Automatic diagnosis method for interpretability-oriented adaptive multi-agent cooperation

Abstract

The invention discloses an automatic diagnosis algorithm for interpretability-oriented adaptive multi-agent cooperation. The method comprises the following steps: firstly, constructing a bidirectional probability network of symptoms and diseases based on a data set, calculating a disease weight through a pre-diagnosis module, and deciding to enter a Doctor Agent module or a Med-Team Agent module; the Doctor Agent simulates single expert diagnosis and judges whether a patient can be diagnosed or not or continues to inquire related symptoms to assist diagnosis in combination with a medical knowledge base and a bidirectional probability network. The Med-Team Agent is used for simulating consultation of multiple experts, and comprises a Manager Agent and a plurality of Disease Agents which respectively represent an attending physician and a plurality of disease experts. And the Manager Agent dynamically manages a plurality of Disease Agents, selects an appropriate expert to generate related symptoms according to the symptom information of the patient, and finally completes diagnosis. The method is suitable for medical automatic diagnosis tasks, through combination of multi-agent cooperation and bidirectional probability network prompting, the interpretability of diagnosis is improved, and efficient inquiry and accurate diagnosis are achieved.

Description

Technical Field

[0001] The present invention belongs to the intersection of artificial intelligence science and medicine, and specifically relates to the field of automatic diagnosis. More specifically, the present invention provides an automatic diagnosis method for adaptive multi-agent cooperation oriented to explainability. Background Art

[0002] Automatic disease diagnosis is both critical and complex in the medical field, and is an important bridge between medical expertise and clinical decision-making. Diagnostic tasks usually rely on the patient's main complaint and the doctor's subsequent supplementary symptoms to predict possible diseases. Doctors should make the most accurate diagnosis within an appropriate number of inquiries to balance efficiency and accuracy.

[0003] At present, automatic disease diagnosis mainly relies on reinforcement learning (RL) technology. Early pioneering research introduced symptom checking methods based on reinforcement learning. With the advancement of technology, some studies have further proposed hierarchical reinforcement learning for diagnosis and situational decision-making, using deep Q networks to extract symptom information from patient interactions, and exploring the framework of integrating medical knowledge into RL policy learning. The development of these technologies, such as two-level hierarchical reinforcement learning, policy gradient methods combined with generative adversarial networks, and RL models optimized by multi-level reward mechanisms and dialogue data, has greatly improved the accuracy of diagnosis. However, most of these models are "black box" and lack interpretability. In the medical field, the transparency of the diagnostic process is crucial to gaining the trust of patients and doctors. Therefore, the interpretability of such technologies has become an urgent problem to be solved.

[0004] Large language models (LLMs) have shown great potential in the field of automated diagnosis with their powerful reasoning and conversation capabilities combined with extensive knowledge reserves. These models are able to handle a variety of diseases and interact efficiently with patients. However, existing studies have shown that although LLMs have excellent language understanding and reasoning capabilities, they are insufficient when asking about key symptoms related to the disease. How to design a reasonable inquiry strategy to improve the accuracy of diagnosis remains a challenging problem.

[0005] In summary, there is an urgent need to design an efficient and interpretable medical automatic diagnosis method to solve the above problems. Summary of the invention

[0006] In view of this, the present invention proposes an automatic diagnosis method for adaptive multi-agent cooperation oriented to interpretability. First, in order to solve the problem of lack of interpretability of traditional "black box" models, the present invention starts from the real diagnosis scenario, models the actual single-department expert diagnosis and multi-department expert consultation, and proposes a multi-agent cooperative diagnosis framework. Each step of the system has a clear function, and each component can be explained separately. The entire system simulates the real diagnosis and can transform the black box decision of the diagnosis into an explainable chain. Second, in order to solve the problem of insufficient efficiency in asking symptoms in large models at this stage, the present invention proposes a consultation strategy based on symptom-disease bidirectional probabilistic network prompts, and is routed by a discriminant large model to distinguish between single-expert diagnosis and multi-expert consultation, so as to achieve high efficiency of "asking" and high accuracy of "diagnosis".

