enter the path as visited graph nodes.

[0027] If the starting graph node of the task-dependent directed edge is detected to be a visited graph node in the current recursive traversal path, it is determined that there is a dependency cycle in the initial topology graph structure. The task-dependent directed edge that causes the dependency cycle is deleted. A global starting node with an in-degree of zero is created in the initial topology graph structure, and the task-dependent directed edge pointing to the original destination map node pointed to by the deleted task-dependent directed edge is re-established by the global starting node.

[0028] The initial topology graph structure after loop detection and correction is defined as a directed acyclic graph topology structure.

[0029] Furthermore, the method for obtaining the feedback-ready topology graph includes:

[0030] Perform topological sorting on the directed acyclic graph topology and extract graph nodes with an in-degree of zero and immediate feedback interaction attributes as initial feedback nodes.

[0031] Based on the predefined output channel type identifier in the initial feedback node, the feedback content is mapped to the corresponding output channel for presentation;

[0032] The preset status flag of the initial feedback node that has been completed is updated from the pending state to the responded state, and is synchronously overwritten to the corresponding graph node in the directed acyclic graph topology to generate a feedback-ready topology graph.

[0033] Furthermore, the method for encapsulating the asynchronous event message includes:

[0034] Traverse the feedback ready topology graph, extract the graph nodes with the status flag set to be executed as the nodes to be dispatched and scheduled, and summarize them into a set of tasks to be dispatched;

[0035] For each scheduled task node in the set of tasks to be dispatched, extract the string number value, category identifier text, and control object content;

[0036] Get the current system clock timestamp, and calculate a hash string based on the clock timestamp and the string number value;

[0037] The globally unique hash string, string number value, category identifier text, and control object content are combined and encapsulated into an asynchronous event message.

[0038] Furthermore, the asynchronous event message also includes:

[0039] Based on the category identifier text in the asynchronous event message, the asynchronous event message is published to the corresponding preset topic channel in the asynchronous message queue;

[0040] Instantiate the context object space during memory allocation;

[0041] The task handle is constructed using a hash string and a memory address pointer to the context object space. The corresponding context object space is located and the contents of the control object are extracted.

[0042] Furthermore, the method for obtaining the state change queue includes:

[0043] Continuously listen to the independent return topic channel pre-configured in the asynchronous message queue to obtain the structured state data packets of the downstream functional agents as return state data;

[0044] The returned status data includes the static cache task number, execution progress value, and vital sign monitoring value.

[0045] The vital sign monitoring value is compared with the preset lower and upper thresholds in the content of the controlled object. If the vital sign monitoring value is less than the lower threshold or greater than the upper threshold, the corresponding returned status data is marked as a pending change status. All returned status data marked as pending change status are extracted, serialized and summarized, and the status change queue is output.

[0046] Furthermore, the execution process of the local correction of node weights includes:

[0047] Retrieve the return status data of the pending status change from the status change queue, and locate the corresponding graph node with the same static cache task number in the directed acyclic graph topology as the current trigger node.

[0048] Traverse downstream along the task-dependent directed edges to form a set of nodes to be corrected, which are all graph nodes reachable from the currently triggered node;

[0049] Based on the status of the change to be processed, define the abnormal feature monitoring value, and define the arithmetic mean of the lower threshold and the upper threshold as the center point of the threshold interval, and calculate the deviation of the abnormal feature monitoring value from the center point of the threshold interval.

[0050] The deviation magnitude is used as a weight enhancement factor to perform differential summation on the graph nodes in the set of nodes to be corrected, resulting in the corrected node weight values.

[0051] Furthermore, the method for obtaining the closed-loop adjustment log includes:

[0052] The local corrected path is obtained by reordering all nodes in the directed acyclic graph topology in descending order according to the corrected node weight values.

[0053] The scheduling control switch is executed according to the local correction path: for graph nodes whose node weight increase is greater than the preset trigger threshold and whose corresponding asynchronous event messages have not yet been sent, the protocol encapsulation is triggered in advance and published to the asynchronous message queue.

[0054] For graph nodes whose node weight has decreased, retract any asynchronous event messages that have not yet been published.

[0055] A single operation record containing the data to be changed, the corrected node weight value, and the local correction path is appended to the log database and defined as a closed-loop adjustment log.

[0056] Secondly, the present invention provides an intelligent agent cooperative scheduling system for multimodal medical intent recognition, which is used to implement the aforementioned intelligent agent cooperative scheduling method for multimodal medical intent recognition, the system comprising:

[0057] The fusion topology module is used to acquire multimodal health consultation data, access the multimodal health consultation data through a centralized intelligent agent, output the intent feature set, obtain the profile data structure, and perform matrix fusion of the profile data structure and intent feature set through a centralized intelligent agent to output the task scheduling sequence. Based on the task scheduling sequence, a directed acyclic graph topology structure for multi-stage task execution is constructed.

[0058] The asynchronous dispatch module is used to extract the initial feedback node from the directed acyclic graph topology, perform output mode mapping and state flag bit overwriting on the initial feedback node to obtain the feedback ready topology graph, extract the scheduling task nodes to be dispatched from the feedback ready topology graph, summarize the scheduling task nodes to be dispatched into a set of tasks to be dispatched and encapsulate them into asynchronous event messages.

[0059] The state monitoring module is used to monitor the return state data generated by asynchronous event messages in real time, output a state change queue, perform local correction of node weights on the directed acyclic graph topology based on the state change queue, and output the local correction path and closed-loop adjustment log.

[0060] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0061] This invention eliminates intent comprehension bias caused by single semantic input through cross-modal feature space alignment and matrix fusion of dynamic health profiles, improving the accuracy of intent recognition and reasoning, and solving the pain point of high error rate in subsequent intervention paths caused by traditional solutions. The global state manager and asynchronous message queue transform the traditional unidirectional task distribution into real-time monitoring of downstream return state data, accurately intercepting asynchronous transition states, effectively reducing the blind spot of timing competition, significantly reducing the probability of state deadlock in the collaborative network, and improving the complete execution rate of long-cycle task flows. The node weight local correction mechanism combined with the state change queue realizes lightweight adaptive switching of scheduling paths, which not only ensures the adaptive health management closed loop, but also eliminates the need to call the vertical large model again for global secondary reasoning, greatly reducing the frequency of concurrent calls and significantly reducing the overall computing power consumption and network bandwidth load. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a flowchart of an intelligent agent collaborative scheduling method for multimodal medical intent recognition provided in an embodiment of the present invention.

[0064] Figure 2 A schematic diagram of the directed acyclic graph topology for multi-stage task execution provided in an embodiment of the present invention;

[0065] Figure 3 This is a diagram of the initial feedback node output modality mapping architecture provided in an embodiment of the present invention.

[0066] Figure 4 This is a schematic diagram of an asynchronous message queue asynchronous distribution architecture provided in an embodiment of the present invention;

[0067] Figure 5 This is a projection diagram of the mathematical solution for the deviation amplitude of abnormal feature monitoring values ​​provided in an embodiment of the present invention;

[0068] Figure 6 This is a functional block diagram of an intelligent agent collaborative scheduling system for multimodal medical intent recognition provided in an embodiment of the present invention. Detailed Implementation

[0069] 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.

[0070] Example 1

[0071] Please see Figure 1 As shown, this embodiment provides an agent cooperative scheduling method for multimodal medical intent recognition, including:

[0072] Step S10: Acquire multimodal health consultation data, access the multimodal health consultation data through a centralized intelligent agent, output the intent feature set, obtain the profile data structure, and perform matrix fusion of the profile data structure and intent feature set through a centralized intelligent agent to output the task scheduling sequence, and construct a directed acyclic graph topology structure for multi-stage task execution based on the task scheduling sequence.

[0073] Further, step S10 includes:

[0074] Step S11: Obtain multimodal health consultation data, access the multimodal health consultation data through a centralized intelligent agent, and output the intent feature set.

[0075] In practical applications of multimodal data processing, users often simultaneously describe their subjective feelings via voice, input past records via text, and upload relevant reports via images. Existing technologies typically feed different modalities into independent recognition engines for sequential parsing, then summarize the intent through simple keyword concatenation. This approach fails to capture the semantic interrelationships between modalities, leading to the loss of crucial semantic information due to dimensionality reduction. Therefore, it is necessary to establish a cross-modal joint recognition mechanism uniformly scheduled by a centralized intelligent agent. This mechanism performs joint representation of separate single-modal inputs within a unified feature space, outputting a structured medical intent with high dimensionality and directionality.

