Emergency method and system based on large language model scheduling and multi-agent cooperation

Abstract

The application discloses an emergency method and system based on large language model scheduling and multi-agent cooperation, belongs to the technical field of medical artificial intelligence and intelligent decision support, and solves the problem that the existing AI diagnosis method is difficult to cope with the clinical demand of the emergency department.The application discloses an emergency method and system based on large language model scheduling and multi-agent cooperation, a layered architecture is constructed with a large language model as a scheduling center and special expert agents as execution units, an emergency clinical path is encoded as a finite state machine for information completeness verification and active questioning, and neural-symbol dual verification and full-link evidence anchoring are fused to generate a traceable diagnosis report, effectively bridging the capability gap between general models and special models, realizing the transformation from passive response to compliant active guidance, eliminating the black box effect, and providing efficient, safe and interpretable intelligent decision support for emergency clinical treatment.

Description

Technical Field

[0001] This invention relates to the fields of medical artificial intelligence and intelligent decision support technology, specifically to an emergency treatment method and system based on large language model scheduling and multi-agent collaboration. Background Technology

[0002] The emergency department is a clinical department in a hospital that provides 24-hour emergency medical services for patients with acute illnesses, accidental injuries, and critical conditions. It implements a standardized triage system, establishing life-saving pathways for patients through rapid diagnosis and emergency intervention, and is a crucial link in the public health emergency and medical treatment system. Therefore, the work of the emergency department can be seen as a microcosm of the overall work of the hospital. As emergency physicians, the work is fast-paced and highly unpredictable, requiring a broad range of medical knowledge and, in most cases, the ability to make independent judgments and solve problems. In the current context of continuously strained medical resources and ever-increasing demand for emergency care, the emergency department faces efficiency bottlenecks: In terms of human resources, qualified emergency physicians need to master a wide range of interdisciplinary medical knowledge and possess the ability to make rapid decisions under high pressure. The training cycle for such multi-skilled personnel is several years, resulting in a long-term shortage of professional emergency physicians. In terms of process design, as a cross-functional hub for various hospital departments, the efficiency of the emergency department is heavily influenced by the collaborative capabilities of other departments. After completing the initial diagnosis, patients often need to wait for imaging examinations, laboratory test results, and specialist consultations. Delays in any of these steps directly impact the emergency department, creating a chain reaction. Therefore, against this backdrop, the intervention of artificial intelligence demonstrates significant advantages. Its core value lies in constructing a new human-machine collaborative diagnosis and treatment model: AI systems use natural language processing technology to quickly analyze patient complaints and medical history, achieving accurate intelligent pre-screening and triage, effectively distinguishing between acute and critical cases and non-urgent cases; in the diagnostic stage, deep learning algorithms can perform millisecond-level analysis of medical images, electrocardiograms, and other data, providing doctors with crucial decision-making references. However, the current application of AI in the emergency medicine field faces many challenges. At the technical architecture level, there is a significant capability gap between general-purpose large models and specialized small models. General-purpose large models refer to large language models pre-trained on massive general-purpose corpora, possessing broad semantic understanding and common-sense reasoning capabilities, such as the GPT series, Wenxin Yiyan, and Tongyi Qianwen. In emergency scenarios, they can handle unstructured patient complaints, conduct multi-turn inquiries in a natural dialogue manner, and leverage their extensive medical knowledge to engage in broad differential diagnostic thinking. However, these models are not designed for high-precision specialized perception tasks; they lack the dedicated capability for deep pattern recognition of pixel-level images or precise numerical indicators, thus failing to achieve the expert-level perception accuracy required in clinical practice. For example, a large language model with comprehensive medical knowledge can infer the possibility of pulmonary embolism based on complaints of "chest pain and difficulty breathing," but it cannot directly input a CT angiography image and accurately segment the location and volume of the embolus in the pulmonary artery. Specialized small models refer to compact deep neural network models that have been fine-tuned on labeled datasets for specific medical subtasks. They possess perceptual accuracy and execution efficiency surpassing human levels within their narrow domains, capable of extracting precise lesion morphology from images or calculating risk scores from structured indicators. However, their core limitation lies in their lack of general semantic understanding and task orchestration capabilities. They cannot grasp the overall picture of a patient's condition, nor can they understand unstructured textual complaints; they can only exist as passively invoked computational units. For example, a cascaded nnU-Net lung nodule detection model specifically trained on the LUNA16 dataset can output the 3D coordinates and segmentation masks of lung nodules with millimeter-level precision, but it completely fails to understand the differential diagnostic needs implied by a complaint like "hemoptysis accompanied by weight loss," let alone proactively initiate collaborative calls to pathological analysis models.

[0003] In summary, the structural contradictions between the two make it difficult for AI systems to cope with the complex and ever-changing clinical needs of the emergency department. At the level of diagnosis and treatment process, existing systems lack systematic diagnostic paths and safety constraint mechanisms that conform to clinical thinking. They cannot effectively handle real emergency scenarios with incomplete information, nor can they complete differential diagnosis through progressive consultation like senior physicians, and they also lack the necessary risk control capabilities. At the level of interactive experience, current tools generally lack personalized patient file management and continuous medical data tracking, and cannot build a complete health profile, making it difficult for them to provide accurate services based on individual medical history characteristics. Summary of the Invention

[0004] To address the aforementioned shortcomings of existing technologies, this invention provides an emergency diagnosis method and system based on large language model scheduling and multi-agent collaboration, which solves the problem that existing AI diagnostic methods are unable to meet the clinical needs of the emergency department.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: S1. Obtain the patient's original multimodal dataset, and preprocess and encode the data of different modalities separately, then perform time alignment and cross-modal fusion to obtain a fused multimodal contextual feature set; wherein, the original multimodal dataset Including unstructured text-based chief complaints and medical record data Raw medical imaging scan data Pre-consultation voice recording data and structured vital signs and biochemical test data .

[0006] S2. Retrieve the patient's longitudinal health record and concatenate it with the fused multimodal contextual feature set to generate a multimodal medical semantic vector; S3. Input the multimodal medical semantic vector into the large language model fine-tuned by the intelligent agent instructions, generate intermediate reasoning paths including differential diagnosis hypotheses, and then generate subtask call instructions for multiple dedicated expert intelligent agents. S4. The large language model outputs the triage level based on the preset emergency severity index standard and the structured vital signs and biochemical test data, as well as the textual complaints and medical record data, from the fusion of multimodal contextual features. It determines whether the triage level is greater than the preset risk level. If so, proceed to step S7; otherwise, proceed to step S5. S5. Determine whether the fused multimodal context feature set has covered the set of key information slots required for the diagnosis of its corresponding clinical state node. If yes, proceed to step S7; otherwise, proceed to step S6. S6. Generate candidate follow-up strategies for the missing information in the set of uncovered key information slots, calculate the information gain of each candidate follow-up strategy, select the follow-up strategy with the largest information gain to guide the patient to input supplementary evidence, update the supplementary evidence to the fused multimodal context feature set, and return to step S2 to update the multimodal medical semantic vector. S7. Distribute the subtask invocation instructions to the corresponding multiple dedicated expert agents, execute multiple specialized perception tasks in parallel, and output structured perception results that include at least the region of interest. S8. Verify all structured perception results. If all pass the verification, aggregate all structured perception results as a diagnostic conclusion and generate a structured diagnostic report; otherwise, use a large language model to feed back the reasons for failing the verification to the multimodal medical semantic vector and return to step S3.

