Multi-agent-based auxiliary treatment decision-making method and system and storage medium

Abstract

The invention provides an auxiliary doctor seeing decision method and system based on multiple agents and a storage medium, and the method comprises the steps: carrying out the structural information collection of a target patient according to a pre-diagnosis agent, and obtaining the structural data of the patient; inputting the structured data of the patient into a diagnostic reasoning agent for diagnostic reasoning to obtain a diagnostic predicted disease; inputting the diagnosed and predicted disease and the current medical resource load information of the doctor-seeing hospital into a resource scheduling agent for diagnosis and treatment analysis to obtain a diagnosis and treatment recommendation scheme; inputting the diagnosis and treatment recommendation scheme into a medical insurance strategy agent for medical insurance analysis to obtain a medical insurance recommendation scheme; and combining the patient structured data, the diagnosis and prediction disease, the diagnosis and treatment recommendation scheme and the medical insurance recommendation scheme to obtain auxiliary treatment information. According to the embodiment of the invention, based on the auxiliary doctor-seeing information, the auxiliary doctor-seeing decision corresponding to the disease diagnosis, the diagnosis and treatment scheme and the medical insurance scheme can be effectively output, and the diversity of the auxiliary doctor-seeing decision function is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare technology, and in particular to a method, system, and storage medium for assisted medical decision-making based on multi-agent systems. Background Technology

[0002] AI-supported smart healthcare decision support technology is a recent emerging research hotspot. Essentially, it applies artificial intelligence and machine learning techniques to summarize and generalize the characteristics of big data in healthcare, generating corresponding diagnostic networks for use in diagnosing patients' clinical diseases and assisting doctors in making clinical decisions. Smart healthcare technology not only improves doctors' work efficiency but also balances medical resources, enabling even areas with underdeveloped healthcare to access high-quality smart medical services.

[0003] In existing auxiliary medical decision-making processes, auxiliary medical decision-making devices generally output a binary conclusion of "yes or no" for the disease, and cannot provide other auxiliary decision-making information, resulting in a single auxiliary medical decision-making function. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-agent-based method, system, and storage medium for assisting medical treatment decision-making, in order to solve the problem of limited functionality in existing assisting medical treatment decision-making technologies.

[0005] The present invention is implemented as follows: a multi-agent-based assisted medical decision-making method, the method comprising:

[0006] The pre-diagnosis intelligent agent collects structured information from target patients to obtain structured patient data;

[0007] The patient's structured data is input into the diagnostic reasoning agent to perform diagnostic reasoning and obtain a diagnosis and predicted disease.

[0008] The system obtains the hospitals visited by the target patient, and inputs the diagnosed and predicted disease and the current medical resource load information of the hospitals into the resource scheduling agent for diagnosis and treatment analysis to obtain a recommended treatment plan.

[0009] The treatment recommendation plan is input into the medical insurance strategy intelligent agent for medical insurance analysis to obtain the medical insurance recommendation plan.

[0010] The patient's structured data, the diagnosed and predicted diseases, the recommended treatment plan, and the medical insurance recommendation plan are combined to obtain auxiliary medical information, which is then sent to the attending physician corresponding to the target patient.

[0011] Preferably, after sending the auxiliary medical information to the treating physician corresponding to the target patient, the method further includes:

[0012] The diagnosis and treatment results of the target patient are obtained, and the diagnosis and treatment results are input into the feedback evolutionary agent for reward and punishment analysis to obtain the agent's reward and punishment information;

[0013] Based on the agent reward and punishment information, the action reward and punishment are adjusted for the pre-diagnosis agent, the diagnosis reasoning agent, the resource scheduling agent, or the medical insurance strategy agent.

[0014] Preferably, after collecting structured information about the target patient based on the pre-diagnosis intelligent agent and obtaining the patient's structured data, the method further includes:

[0015] When a collaborative concurrent task is received from any intelligent agent, the collaborative concurrent task is input into the collaborative control intelligent agent for collaborative analysis to obtain the collaborative object;

[0016] Send a synchronization command to the collaborative object to control the collaborative object to execute the collaborative concurrent task.

[0017] Preferably, after sending the auxiliary medical information to the treating physician corresponding to the target patient, the method further includes:

[0018] Based on the diagnosis and treatment results of the target patient and the marking information of the treating physician, identify the cases of assisted medical decision-making failure in the historical assisted medical decision-making data, and label the root causes of misdiagnosis in the cases of assisted medical decision-making failure;

[0019] A failure assistance library is constructed based on the auxiliary failure cases labeled with the root causes of misdiagnosis.

[0020] When new patient structured data is collected, the new patient structured data is matched with the failure assistance library by label;

[0021] If the new patient structured data successfully matches the tags in the failure assistance library, the transmission of the synchronization command is stopped, and an assistance error message is sent.

[0022] Preferably, the method further includes:

[0023] The medical record image sent by the target patient is obtained, and the image is recognized according to the target detection model to obtain the text region;

[0024] The text region is converted into text to obtain medical record text, and the medical record text is input into a pre-trained named entity recognition model for entity recognition to obtain a target entity set.

[0025] The target entity set is matched with a medical knowledge graph, and a medication recommendation report is generated based on the entity matching results.

[0026] The medication recommendation report will be sent to the target patient.