[0076] Specifically, the centralized intelligent agent refers to a centralized scheduling logic node deployed on a cloud server cluster, used to uniformly receive multimodal input data from users and distribute scheduling instructions to downstream vertical large models and various functional intelligent agents. The centralized intelligent agent accesses the multimodal health consultation data input by users through a client module deployed on the user terminal; wherein, the client module refers to a software development kit module embedded in the user terminal application, responsible for authorized collection, encrypted transmission, and rendering and display of the final results of the multimodal data. The multimodal health consultation data includes audio streams, text strings, and image bitmaps. The audio stream refers to the audio file used by the user to describe their complaints, collected via the terminal microphone. This audio stream is digitized according to preset audio acquisition parameters. These preset audio acquisition parameters include a sampling rate and a quantization depth, set to ensure that the fidelity of the acoustic features meets the minimum requirements for model recognition without causing excessive network transmission latency. For example, the sampling rate is 44.1 kHz, and the quantization depth is 16 bits. The text string refers to a sequence of characters entered by the user via the terminal keyboard to describe historical records, such as past medical history, medication history, or current subjective feelings. The text string is acquired according to a preset text acquisition rule; the preset text acquisition rule refers to limiting the upper limit of the number of characters input at one time, which is set to prevent excessively long and invalid text from causing the input context of a large model to overflow. For example, the upper limit of the number of characters is set to 500 characters. The image bitmap refers to an image file containing feature expression or report information, such as a photo of vital signs or a scanned copy of an examination report, uploaded by the user through the terminal camera or photo album. The image bitmap is acquired according to a preset image acquisition specification; the image acquisition specification includes requirements for image format and clarity, which is set to balance the preservation of image details and network transmission overhead. For example, the uploaded format is a mainstream bitmap format such as JPEG. It should be noted that, regarding the acquisition and processing of the above-mentioned audio streams, text strings, and image bitmaps, the user information and data involved in the embodiments of this invention are all information and data that have been explicitly authorized by the user. All data collected in this disclosure has been explicitly authorized by the data subject. For example, the user confirms the data use authorization agreement through the APP pop-up window. The scope of authorization covers the application links such as data collection, feature analysis and system scheduling required for the implementation of the technical solution of this invention, and does not exceed the usage scenario authorized by the user. At the same time, the collected user personal information is de-identified using technologies such as K-anonymity and differential privacy, and only the de-identified data used for system reasoning and feature analysis is retained. It does not involve any discriminatory applications, and the data collection and processing strictly comply with the relevant laws and regulations on personal information protection.

[0077] The centralized intelligent agent performs format preprocessing on the multimodal health consultation data. Specifically, for audio streams, audio transcoding tools, such as the open-source FFmpeg, are used to convert them into a unified standard audio data format, such as 16kHz mono PCM format. For bitmap images, they are scaled to a preset unified pixel resolution, which is set to adapt to the standard input matrix dimensions of the downstream vertical large model base; for example, the unified pixel resolution is set to 1024×1024 pixels. For text strings, a character conversion module performs standardized parsing and redundant character removal according to a preset standard character encoding specification. The character conversion module is a conventional text processing script or API component in existing technology, such as a built-in string processing function based on Python. The standard character encoding specification is a universal encoding standard globally configured during system initialization, such as UTF-8 encoding, which is set to eliminate garbled characters caused by differences in default encoding across different terminal devices, ensuring the accuracy of cross-platform transmission of text feature data. The centralized intelligent agent then transmits the preprocessed, standardized multimodal health consultation data to the vertical large model. The aforementioned vertical large-scale model refers to a commercial large language model service deployed on a cloud-based GPU server cluster, providing multimodal semantic understanding capabilities in the medical field through API interfaces. This commercial large language model service uses a large language model with multimodal understanding capabilities as its foundation, serving as a vertical large-scale model engine specifically for handling domain knowledge reasoning. For example, it uses GPT-4o or Qwen-VL as its base model and completes medical domain knowledge injection training on publicly available datasets such as MIMIC-IV electronic medical record dataset, MedQA medical question-answering dataset, and NIH ChestX-ray14 chest X-ray dataset. It possesses the ability to standardize and map specific domain terminology, jointly recognize cross-modal multidimensional features, and reason about the state transitions of business nodes. Simultaneously, multiple independent functional agents are configured downstream. Each functional agent is an independent software program module that encapsulates specific execution logic and interface calls, used to receive asynchronous event instructions issued by the centralized agent and execute corresponding system interaction tasks.

[0078] When a centralized agent invokes a vertical large-scale model, it constructs a request payload containing standardized multimodal health consultation data. The request payload consists of four sequentially concatenated parts: system instructions, task description, context variable injection, and output format constraints. Specifically: the system instructions constrain the role and behavioral boundaries of the vertical large-scale model, guiding it to process input data using the logic of a professional inference engine. For example, the instructions require the model to perform joint semantic parsing of input features as a domain knowledge inference engine. The task description clarifies the actions to be performed. For example, the instructions require the model to perform cross-modal feature space alignment and extraction, and output structured intent features and demand orientations. Context variable injection refers to injecting preprocessed audio stream data, character sequences, and image bitmaps into the corresponding input parameter fields of the request payload in the form of encoded strings or text, respectively. The output format constraints restrict the data structure returned by the vertical large-scale model, ensuring the availability of subsequent scheduling logic. Specifically, this embodiment constrains the vertical large-scale model to return a standardized JSON structure, and the JSON structure is pre-defined to contain four core feature fields. These are: the `intent_category` field representing the category of medical intent; the `symptom_entity` field representing the list of characteristic entities; the `severity_level` field representing the severity level; and the `health_demand` field representing the specific demand. The `intent_category` field is a data identifier representing the business processing route direction of the user's current multimodal session request, and its value guides the top-level task branch that the system should prioritize activating in the subsequent multi-stage task topology. The `symptom_entity` field is a set of specific business domain feature labels extracted from the user's multimodal input through semantic association and feature alignment, or from report information. The `severity_level` field is a quantitative grading indicator of the urgency or priority level of the feature corresponding to the `symptom_entity` field. The `health_demand` field is a structured text description generated for the information feedback the user ultimately expects to obtain or the terminal service the system expects to schedule.

[0079] When a centralized agent calls the vertical large model API, it configures the following calling parameters: temperature parameter, kernel sampling parameter, and maximum output length parameter. The temperature parameter is a scalar parameter that controls the randomness of the large model's sampling; its setting is based on the requirement that professional intent recognition tasks demand extremely high determinism and stability in the output results, and the need to strictly suppress the divergent "illusion" of the large model during inference; for example, the temperature parameter is set to 0.2. The kernel sampling parameter is a scalar parameter that controls the cumulative probability threshold of the large model's sampled candidate words; its setting is based on the need to strike a balance between preserving the diversity of feature extraction vocabulary and suppressing low-frequency noise output to ensure the complete coverage of entity recognition; for example, the kernel sampling parameter is set to 0.9. The maximum output length parameter refers to an integer parameter that limits the upper limit of the character sequence returned by the large model in a single instance. Its setting is based on ensuring complete coverage of the typical upper bound of the structured return result in a single consultation scenario, avoiding the loss of structured fields due to output truncation. For example, the maximum output length parameter is set to 2048. The centralized agent receives the structured JSON string returned by the vertical large model API, parses the JSON string into a structured data object using a standard JSON deserializer, and extracts the values ​​of the intent_category, symptom_entity, severity_level, and health_demand fields. The values ​​of these four fields are then combined to define the intent feature set.

[0080] Step S12: Obtain the profile data structure, and perform matrix fusion of the profile data structure and the intent feature set through a centralized intelligent agent to output the task scheduling sequence.

[0081] When collecting multimodal health consultation data, the client module of the user terminal simultaneously encapsulates a unique user identifier. This unique user identifier is a de-identified hash string assigned by the system after the user completes authorized registration, used to uniquely map the user's identity. After parsing the unique user identifier, the centralized intelligent agent initiates a query request to a cloud-based key-value storage database. This cloud-based key-value storage database is a non-relational storage system deployed in a cloud server cluster, using a key-value pair structure for high-concurrency data read and write, used for persistent storage of user-authorized de-identified historical data. The centralized intelligent agent retrieves the profile data structure corresponding to the unique user identifier from the cloud-based key-value storage database. The profile data structure refers to a structured data object deployed in a cloud-based key-value storage database, used to maintain a single user's physiological baseline parameters, past medical history tags, and follow-up indicator time series in real time. It includes three sub-fields: the first sub-field is the physiological baseline vector, containing seven scalar values: age, gender, height, weight, baseline blood pressure, baseline heart rate, and fasting blood glucose; the second sub-field is the historical business status tag sequence, containing historical status code groups arranged in reverse chronological order. These historical status code groups are structured historical business tags extracted from the user's historical electronic medical records, constructed based on the International Classification of Diseases (ICD). Each code in this string array serves as an independent discrete feature tag; the third sub-field is the follow-up indicator time series, containing four types of time-series numerical matrices: daily steps, sleep duration, average heart rate, and blood oxygen saturation, transmitted from the user's bound wearable device within the last 30 days.

[0082] The centralized agent performs vectorized encoding on the intent feature set: it calls the feature embedding interface of the vertical large model API, concatenates the text content of the symptom_entity field and health_demand field in the intent feature set, and inputs it into the feature embedding interface, outputting an intent embedding vector. The intent embedding vector refers to a numerical array generated by the feature embedding interface that represents the user's current session request. The dimension of the numerical array of the intent embedding vector output by the feature embedding interface is strictly consistent with the spatial dimension during the pre-training of the vertical large model to ensure that the feature matrix can be correctly parsed and dimensionality reduced by the model; for example, the dimension is set to 1536. The centralized agent performs vectorization encoding on the individual physiological baseline vector, past medical history label sequence, and follow-up indicator time series sequence in the portrait data structure; specifically, the physiological baseline vector containing 7 scalar values ​​is expanded to the dimension of the numerical array by zero-padding, and defined as the physiological baseline feature vector; the historical state encoding group is input to the feature embedding interface of the vertical large model API for feature extraction, and the output feature vector with the same dimension as the intent embedding vector is defined as the historical state embedding vector, which refers to the numerical array used to represent the user's historical disease label context in the feature matrix calculation; for the follow-up indicator time series sequence Four time-series numerical matrices were constructed, comprising daily steps, sleep duration, average heart rate, and blood oxygen saturation. Along the time step dimension, scalar numerical sets for each time-series value over a continuous 30-day period were extracted. The arithmetic mean and standard deviation of each scalar numerical set were calculated, yielding four mean values ​​and four standard deviation values. Following the sequential concatenation of the mean and standard deviation of daily steps, sleep duration, average heart rate, and blood oxygen saturation, these four mean values ​​and four standard deviation values ​​were combined to construct a row vector containing eight scalar elements. This vector was defined as a statistical feature vector representing the recent dynamic fluctuation trend of vital signs. Subsequently, this statistical feature vector containing eight scalar elements was serialized into a floating-point string and input into the feature embedding interface of a vertical large model for dimensionality upscaling projection, expanding its mapping to the dimension of the numerical array to obtain the follow-up indicator time-series feature vector.