[0007] On the other hand, an emergency system based on a large language model scheduling and multi-agent collaboration method is provided, including: The multimodal perception module receives the patient's original multimodal dataset, standardizes and concatenates the original multimodal dataset to obtain a fused multimodal contextual feature set, and calls the patient's longitudinal health record to concatenate the fused multimodal contextual feature set with the longitudinal health record to generate a multimodal medical semantic vector. The agent scheduling module is used to input multimodal medical semantic vectors into a large language model fine-tuned by agent instructions, generate intermediate reasoning paths including differential diagnosis hypotheses, and generate subtask calling instructions pointing to multiple dedicated expert agents based on the intermediate reasoning paths. The clinical pathway state machine module is used to receive intermediate inference paths, match the corresponding clinical state nodes using pre-embedded clinical pathway codes, and determine whether the fused multimodal context feature set has covered the predefined key information slot set in the clinical state node. If it has, the expert execution module is triggered; otherwise, the active guidance module is triggered. The triage level module is used to output the triage level based on the large language model with the preset emergency severity index standard, and according to the structured vital signs and biochemical test data, as well as the textual complaints and medical record data in the fusion multimodal context feature set. It determines whether the triage level is greater than the preset risk level. If so, the expert execution module is executed; otherwise, the active guidance module is executed. The active guidance module is used to generate candidate follow-up questioning strategies based on the missing information not covered in the key information slot set. Based on the preset candidate disease prior probability distribution corresponding to the differential diagnosis hypothesis, the information gain of each candidate follow-up questioning strategy is calculated, and the follow-up questioning strategy with the largest information gain is selected to guide the patient to complete the evidence. The completed evidence is updated to the fused multimodal context feature set and returned to the agent scheduling module. The expert execution module is used to distribute subtask invocation instructions to multiple corresponding dedicated expert agents to execute specialist perception tasks, including image segmentation, pathological analysis and risk scoring, in parallel, and output structured perception results with confidence. The neural symbol verification and report generation module is used to verify the structured perception results through pre-set logical rules based on hard-coded emergency clinical guidelines. The verification includes at least medical logic consistency review and contraindication comparison. If the verification passes, the structured perception result is determined as a diagnostic conclusion, and the diagnostic conclusion is mapped and anchored to the region of interest of the corresponding medical image and the clauses in the emergency clinical guidelines on which it is based. The mapping and anchoring results are used as the anchoring basis to generate a traceable structured diagnostic report containing the reasoning path and the anchoring basis. Otherwise, the large language model regenerates the reasoning path based on the reason for the verification failure and returns it to the agent scheduling module.

[0008] Compared with the prior art, the present invention has the following significant advantages: 1. This invention adopts a hierarchical heterogeneous architecture with a large language model as the cognitive scheduling center and multiple specialized expert agents as execution units. The large language model is responsible for generating differential diagnostic hypotheses from multimodal data based on thought chain reasoning and dynamically decomposing them into parallel executable sub-task calling instructions. Each specialized expert agent executes specialized perception tasks such as image segmentation, pathological analysis, and risk scoring in parallel on independent computing nodes. This organically integrates the general reasoning breadth of the large language model with the expert-level perception accuracy of the specialized small models, effectively solving the structural contradiction of the general large model lacking professional depth and the specialized small models being unable to integrate into the overall diagnosis and treatment process, thus improving the diagnostic efficiency and accuracy in high-load and complex multi-disease environments of emergency departments.

[0009] 2. This invention, on the one hand, encodes emergency clinical pathways into digital rules of a finite state machine and uses them as hard sandbox boundaries for reasoning in a large language model to enforce information completeness verification; on the other hand, when key discriminative points are detected as missing, it selects the optimal follow-up questioning strategy that maximizes the reduction of diagnostic entropy from candidate follow-up strategies based on an information gain algorithm, guiding patients to supplement diagnostic evidence. This mechanism achieves a paradigm shift from passive response to compliant proactive guidance, effectively eliminating the risk of misdiagnosis caused by skipped consultations and information omissions. Furthermore, through pre-built logical rules based on hard-coded emergency clinical guidelines in the neural symbolic engine, it performs medical logical consistency checks and contraindication comparisons on diagnostic conclusions, forcibly triggering replanning when evidence conflicts or violations of contraindications are detected, thereby establishing a safety control system that conforms to emergency clinical standards.

[0010] 3. This invention utilizes the patient's longitudinal health records to semantically align and stitch together longitudinal data such as the patient's historical medical records, long-term medication history, allergy history, and family history with the current emergency department cross-sectional data, constructing a panoramic health context that includes the time dimension. This enables the system to make individualized and accurate judgments based on the patient's entire life cycle medical history, transforming from a single diagnosis to personalized and accurate decision-making, and improving the individual specificity and clinical applicability of diagnostic conclusions.

[0011] 4. This invention uses a full-link evidence anchoring mechanism to map and anchor diagnostic conclusions to the region of interest in medical images and the relevant clauses in emergency clinical guidelines, generating a traceable structured diagnostic report containing a complete reasoning path and anchoring evidence. This provides clinicians with decision-making basis that can be traced and verified item by item, eliminating the black box effect of AI-assisted diagnosis and rebuilding trust between doctors and patients.

[0012] 5. Step S4 of this invention calculates a comprehensive priority index by combining the objective scoring of structured vital signs with the semantic parsing of the chief complaint text by a large language model, and performs path triage based on a preset threshold: when determined to be high-risk, the system directly triggers parallel intelligent agents to execute, completing the specialist perception and diagnostic output with the shortest link, ensuring that critically ill patients receive the fastest treatment response, fundamentally eliminating the risk of delaying rescue time due to the system's cumbersome consultation process; when determined to be non-high-risk, it enters the information completeness verification and proactive questioning stage, ensuring that the differential diagnosis evidence for regular patients is sufficient and complete before processing. This design realizes precise scheduling of "triage of urgent and non-urgent patients" in emergency scenarios, enhancing the rationality of triage and early warning capabilities through the complementarity of objective data and semantic understanding, and constraining the reasoning range of the large language model in high-risk states with clear hierarchical boundaries, reducing computational consumption and potential hallucination interference, while ensuring the timeliness of treatment for critically ill patients and the rigor of diagnosis for regular patients.