[0027] Preferably, before inputting the medical record text into the pre-trained named entity recognition model for entity recognition, the method further includes:

[0028] Obtain training samples, determine annotation information based on different types of sample entities in the training samples, and label the training samples according to the different annotation information to obtain model samples;

[0029] Random prompt templates are randomly selected from the prompt template library to obtain random prompt templates. The named entity recognition large model is then trained based on the random prompt templates and the model samples until the named entity recognition large model converges, thus obtaining the pre-trained named entity recognition large model.

[0030] Preferably, training the named entity recognition large model based on the random prompt template and the model samples includes:

[0031] The named entity recognition model is used to segment the model samples to obtain sample words, and the sample words are then subjected to word embedding processing to obtain sample vectors.

[0032] The target entity type is determined based on the random prompt template, and the target vector is determined based on the target entity type;

[0033] The output vector in the sample vector is determined based on the target vector, and the model loss is determined based on the output vector;

[0034] The parameters of the named entity recognition model are updated based on the model loss.

[0035] Another objective of this invention is to provide a multi-agent-based assisted medical decision-making system, the system comprising:

[0036] The information collection module is used to collect structured information about the target patient based on the pre-diagnosis intelligent agent, and obtain structured patient data.

[0037] The diagnostic reasoning module is used to input the patient's structured data into the diagnostic reasoning agent to perform diagnostic reasoning and obtain a diagnosis and prediction of the disease.

[0038] The diagnosis and treatment analysis module is used to obtain the hospital visited by the target patient, and input the diagnosed and predicted disease and the current medical resource load information of the hospital into the resource scheduling agent for diagnosis and treatment analysis to obtain a recommended diagnosis and treatment plan;

[0039] The medical insurance analysis module is used to input the treatment recommendation plan into the medical insurance strategy intelligent agent for medical insurance analysis to obtain the medical insurance recommendation plan;

[0040] The decision support module is used to combine the patient's structured data, the diagnosed and predicted diseases, the recommended treatment plan, and the medical insurance recommendation plan to obtain auxiliary medical information, and send the auxiliary medical information to the attending physician corresponding to the target patient.

[0041] Preferably, the multi-agent-based assisted medical decision-making system further includes:

[0042] The reward and punishment adjustment module is used to obtain the diagnosis and treatment results of the target patient and input the diagnosis and treatment results into the feedback evolutionary agent for reward and punishment analysis to obtain the agent's reward and punishment information.

[0043] Based on the agent reward and punishment information, the action reward and punishment are adjusted for the pre-diagnosis agent, the diagnosis reasoning agent, the resource scheduling agent, or the medical insurance strategy agent.

[0044] In this embodiment of the invention, the pre-diagnosis intelligent agent can effectively collect structured information from target patients, and the diagnostic reasoning intelligent agent can effectively perform diagnostic reasoning on the patient's structured data to obtain a diagnosis and disease prediction. By inputting the diagnosis and disease prediction and the current medical resource load information into the resource scheduling intelligent agent for diagnosis and treatment analysis, the allocation of diagnosis and treatment resources can be effectively carried out to obtain a recommended diagnosis and treatment plan. By inputting the recommended diagnosis and treatment plan into the medical insurance strategy intelligent agent for medical insurance analysis, the optimal path medical insurance recommendation plan can be effectively output. Based on auxiliary medical treatment information, the auxiliary medical treatment decision corresponding to the disease diagnosis, diagnosis and treatment plan and medical insurance plan can be effectively output, ensuring the diversity of auxiliary medical treatment decision function. Attached Figure Description

[0045] Figure 1 This is a flowchart of the multi-agent-based assisted medical decision-making method provided in the first embodiment of the present invention;

[0046] Figure 2 This is a schematic diagram of the structure of the multi-agent-based assisted medical decision-making system provided in the second embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of the structure of the terminal device provided in the third embodiment of the present invention. Detailed Implementation

[0048] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0049] To illustrate the technical solution described in this invention, specific embodiments are described below.

[0050] This invention proposes a medical collaborative decision-making method based on multi-agent reinforcement learning. Its core lies in constructing a dynamic role division architecture, which analyzes a large and fuzzy "intelligent pre-triage" problem and organizes it into six clearly defined and goal-oriented roles. Six types of decoupled agents are deployed to address the patient's situation and needs. Each agent is trained using the mature QMIX algorithm (Q-decomposition Multi-agent Independent eXtension) with "centralized training with decentralized execution" (CTDE).

[0051] We construct pre-diagnosis agents, diagnostic reasoning agents, resource scheduling agents, medical insurance strategy agents, collaborative control agents, and feedback evolution agents, respectively. We independently train local Q-value networks for each agent, evaluate the value of its specific actions, construct a super network, input the global medical state feature vector, and output nonlinear combination parameters. In the hybrid network, we fuse the local Q-values ​​of all agents into the global Q-value.

[0052] The training process involves assigning roles to each agent: collecting historical electronic medical records, misdiagnosis cases, resource scheduling logs, medical insurance refusal cases, process delay cases, and patient outcome data; manually labeling the validity or compliance of actions; initializing the parameters of the local Q-network and supernetwork; in a simulated environment, each agent receives state input, outputs its own action, calculates the global reward based on preset indicators, updates the parameters of the local Q-network and supernetwork through backpropagation, iterates training until each agent's indicators meet the standards, and outputs the results for each agent.