[0083] The centralized agent constructs a fusion input matrix using the intent embedding vector, the expanded physiological baseline feature vector, the historical state embedding vector, and the follow-up index time-series feature vector. Specifically, the centralized agent allocates a two-dimensional numerical matrix object with 4 rows and the number of columns equal to the dimension of a numerical array in memory space. The intent embedding vector, the expanded physiological baseline feature vector, the historical state embedding vector, and the follow-up index time-series feature vector are each treated as an independent row vector and written sequentially into the corresponding rows of the two-dimensional numerical matrix object according to a preset stacking order. This stacking order refers to a fixed correspondence rule in the row index allocation of each feature vector in the two-dimensional matrix, set to ensure that the geometric arrangement of the input features strictly matches the reading rules of the downstream feature fusion operator. For example, row indices 1 to 4 are respectively written to the intent embedding vector, the expanded physiological baseline feature vector, the historical state embedding vector, and the follow-up index time-series feature vector. The centralized agent performs matrix fusion on the fusion input matrix. Specifically, the four rows of feature data in the fusion input matrix are reduced in dimension and compressed into a fusion encoding vector of the same dimension as the intent embedding vector by element-wise weighted summation. During the weighted summation process, the centralized agent assigns preset feature fusion weights to each row vector. These preset feature fusion weights are floating-point coefficients used to control the proportion of information from different sources in the final fused vector. Their setting is based on the fact that in multimodal interaction scenarios, the current session's primary intent has a higher real-time processing priority than historical profile data, requiring both highlighting the dominance of the current intent and considering the correction role of the historical baseline. For example, the preset feature fusion weight allocation ratio is: 0.4 for the intent embedding vector, 0.2 for the expanded physiological baseline feature vector, 0.2 for the historical state embedding vector, and 0.2 for the follow-up indicator time-series feature vector. The centralized agent defines the finally calculated fused encoding vector as the intent profile fusion vector.

[0084] When a centralized agent invokes a vertical large-scale model, it constructs a secondary inference request containing the intent profile fusion vector and the intent feature set. This secondary inference request is also composed of four sequentially concatenated parts: system instructions, task description, context variable injection, and output format constraints. Specifically: the system instructions constrain the role and logical boundaries of the vertical large-scale model in the secondary inference process, guiding the model to process input data logically using the system scheduling engine. For example, the instructions require the model to strictly read the intent feature set and intent profile fusion vector injected in the request payload and perform associated logical inference. The task description clarifies the business scope and topological data structure of the actions to be performed and the output results. For example, the instructions require the model to perform domain logical inference based on the input fusion features, generating a sequence of scheduling task nodes including three categories: state monitoring, behavioral intervention, and business recommendation, and calculating and identifying the concurrent or serial pre-dependencies between each scheduling task node. The context variable injection refers to injecting the intent profile fusion vector as a floating-point array into the `fusion_embedding` parameter field corresponding to the secondary inference request. The term "segment" refers to the predefined input parameter key name in the vertical large model API interface, specifically designed to carry multi-temporal and multi-modal joint encoding features. Its setting is based on providing high-dimensional dense vector input for the hierarchical attention mechanism network of the large model. For example, in the secondary inference request, it manifests as the key name carrying the intent profile fusion vector. The intent feature set is injected into the corresponding intent_feature input parameter field of the secondary inference request in the form of structured text. The intent_feature field refers to the predefined input parameter key name in the vertical large model API interface, specifically designed to carry the explicit intent features of the current session. Its setting is based on enabling the large model to directly read the explicit structured text features of the current session. For example, in the secondary inference request, it manifests as the key name carrying the intent feature set. The output format constraint is used to limit the data structure returned by the vertical large model, ensuring the availability of the long-cycle scheduling network topology. Specifically, the vertical large model is constrained to return a standardized JSON structure, and the JSON structure must pre-include a task_list field. The task_list field refers to a structured array field generated by the vertical large model inference, which contains multiple specific scheduling task node objects. Its setting is based on providing a unified task node container to persistently record a collection of all scheduling tasks and their associated information that the system needs to trigger subsequently.

[0085] When a centralized agent initiates secondary inference by calling the vertical large model API, it configures the following call parameters, including: a secondary inference temperature parameter and a secondary inference maximum output length parameter. The secondary inference temperature parameter is a scalar parameter that controls the sampling randomness of the large model when generating task parameters and topology; its setting is based on ensuring the stability of the global business scheduling logic while allowing the model to have reasonable parameter divergence and fine-tuning capabilities when generating specific task parameters; for example, the secondary inference temperature parameter is set to 0.3. The secondary inference maximum output length parameter is an integer parameter that limits the upper limit of the structured character sequence returned by the large model in a single operation; its setting is based on the fact that long-cycle scheduling schemes typically involve complex graph structure descriptions of multiple different task nodes, requiring that the complete structured JSON return result not be truncated by the system, resulting in missing fields; for example, the secondary inference maximum output length is set to 4096. After receiving the secondary inference request, the vertical large model calls its built-in base large language model to perform domain logic inference. The aforementioned large language model refers to the conventional open-source or commercial multimodal large language model foundation upon which the vertical large model relies, such as the open-source Qwen-VL base model or the commercial GPT-4o base model, which runs by directly calling its underlying forward propagation inference interface.Specifically, the vertical large model parses the secondary inference request, extracts the intent profile fusion vector injected into the `fusion_embedding` input parameter field, and the intent feature set injected into the `intent_feature` input parameter field; it uses the intent profile fusion vector and the intent feature set as initial condition variables, inputs them into the base large language model, and outputs a structured JSON string that meets the output format constraints; the centralized agent receives the structured JSON string returned by the vertical large model API due to the execution of domain logic inference, parses the JSON string into a structured data object using a standard JSON deserializer, and extracts the value of the `task_list` field, defining it as a task scheduling sequence; the task scheduling sequence specifically includes an array structure composed of multiple scheduling task nodes, where each scheduling task node is preset to include five attribute fields: `task_id` field, which is the unique task number of the node, used to uniquely identify the scheduling task in the graph structure; `task_type` field, which is the business type identifier of the task, whose value is... The value range is strictly set to one of three categories: status monitoring, behavioral intervention, or business recommendation. The status monitoring category refers to a data collection control task distributed by the system to downstream systems to trigger external wearable devices to periodically collect and asynchronously report physiological data. The behavioral intervention category refers to a logical control task distributed by the system to downstream systems to trigger terminals to unidirectionally push standard health behavior guidance information or timed reminder events to users. The business recommendation category refers to a data interaction and docking task distributed by the system to downstream systems to trigger external medical institution information systems to perform specific department information retrieval or medical route guidance. The `task_param` field is a task parameter control object, encapsulating the control threshold, collection frequency, or triggering parameters required for downstream execution of the task. The control threshold is an individualized safety control interval boundary generated by the vertical large model based on the user's profile data structure, used to determine whether the vital sign monitoring value is abnormal. The vital sign monitoring value refers to a quantitative scalar value that reflects the user's current objective physiological state or behavioral indicators, actually collected by the downstream functional agent through external hardware devices. The control threshold specifically includes a lower threshold and an upper threshold, and the lower threshold and the upper threshold are both general terms for dynamic safety boundaries set for different categories of underlying feature data.Specifically, the lower threshold refers to the minimum safe limit for fluctuations in the vital sign monitoring values ​​of a specific category, set to prevent health risks caused by excessively low vital sign indicators in users; for example, for fasting blood glucose monitoring scenarios, the lower threshold is set to 3.9. The upper threshold refers to the maximum safe limit for fluctuations in the vital sign monitoring values ​​of a specific category, set to prevent the risk of acute warnings triggered by excessively high vital sign indicators; for example, for resting heart rate monitoring scenarios, the upper threshold is set to 100. The acquisition frequency refers to the time sampling interval period during which the system schedules downstream functional agents to control external hardware devices to acquire user underlying feature data when distributing status monitoring tasks; it is used to standardize the data sampling density during long-cycle operation; for example, the acquisition frequency includes a cycle of collecting heart rate data once per hour. The system uses a timed cycle, or a daily period, to collect sleep duration data. The triggering parameters refer to contextual condition variables used by the system to precisely control the activation timing, execution conditions, and delivery method of business recommendation tasks or specific behavioral intervention tasks. For example, the triggering parameters include the target department identification code when performing medical recommendations, the specific delivery timestamp of the timed medication reminder, and the channel identifier code of the designated push terminal. The `precondition_task_ids` field is a set of preceding tasks used to store the `task_id` of the preceding tasks directly dependent on by the current task, indicating which preceding tasks the current task must wait for to be activated. The `estimated_duration` field is the estimated execution time window length of the task, representing the task's lifecycle on the timeline. Through the parameter combination of the above five attribute fields, the task scheduling sequence is characterized by the ability of the same preceding task to fan out and trigger multiple subsequent tasks, and multiple preceding tasks to fan in and converge into a single subsequent task.