[0013] 6. By constructing a six-dimensional structured diagnostic report that includes triage level, candidate diagnosis list, critical illness screening conclusions, recommended supplementary examination items, contraindications and risk warnings, and treatment priority suggestions, emergency decision support has achieved a leap from single diagnostic output to a full-dimensional clinical decision matrix. Specifically, the triage level adopts a weighted fusion mechanism of objective scoring and semantic understanding, integrating improved early warning objective scores of structured vital signs with the pathophysiological semantic analysis of the chief complaint by a large language model, overcoming the one-sidedness of triage relying solely on numerical scores or text understanding; the candidate diagnosis list is arranged in order of confidence, providing doctors with a clear view of differential diagnosis priorities; the critical illness screening conclusions use a neural symbol engine with a pre-built high-risk and fatal emergency knowledge base for hard rule matching, forcibly screening for emergencies with hidden symptoms but high risk of death and outputting risk warning words, effectively avoiding the risk of AI models missing rare high-risk emergencies; the recommended supplementary examination items set identifies candidate diagnoses... The system identifies disease types with confidence levels below a threshold in the list and maps them back to the key information slots required for diagnosis, automatically outputting missing evidence to facilitate further investigation. Contraindications and risk warnings are generated by retrieving allergy and chronic disease history from the patient's longitudinal health record, using a symbolic rule engine to force the output of absolute drug contraindications and warnings, blocking the generation of unauthorized prescriptions at the source and building a personalized safety medication defense. Treatment priority suggestions are generated by matching the highest-confidence diagnosis with an authoritative emergency clinical guideline vector library, outputting early rescue measures or specialist admission recommendations prioritized according to time urgency, ensuring that clinical treatment conforms to evidence-based medicine principles while also meeting the timeliness requirements of emergency care. Overall, this six-dimensional reporting structure organically integrates diagnostic reasoning, risk assessment, evidence completion, safety control, and treatment guidance, forming a structured decision output covering the entire emergency care chain. It provides clinicians with transparent decision-making basis that can be traced and verified item by item, and through neurosymbolic collaboration and personalized record retrieval, it achieves an upgrade from general diagnosis to individualized and precise decision-making. Attached Figure Description

[0014] Figure 1 The flowchart shows an emergency treatment method and system based on large language model scheduling and multi-agent collaboration. Detailed Implementation

[0015] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0016] One aspect, reference Figure 1This embodiment provides an emergency treatment method based on large language model scheduling and multi-agent collaboration, including the following steps: S1. Obtain the patient's original multimodal dataset, and preprocess and encode the data of different modalities respectively. Then, perform time alignment and cross-modal fusion to obtain a fused multimodal contextual feature set.

[0017] In this embodiment, the original multimodal dataset ,in This consists of unstructured text-based chief complaints and medical record data. Raw medical imaging scan data, For pre-consultation voice recording data, This is structured vital signs and biochemical test data.

[0018] The system performs standardization processing on data from different modalities: For unstructured text-based chief complaints and medical record data Unstructured text in scanned medical records is extracted using text recognition models such as PaddleOCR. Rule-based text cleaning algorithms are then used to remove irrelevant characters and noise. Pre-trained medical named entity recognition models (such as BioBERT) are then used to extract key medical phenotypic entities, which are matched against standard medical ontology libraries (such as UMLS and SNOMED CT) to confirm the standardized medical concept descriptions of the entities. Finally, text feature vectors are generated. ; For medical imaging data The original DICOM images from different sources were isotropically resampled to a fixed voxel spacing (e.g., 1mm×1mm×1mm) using trilinear interpolation. Based on the observation requirements of the target disease, a corresponding Henle unit threshold was set to adjust the target window width and level. Subsequently, Z-score normalization was used to normalize the grayscale values ​​of the cropped pixels, resulting in a standardized image feature vector. ; For pre-consultation voice recording data An automatic speech recognition (ASR) acoustic model is used to transcribe the pre-consultation speech into a text sequence. Environmental noise is filtered by combining acoustic features in the medical field, and a speech feature vector is generated after semantic encoding. ; For structured vital signs and biochemical test data The terminology of heterogeneous test indicators from different sources was standardized and mapped. KNN interpolation was used to fill in missing data during the standardization process. After numerical normalization, structured feature vectors were generated. .

[0019] After standardizing the data for each modality, a vision-language contrastive pre-trained model architecture was adopted. Within a unified emergency medicine semantic space, the contrastive loss function was used to calculate the... , , ,and The feature similarity between them is used to complete cross-modal alignment. Cross-modal feature alignment is used to spatially map and fuse currently acquired different modal data. The fusion methods include: for , and The embedding layer directly maps to tensors of the corresponding dimensions; for Its original pixel feature dimension differs from the text embedding space dimension, requiring specialized cross-modal projection processing. Specifically, firstly, a pre-trained medical vision encoder with frozen parameters, trained offline, is loaded, and then the pre-processed... The image is divided into fixed-size patches, and a visual feature matrix of shape (N, 768) is extracted, where N is the number of image patches and 768 is the feature dimension output by the visual encoder. Next, a cross-modal projection head is constructed, consisting of two multilayer perceptron layers concatenated with a GELU activation function. This projection head linearly maps the 768-dimensional visual features to a 4096-dimensional native embedding vector space supported by a large language model, ensuring that the mapped image features are dimensionally similar to those output by the text embedding layer. , and The feature vectors remain consistent. Finally, during the concatenation stage, the system inserts special modality delimiters (such as [IMG] and [ / IMG]) before and after the mapped visual vector sequence, respectively, and combines them with... , and The sequences are directly concatenated along the length dimension to form a unified input sequence, which ultimately generates a fused multimodal contextual feature set.

[0020] S2. Retrieve the patient's longitudinal health record and concatenate it with the fused multimodal contextual feature set to generate a multimodal medical semantic vector.

[0021] After generating the fused multimodal contextual feature set, a personalized medical record graph database is established based on the FHIR standard. This database automatically retrieves and extracts longitudinal health record data corresponding to the current patient, including historical medical records, long-term medication history, allergy history, and family history. A temporal correlation algorithm is used to semantically align the patient's historical longitudinal data with the current emergency room cross-sectional data. Finally, the longitudinal health record data is concatenated with the fused multimodal contextual feature set to generate a multimodal medical semantic vector containing a temporal panoramic health context. .

[0022] S3. Input the multimodal medical semantic vector into the large language model fine-tuned by the intelligent agent instructions, generate intermediate reasoning paths including differential diagnosis hypotheses, and then generate subtask call instructions for multiple dedicated expert intelligent agents.