[0053] Example 1

[0054] Please see Figure 1 This is a flowchart of a multi-agent-based assisted medical decision-making method provided in the first embodiment of the present invention. This multi-agent-based assisted medical decision-making method can be applied to any device or system, and includes the following steps:

[0055] Step S10: Collect structured information of the target patient based on the pre-diagnosis intelligent agent to obtain structured patient data;

[0056] Among them, the pre-diagnosis intelligent agent (Nurse Agent) is used to achieve structured collection of symptoms and classification of disease complexity through multi-turn dialogue state tracking;

[0057] Local Q-value definition: The local q-value of each pre-diagnosis agent corresponds to the value of "symptom collection, condition inquiry, and demand collection actions", reflecting the value of the action to "self-collection task completion degree".

[0058] Global Q-value synthesis: The hypernetwork is input with global medical status (such as the patient's current department, preliminary symptom labels, and hospital resource load) to generate nonlinear combination parameters, and the local Q-values ​​of all pre-diagnosis agents are merged into a global Q-value.

[0059] For example, if a patient is suspected of having a myocardial infarction (the global status includes the labels "chest pain + profuse sweating"), the supernetwork increases the local Q value weight of the "inquire about ECG history" action, making the global Q value more inclined to promote this key acquisition action.

[0060] The pre-training process of the pre-diagnosis agent includes:

[0061] Collect data on the "nurse consultation-symptom collection" process from historical electronic medical records (EMRs) and annotate key actions (such as effective follow-up questions and redundant operations). For example, extract effective actions such as nurses asking "location of chest pain" and "whether it is accompanied by palpitations" from 1000 myocardial infarction medical records as positive samples;

[0062] Initialize the local Q-network and supernetwork parameters of the pre-diagnosis agent; simulate the diagnosis and treatment scenario, the pre-diagnosis agent outputs the collection action, triggers symptom structuring (such as generating the label "chest pain + radiating to the left arm"); calculate the global reward (such as the score of "whether the symptoms required for the diagnosis of myocardial infarction are fully covered"); and update the local Q-network and supernetwork in reverse.

[0063] The system undergoes numerous iterations of training until the completeness of symptom collection stabilizes at a high level. The pre-trained pre-diagnosis agent then engages in multiple rounds of dialogue with the target patient to obtain dialogue information. This dialogue information is then processed in a structured manner to obtain structured patient data.

[0064] Step S20: Input the patient's structured data into the diagnostic reasoning agent to perform diagnostic reasoning and obtain a diagnostic prediction of the disease;

[0065] Among them, the Doctor Agent is used to make diagnostic decisions based on a clinical guideline knowledge base and a graph neural network (GNN);

[0066] Local Q-value definition: The local Q-value of a diagnostic reasoning agent corresponds to the value of a "diagnostic action (such as invoking guideline clauses or triggering differential checks)," reflecting the value of that action to the "accuracy of its own diagnosis."

[0067] Global Q-value synthesis: The hypernetwork takes the global medical state as input and generates local Q-values ​​for the diagnostic reasoning agent based on parameter combinations. For example:

[0068] If no relevant disease is found, the super network reduces the local Q-value weight of the "common disease search" action and increases the weight of the "rare disease analysis" action to ensure overall diagnostic efficiency.

[0069] The pre-training process of the diagnostic reasoning agent includes:

[0070] Organize historical misdiagnosis cases (such as misdiagnosing myocardial infarction as angina) and guideline-compliant cases, and mark the "effectiveness" of diagnostic actions. For example, from 500 misdiagnosis medical records, mark errors such as "failure to perform Holter monitoring in a timely manner" as negative samples;

[0071] Initialize the local Q-network and hypernetwork parameters of the diagnostic reasoning agent; input the structured symptoms output by the pre-diagnosis agent (such as "chest pain + troponin threshold"), and the diagnostic reasoning agent outputs diagnostic actions; calculate the global reward by combining diagnostic accuracy and treatment efficiency; update the network parameters in reverse and iteratively train until the diagnostic accuracy (such as the accuracy of identifying myocardial infarction) stabilizes at a high level; use the pre-trained diagnostic reasoning agent to perform diagnostic reasoning on the patient's structured data to obtain a diagnostic prediction of the disease.

[0072] Step S30: Obtain the hospital visited by the target patient, and input the diagnosed and predicted disease and the current medical resource load information of the hospital into the resource scheduling agent for diagnosis and treatment analysis to obtain a recommended diagnosis and treatment plan;

[0073] Among them, the resource scheduling intelligent agent (Hospital Agent) is used to access the hospital-related information data stream in real time, combine the hidden Markov model to predict the load of medical resources and recommend treatment plans.

[0074] The local Q-value is defined as the value of a resource allocation action (such as bed scheduling or equipment reservation).

[0075] Global Q-value synthesis: The global medical state is input into the hypernetwork, and the local Q-values ​​of the scheduling agents are combined.

[0076] For example:

[0077] If there are two patients with myocardial infarction at the same time (the global status includes "only 1 emergency bed left"), the super network increases the local Q value weight of the "coordinate temporary addition of an operating room bed" action to avoid resource conflicts.

[0078] The pre-training process of the resource scheduling agent includes:

[0079] Data preparation: Extract data such as resource scheduling logs from the hospital's HIS system and mark "efficient allocation" and "conflicting allocation" actions. For example, from 1000 emergency bed allocations, mark controversial actions such as "prioritizing general beds for myocardial infarction patients" as optimization samples;

[0080] Simulate concurrent scenarios with multiple patients (e.g., 1 myocardial infarction + 1 stroke + 1 general emergency room visit), the resource scheduling agent outputs resource allocation actions; global rewards are calculated based on resource utilization (e.g., CT room idling rate) and patient waiting time; the network is updated in reverse, and iteratively trained until the resource conflict rate is reduced to a low level. Based on the pre-trained resource scheduling agent, diagnostic and treatment analysis is performed on the diagnosed and predicted diseases and the current medical resource load information to obtain recommended treatment plans.