[0086] Step S13: The centralized agent constructs a directed acyclic graph topology for multi-stage task execution based on the task scheduling sequence.

[0087] After obtaining the task scheduling sequence, in order to transform the task scheduling sequence into an execution control structure that can be triggered and scheduled sequentially by downstream functional agents, the centralized agent constructs a directed acyclic graph topology structure for multi-stage task execution based on the task scheduling sequence.

[0088] Specifically, the directed acyclic graph topology refers to a graph theory data structure composed of a set of scheduled task nodes, a set of task-dependent directed edges, and a node weight mapping table. Its underlying data source relies on the `task_list` field within the task scheduling sequence. Specifically, the centralized agent traverses each scheduled task node in the `task_list` field and performs the following three operations for each node. The first operation is node registration, which uses the `task_id` contained in the scheduled task node as a unique identifier and `task_type`, `task_param`, and `estimated_duration` as node attributes, registering it as a graph node in the directed acyclic graph topology, thus forming the set of scheduled task nodes. The second operation is dependency edge construction, which traverses each precondition_task_ids array in the scheduled task node, and for each precondition_task_id, establishes a directed edge between the existing graph node corresponding to the precondition_task_id and the current graph node. The direction of the directed edge is from the precondition_task_id to the current graph node, thus forming the set of task-dependent directed edges. The third operation is node weight initialization, which initializes a node weight value for the graph node, thereby generating the corresponding node weight mapping table. The node weight value is a relative priority scalar representing the priority of the scheduled task node in the current scheduling time. The initialization calculation process for the node weight value is as follows: extract the number of elements contained in the `precondition_task_ids` array of the scheduled task node, defining it as the in-degree value of the graph node, which corresponds to the number of preceding dependent tasks of the graph node; extract the `estimated_duration` value from the graph node's attributes; add the reciprocal of the in-degree value to the reciprocal of the `estimated_duration` value and normalize the result to the range of 0 to 1 to obtain the node weight value; the node weight value is a dynamically quantified value that changes in real time with the system's operating state.

[0089] After completing the registration of graph nodes for all scheduled task nodes in the `task_list` field, constructing task-dependent directed edges, and initializing graph node weights, the centralized agent generates an initial topology graph structure containing the set of scheduled task nodes and the set of task-dependent directed edges. The centralized agent then performs loop detection on the initial topology graph structure. Specifically, a depth-first search algorithm is invoked to traverse the initial topology graph. During the traversal, the sequence of graph nodes visited sequentially is defined as the current recursive traversal path, and the graph nodes already within the current recursive traversal path are defined as visited graph nodes. If, during the traversal, the starting node of any task-dependent directed edge is detected to be a visited graph node within the current recursive traversal path, a dependency cycle is determined to exist in the initial topology graph. The centralized agent deletes the task-dependent directed edge that caused the dependency cycle and creates a global starting node with an in-degree of zero in the initial topology graph. Between the global starting node and the destination map node originally pointed to by the deleted task-dependent directed edge, a new task-dependent directed edge pointing to that destination map node is re-established by the global starting node. If no dependency cycle is detected during the traversal, the initial topology graph is determined to satisfy the directed acyclic graph property. The initial topology graph structure after cycle detection and correction is formally defined as a directed acyclic graph topology structure. See also... Figure 2 This is a schematic diagram of the directed acyclic graph topology for multi-stage task execution provided in an embodiment of the present invention. Figure 2 As shown in the diagram, each circular node represents a scheduling task node, and their set constitutes the scheduling task node set. The circular nodes are labeled with task_id and task_type to uniquely identify and distinguish the business task category. The task_id is followed by a different numerical number, representing a different scheduling task node number, which is used to uniquely identify the corresponding scheduling task in the graph structure. Each directed arrow in the diagram represents a task dependency directed edge, and their set constitutes the task dependency directed edge set. The direction of each directed arrow indicates the dependency propagation direction from the preceding graph node to the current graph node.

[0090] Step S10 addresses the technical challenges of intent misunderstanding caused by traditional single semantic input and the mismatch between scheduling schemes and individual user real-time business states by utilizing multimodal health consultation data, intent feature sets, profile data structures, intent profile fusion vectors, task scheduling sequences, and directed acyclic graph topology. It achieves unified alignment across modal feature spaces, personalized and accurate inference, and structured modeling of multi-stage task dependencies. Specifically, the multimodal health consultation data integrates audio streams, text strings, and image bits... Figure 3The heterogeneous raw inputs are uniformly accessed; the intent feature set is output by a vertical large model based on cross-modal joint representation, eliminating the loss of semantic dimensionality reduction in single-modal mode; the profile data structure provides individualized reasoning context; the intent profile fusion vector jointly encodes the intent embedding vector, the expanded physiological baseline feature vector, the historical state embedding vector, and the follow-up index time-series feature vector in a unified embedding space through matrix fusion; the task scheduling sequence output contains the mesh task dependency relationship of multiple scheduling task nodes; the directed acyclic graph topology structure transforms the mesh dependency into a schedulable graph theory control base.

[0091] Step S20: Extract the initial feedback node from the directed acyclic graph topology, perform output mode mapping and state flag bit overwriting on the initial feedback node to obtain the feedback ready topology graph, extract the scheduling task nodes to be distributed from the feedback ready topology graph, summarize the scheduling task nodes to be distributed into a task set to be distributed and encapsulate and transform it into an asynchronous event message.

[0092] Further, step S20 includes:

[0093] Step S21: Extract the initial feedback node from the directed acyclic graph topology, perform output mode mapping and state flag bit overwriting on the initial feedback node to obtain the feedback-ready topology graph.

[0094] After obtaining the directed acyclic graph topology, in order to enable users to obtain intuitive results corresponding to the multimodal health consultation data input immediately after initiating a consultation session—that is, the multimodal input set of voice audio streams, text strings, and image bitmaps synchronously collected and uploaded during the consultation session—and to avoid the experience gap in traditional scheduling systems where users submit requests and then wait for downstream functional agents to complete serial execution before receiving output, an instant feedback mechanism based on initial feedback node parsing and output modality mapping is established. The purpose is to prioritize mapping graph nodes located at the topology starting point in the directed acyclic graph topology, and whose business attributes belong to the feedback category, to the corresponding output channels, that is, the logical components or hardware interfaces on the user terminal used to present multimedia or text, for instant presentation.

[0095] Specifically, the initial feedback node refers to a graph node in the directed acyclic graph (DAG) topology with an in-degree of zero and whose `task_type` field is not of the "status monitoring" class, i.e., it belongs to the "behavioral intervention" or "business recommendation" categories and possesses an instant feedback interaction attribute. The centralized agent performs topological sorting on the DAG topology, calling the classic Kahn topological sorting algorithm. The Kahn topological sorting algorithm uses all graph nodes and task-dependent directed edges in the DAG topology as input to construct a graph node set. By traversing all graph nodes in the DAG topology, based on the in-degree value of each graph node, it filters out graph nodes with an in-degree of zero, summing them into a set of graph nodes with an in-degree of zero. Iterative out-degree elimination is then performed on this graph node set. Specifically, in each iteration, a preset node selection rule is used. This rule states that when multiple graph nodes exist in the set of graph nodes with an in-degree of zero, they are selected according to their node weights. The nodes are selected sequentially in descending order of their numerical values. If the node weights are the same, they are further selected in ascending order of the lexicographical order of their task_id strings. A graph node with an in-degree of zero is selected from the set of graph nodes and appended to the output graph node sequence. The selected graph node and all its directed edges that are dependent on it are virtually removed from the directed acyclic graph topology. For the downstream graph nodes pointed to by the virtually removed task-dependent directed edges, their current dynamically cached in-degree is decremented by one. The graph nodes whose dynamically cached in-degree becomes zero after decrementing are added to the set of graph nodes with an in-degree of zero. The above iterative selection and removal process is repeated until the graph node set is empty, and finally a unique graph node sequence is output. The graph node sequence refers to a linear one-dimensional array structure formed by arranging all graph nodes in the directed acyclic graph topology according to topological dependencies and the node selection rules, where a preceding dependent graph node must be placed before a subsequent dependent graph node. The centralized agent extracts all matching graph nodes from the graph node sequence in order of arrangement, and summarizes them to form the initial feedback node set.