[0023] In this embodiment, the large language model is θ represents the model parameters, specifically including the base attention weight matrix and feedforward network weights of the pre-trained language model, as well as the low-rank adapter parameters injected and updated during the agent instruction fine-tuning stage using low-rank adaptive techniques. These parameters collectively control the mapping process from multimodal inputs to subtask API probability distributions. The agent instruction fine-tuning method for the large language model involves training the model using strongly supervised data of "observation-thinking-action" triples extracted by heuristic and feedback-enhanced synthesis (FAS+HAS) techniques. This process, by minimizing the autoregressive cross-entropy loss, guides the model to internalize the ability to map multimodal medical inputs to tool-calling logic conforming to emergency clinical guidelines within the underlying weights. The method for fine-tuning the large language model using agent instructions includes: obtaining multimodal medical semantic vector samples as input samples for the large language model, intermediate inference path samples for differential diagnostic hypotheses, and subtask call instruction samples pointing to a dedicated expert agent as output samples for the large language model; inputting the samples with input and output samples into the large language model, and fine-tuning the large language model by minimizing the autoregressive cross-entropy loss.

[0024] Specifically, a strongly supervised training sample consisting of triples of observation data, thought data, and action data is obtained. The observation data is used as the input to the large language model, and the thought data and action data are used as supervision signals. The large language model is fine-tuned by minimizing the autoregressive cross-entropy loss so that the large language model can generate inference paths and subtask invocation instructions based on the input multimodal medical semantic vectors. Among them, the observation data is the multimodal medical semantic vector sample, the thought data is the intermediate inference path sample including differential diagnosis hypothesis, and the action data is the subtask invocation instruction sample pointing to the dedicated expert agent.

[0025] S4. The large language model outputs a triage level based on a preset emergency severity index standard and the structured vital signs and biochemical test data, as well as the textual complaints and medical record data, from the fusion of multimodal contextual feature sets. It then determines whether the triage level is greater than a preset risk level. If so, proceed to step S7; otherwise, proceed to step S5. S5. Determine whether the fused multimodal context feature set has covered the set of key information slots required for the diagnosis of its corresponding clinical state node. If yes, proceed to step S7; otherwise, proceed to step S6. Specifically, the intermediate reasoning path is input into a pre-constructed finite state machine embedded with clinical path codes. Based on the intermediate reasoning path and using the pre-embedded clinical path codes, the machine matches the corresponding clinical state node and determines whether the fused multimodal context feature set has covered the predefined set of key information slots in the clinical state node. If so, proceed to step S6; otherwise, proceed to step S5. The set of key information slots includes the information required to complete the diagnosis of the corresponding clinical state node.

[0026] S6. Generate candidate follow-up strategies for the missing information in the set of uncovered key information slots, calculate the information gain of each candidate follow-up strategy, select the follow-up strategy with the largest information gain to guide the patient to input supplementary evidence, update the supplementary evidence to the fused multimodal context feature set, and return to step S2 to update the multimodal medical semantic vector. Specifically, subtask invocation instructions are distributed to multiple dedicated expert agents to perform specialist perception tasks, including image segmentation, pathological analysis, and risk scoring, in parallel, and output structured perception results with confidence; wherein, the structured perception results include at least the region of interest extracted from medical images.

[0027] S7. Distribute the subtask invocation instructions to the corresponding multiple dedicated expert agents, execute multiple specialized perception tasks in parallel, and output structured perception results that include at least the region of interest. S8. Verify all structured perception results. If all pass the verification, aggregate all structured perception results as a diagnostic conclusion and generate a structured diagnostic report; otherwise, use a large language model to feed back the reasons for failing the verification to the multimodal medical semantic vector and return to step S3.

[0028] Specifically, the structured perception results are verified using pre-defined logical rules in the neural symbol engine, which are hard-coded based on emergency clinical guidelines. The verification includes at least a medical logic consistency check and contraindication comparison. If the verification passes, the structured perception results are determined as diagnostic conclusions, and the diagnostic conclusions are mapped and anchored to the regions of interest in the structured perception results and the clauses in the emergency clinical guidelines on which they are based. The mapped and anchored results are used as the anchoring basis to generate a traceable structured diagnostic report containing the reasoning path and the anchoring basis. Otherwise, the large language model regenerates the reasoning path based on the reason for the verification failure and returns to step S3.

[0029] In this embodiment, the structured diagnostic report Including triage levels Candidate Diagnosis List Conclusions of the screening for critical illnesses Recommended Supplementary Checklist Contraindications and Risk Warnings Recommendations on handling priorities , represented as ; Triage Level The acquisition method includes: extracting the quantitative values ​​of vital signs from the structured vital signs and biochemical test data from the fused multimodal context feature set, and scoring the quantitative values ​​of vital signs according to the emergency severity index standard to obtain the objective basic risk score; then using a large language model to perform pathophysiological semantic analysis on the unstructured text complaints and medical record data extracted from the fused multimodal context feature set to obtain the symptom semantic risk weights, and weighting and fusing them with the objective basic risk score to obtain the comprehensive emergency priority index, and obtaining the corresponding triage level through the emergency preset grading threshold.

[0030] Comprehensive Emergency Priority Index The calculation expression is: ; An improved objective early warning score is calculated based on vital sign characteristics (such as heart rate, blood pressure, and blood oxygen) extracted from structured vital sign and biochemical test data, serving as an objective basis for risk assessment. The symptom risk weight coefficients are the outputs of semantic parsing of multimodal complaints using a large language model; α and β are the preset modality fusion adjustment weights, respectively.

[0031] Candidate Diagnosis List The acquisition methods include: the large language model extracts the disease type and confidence level corresponding to each structured perception result, and arranges all disease types in descending order of confidence level.

[0032] Critical Illness Screening Conclusion The acquisition methods include: judging the candidate diagnosis list through a neural symbolic engine with a pre-built knowledge base of high-risk and life-threatening acute illnesses. If a high-risk, life-threatening emergency exists, the system outputs a conclusion confirming the presence of that emergency and a corresponding risk warning term; otherwise, it outputs a conclusion confirming the absence of such an emergency. Specifically, the large language model has a pre-built knowledge base of high-risk, life-threatening emergencies. After generating a candidate diagnosis list, the system uses a rule engine to compare the previously output list with specific abnormal features. For emergencies that present with subtle symptoms but pose a life-threatening risk, the system forces a clear conclusion and a warning of remaining risks.

[0033] Recommended Supplementary Checklist The acquisition method includes: acquiring all disease types with confidence levels below a preset threshold in the candidate diagnosis list and the set of key information slots required for their diagnosis; outputting the missing information in the set of key information slots not covered in the fusion multimodal context feature set; and aggregating all missing information as a set of recommended supplementary examination items.

[0034] Contraindications and Risk Warnings The acquisition methods include: based on the patient's longitudinal health record information in the multimodal medical semantic vector, outputting absolute contraindications and contraindication warnings for drugs. Specifically, using an asynchronous data delivery mechanism, the preliminary diagnostic and treatment plan is formatted as entity objects and delivered to the working memory of the neural-symbolic rule engine. The engine performs pattern matching based on dynamically retrieved patient longitudinal records (such as allergy history and chronic disease history), forcibly outputting absolute and relative contraindications warnings for specific drugs or examinations, and blocking the generation of violations.