[0081] Step S40: Input the treatment recommendation plan into the medical insurance strategy intelligent agent for medical insurance analysis to obtain the medical insurance recommendation plan;

[0082] Among them, the Insurance Agent is used to calculate the optimal reimbursement path by relying on the medical insurance policy knowledge graph and using dynamic programming algorithm.

[0083] Local Q-value is defined as the value of "medical insurance intervention actions (such as medication replacement and examination item screening)".

[0084] Global Q-value synthesis: The global medical status is input into the hypernetwork, and the local Q-values ​​of the medical insurance agent are combined.

[0085] For example:

[0086] If the patient is critically ill and has sufficient medical insurance balance (global status includes "medical insurance reimbursement limit remaining 100,000"), the super network reduces the local Q value weight of the "forced replacement of high-priced drugs" action to prioritize the treatment effect.

[0087] The pre-training process of the healthcare strategy agent includes:

[0088] Collect cases of medical insurance refusal to pay (such as refusal to pay due to use of drugs outside the medical insurance catalog) and compliant treatment cases, and mark the "compliance" of medical insurance intervention actions. For example, from 500 medical insurance refusal medical records, extract violations such as "use of unreimbursed consumables" as negative samples;

[0089] Input the treatment plan (e.g., "imported stent + high-end anticoagulant" for myocardial infarction patients), and the medical insurance intelligent agent outputs intervention actions (e.g., "replace with medical insurance stent"); global rewards are calculated based on medical insurance refusal rate and patient treatment effect (e.g., postoperative complication rate);

[0090] The network is updated in reverse and iteratively trained until the medical insurance rejection rate stabilizes at a low level. The pre-trained medical insurance strategy agent then performs medical insurance analysis on the treatment recommendation plan to obtain the medical insurance recommendation plan.

[0091] Step S50: Combine the patient's structured data, the diagnosed and predicted disease, the treatment recommendation plan, and the medical insurance recommendation plan to obtain auxiliary medical information, and send the auxiliary medical information to the doctor corresponding to the target patient.

[0092] Among them, based on auxiliary medical information, it can effectively output auxiliary medical decision-making corresponding to disease diagnosis, treatment plan and medical insurance plan, thus ensuring the diversity of auxiliary medical decision-making functions.

[0093] Optionally, after sending the auxiliary medical information to the treating physician corresponding to the target patient, the method further includes:

[0094] The diagnosis and treatment results of the target patient are obtained, and the diagnosis and treatment results are input into the feedback evolutionary agent for reward and punishment analysis to obtain the agent's reward and punishment information;

[0095] Based on the agent reward and punishment information, the action reward and punishment of the pre-diagnosis agent, the diagnosis reasoning agent, the resource scheduling agent, or the medical insurance strategy agent are adjusted.

[0096] Among them, the Feedback Agent is used to deploy a reinforcement learning reward mechanism to quantify patient satisfaction and resource utilization.

[0097] The local Q-value is defined as the value of a "reward-shaping action" (such as imposing a penalty on a pre-diagnostic agent or a reward on a diagnostic reasoning agent).

[0098] Global Q-value synthesis: The hypernetwork takes the global medical status as input and combines it to generate the global Q-value of the agent. For example, if a patient dies due to a missed diagnosis (the global status includes "serious adverse event"), the hypernetwork increases the local Q-value weight of the actions of "reducing the decision-making weight of the diagnostic agent + strengthening the nurse's data collection training", thus promoting the overall evolution of the system.

[0099] The pre-training process of a feedback evolutionary agent includes:

[0100] Collect patient outcome data (recovery / death / complications) and agent historical strategies (such as nurse consultation records and doctor diagnostic pathways), and label them with "decision effectiveness." For example, from 1000 recovery medical records, mark effective strategies such as "detailed nurse consultation of family medical history" as positive samples;

[0101] Input the treatment outcome and the agent outputs reward / punishment actions; global reward is calculated by combining the patient's long-term prognosis and the diversity of agent strategies; the network is updated in reverse and iterated until the patient's recovery rate continues to improve to a stable level after the system strategy is iterated.

[0102] Each agent operates in coordination with an event-driven workflow, triggering chain collaboration with the event of disease onset to recovery as the main event.

[0103] Furthermore, after the pre-diagnosis intelligent agent collects structured information about the target patient and obtains the patient's structured data, it also includes:

[0104] When a collaborative concurrent task is received from any agent, the collaborative concurrent task is input into the collaborative control agent for collaborative analysis to obtain a collaborative object; a synchronization command is sent to the collaborative object to control the collaborative object to execute the collaborative concurrent task.

[0105] Among them, the Coordinate Agent is used to achieve asynchronous communication and conflict resolution between agents through the process engine, so as to realize collaborative cooperation;

[0106] Local Q-value is defined as the value of a "synchronous instruction action (such as triggering the next step of diagnosis or resolving conflicts)".

[0107] Global Q-value synthesis: The hypernetwork inputs the global medical status and combines the global Q-values ​​of the collaborative agents. For example, if the nurse has completed symptom collection and the CT scan has been scheduled (the global status includes "preparing for diagnosis"), the hypernetwork increases the local Q-value weight of the action "triggering the doctor's diagnosis process" to accelerate the diagnosis and treatment loop.