[0096] The centralized agent performs output modality mapping for each initial feedback node in the initial feedback node set. This output modality mapping refers to mapping the feedback content of the initial feedback node to the corresponding output channel based on the value of the predefined `feedback_modality` subfield in the `task_param` field of the initial feedback node. The `feedback_modality` subfield is a string identifier pre-encapsulated in the `task_param` field, indicating the type of the output media channel corresponding to the current initial feedback node; its value range includes at least the "voice" identifier for voice output channels, the "video" identifier for multimedia video output channels, and the "text_image" identifier for multimodal image / text output channels. Specifically, when the value of the feedback_modality subfield is "voice", it instructs the system to extract the text field from the task_param, specifically the feedback_content text field; when the value of the feedback_modality subfield is "video", it instructs the system to extract the multimedia resource identifier contained in the feedback_content field. The multimedia resource identifier refers to the index code or string identifier tag pre-encapsulated in the task_param, used to uniquely retrieve and match the target video digital file in the cloud or local media material library; when the value of the feedback_modality subfield is "text_image", it instructs the system to extract the text field and image address contained in the feedback_content field. The image address refers to the standard network address string pre-encapsulated in the task_param, used to indicate the addressing path and access address of the target static image data on the network server.The output modality mapping rule is as follows: If the value of the `feedback_modality` field is "voice", then the `feedback_content` text field in the `task_param` of the initial feedback node is input to the text-to-speech engine. For example, an open-source TTS text-to-speech engine is used to generate a unified format voice response stream, which is then broadcast to the user terminal through the speaker interface of the client module. If the value of the `feedback_modality` field is "text_image", then the centralized agent requests and parses the image address through the network to obtain the image data it points to, and combines the text field in the `feedback_content` field with the image data to render a text-image display page, which is then presented to the user terminal through the built-in browsing control of the client module. The text-image display page refers to the page rendered by combining and rendering the text characters parsed from the text field with the image data. A structured text page for visual interactive display on the terminal screen; the built-in browsing control refers to a software component embedded in the client module for parsing, loading, and local rendering of the structured text page, for example, a WebView control; if the value of the feedback_modality field is "video", then according to the multimedia resource identifier in the feedback_content field, a video guidance file in the pre-made media material library is retrieved and presented to the user terminal through the video playback control of the client module; wherein, the media material library refers to a data storage set deployed in a cloud server cluster or cached locally on the user terminal for persistent storage of various standardized multimedia digital resources; the video guidance file refers to a multimedia video digital file stored in the media material library that encapsulates specific business guidance or interactive demonstration content and can be streamed or downloaded by the terminal.

[0097] In a directed acyclic graph (DAG) topology, each graph node is configured with a state flag bit representing the current task scheduling activation state. The task scheduling activation state refers to an enumerated attribute variable representing the operational stage of the scheduling task graph node in the asynchronous collaborative execution lifecycle of multiple agents and the degree to which it can be triggered by downstream functional agents. The state flag bit is initialized to a first state value ("pending execution") during graph node registration, indicating that the graph node has been constructed but has not yet triggered a response. After completing the mapping and presentation of the output modality from the initial feedback node to the user, the centralized agent updates the state flag bit of the initial feedback node from the first state value to a second state value ("responded"), indicating that the immediate feedback corresponding to the graph node has been successfully rendered and received by the user terminal. The agent also synchronously overwrites the change in the state flag bit of the initial feedback node from the first state value to the second state value in the corresponding graph node of the DAG topology. The centralized agent defines the DAG topology after the synchronous overwriting of the value change as a feedback-ready topology. (See also...) Figure 3 This is a diagram of the initial feedback node output modal mapping architecture provided in an embodiment of the present invention. Figure 3 As shown, when performing output modality mapping for the initial feedback node, the core relies on the feedback_modality subfield pre-encapsulated in the task_param field. This architecture diagram illustrates three parallel cross-modal rendering chains: when the identifier is "voice", the system calls the text-to-speech engine to generate a speech response stream and plays it through the speaker interface; when the identifier is "text_image", the system uses a built-in browsing control to combine the extracted text field with image data obtained from a network request and render it into a text and image display page; when the identifier is "video", the system uses a multimedia resource identifier to address the media material library and finally presents the video guidance file through a video playback control.

[0098] Step S22: Extract the nodes of the scheduled tasks to be distributed from the feedback ready topology graph, summarize the nodes of the scheduled tasks to be distributed into a set of tasks to be distributed, and encapsulate the set of tasks to be distributed into asynchronous event messages.

[0099] After receiving the feedback ready topology, in order to avoid scheduling blockage caused by the centralized agent waiting for the downstream functional agents to complete their serial execution in long-cycle business scheduling scenarios, an asynchronous event distribution mechanism based on an asynchronous message queue is established. The purpose is to completely decouple the main scheduling thread from the downstream execution thread on the communication link and achieve concurrent scheduling.

[0100] Specifically, the centralized intelligent agent performs asynchronous distribution through an asynchronous message queue deployed on a cloud server cluster. This asynchronous message queue refers to a communication service component built on message middleware, exemplarily using the open-source Apache Kafka middleware, to provide asynchronous data buffering and routing decoupling between message producers (centralized intelligent agents) and message consumers (downstream functional intelligent agents). The asynchronous message queue creates and configures three independent topic channels during system initialization, named the status monitoring channel, behavior intervention channel, and business recommendation channel. Specifically, the status monitoring channel is a data pipeline deployed in the asynchronous message queue, specifically used for buffering and routing data collection control commands sent to the status monitoring intelligent agent; the behavior intervention channel is a data pipeline deployed in the asynchronous message queue, specifically used for buffering and routing logic control commands sent to the behavior intervention intelligent agent; and the business recommendation channel is a data pipeline deployed in the asynchronous message queue, specifically used for buffering and routing data interaction commands sent to the business recommendation intelligent agent.

[0101] The downstream functional intelligent agents are specifically classified and configured as state monitoring intelligent agents, behavior intervention intelligent agents, and business recommendation intelligent agents. The state monitoring intelligent agent refers to a data acquisition control software module that receives asynchronous event messages routed from an asynchronous message queue and triggers external hardware devices, such as wearable devices, to periodically collect and asynchronously report underlying feature data. The behavior intervention intelligent agent refers to a logic control software module that receives the asynchronous event messages and triggers a terminal to unidirectionally push standard behavior guidance information or timed reminder events to the user. The business recommendation intelligent agent refers to a data interaction and docking software module that receives the asynchronous event messages and triggers an external business information system to perform specific business information retrieval or routing guidance.

[0102] The centralized agent traverses all graph nodes in the feedback-ready topology graph, extracts graph nodes whose state flag is the first state value, and aggregates the extracted graph nodes to form a set of tasks to be distributed, and defines the extracted graph nodes as task nodes to be distributed and scheduled. For each task node to be distributed and scheduled in the set of tasks to be distributed, the centralized agent performs protocol encapsulation and transforms it into a standardized asynchronous event message. The data payload structure of the asynchronous event message is a standardized JSON string. Its encapsulation and transformation process specifically includes assigning and generating values ​​to the following five preset fields: `event_id` field: The centralized agent extracts the string number value from the `task_id` field of the task node to be dispatched and the obtained current system clock timestamp, concatenates the string number value with the clock timestamp to obtain the event signature code; inputs the event signature code into the system's built-in hash generation algorithm, for example, using the SHA-256 digest algorithm, to generate a globally unique hash string, and assigns the globally unique hash string to the `event_id` field as a unique tracking identifier for the asynchronous event message during concurrent system flow; `source_task_id` field: The centralized agent extracts the string number value from the `task_id` field of the task node to be dispatched and assigns it to the `source_task_id` field to maintain a unique mapping relationship between the dispatched message and the underlying directed acyclic graph structure nodes; `targ`; The `et_agent_type` field is where the centralized agent extracts the category identifier text from the `task_type` field of the task node's attributes, specifically extracting text related to status monitoring, behavioral intervention, or business recommendation. This text is then directly assigned to the `target_agent_type` field to clearly identify the target receiving agent type of the message. The `task_param` field is where the centralized agent extracts pre-encapsulated control object content, such as control thresholds, collection frequencies, or trigger parameters, from the `task_param` field of the task node's attributes. This extracted control object content is serialized into a JSON object and assigned to the `task_param` field within the asynchronous event message to transmit the context condition variables required for task execution downstream. The `dispatch_timestamp` field is where the centralized agent calls the operating system clock API of the currently running server to obtain the specific timestamp value of the moment the message was encapsulated in the execution protocol and assigns it to the `dispatch_timestamp` field to record the system distribution time.The centralized intelligent agent, acting as the message producer, executes precise routing rules based on the text value of the `target_agent_type` field in the asynchronous event message and publishes the asynchronous event message to the corresponding topic channel of the asynchronous message queue. Specifically, if the value of the `target_agent_type` field is "status monitoring," the asynchronous event message is published to the "status monitoring" channel; if the value of the `target_agent_type` field is "behavioral intervention," the asynchronous event message is published to the "behavioral intervention" channel; and if the value of the `target_agent_type` field is "business recommendation," the asynchronous event message is published to the "business recommendation" channel.

[0103] The state monitoring agent, behavior intervention agent, and business recommendation agent continuously listen to their respective topic channels as consumers. After receiving an asynchronous event message from the topic channel, the corresponding downstream functional agent performs task handle registration. Specifically, the downstream functional agent instantiates a context object space in the allocated memory of its cloud container cluster to isolate concurrent task data; extracts the globally unique hash string from the event_id field as the mapping key, and uses the memory address pointer of the context object space as the mapping value, writing it into the task lifecycle registry maintained internally by the downstream functional agent. This task lifecycle registry is maintained in the local allocated memory of the corresponding downstream functional agent, organized using a concurrency-safe hash mapping table data structure, and used for dynamic registration, tracking, and maintenance of the entire lifecycle state and corresponding memory mapping relationships of all concurrent asynchronous events carried by the current downstream functional agent from activation, execution to termination. The system control table; a logical control identifier structure containing the globally unique hash string and the memory addressing pointer, used to indirectly reference the underlying memory resources, is defined as a task handle bound to the current asynchronous event to ensure the security and independence of memory data for different concurrent tasks; the downstream functional agent reads the task handle, performs memory pointer addressing to accurately locate the corresponding context object space; the task_param field is deserialized using a standard JSON parsing library to extract specific control thresholds, collection frequencies, or trigger parameters, and the extracted parameters are passed as call input parameters to the downstream functional agent to complete the actual interaction with external hardware devices or information systems. See also Figure 4 This is a schematic diagram of the asynchronous message queue asynchronous distribution architecture provided in an embodiment of the present invention. Figure 4As shown in the diagram, the centralized agent on the left acts as the message producer, publishing asynchronous event messages to the asynchronous message queue in the middle. The three parallel channels in the middle represent the status monitoring channel, the behavior intervention channel, and the business recommendation channel, respectively. These three independent topic channels together constitute the asynchronous data buffer and routing decoupled data pipeline of the asynchronous message queue. The status monitoring agent, behavior intervention agent, and business recommendation agent on the right act as message consumers, each continuously listening to their corresponding topic channels to receive asynchronous event messages forwarded by routing and to execute subsequent task handle registration and actual interaction actions.