[0035] Recommendations for handling priorities The acquisition methods include: based on the disease type with the highest confidence in the candidate diagnosis list, matching the intervention plan for that disease type using a large language model with a pre-built authoritative emergency clinical guideline vector library, and outputting early rescue measures or specialist admission recommendations prioritized according to time urgency. Specifically, the preferred diagnosis with the highest confidence in the candidate diagnosis list is extracted, or based on the highest triage level, the corresponding entity is input into the RAG-based retrieval module, and matching intervention plans are retrieved from the authoritative emergency clinical guideline vector library, outputting early rescue measures or specialist admission recommendations prioritized according to time urgency.

[0036] As a further aspect of this embodiment, step S3 includes: S31. Input the multimodal medical semantic vector into a large language model fine-tuned by the intelligent agent's instructions, and use thought chain reasoning to generate a probability model containing differential diagnosis hypotheses and satisfying autoregressive conditional probabilities. The intermediate reasoning path, The calculation expression is:

[0037] in, For the first Step-by-step reasoning logic state; This represents the number of reasoning steps. These are the training parameters for the large language model; This is a multimodal medical semantic vector.

[0038] in, Using a large language model as a scheduler, given multimodal medical semantic vectors, it autoregressively generates a complete intermediate reasoning path consisting of a series of continuous logical deduction states based on its own network model parameters. The joint conditional probability; ,in, , and They are the first , No. and the The logical state of the step-by-step reasoning; The format is strictly standardized as a semi-structured text sequence containing necessary fields such as "symptom extraction," "differential diagnosis hypothesis," and "examination indication analysis." To ensure the absolute medical correctness of the inference path R and suppress model illusion, a reward model trained on real evaluation data from emergency room doctors is introduced at the decoding end of the generation process. Dynamic penalty weights are applied to dangerous inference directions that deviate from clinical pathway constraints or involve medical contraindications through a human feedback reinforcement learning strategy. For the first Step-by-step reasoning logic state, For the number of reasoning steps, These are the training parameters for a large language model that has been fine-tuned by instructions from the intelligent agent. This is a multimodal medical semantic vector; S32. Based on the intermediate reasoning path Using multimodal medical semantic vectors, a large language model is employed to generate a set of execution subtasks pointing to multiple specialized expert agents, and the optimal set of execution subtasks is obtained by maximizing conditional probabilities. :

[0039] in, To be the optimal set of subtasks to execute. ,in, , , and The first , No. , No. and the Each subtask is an optimal execution subtask; each subtask corresponds to an API call instruction (subtask call instruction) for a specific expert agent, which includes the target agent number and the corresponding input parameters. For example, Represented as key-value pair combination , This represents the identifier number of the target specialist agent that is uniquely pointed to and activated by the core scheduler in the expert execution layer. This refers to the dynamic input parameters that the target specialty intelligent agent needs to receive when performing specific feature extraction and perception tasks, thereby achieving accurate parsing from macroscopic diagnostic intentions to microscopic parallel scheduling instructions. This process must undergo JSON Schema format validation and parameter validity review by an independent digital symbol parser. If parameters are out of bounds or violate pre-set hard rules, the transmission will be immediately blocked and a forced replanning will be triggered. Indicates the execution of a set of subtasks The optimization operation that finds the maximum posterior probability; To provide a multimodal medical semantic vector and the generated intermediate inference paths Under these conditions, large language models generate specific subtask sets. The conditional probabilities, which are constrained by the OpenAPI Schema template predefined in the model vocabulary, achieve an exact mapping from natural language inference to a predefined specialist agent API function library by forcing the large language model to output lexical sequences in JSON format.

[0040] As a further embodiment, step S5 includes: S51, intermediate reasoning path The input is fed into a pre-constructed finite state machine, which embeds the encoding of the emergency clinical pathway; its state set represents discrete standard diagnostic stages in the emergency clinical pathway, and the input set represents the reasoning path from the intermediate points. The standardized medical entities extracted from them have state transition functions that define the flow logic of transitioning from one clinical state node to the next clinical state node when a preset threshold condition is met.

[0041] In this embodiment, the method for constructing a finite state machine is as follows: A mathematical model of a finite state machine is established, comprising a state set, an input set, and transition logic. The state set For discrete standard diagnostic stages in the emergency clinical pathway, the input set E is a standardized medical entity dictionary extracted from multimodal features by a large language model and mapped by ontology. The state transition function determines the flow path. This represents the initial registration status of a patient upon admission. A structured data description language is used to digitize and statically encode expert rules from authoritative clinical guidelines (Emergency Internal Medicine Treatment Guidelines). Each clinical diagnostic stage is declared as a node object with attribute constraints; the core attributes of the node must explicitly define the set of key information slots necessary for the state transition, as well as the corresponding transition target state and condition thresholds. The rule engine module in the safety decision and proactive guidance layer performs lexical and syntactic parsing on the above structured JSON encoded file, deserializing it in memory during system runtime and instantiating it into a directed graph structure global monitoring object, serving as the hard sandbox boundary for large language model inference.

[0042] S52. Extracting intermediate reasoning paths using finite state machines. The standardized medical entities extracted are matched with clinical pathway codes to determine the current clinical status node of the standardized medical entity. ; S53. Obtain Clinical Status Nodes The predefined set of key information slots And obtain the current multimodal medical semantic vector. The set of valid information slots that have been successfully extracted Calculate the set of missing information Its calculation expression is:

[0043] S54, Judgment Does it meet the requirements? If yes, proceed to step S7; otherwise, proceed to step S6.

[0044] As a further embodiment, step S6 includes: S61. Based on the missing information set Missing information slots (such as missing "nature of chest pain") are mapped from a pre-defined medical knowledge graph to generate a set of candidate follow-up questioning strategies. ,in, Each candidate follow-up question strategy corresponds to a key information slot to be completed and a corresponding natural language guiding statement for output to the patient, such as "Is your chest pain tearing, squeezing, or stabbing?". ,in, , , and The first , No. , No. and the Individual candidate follow-up questioning strategy.

[0045] S62, Based on intermediate inference path Differential diagnostic hypotheses are used to obtain pre-defined candidate diseases. Prior probability distribution And calculate the entropy of current diagnostic uncertainty information. Its calculation expression is: ; in, For the number of candidate diseases, Candidate diseases The Middle One candidate disease; S63, Set of Follow-up Questioning Strategies for Candidates For each candidate follow-up question strategy, obtain its corresponding set of potential effective response values, and combine this with the conditional probability relationship between diseases and symptoms in the medical knowledge graph to calculate the expected conditional information entropy under the assumption of obtaining an effective response. Its calculation expression is: ; ; ; in, for The Middle One candidate follow-up questioning strategy; In response to The set of valid responses, ,in, , and They are the 1st, 2nd and 3rd respectively. Valid responses, such as responses to "nature of chest pain" including "tearing", "squeezing", "needle-like", etc.; To propose a strategy of probing questions Get a reply under the following conditions The marginal probability, ;; In order to get a reply Under the condition of updating the posterior probability of the disease using Bayes' theorem; S64. Calculate the information gain of all candidate follow-up questioning strategies. The calculation expression is as follows: ; in, for Information gain; S65. Select the candidate follow-up questioning strategy with the highest information gain. Output The corresponding natural language guidance script is used to guide patients to fill in the missing information. The calculation expression is: ; S66, Targeting patients The missing information in the input is used as completion evidence to update the fused multimodal contextual feature set. Then return to step S2.