[0108] The pre-training process of the collaborative control agent includes:

[0109] Data preparation: Identify cases of delays in the medical process (such as test results not triggering a diagnosis in a timely manner) and mark them as "effective synchronized actions".

[0110] Simulate concurrent tasks of multiple agents, with the coordinating agents outputting synchronized instructions; calculate global rewards based on the time consumption of the diagnosis and treatment process and the conflict rate; update the network in reverse, iteratively training until the process delay rate is reduced to a low level.

[0111] Furthermore, after sending the auxiliary medical information to the treating physician corresponding to the target patient, the process also includes:

[0112] Based on the diagnosis and treatment results of the target patient and the marking information of the treating physician, identify the cases of assisted medical decision-making failure in the historical assisted medical decision-making data, and label the root causes of misdiagnosis in the cases of assisted medical decision-making failure;

[0113] A failure assistance library is constructed based on the auxiliary failure cases labeled with the root cause of misdiagnosis. When new patient structured data is collected, the new patient structured data is matched with the failure assistance library for label matching.

[0114] If the new patient structured data successfully matches the tags in the failure assistance library, the transmission of the synchronization command is stopped, and an assistance error message is sent.

[0115] In this embodiment, the event sequence from disease onset to recovery is used as the main event. Each agent triggers chain-like collaboration and asynchronous communication by publishing / subscribing to specific events. A memory and agent evolution mechanism is deployed: a successful case library storing structured decision trajectories is constructed; a failure experience library storing failed cases and labeling root causes is constructed; in the processing of new cases, failure mode labels are matched in real time and erroneous decision paths are interrupted; each agent, combined with the updated case library, uses a meta-learning optimizer to fine-tune policy parameters online to optimize the global reward function.

[0116] Autonomous agent evolution mechanism based on memory mechanism:

[0117] A "successful triage case library" is constructed to store structured trajectories, and a "failure experience library" is constructed to label the root causes of misdiagnosis (such as missing symptom associations and misjudgment of high-risk groups). The strategy is updated online through a meta-learning optimizer: when a new case matches the failure label, the erroneous triage path is blocked in real time. Each agent combines the two continuously enriched databases to perform self-evolution, with the evolution goal of maximizing patient satisfaction and medical resource matching.

[0118] In this embodiment, the method further includes:

[0119] The medical record image sent by the target patient is obtained, and the image is recognized according to the target detection model to obtain the text region; wherein, the text in the medical record image is located by the target detection model to obtain the text region.

[0120] The text region is converted into text to obtain medical record text, and the medical record text is input into a pre-trained named entity recognition model for entity recognition to obtain a target entity set.

[0121] The target entity set is matched with a medical knowledge graph, a medication recommendation report is generated based on the entity matching results, and the medication recommendation report is sent to the target patient.

[0122] This embodiment also includes a medical record intelligent parsing module: training a high-performance Chinese medical NER model (named entity recognition large model), the process of which includes constructing a large-scale labeled dataset, adopting a decomposed question-and-answer labeling paradigm, and implementing a dynamic prompting strategy for randomized training; executing a three-level processing flow: after receiving the medical record image, using an object detection model to identify text regions; calling a medical optimization OCR engine to convert the text; inputting the text into the NER model to identify structured medical entities; associating the entities with the medical knowledge graph of multi-source heterogeneous data in the background, performing information completion and logical verification; and finally generating a personalized medication recommendation report by comprehensively analyzing and verifying the results.

[0123] Specifically, intelligent parsing is based on a three-tiered processing architecture of "medical OCR + Chinese medical NER + knowledge graph reasoning":

[0124] The system receives medical record images (such as discharge summaries, medication records, and examination reports) uploaded by target patients. First, it uses an object detection model to identify key areas such as text blocks and tables in the image, eliminating interfering information. Then, it calls a medically optimized OCR engine to perform high-precision text conversion on the identified text areas, outputting unstructured medical record text. The OCR-output medical record text is then input into a pre-trained named entity recognition (NAME) model. This pre-trained model leverages its powerful semantic understanding and generalization capabilities, as well as robustness gained through training with diverse prompts, to accurately identify and extract multiple categories of medical entities from the text, which may have nested relationships. These entities include, but are not limited to, drug information, medication instructions, clinical findings, laboratory indicators, and contraindications. The extracted structured entity information is then used as input and dynamically correlated with a medical knowledge graph built in the background. This knowledge graph integrates multi-source heterogeneous data such as patient historical health records, medical insurance drug catalogs, authoritative clinical guidelines, and drug instructions. Through graph reasoning, the system automatically completes the following information completion and logical verification: personalized information matching and security risk verification. By synthesizing all analysis results, a clear, personalized, and safe medication recommendation report is generated for the user. This report not only includes the medication plan extracted from medical record images but also adds risk warnings based on the patient's individual circumstances, medical insurance reimbursement estimates, and relevant health guidance. This overcomes the fundamental limitation of traditional OCR technology, which can only convert text but cannot perform deep semantic understanding and logical verification.

[0125] Preferably, before inputting the medical record text into the pre-trained named entity recognition model for entity recognition, the method further includes:

[0126] Obtain training samples, determine annotation information based on different types of sample entities in the training samples, and label the training samples according to the different annotation information to obtain model samples;

[0127] Random prompt templates are randomly selected from the prompt template library to obtain random prompt templates. The named entity recognition large model is trained based on the random prompt templates and the model samples until the named entity recognition large model converges, thus obtaining the pre-trained named entity recognition large model.