[0104] Step S20 addresses the shortcomings of traditional scheduling systems, such as the long wait times after users submit multimodal requests for downstream functional agents to complete serial execution before receiving output, and scheduling blockages caused by synchronous waiting in long-cycle business scheduling scenarios. It utilizes an initial feedback node, a feedback-ready topology graph, asynchronous event messages, and an asynchronous message queue to achieve real-time presentation of multimodal targeted feedback and concurrent distribution of asynchronous events from multiple agents. Specifically, the initial feedback node extracts graph nodes with an in-degree of zero and whose business attribute belongs to the feedback class from the directed acyclic graph topology; the feedback-ready topology graph synchronously covers and locks the status flag bits of the graph nodes that have already responded; and the asynchronous event message encapsulates the scheduling task nodes to be distributed into standardized data payloads containing fields such as globally unique tracking identifiers.

[0105] Step S30: Monitor the return status data generated by the asynchronous event message in real time, output the status change queue, perform local correction of node weights on the directed acyclic graph topology based on the status change queue, and output the local correction path and closed-loop adjustment log.

[0106] Further, step S30 includes:

[0107] Step S31: Monitor the status data returned by the asynchronous event messages in real time and output the status change queue.

[0108] After the centralized agent completes the distribution of asynchronous event messages and each downstream functional agent starts its corresponding local asynchronous execution thread, in order to perceive the real-time state changes and data feedback generated by the state monitoring agent, behavior intervention agent, and business recommendation agent during long-term operation, and to avoid the one-time task distribution based on the static inference results at a specific moment by the centralized agent losing its timeliness due to the delayed changes in the downstream execution state, a feedback state interception mechanism based on real-time monitoring by a global state manager is established. The purpose is to incorporate the state changes of each downstream asynchronous functional agent into the global scheduling observable field of view of the centralized agent.

[0109] Specifically, the global state manager refers to a control module deployed within the centralized intelligent agent, used to acquire and maintain real-time data feedback from all downstream functional intelligent agents. The global state manager continuously monitors the pre-configured independent feedback topic channels in the asynchronous message queue to acquire real-time feedback state data asynchronously transmitted by the state monitoring intelligent agent, behavior intervention intelligent agent, and business recommendation intelligent agent. All data contained in this feedback state data falls within the scope of the user's pre-defined authorized control object content rights data, and data analysis and caching strictly comply with relevant personal information protection compliance policies. The feedback state data refers to the structured state data packets published to the independent feedback topic channels by the state monitoring intelligent agent, behavior intervention intelligent agent, and business recommendation intelligent agent during the asynchronous flow execution process carrying out their data collection, behavior guidance, or data interaction. Each returned status data payload is a standardized JSON string containing five core fields: `callback_event_id`, a globally unique string generated by the downstream functional agent when generating the status data packet, calculated by concatenating the current return system timestamp with the memory address pointer of the context object space using a hash algorithm; `source_agent_type`, a string directly recording the category of the downstream functional agent sending the status data, i.e., status monitoring, behavioral intervention, or business recommendation text, clearly identifying the source agent category of the current returned status data; and `reference_task_id`. The `source_task_id` field refers to the string number value of the statically cached `source_task_id` field, which is directly extracted and transmitted by the corresponding downstream functional agent by reading the context object space corresponding to the task handle. This is used to maintain a unique mapping relationship between the current feedback status data and the corresponding graph node in the feedback ready topology graph. The `status_metric` field is used to encapsulate the execution progress value of the corresponding scheduled task graph node, as well as the vital sign monitoring value that reflects the current objective physiological state or behavioral indicators, which is actually collected by the status monitoring agent through external hardware devices. The execution progress value is a numerical scalar used to quantify the percentage of execution progress of the corresponding downstream functional agent on the currently assigned scheduled task node, and its value range is set to an integer between 0 and 100. The `callback_timestamp` field refers to the system clock timestamp value recorded when the downstream functional agent generates and sends the feedback status data packet to the independent feedback topic channel by calling the system clock API.To achieve standardized management of status text labels, the downstream functional agent internally has pre-set status text labels to characterize its current task execution stage. These include a "running" text label indicating that the current task processing is active, and a "completed" text label indicating that the current processing has reached the final state. The trigger condition for the "running" text label is as follows: when the corresponding downstream functional agent successfully receives and parses the `task_param` field in the asynchronous event message, and successfully requests and starts a local asynchronous execution thread, the local asynchronous execution thread initiates a low-level communication handshake request to the target external hardware device or external business information system, and receives a response confirmation message for the first time indicating successful connection establishment. For example, such as an ACK confirmation packet for TCP connection establishment or a Bluetooth pairing success signal, it indicates that the underlying data link has been established. Qualitative connection is established. The downstream functional agent updates the current status text label locally and assigns it to the "Execution in Progress" text label. The trigger condition for the "Execution Completed" text label is as follows: when the local asynchronous execution thread receives a complete feature data packet with an end identifier (e.g., a data stream with an EOF end identifier) ​​or receives a success status code response containing target business data from an external business information system (e.g., an HTTP 200 OK status message), and successfully serializes and writes all the acquired data into the context object space, it indicates that the data acquisition action for the task's lifecycle has been completely terminated. The downstream functional agent updates the current status text label and assigns it to the "Execution Completed" text label, then encapsulates and sends back the status data. After receiving the back status data, the global status manager compares each vital sign monitoring value in the status_metric field with the control threshold execution value pre-encapsulated in the task_param field of the corresponding scheduling task graph node.The global state manager extracts the lower and upper bound thresholds from the control thresholds. Specifically, if the vital sign monitoring value in the status_metric field is less than the lower bound threshold or greater than the upper bound threshold, it is determined that the returned status data has triggered the control threshold, and the returned status data is marked as a pending change status. If the execution progress value or status text label in the status_metric field, compared to the historical execution progress value or historical status text label of the reference_task_id in the state cache table maintained by the global state manager in memory, undergoes a state transition, it is determined that the returned status data has undergone a state change, and the global state manager also marks the returned status data as a pending change status. Here, the state transition refers to the current execution progress value or the current status text label, within a preset discrete control threshold. If a numerical boundary jump or text switch occurs between control nodes, specifically, the execution progress value changes from any value less than 100 to a value equal to 100, or the corresponding status text label changes from "in execution" to "execution completed," this indicates that the scheduling task node has crossed a substantial control phase in the downstream entity's operation. If all the vital sign monitoring values ​​in the status_metric field are strictly within the corresponding lower and upper threshold ranges, and no state transition occurs, it is determined to be a normal feedback state. The global state manager only archives this feedback state data to the historical state log database. The historical state log database refers to a historical storage system deployed in a cloud server cluster, built using a relational database management system, used to persistently store all normal feedback state data that has not triggered the control threshold and has not undergone the state transition. All feedback state data marked as pending change states are extracted, serialized, and summarized to output a state change queue, which is then buffered in the local memory of the centralized intelligent agent.

[0110] Step S32: Perform local correction of node weights on the directed acyclic graph topology based on the state change queue, and output the local correction path and closed-loop adjustment log.

[0111] After obtaining the state change queue, in order to respond instantly to the abnormal data returned by each downstream functional agent and dynamically adjust the distribution order of subsequent tasks, thereby avoiding the additional large model computing power overhead caused by re-calling the vertical large model to perform global secondary inference every time a state transition is captured, a local node weight adaptive correction mechanism is established. The centralized agent directly reads the state change queue through the global state manager, and while preserving the overall mesh topological dependency of the directed acyclic graph topology, only recalculates the weights of the affected local graph nodes to achieve lightweight adaptive adjustment of the cooperative scheduling path.