[0046] As a further embodiment, step S7 specifically includes: S71. Based on the optimal set of subtasks to be executed. The identifier number of each subtask will be The subtask invocation instructions are distributed to the corresponding target specialist agents in multiple dedicated expert agents.

[0047] S72. Each target specialty agent executes specialty perception tasks in parallel on independent computing nodes and outputs structured perception results with confidence scores. Specifically, when the subtask instruction points to lung medical image analysis, the system invokes a lung disease diagnosis agent. Its underlying layer employs a cascaded nnU-Net framework finely tuned using the LUNA16 and LIDC-IDRI datasets. It sequentially performs coarse lesion localization via a low-resolution network and fine segmentation via a high-resolution network, outputting structured results including the three-dimensional center coordinates, maximum diameter, and lesion mask tensor of the lung nodules. When the subtask instruction points to whole-slice pathological image analysis, the system invokes a pathological analysis agent. Its underlying layer uses a Clustering Constrained Attention Multiple Instance Learning (CLAM) model trained on the Camelyon16 dataset. It segments the whole-slice image into a set of image patches and calculates the attention weights of each patch for malignant features, outputting the localization of high-risk areas. The system generates a region of interest (ROI) heatmap. When the subtask instruction points to joint evaluation of cross-modal images, the system invokes a multimodal fusion agent. Its underlying layer uses a bidirectional progressive feature alignment network (BSAFusion) to predict the bidirectional deformation field by calculating spatial mapping rules. This completes the non-rigid feature-level weighted fusion of unaligned MRI and PET-CT images within the feature space, outputting a fused multimodal feature map. When the subtask instruction points to verification of vital signs and biochemical indicators, the system invokes an emergency risk scoring agent. Its underlying layer uses an XGBoost model trained on the MIMIC-IV ICU dataset. This model performs non-linear feature cross-mapping on the input structured indicators, outputting an emergency early warning score (such as the MEWS score) and the probability of occurrence of specific acute and critical illnesses. The above structured perception results include at least the ROI extracted from medical images.

[0048] S73. Encapsulate the structured perception results of all target specialty intelligent agents into standard structured perception results, which are in accordance with the JSON specification.

[0049] As a further embodiment, step S8 includes: S81. The standard structured perception results are verified using pre-built logical rules in the neural symbol engine and hard-coded based on emergency clinical guidelines. The verification includes at least medical logic consistency review and contraindication comparison. The neural symbol engine adopts an asynchronous interception architecture, and its symbol end deploys a rule engine and hard-codes absolute contraindications and logical predicates in the emergency clinical guidelines.

[0050] In this embodiment, an asynchronous interception architecture of "large model generation (neural computation) + rule engine verification (symbolic computation)" is adopted by the neural symbol engine. The symbol engine deploys an open-source rule engine, which hard-codes absolute contraindications and logical predicates from emergency clinical guidelines. After outputting the preliminary treatment plan, the core scheduler of the large language model on the neural side must first deserialize its output JSON-formatted subtask instructions into entity objects and submit them to the rule engine's working memory for pattern matching. Once a blocking rule is triggered, the rule engine immediately throws an exception callback. The large language model receives a system-level negative prompt containing the specific reason for the violation and is forced to regenerate the planned path under that constraint. Only after the rule engine returns a pass signal is the instruction sent to the expert execution layer.

[0051] S82. If the structured perception result passes the verification, the structured perception result is determined as the diagnostic conclusion; otherwise, the neural symbol engine generates negative prompt words containing specific reasons for the violation and feeds them back to the large language model, forcing the large language model to regenerate the intermediate reasoning path based on the negative prompt words and return to step S3.

[0052] In this implementation, the finite state machine is used to check whether the current patient's multimodal characteristics meet the necessary information slots for the corresponding clinical state (such as "suspected myocardial infarction"). If key identification points (such as the nature and cause of pain) are missing, the system uses the information gain algorithm to generate the optimal follow-up questioning path, which drives the large language model to generate targeted follow-up questioning strategies that conform to medical logic to guide the patient to complete the information. During this process, any logical jump that violates the flow path defined by the FSM will be blocked by hard rules.

[0053] S83. Map and anchor the verified diagnostic conclusions to the corresponding regions of interest in the medical images and the relevant authoritative guidelines to generate anchoring criteria. For visual perception criteria, use a cross-modal spatial mapping algorithm to align the diagnostic conclusions to the heatmap of the regions of interest or lesion segmentation mask of the medical images generated by the target specialty agent. For text and knowledge-based criteria, use retrieval enhancement generation technology to retrieve and cite specific clauses of authoritative guidelines from external medical literature and guideline vector databases.

[0054] S84, Generate a path containing intermediate reasoning paths With anchored evidence, a traceable, structured diagnostic report is used by clinicians to trace and verify diagnostic conclusions item by item.

[0055] In this implementation, when generating interpretability reports, the system establishes a heterogeneous evidence classification and anchoring mechanism: for textual and knowledge-based evidence, semantic retrieval is performed from external medical literature and guideline vector databases using retrieval enhancement generation technology, mandating that the output diagnostic conclusions must cite specific clauses from authoritative guidelines; for visual perception-based evidence, a cross-modal space mapping algorithm is used to precisely align diagnostic inferences to medical image ROI heatmaps or lesion segmentation masks generated by the expert agent's internal reasoning. By combining external retrieval knowledge with internal perception features, a structured decision output with full-link traceability is formed.