[0128] The high performance and robustness of the large-scale Chinese medical named entity recognition model used in this embodiment are mainly achieved through the following innovative training methods:

[0129] Construction of a large-scale, multi-scenario medical dataset: To address the problem of scarce training data, a Chinese medical NER dataset covering eight major medical scenarios was first constructed. This dataset contains over 40,000 samples and over 3,400 entity types, with data sources including anonymized real-world online medical data, online doctor-patient consultation texts, authoritative medical books, and drug instructions. The scale and scenario diversity of this dataset provide a solid foundation for the model to learn complex medical linguistic phenomena.

[0130] An innovative decomposed question-answering annotation method addresses the low recall rate of traditional direct annotation methods when handling complex or nested entities. This method employs a decomposed question-answering data annotation and selection paradigm. Specifically, for a text, the annotation process is broken down into a series of questions targeting different entity types. This approach guides the model to analyze the text step-by-step, much like a human expert, significantly improving its ability to recognize complex and nested entities. Experiments show that, compared to direct annotation, the model trained using this method achieves a 6% improvement in F1 score.

[0131] Robust Training Strategy Based on Dynamic Prompt Randomization: To enhance the large model's ability to handle diverse and non-standard user input in real-world applications (i.e., response sensitivity and robustness), diverse prompt templates were designed. During model training, Dynamic Prompt Strategy Randomization was implemented. In each training batch, the system randomly selects one or more combinations from a pre-defined prompt template library to construct the input, forcing the model to learn the commonalities of entity extraction under different question formats, rather than simply relying on a fixed prompt format.

[0132] The model trained using the above method shows a 5% performance improvement on the comprehensive test set compared to conventional training methods. More importantly, this training strategy demonstrates exceptional efficiency, enabling a 7 billion (7B) parameter model to outperform a 32 billion (32B) parameter model trained without this method on the NER task.

[0133] Further, the named entity recognition large model is trained based on the random prompt template and the model samples, including:

[0134] The named entity recognition model is used to segment the model samples to obtain sample words, and then the sample words are embedded to obtain sample vectors; wherein, by performing word embedding on the sample words, the word features of the sample words can be effectively extracted;

[0135] The target entity type is determined based on the random prompt template, and the target vector is determined based on the target entity type; wherein, the template identifier of the random prompt template is obtained, the template identifier is matched with the type lookup table to obtain the target entity type, the type lookup table stores the correspondence between different template identifiers and corresponding target entity types, and the vector corresponding to the target entity type is obtained to obtain the target vector;

[0136] The output vector in the sample vector is determined based on the target vector, the model loss is determined based on the output vector, and the parameters of the named entity recognition model are updated based on the model loss. Specifically, the similarity between the sample vector and the target vector is calculated to obtain the vector similarity. The sample vector with the vector similarity and the similarity threshold is determined as the output vector. The loss is calculated based on the output vector and the annotation information to obtain the model loss.

[0137] In this embodiment, the pre-diagnosis intelligent agent can effectively collect structured information from target patients, and the diagnostic reasoning intelligent agent can effectively perform diagnostic reasoning on the patient's structured data to obtain a diagnosis and disease prediction. By inputting the diagnosis and disease prediction and current medical resource load information into the resource scheduling intelligent agent for treatment analysis, treatment resources can be effectively allocated to obtain treatment recommendations. By inputting the treatment recommendations into the medical insurance strategy intelligent agent for medical insurance analysis, the optimal path medical insurance recommendation can be effectively output. Based on auxiliary medical information, auxiliary medical decisions corresponding to disease diagnosis, treatment plan, and medical insurance plan can be effectively output, ensuring the diversity of auxiliary medical decision-making functions. In this embodiment, a multi-agent architecture with dynamic role division is adopted. Cross-platform collaboration of patient, hospital, and medical insurance data is realized through event-driven workflow to improve the accuracy of decision-making. Relying on reinforcement learning reward mechanism and memory bank evolution mechanism, resource scheduling strategy is optimized in real time and iterates itself to solve the problems of poor dynamic adaptability and lack of evolutionary ability.

[0138] Example 2

[0139] Please see Figure 2 This is a schematic diagram of the structure of a multi-agent-based assisted medical decision-making system 100 provided in the second embodiment of the present invention, including:

[0140] The information collection module 10 is used to collect structured information about the target patient based on the pre-diagnosis intelligent agent, and obtain structured patient data.

[0141] The diagnostic reasoning module 11 is used to input the patient's structured data into the diagnostic reasoning agent to perform diagnostic reasoning and obtain a diagnosis and prediction of the disease.

[0142] The diagnosis and treatment analysis module 12 is used to obtain the hospital where the target patient has been treated, and input the diagnosed and predicted disease and the current medical resource load information of the hospital into the resource scheduling agent for diagnosis and treatment analysis to obtain a recommended treatment plan.

[0143] The medical insurance analysis module 13 is used to input the treatment recommendation plan into the medical insurance strategy intelligent agent for medical insurance analysis to obtain the medical insurance recommendation plan.

[0144] The auxiliary decision-making module 14 is used to combine the patient's structured data, the diagnosed and predicted disease, the treatment recommendation plan, and the medical insurance recommendation plan to obtain auxiliary medical information, and send the auxiliary medical information to the doctor corresponding to the target patient.