[0112] Specifically, the global state manager sequentially retrieves each piece of returned state data marked as pending change status from the state change queue and defines it as pending change data. For each piece of pending change data, the global state manager reads the `reference_task_id` field, locates the corresponding graph node with the same `task_id` in the directed acyclic graph topology, and defines this corresponding graph node as the current trigger node. Following the arrow pointing to the task-dependent directed edges, the global state manager traverses downstream all graph nodes directly or indirectly reachable from the current trigger node, and summarizes the traversed downstream graph nodes into a set, defined as the set of nodes to be corrected. Each graph node in the set of nodes to be corrected is defined as a node to be corrected. For each node to be corrected, the global state manager performs local node weight correction, that is, keeping the node weight values ​​of all graph nodes in the directed acyclic graph topology except for the set of nodes to be corrected unchanged, and only performing update calculations on the node weight values ​​of each node to be corrected. During the update calculation process, the global state manager first extracts the triggering factors that led to the current judgment result. Specifically, if the change in the data to be processed is triggered by a control threshold being set due to a vital sign monitoring value being less than the corresponding lower threshold or greater than the corresponding upper threshold, the global state manager defines the out-of-bounds vital sign monitoring value as an abnormal feature monitoring value. The global state manager calculates the deviation of the abnormal feature monitoring value from the center point of the corresponding threshold interval. The deviation amplitude refers to a dimensionless scalar coefficient used to quantify the relative distance ratio of the abnormal feature monitoring value deviating from the center point of the threshold interval, and its calculation formula is as follows: Where m represents the abnormal feature monitoring value; v represents the lower bound threshold; and u represents the upper bound threshold. This refers to the value obtained by arithmetically averaging the lower and upper thresholds, corresponding to the center point of the threshold interval. The calculation formula is based on the following: by calculating the algebraic distance between the abnormal feature monitoring value and the center point of the threshold interval, and dividing it by the total width of the threshold interval represented by the difference between the upper and lower thresholds, the numerical differences of different physical dimensions are eliminated, enabling heterogeneous feature data to be compared laterally on a unified scale.

[0113] The global state manager uses the deviation magnitude as a weighting factor and executes differentiated node weight correction rules based on the business association type between the node to be corrected and the current triggering node. Specifically, if the task_type field of the node to be corrected and the task_type field of the current triggering node have the same value, they are determined to be of the same origin business association. The global state manager sums the current node weight value of the node to be corrected with the deviation magnitude to obtain the corrected node weight value. When there is a direct task dependency directed edge between two task nodes and their task_type field values ​​are different, they are determined to be of adjacent business association. The global state manager sums the current node weight value of the node to be corrected with half of the deviation magnitude to obtain the corrected node weight value. If the node to be corrected... Although the `task_type` field of the current triggering node and the `task_type` field of the current triggering node are topologically reachable in the directed acyclic graph topology, they do not have a cascading evolution or direct causal triggering relationship in terms of business logic. Specifically, the execution result of the current triggering node is not used as the boundary triggering basis for the execution of the node to be corrected. For example, if the `task_type` of the current triggering node is a state monitoring category, while the `task_type` of the node to be corrected is a parallel behavior intervention category, and the behavior intervention category task runs independently in terms of business logic and does not rely on the abnormal feature monitoring value returned by the state monitoring category task to change its own logical control, then it is determined that there is no business association, and the global state manager keeps the node weight value of the node to be corrected unchanged.

[0114] After the global state manager completes the weight correction of all the nodes to be corrected in the set of nodes to be corrected, it performs a descending order rearrangement of all graph nodes in the directed acyclic graph topology according to the corrected node weight values, and obtains the rearranged graph node sequence, which is defined as the local correction path. The centralized agent performs scheduling control switching based on the local correction path: For graph nodes whose weight increase exceeds a preset trigger threshold, if their corresponding asynchronous event messages have not yet been sent, the centralized agent triggers their protocol encapsulation and publication to the asynchronous message queue in advance. The trigger threshold is set to ensure that the system is sufficiently sensitive to sudden anomalies or critical state evolutions while effectively filtering out frequent and ineffective rescheduling overhead caused by minor fluctuations in underlying data, preventing the multi-agent collaborative network from falling into scheduling jitter and communication congestion. For example, the trigger threshold is set to 0.3. For graph nodes whose weight decreases, the centralized agent withdraws their unpublished asynchronous event messages from the system's pending publication buffer. After completing the local correction and scheduling path switching, the global state manager calls the database system's write interface to sequentially append the operation record of this local correction to the log database deployed on the cloud server cluster, and defines this single operation record as a closed-loop adjustment log. The historical state log database is used to persistently store normal state data without anomalies. Each record specifically includes: the pending change data that triggered this correction, the corrected node weight values ​​of each node, and the local correction path. See Figure 5 This is a projection diagram of the mathematical solution for the deviation amplitude of abnormal feature monitoring values ​​provided in an embodiment of the present invention. Figure 5 As shown, when the global state manager intercepts out-of-bounds abnormal feature monitoring values, it does not perform a simple yes / no judgment, but instead substitutes them into a one-dimensional mathematical coordinate axis for dimensionless processing. The underlying geometric meaning of the deviation magnitude formula mapped in this figure is as follows: the horizontal axis defines the absolute safety interval formed by the lower and upper bound thresholds. The system first extracts the center point of this threshold interval. Subsequently, it calculates the algebraic distance from the abnormal feature monitoring values ​​outside the interval to this center point and divides it by the total width of the threshold interval. Through this geometrical algebraic mapping, the system solves for the deviation magnitude of the purely physical numerical scalar, thereby eliminating the dimensional differences in heterogeneous feature data and ensuring that multiple agents can perform fair local weight adjustments in subsequent topology propagation.

[0115] Step S30 solves the technical problems of state synchronization competition between centralized agents and downstream asynchronous functional components, interface call blind spots, multi-agent collaborative network state deadlock, task flow execution interruption, and loss of data transmission information due to interface call blind spots by transmitting back state data, state change queue, abnormal feature monitoring values, deviation magnitude, corrected node weight values, local correction paths, and closed-loop adjustment logs. It achieves global observability of asynchronous execution state, elimination of scheduling deadlock and task interruption, and significant reduction of computing power overhead and network bandwidth load of vertical large model. Specifically, the back-transmission status data incorporates the asynchronous execution progress and vital sign monitoring values ​​of downstream functional agents into the observable field of centralized scheduling; the state change queue aggregates all state data that triggers control thresholds or causes state transitions; the abnormal feature monitoring values ​​accurately pinpoint key data items that trigger out-of-bounds conditions; the deviation amplitude is dimensionless by comparing the center point of the threshold interval, enabling the horizontal comparison of vital sign monitoring values ​​with different physical dimensions within a unified correction logic; the corrected node weight values ​​reflect downstream state changes in real time to the scheduling decision of the directed acyclic graph topology based on the business association type; the local correction path drives the system to switch scheduling paths, completing an efficient adaptive scheduling closed loop without re-triggering the global inference of the vertical large model; and the closed-loop adjustment log enables the structured and persistent retention of the dynamic adjustment trajectory of the system topology.

[0116] Example 2

[0117] This embodiment, based on Embodiment 1, provides an intelligent agent collaborative scheduling system for multimodal medical intent recognition, such as... Figure 6 As shown, it includes:

[0118] The fusion topology module is used to acquire multimodal health consultation data, access the multimodal health consultation data through a centralized intelligent agent, output the intent feature set, obtain the profile data structure, and perform matrix fusion of the profile data structure and intent feature set through a centralized intelligent agent to output the task scheduling sequence. Based on the task scheduling sequence, a directed acyclic graph topology structure for multi-stage task execution is constructed.

[0119] The asynchronous dispatch module is used to extract the initial feedback node from the directed acyclic graph topology, perform output mode mapping and state flag bit overwriting on the initial feedback node to obtain the feedback ready topology graph, extract the scheduling task nodes to be dispatched from the feedback ready topology graph, summarize the scheduling task nodes to be dispatched into a set of tasks to be dispatched and encapsulate them into asynchronous event messages.

[0120] The state monitoring module is used to monitor the return state data generated by asynchronous event messages in real time, output a state change queue, perform local correction of node weights on the directed acyclic graph topology based on the state change queue, and output the local correction path and closed-loop adjustment log.

[0121] In the fusion topology module, the process of acquiring multimodal health consultation data, accessing the multimodal health consultation data through a centralized intelligent agent, outputting an intent feature set, acquiring a profile data structure, and performing matrix fusion of the profile data structure and the intent feature set through a centralized intelligent agent to output a task scheduling sequence, and constructing a directed acyclic graph topology structure for multi-stage task execution based on the task scheduling sequence, includes:

[0122] Step S11: Acquire multimodal health consultation data, access the multimodal health consultation data through a centralized intelligent agent, and output the intent feature set;

[0123] Step S12: Obtain the profile data structure, and perform matrix fusion of the profile data structure and the intent feature set through a centralized intelligent agent to output the task scheduling sequence;

[0124] Step S13: The centralized agent constructs a directed acyclic graph topology for multi-stage task execution based on the task scheduling sequence.

[0125] In the asynchronous distribution module, the process involves extracting initial feedback nodes from the directed acyclic graph topology, performing output mode mapping and state flag bit overwriting on the initial feedback nodes to obtain a feedback-ready topology graph, extracting task nodes to be distributed from the feedback-ready topology graph, and summarizing the task nodes to be distributed into a task set and encapsulating it into an asynchronous event message, including:

[0126] Step S21: Extract the initial feedback node from the directed acyclic graph topology, perform output mode mapping and state flag bit overwriting on the initial feedback node to obtain the feedback ready topology graph;

[0127] Step S22: Extract the nodes of the scheduled tasks to be distributed from the feedback ready topology graph, summarize the nodes of the scheduled tasks to be distributed into a set of tasks to be distributed, and encapsulate the set of tasks to be distributed into asynchronous event messages.

[0128] In the state monitoring module, the real-time monitoring executes asynchronous event messages to generate feedback state data, outputs a state change queue, performs local correction of node weights on the directed acyclic graph topology based on the state change queue, and outputs the local correction path and closed-loop adjustment log, including:

[0129] Step S31: Monitor the status data returned by the asynchronous event messages in real time and output the status change queue.