[0056] On the other hand, this embodiment provides an emergency system based on a large language model scheduling and multi-agent collaboration emergency method, including: The multimodal perception module receives the patient's original multimodal dataset, standardizes and concatenates the original multimodal dataset to obtain a fused multimodal contextual feature set, and calls the patient's longitudinal health record to concatenate the fused multimodal contextual feature set with the longitudinal health record to generate a multimodal medical semantic vector. The agent scheduling module is used to input multimodal medical semantic vectors into a large language model fine-tuned by agent instructions, generate intermediate reasoning paths including differential diagnosis hypotheses, and generate subtask calling instructions pointing to multiple dedicated expert agents based on the intermediate reasoning paths. The clinical pathway state machine module is used to receive intermediate inference paths, match the corresponding clinical state nodes using pre-embedded clinical pathway codes, and determine whether the fused multimodal context feature set has covered the predefined key information slot set in the clinical state node. If it has, the expert execution module is triggered; otherwise, the active guidance module is triggered. The triage level module is used to output the triage level based on the large language model with the preset emergency severity index standard, and according to the structured vital signs and biochemical test data, as well as the textual complaints and medical record data in the fusion multimodal context feature set. It determines whether the triage level is greater than the preset risk level. If so, the expert execution module is executed; otherwise, the active guidance module is executed. The active guidance module is used to generate candidate follow-up questioning strategies based on the missing information not covered in the key information slot set. Based on the preset candidate disease prior probability distribution corresponding to the differential diagnosis hypothesis, the information gain of each candidate follow-up questioning strategy is calculated, and the follow-up questioning strategy with the largest information gain is selected to guide the patient to complete the evidence. The completed evidence is updated to the fused multimodal context feature set and returned to the agent scheduling module. The expert execution module is used to distribute subtask invocation instructions to multiple corresponding dedicated expert agents to execute specialist perception tasks, including image segmentation, pathological analysis and risk scoring, in parallel, and output structured perception results with confidence. The neural symbol verification and report generation module is used to verify the structured perception results through pre-set logical rules based on hard-coded emergency clinical guidelines. The verification includes at least medical logic consistency review and contraindication comparison. If the verification passes, the structured perception result is determined as a diagnostic conclusion, and the diagnostic conclusion is mapped and anchored to the region of interest of the corresponding medical image and the clauses in the emergency clinical guidelines on which it is based. The mapping and anchoring results are used as the anchoring basis to generate a traceable structured diagnostic report containing the reasoning path and the anchoring basis. Otherwise, the large language model regenerates the reasoning path based on the reason for the verification failure and returns it to the agent scheduling module.

[0057] In summary, this invention constructs a hierarchical heterogeneous architecture with a large language model as the cognitive scheduling center and multiple specialized expert agents as execution units. It encodes the emergency clinical pathway into a hard sandbox boundary of a finite state machine for information completeness verification and proactive questioning guidance based on information gain. Furthermore, it employs a neural symbol dual verification mechanism and full-link evidence anchoring technology to ultimately generate a traceable structured diagnostic report containing a complete reasoning path and anchoring basis. This effectively overcomes the capability gap between general-purpose large models and specialized small models, achieving a paradigm shift from passive response to compliant proactive guidance. It eliminates the black-box effect of AI-assisted diagnosis and provides an intelligent comprehensive solution for emergency clinical decision-making that combines cognitive breadth and perceptual accuracy with diagnostic efficiency and safe traceability.

Claims

1. An emergency treatment method based on large language model scheduling and multi-agent collaboration, characterized in that, Including the following steps: S1. Obtain the patient's original multimodal dataset, and preprocess and encode the data of different modalities separately, then perform time alignment and cross-modal fusion to obtain a fused multimodal contextual feature set; wherein, the original multimodal dataset Including unstructured text-based chief complaints and medical record data Raw medical imaging scan data Pre-consultation voice recording data and structured vital signs and biochemical test data ; S2. Retrieve the patient's longitudinal health record and concatenate it with the fused multimodal contextual feature set to generate a multimodal medical semantic vector; S3. Input the multimodal medical semantic vector into the large language model fine-tuned by the intelligent agent instructions, generate intermediate reasoning paths including differential diagnosis hypotheses, and then generate subtask call instructions for multiple dedicated expert intelligent agents. S4, the large language model is based on a pre-defined emergency severity index standard, and according to the fusion of multimodal contextual features... as well as Output the triage level and determine whether the triage level is greater than the preset risk level. If it is, proceed to step S7; otherwise, proceed to step S5. S5. Determine whether the fused multimodal context feature set has covered the set of key information slots required for the diagnosis of its corresponding clinical state node. If yes, proceed to step S7; otherwise, proceed to step S6. S6. Generate candidate follow-up strategies for the missing information in the set of uncovered key information slots, calculate the information gain of each candidate follow-up strategy, select the follow-up strategy with the largest information gain to guide the patient to input supplementary evidence, update the supplementary evidence to the fused multimodal context feature set, and return to step S2 to update the multimodal medical semantic vector. S7. Distribute the subtask invocation instructions to the corresponding multiple dedicated expert agents, execute multiple specialized perception tasks in parallel, and output structured perception results that include at least the region of interest. S8. Verify all structured perception results. If all pass the verification, aggregate all structured perception results as a diagnostic conclusion and generate a structured diagnostic report; otherwise, use a large language model to feed back the reasons for failing the verification to the multimodal medical semantic vector and return to step S3.

2. The emergency treatment method based on large language model scheduling and multi-agent collaboration according to claim 1, characterized in that, Structured Diagnostic Report Including triage levels Candidate Diagnosis List Conclusions of the screening for critical illnesses Recommended Supplementary Checklist Contraindications and Risk Warnings Recommendations on handling priorities , represented as ; Triage Level The acquisition method includes: extracting the quantitative values ​​of vital signs from the structured vital signs and biochemical test data from the fused multimodal context feature set, and scoring the quantitative values ​​of vital signs according to the emergency severity index standard to obtain the objective basic risk score; then using a large language model to perform pathophysiological semantic analysis on the unstructured text complaints and medical record data extracted from the fused multimodal context feature set to obtain the symptom semantic risk weight, and weighting and fusing it with the objective basic risk score to obtain the comprehensive emergency priority index, and obtaining the corresponding triage level through the emergency preset grading threshold; Candidate Diagnosis List The acquisition methods include: the large language model extracts the disease type and confidence level corresponding to each structured perception result, and arranges all disease types in descending order of confidence level; Critical Illness Screening Conclusion The acquisition methods include: judging the candidate diagnosis list through a neural symbolic engine with a pre-built knowledge base of high-risk and life-threatening acute illnesses. If a high-risk, life-threatening emergency exists, output the conclusion that the high-risk, life-threatening emergency exists and the corresponding risk warning words; otherwise, output the conclusion that no emergency exists. Recommended Supplementary Checklist The acquisition method includes: acquiring all disease types with confidence levels below a preset threshold in the candidate diagnosis list and the set of key information slots required for their diagnosis; outputting the missing information in the set of key information slots not covered in the fusion multimodal context feature set; and aggregating all missing information as a set of recommended supplementary examination items. Contraindications and Risk Warnings The acquisition methods include: based on the patient's longitudinal health record information in the multimodal medical semantic vector, outputting absolute drug contraindications and contraindication warnings; Recommendations for handling priorities The acquisition methods include: based on the disease type with the highest confidence in the candidate diagnosis list, matching the intervention plan for the disease type with a large language model that has a pre-set authoritative emergency clinical guideline vector library, and outputting early rescue measures or specialist admission recommendations that follow the order of time urgency.

3. The emergency treatment method based on large language model scheduling and multi-agent collaboration according to claim 1, characterized in that, Methods for fine-tuning large language models using agent instructions include: The multimodal medical semantic vector samples are used as the sample input of the large language model, the intermediate reasoning path samples of differential diagnosis hypotheses and the subtask call instruction samples pointing to the dedicated expert agent are used as the sample output of the large language model. We will use a large language model with sample inputs and sample outputs, and fine-tune the large language model by minimizing the autoregressive cross-entropy loss.