[0145] The reward and punishment adjustment module 15 is used to obtain the diagnosis and treatment results of the target patient, and input the diagnosis and treatment results into the feedback evolutionary agent for reward and punishment analysis to obtain the agent's reward and punishment information;

[0146] Based on the agent reward and punishment information, the action reward and punishment are adjusted for the pre-diagnosis agent, the diagnosis reasoning agent, the resource scheduling agent, or the medical insurance strategy agent.

[0147] The collaborative control module 16 is used to input the collaborative concurrent task sent by any intelligent agent into the collaborative control intelligent agent for collaborative analysis to obtain the collaborative object when it receives the collaborative concurrent task sent by any intelligent agent.

[0148] Send a synchronization command to the collaborative object to control the collaborative object to execute the collaborative concurrent task.

[0149] The autonomous evolution module 17 is used to identify cases of assisted medical decision failure in historical assisted medical decision data based on the diagnosis and treatment results of the target patient and the marking information of the treating physician, and to label the root causes of misdiagnosis in the cases of assisted medical decision failure.

[0150] A failure assistance library is constructed based on the auxiliary failure cases labeled with the root causes of misdiagnosis.

[0151] When new patient structured data is collected, the new patient structured data is matched with the failure assistance library by label;

[0152] If the new patient structured data successfully matches the tags in the failure assistance library, the transmission of the synchronization command is stopped, and an assistance error message is sent.

[0153] The medical record intelligent parsing module 18 is used to acquire the medical record image sent by the target patient, and to perform image recognition on the medical record image according to the target detection model to obtain the text area;

[0154] The text region is converted into text to obtain medical record text, and the medical record text is input into a pre-trained named entity recognition model for entity recognition to obtain a target entity set.

[0155] The target entity set is matched with a medical knowledge graph, and a medication recommendation report is generated based on the entity matching results.

[0156] The medication recommendation report will be sent to the target patient.

[0157] Optionally, the medical record intelligent parsing module 18 is further configured to: acquire training samples, determine annotation information based on different types of sample entities in the training samples, and label the training samples according to different annotation information to obtain model samples;

[0158] Random prompt templates are randomly selected from the prompt template library to obtain random prompt templates. The named entity recognition large model is then trained based on the random prompt templates and the model samples until the named entity recognition large model converges, thus obtaining the pre-trained named entity recognition large model.

[0159] Furthermore, the medical record intelligent parsing module 18 is also used to: segment the model samples according to the named entity recognition big model to obtain sample segments, and perform word embedding processing on the sample segments to obtain sample vectors;

[0160] The target entity type is determined based on the random prompt template, and the target vector is determined based on the target entity type;

[0161] The output vector in the sample vector is determined based on the target vector, and the model loss is determined based on the output vector;

[0162] The parameters of the named entity recognition model are updated based on the model loss.

[0163] In this embodiment, the pre-diagnosis intelligent agent can effectively collect structured information from the target patient, and the diagnostic reasoning intelligent agent can effectively perform diagnostic reasoning on the patient's structured data to obtain a diagnosis and disease prediction. By inputting the diagnosis and disease prediction and the current medical resource load information into the resource scheduling intelligent agent for diagnosis and treatment analysis, the allocation of diagnosis and treatment resources can be effectively carried out to obtain a recommended diagnosis and treatment plan. By inputting the recommended diagnosis and treatment plan into the medical insurance strategy intelligent agent for medical insurance analysis, the optimal path medical insurance recommendation plan can be effectively output. Based on the auxiliary medical treatment information, the auxiliary medical treatment decision corresponding to the disease diagnosis, diagnosis and treatment plan and medical insurance plan can be effectively output, ensuring the diversity of the auxiliary medical treatment decision function.

[0164] Example 3

[0165] Figure 3This is a structural block diagram of a terminal device 2 provided in the third embodiment of this application. For example... Figure 3 As shown, the terminal device 2 in this embodiment includes a processor 20, a memory 21, and a computer program 22 stored in the memory 21 and executable on the processor 20, such as a program for a multi-agent-based assisted medical decision-making method. When the processor 20 executes the computer program 22, it implements the steps in the various embodiments of the multi-agent-based assisted medical decision-making methods described above.

[0166] For example, the computer program 22 may be divided into one or more modules, which are stored in the memory 21 and executed by the processor 20 to complete this application. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program 22 in the terminal device 2. The terminal device may include, but is not limited to, the processor 20 and the memory 21.

[0167] The processor 20 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0168] The memory 21 can be an internal storage unit of the terminal device 2, such as a hard drive or memory of the terminal device 2. The memory 21 can also be an external storage device of the terminal device 2, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the terminal device 2. Furthermore, the memory 21 can include both internal and external storage units of the terminal device 2. The memory 21 is used to store the computer program and other programs and data required by the terminal device. The memory 21 can also be used to temporarily store data that has been output or will be output.

[0169] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0170] If an integrated module is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. This computer-readable storage medium can be non-volatile or volatile. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable storage medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the contents of a computer-readable storage medium may be appropriately added to or subtracted from the contents as required by the legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, a computer-readable storage medium may not include electrical carrier signals and telecommunication signals.