[0130] Step S32: Perform local correction of node weights on the directed acyclic graph topology based on the state change queue, and output the local correction path and closed-loop adjustment log.

[0131] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0132] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An agent-based cooperative scheduling method for multimodal medical intent recognition, characterized in that, The method includes: Acquire multimodal health consultation data, access the multimodal health consultation data through a centralized intelligent agent, output intent feature set, obtain profile data structure, and perform matrix fusion of profile data structure and intent feature set through a centralized intelligent agent to output task scheduling sequence, and construct a directed acyclic graph topology structure for multi-stage task execution based on task scheduling sequence. The initial feedback node is extracted from the directed acyclic graph topology. Output mode mapping and state flag bit overwriting are performed on the initial feedback node to obtain the feedback ready topology. The task nodes to be dispatched are extracted from the feedback ready topology. The task nodes to be dispatched are aggregated into a task set to be dispatched and encapsulated into asynchronous event messages. Real-time monitoring of the status data returned by asynchronous event messages, outputting a status change queue, performing local correction of node weights on the directed acyclic graph topology based on the status change queue, and outputting the local correction path and closed-loop adjustment log.

2. The agent-based collaborative scheduling method for multimodal medical intent recognition according to claim 1, characterized in that, The method for obtaining the intent feature set includes: Multimodal health consultation data is preprocessed to generate standardized multimodal health consultation data. Construct a request payload containing standardized multimodal health consultation data and output format constraints, and send the request payload to the preset vertical large model, outputting a structured JSON string; Receive the structured JSON string returned by the vertical large model and parse it into a structured data object using a standard JSON deserializer; Extract medical intent categories, feature entity lists, severity levels, and specific needs from structured data objects, and combine them to define an intent feature set.

3. The agent-based cooperative scheduling method for multimodal medical intent recognition according to claim 1, characterized in that, The method for obtaining the task scheduling sequence includes: Obtain the profile data structure, including physiological baseline parameters, past medical history labels, and time series of follow-up indicators; Vectorization encoding is performed on the intent feature set, physiological baseline vector, historical business status label sequence and follow-up indicator time series sequence to generate the intent embedding vector, physiological baseline feature vector, historical status embedding vector and follow-up indicator time series feature vector respectively. The intent embedding vector, physiological baseline feature vector, historical state embedding vector and follow-up index time series feature vector are used to construct a fusion input matrix and perform element-wise weighted summation to reduce the dimension and compress it into an intent profile fusion vector. A secondary inference request containing an intent profile fusion vector and an intent feature set is sent to the vertical large model, and a structured data object is output from which a task scheduling sequence is extracted. The task scheduling sequence contains multiple scheduling task nodes.

4. The agent-based collaborative scheduling method for multimodal medical intent recognition according to claim 1, characterized in that, The method for constructing the directed acyclic graph topology includes: Traverse the scheduled task nodes in the task scheduling sequence and register them as graph nodes to form a set of scheduled task nodes. Extract the predecessor dependency task numbers of the scheduled task nodes, and establish directed edges between the graph nodes corresponding to the predecessor dependency task numbers and the current graph node to form a set of task dependency directed edges. The number of preceding dependent task IDs is counted as the in-degree value, and the node weight value is obtained using the in-degree value. An initial topology graph structure is generated from the set of scheduled task nodes and the set of directed edges that depend on the tasks. If a dependency cycle exists, the directed edge that causes the cycle is deleted. The directed edges pointing to the destination graph nodes are reconstructed from the created global starting node, generating a directed acyclic graph topology structure.

5. The agent-based cooperative scheduling method for multimodal medical intent recognition according to claim 4, characterized in that, The method for constructing the directed acyclic graph topology also includes: Perform loop detection on the initial topology graph structure, define the node sequence formed by the graph nodes traversed in sequence as the current recursive traversal path, and define the graph nodes that enter the path as visited graph nodes. If the starting graph node of the task-dependent directed edge is detected to be a visited graph node in the current recursive traversal path, it is determined that there is a dependency cycle in the initial topology graph structure. The task-dependent directed edge that causes the dependency cycle is deleted. A global starting node with an in-degree of zero is created in the initial topology graph structure, and the task-dependent directed edge pointing to the original destination map node pointed to by the deleted task-dependent directed edge is re-established by the global starting node. The initial topology graph structure after loop detection and correction is defined as a directed acyclic graph topology structure.

6. The agent-based cooperative scheduling method for multimodal medical intent recognition according to claim 5, characterized in that, The method for obtaining the feedback-ready topology graph includes: Perform topological sorting on the directed acyclic graph topology and extract graph nodes with an in-degree of zero and immediate feedback interaction attributes as initial feedback nodes. Based on the predefined output channel type identifier in the initial feedback node, the feedback content is mapped to the corresponding output channel for presentation; The preset status flag of the initial feedback node that has been completed is updated from the pending state to the responded state, and is synchronously overwritten to the corresponding graph node in the directed acyclic graph topology to generate a feedback-ready topology graph.

7. The agent-based cooperative scheduling method for multimodal medical intent recognition according to claim 1, characterized in that, The method for encapsulating the asynchronous event message includes: Traverse the feedback ready topology graph, extract the graph nodes with the status flag set to be executed as the nodes to be dispatched and scheduled, and summarize them into a set of tasks to be dispatched; For each scheduled task node in the set of tasks to be dispatched, extract the string number value, category identifier text, and control object content; Get the current system clock timestamp, and calculate a hash string based on the clock timestamp and the string number value; The globally unique hash string, string number value, category identifier text, and control object content are combined and encapsulated into an asynchronous event message.

8. The agent-based cooperative scheduling method for multimodal medical intent recognition according to claim 7, characterized in that, The asynchronous event message also includes: Based on the category identifier text in the asynchronous event message, the asynchronous event message is published to the corresponding preset topic channel in the asynchronous message queue; Instantiate the context object space during memory allocation; The task handle is constructed using a hash string and a memory address pointer to the context object space. The corresponding context object space is located and the contents of the control object are extracted.

9. The agent-based cooperative scheduling method for multimodal medical intent recognition according to claim 8, characterized in that, The method for obtaining the state change queue includes: Continuously listen to the independent return topic channel pre-configured in the asynchronous message queue to obtain the structured state data packets of the downstream functional agents as return state data; The returned status data includes the static cache task number, execution progress value, and vital sign monitoring value. The vital sign monitoring value is compared with the preset lower and upper thresholds in the content of the controlled object. If the vital sign monitoring value is less than the lower threshold or greater than the upper threshold, the corresponding returned status data is marked as a pending change status. All returned status data marked as pending change status are extracted, serialized and summarized, and the status change queue is output.

10. The agent-based cooperative scheduling method for multimodal medical intent recognition according to claim 9, characterized in that, The execution process of the local adjustment of node weights includes: Retrieve the return status data of the pending status change from the status change queue, and locate the corresponding graph node with the same static cache task number in the directed acyclic graph topology as the current trigger node. Traverse downstream along the task-dependent directed edges to form a set of nodes to be corrected, which are all graph nodes reachable from the currently triggered node; Based on the status of the change to be processed, define the abnormal feature monitoring value, and define the arithmetic mean of the lower threshold and the upper threshold as the center point of the threshold interval, and calculate the deviation of the abnormal feature monitoring value from the center point of the threshold interval. The deviation magnitude is used as a weight enhancement factor to perform differential summation on the graph nodes in the set of nodes to be corrected, resulting in the corrected node weight values.

11. The agent-based cooperative scheduling method for multimodal medical intent recognition according to claim 10, characterized in that, The method for obtaining the closed-loop adjustment log includes: The local corrected path is obtained by reordering all nodes in the directed acyclic graph topology in descending order according to the corrected node weight values. The scheduling control switch is executed according to the local correction path: for graph nodes whose node weight increase is greater than the preset trigger threshold and whose corresponding asynchronous event messages have not yet been sent, the protocol encapsulation is triggered in advance and published to the asynchronous message queue. For graph nodes whose node weight has decreased, retract any asynchronous event messages that have not yet been published. A single operation record containing the data to be changed, the corrected node weight value, and the local correction path is appended to the log database and defined as a closed-loop adjustment log.

12. An agent-based collaborative scheduling system for multimodal medical intent recognition, used to implement the agent-based collaborative scheduling method for multimodal medical intent recognition as described in any one of claims 1-11, characterized in that the system comprises: The fusion topology module is used to acquire multimodal health consultation data, access the multimodal health consultation data through a centralized intelligent agent, output the intent feature set, obtain the profile data structure, and perform matrix fusion of the profile data structure and intent feature set through a centralized intelligent agent to output the task scheduling sequence. Based on the task scheduling sequence, a directed acyclic graph topology structure for multi-stage task execution is constructed. The asynchronous dispatch module is used to extract the initial feedback node from the directed acyclic graph topology, perform output mode mapping and state flag bit overwriting on the initial feedback node to obtain the feedback ready topology graph, extract the scheduling task nodes to be dispatched from the feedback ready topology graph, summarize the scheduling task nodes to be dispatched into a set of tasks to be dispatched and encapsulate them into asynchronous event messages. The state monitoring module is used to monitor the return state data generated by asynchronous event messages in real time, output a state change queue, perform local correction of node weights on the directed acyclic graph topology based on the state change queue, and output the local correction path and closed-loop adjustment log.

Images