4. The emergency treatment method based on large language model scheduling and multi-agent collaboration according to claim 1, characterized in that, Step S3 includes: S31. Input the multimodal medical semantic vector into a large language model fine-tuned by the intelligent agent's instructions, and use thought chain reasoning to generate a probability model containing differential diagnosis hypotheses and satisfying autoregressive conditional probabilities. The intermediate reasoning path, The calculation expression is: in, For the first Step-by-step reasoning logic state; This represents the number of reasoning steps. These are the training parameters for the large language model; This is a multimodal medical semantic vector; S32. Based on the intermediate reasoning path Using multimodal medical semantic vectors, a large language model is employed to generate a set of execution subtasks pointing to multiple specialized expert agents, and the optimal set of execution subtasks is obtained by maximizing conditional probabilities. : in, To execute a set of subtasks The optimization operation that finds the maximum posterior probability; To provide a multimodal medical semantic vector and the generated intermediate inference paths Under the given conditions, the large language model generates the optimal set of subtasks. The conditional probability.

5. The emergency treatment method based on large language model scheduling and multi-agent collaboration according to claim 1, characterized in that, Step S5 includes: S51, intermediate reasoning path The input is a pre-built finite state machine that encodes emergency clinical pathways; S52. Extracting intermediate reasoning paths using finite state machines. The standardized medical entities are matched with clinical pathway codes to determine the corresponding clinical status nodes. ; S53. Obtain Clinical Status Nodes Predefined set of key information slots And obtain the current multimodal medical semantic vector. The set of valid information slots that have been successfully extracted Calculate the set of missing information ; S54, Judgment Does it meet the requirements? If yes, proceed to step S7; otherwise, proceed to step S6.

6. The emergency treatment method based on large language model scheduling and multi-agent collaboration according to claim 5, characterized in that, Step S6 includes: S61. Based on the missing information set The missing information slots in the database are mapped from a pre-defined medical knowledge graph to generate a set of candidate follow-up questioning strategies. ,in, Each candidate follow-up question strategy corresponds to a key information slot to be filled and a natural language guiding statement for outputting to the patient. S62, Based on the intermediate reasoning path Differential diagnostic hypotheses are used to obtain pre-defined candidate diseases. Prior probability distribution And calculate the entropy of current diagnostic uncertainty information. : ； in, For the number of candidate diseases, For the first One candidate disease; S63, Set of Follow-up Questioning Strategies for Candidates For each candidate follow-up question strategy, obtain its corresponding set of potential effective response values, and combine this with the conditional probability relationship between diseases and symptoms in the medical knowledge graph to calculate the expected conditional information entropy under the assumption of obtaining an effective response. : ； ； ； in, for The Middle One candidate follow-up questioning strategy; A set of valid responses; To propose a strategy of probing questions Get a reply under the following conditions The marginal probability; In order to get a reply Under the condition of updating the posterior probability of the disease using Bayes' theorem; S64. Calculate the information gain of all candidate follow-up strategies: ； in, for Information gain; S65. Select the candidate follow-up questioning strategy with the highest information gain. Output Corresponding natural language guidance scripts are used to guide patients to input missing information; S66, The patient's... The missing information in the input is used as completion evidence to update the fused multimodal contextual feature set. Then return to step S2.

7. The emergency treatment method based on large language model scheduling and multi-agent collaboration according to claim 4, characterized in that, Step S7 specifically includes: S71. Based on the optimal set of subtasks to be executed. The identifier number of each subtask will be The subtask invocation instructions are distributed to the corresponding target specialty agents in multiple dedicated expert agents; S72. Each target specialty agent executes specialty perception tasks in parallel on independent computing nodes and outputs structured perception results. S73. Encapsulate the structured perception results of all target specialty agents into standard structured perception results.

8. The emergency treatment method based on large language model scheduling and multi-agent collaboration according to claim 1, characterized in that, Step S8 includes: S81. The standard structured perception results are verified using pre-built logical rules in the neural symbol engine and hard-coded based on emergency clinical guidelines. The verification includes at least medical logic consistency review and contraindication comparison. The neural symbol engine adopts an asynchronous interception architecture, and its symbol end deploys a rule engine and hard-codes absolute contraindications and logical predicates in the emergency clinical guidelines. S82. If the structured perception result passes the verification, the structured perception result is determined as the diagnostic conclusion; otherwise, the neural symbol engine generates a negative prompt word containing the specific reason for the violation and feeds it back to the large language model, forcing the large language model to regenerate the intermediate reasoning path based on the negative prompt word and return to step S3. S83. Map and anchor the verified diagnostic conclusions to the corresponding regions of interest in the medical images and the authoritative guidelines on which they are based, respectively, to generate anchoring basis; S84. Generate the intermediate inference path. Traceable, structured diagnostic reports with anchoring basis.

9. An emergency medical system based on large language model scheduling and multi-agent collaboration, characterized in that, include: The multimodal perception module is used to acquire the patient's original multimodal dataset, perform standardization processing, and then concatenate them to obtain a fused multimodal contextual feature set; The patient's longitudinal health record is retrieved and concatenated with a fused multimodal contextual feature set to generate a multimodal medical semantic vector; The agent scheduling module is used to input multimodal medical semantic vectors into a large language model that is fine-tuned by agent instructions, generate intermediate reasoning paths including differential diagnosis hypotheses, and generate subtask calling instructions that point to multiple dedicated expert agents. The triage level module is used to output the triage level based on the large language model with the preset emergency severity index standard, and according to the structured vital signs and biochemical test data, as well as the textual complaints and medical record data in the fusion multimodal context feature set. It determines whether the triage level is greater than the preset risk level. If so, the expert execution module is executed; otherwise, the active guidance module is executed. The clinical pathway state machine module is used to input the intermediate reasoning path into a pre-built finite state machine embedded with clinical pathway codes, and determine whether the fused multimodal context feature set is the set of key information slots required for the diagnosis of the corresponding clinical state node. If it is covered, the expert execution module is triggered; otherwise, the active guidance module is triggered. The active guidance module is used to generate candidate follow-up questioning strategies based on the missing information in the set of uncovered key information slots, calculate the information gain of each candidate follow-up questioning strategy, select the follow-up questioning strategy with the largest information gain to guide the patient to input supplementary evidence, update the supplementary evidence to the fused multimodal context feature set and return it to the multimodal perception module. The expert execution module is used to distribute subtask invocation instructions to multiple corresponding dedicated expert agents, execute multiple specialized perception tasks in parallel, and output structured perception results that include at least the region of interest. The neural symbol verification and report generation module is used to verify the structured perception results. If the verification passes, the structured perception result is determined as the diagnostic conclusion. Otherwise, the reason for failing the verification is fed back to the multimodal medical semantic vector using a large language model and returned to the agent scheduling module.

Images