[0171] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. A multi-agent-based assisted medical decision-making method, characterized in that, The method includes: The pre-diagnosis intelligent agent collects structured information from target patients to obtain structured patient data; The patient's structured data is input into the diagnostic reasoning agent to perform diagnostic reasoning and obtain a diagnosis and predicted disease. The system obtains the hospitals visited by the target patient, and inputs the diagnosed and predicted disease and the current medical resource load information of the hospitals into the resource scheduling agent for diagnosis and treatment analysis to obtain a recommended treatment plan. The treatment recommendation plan is input into the medical insurance strategy intelligent agent for medical insurance analysis to obtain the medical insurance recommendation plan. The patient's structured data, the diagnosed and predicted diseases, the recommended treatment plan, and the medical insurance recommendation plan are combined to obtain auxiliary medical information, which is then sent to the attending physician corresponding to the target patient.

2. The multi-agent-based assisted medical decision-making method as described in claim 1, characterized in that, After sending the auxiliary medical information to the attending physician corresponding to the target patient, the process also includes: The diagnosis and treatment results of the target patient are obtained, and the diagnosis and treatment results are input into the feedback evolutionary agent for reward and punishment analysis to obtain the agent's reward and punishment information; Based on the agent reward and punishment information, the action reward and punishment are adjusted for the pre-diagnosis agent, the diagnosis reasoning agent, the resource scheduling agent, or the medical insurance strategy agent.

3. The multi-agent-based assisted medical decision-making method as described in claim 1, characterized in that, After the pre-diagnosis intelligent agent collects structured information from the target patient and obtains the patient's structured data, it also includes: When a collaborative concurrent task is received from any intelligent agent, the collaborative concurrent task is input into the collaborative control intelligent agent for collaborative analysis to obtain the collaborative object; Send a synchronization command to the collaborative object to control the collaborative object to execute the collaborative concurrent task.

4. The multi-agent-based assisted medical decision-making method as described in claim 3, characterized in that, After sending the auxiliary medical information to the attending physician corresponding to the target patient, the process also includes: Based on the diagnosis and treatment results of the target patient and the marking information of the treating physician, identify the cases of assisted medical decision-making failure in the historical assisted medical decision-making data, and label the root causes of misdiagnosis in the cases of assisted medical decision-making failure; A failure assistance library is constructed based on the auxiliary failure cases labeled with the root causes of misdiagnosis. When new patient structured data is collected, the new patient structured data is matched with the failure assistance library by label; If the new patient structured data successfully matches the tags in the failure assistance library, the transmission of the synchronization command is stopped, and an assistance error message is sent.

5. The multi-agent-based assisted medical decision-making method as described in claim 1, characterized in that, The method further includes: The medical record image sent by the target patient is obtained, and the image is recognized according to the target detection model to obtain the text region; The text region is converted into text to obtain medical record text, and the medical record text is input into a pre-trained named entity recognition model for entity recognition to obtain a target entity set. The target entity set is matched with a medical knowledge graph, and a medication recommendation report is generated based on the entity matching results. The medication recommendation report will be sent to the target patient.

6. The multi-agent-based assisted medical decision-making method as described in claim 5, characterized in that, Before inputting the medical record text into the pre-trained named entity recognition model for entity recognition, the following steps are also included: Obtain training samples, determine annotation information based on different types of sample entities in the training samples, and label the training samples according to the different annotation information to obtain model samples; Random prompt templates are randomly selected from the prompt template library to obtain random prompt templates. The named entity recognition large model is then trained based on the random prompt templates and the model samples until the named entity recognition large model converges, thus obtaining the pre-trained named entity recognition large model.

7. The multi-agent-based assisted medical decision-making method as described in claim 6, characterized in that, Training a large named entity recognition model based on the random prompt template and the model samples includes: The named entity recognition model is used to segment the model samples to obtain sample words, and the sample words are then subjected to word embedding processing to obtain sample vectors. The target entity type is determined based on the random prompt template, and the target vector is determined based on the target entity type; The output vector in the sample vector is determined based on the target vector, and the model loss is determined based on the output vector; The parameters of the named entity recognition model are updated based on the model loss.

8. A multi-agent-based assisted medical decision-making system, characterized in that, The system includes: The information collection module is used to collect structured information about the target patient based on the pre-diagnosis intelligent agent, and obtain structured patient data. The diagnostic reasoning module is used to input the patient's structured data into the diagnostic reasoning agent to perform diagnostic reasoning and obtain a diagnosis and prediction of the disease. The diagnosis and treatment analysis module is used to obtain the hospital visited by the target patient, and input the diagnosed and predicted disease and the current medical resource load information of the hospital into the resource scheduling agent for diagnosis and treatment analysis to obtain a recommended diagnosis and treatment plan; The medical insurance analysis module is used to input the treatment recommendation plan into the medical insurance strategy intelligent agent for medical insurance analysis to obtain the medical insurance recommendation plan; The decision support module is used to combine the patient's structured data, the diagnosed and predicted diseases, the recommended treatment plan, and the medical insurance recommendation plan to obtain auxiliary medical information, and send the auxiliary medical information to the attending physician corresponding to the target patient.

9. The multi-agent-based assisted medical decision-making system as described in claim 8, characterized in that, The multi-agent-based assisted medical decision-making system also includes: The reward and punishment adjustment module is used to obtain the diagnosis and treatment results of the target patient and input the diagnosis and treatment results into the feedback evolutionary agent for reward and punishment analysis to obtain the agent's reward and punishment information. Based on the agent reward and punishment information, the action reward and punishment are adjusted for the pre-diagnosis agent, the diagnosis reasoning agent, the resource scheduling agent, or the medical insurance strategy agent.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images