[0007] The present invention consists of a pre-diagnosis module, a Doctor Agent module and a Med-Team Agents module. The pre-diagnosis module constructs a unidirectional probability network from symptoms to diseases, calculates the disease weight by comparing the patient's dominant symptoms with the known condition, and a large discriminant model judges whether to enter the Doctor Agent module or the Med-Team Agents module according to the result. The Doctor Agent module simulates a single expert diagnosis scenario, can combine the medical knowledge base and the bidirectional probability network, and call the external knowledge base and predefined tools according to the strategy, based on the patient's symptoms, determine whether to confirm the diagnosis, or further inquire about related symptoms to assist in diagnosis. The Med-Team Agents module simulates the consultation scenario of multiple disease experts, constructs a Manager Agent and multiple Disease Agents, which simulate the attending physician and multiple disease experts respectively. The ManagerAgent dynamically manages multiple Disease Agents, decides which Disease Agents to use to generate related symptoms based on the patient's symptom information, and finally makes a diagnosis.

[0008] Specifically, the present invention constructs a disease-to-symptom directed probability matrix M D-S and the symptom-to-disease matrix M S-D . Use D = [d1, d2, ..., d m ] represents the disease set, m is the number of diseases in the disease set; S = [s1, s2, ..., s n ] represents the symptom set, and n represents the number of symptoms in the symptom set. In the automatic diagnosis scenario, each patient has only one disease, so the disease state can only be present or absent. According to the distribution of the data set, the probability of each disease is calculated. Assuming that the patient has a specific disease i In the case of j The probability of symptoms P(sj |d i ), construct the disease-to-symptom matrix M D-S :

[0009]

[0010] Among them, M S-D ∈R m×n In addition, given the symptom status information, the conditional probability of each disease can be obtained as follows:

[0011]

[0012] in

[0013] P(s j )=∑ d∈D P(s j |d i )*P(d i ),

[0014] P(d i |s j ) indicates that the patient has s j Symptoms, disease i The probability of constructing the matrix M from symptoms to diseases S-D :

[0015]

[0016] Among them, M S-D ∈R n×m .

[0017] In the pre-diagnosis module, a symptom extraction prompt is designed to guide the LLM to extract the patient's symptom information from the natural sentence of the patient's complaint and represent it as an n-dimensional vector S o Among them, if S o The element value in is 1, which means the symptom exists; 0 means the symptom is unknown; -1 means the symptom is denied. S-D , and use the product matrix method to obtain the disease weight list D p :

[0018] D p =S o ×M S-D ,

[0019] Where D pis an m-dimensional vector, in which the element value represents the weight corresponding to each disease. The higher the weight, the more similar the patient's symptoms are to the disease. In addition, to achieve an adaptive medical decision-making framework, a discriminative LLM is designed as a generalist practitioner to determine the complexity of medical diagnosis. The discriminative LLM acts as a classifier, obtains information about how to define the complexity of the patient's symptoms, and classifies the patient's subsequent diagnosis into a single expert diagnosis or a multi-expert consultation according to the complexity.

[0020] The Doctor Agent module simulates a single attending specialist consultation scenario and has a simple clinical decision path. For a given patient's inquiry record, the Doctor Agent observes the patient's symptoms and makes the following decisions based on the tool call results: continue to ask the patient about their symptoms and repeat the process, or directly diagnose a disease:

[0021] A next =Doctor_Agent(prompt,historytool),

[0022] Among them A next Indicates a symptom or a disease, and Doctor Agent makes decisions based on prompts and history. Specifically, in the present invention, the LLM is assigned a specialist role, and its medical professional capabilities are enhanced by combining with an external medical knowledge base, thereby transforming it into a Doctor Agent. In addition, a M-based S-D Get the disease weight ranking tool1 and a M-based D-S Obtain tool2 of the relevant symptom list to assist the model in making more accurate symptom selection. In a patient consultation, Doctor Agent first calls tool1 from the tool pool to obtain the disease weight ranking ranked_diseases according to the patient's manifested symptoms, and decides whether a diagnosis can be made based on the disease weight result. If a diagnosis can be made, the consultation ends; otherwise, call tool2 to obtain the symptom ranking corresponding to the disease with the highest weight, and return the top 20% of the symptom list symptom_list:

[0023] symptom_list=Top 20％ (tool2(Max(ranked_diseases))),

[0024] Doctor Agent can combine its own knowledge and external medical knowledge base to make the best symptom selection, rewrite the selected symptoms into user-friendly natural language for inquiry, send it to the patient, and wait for a reply. Doctor Agent simulates a specialist consultation and has the following three capabilities. First, it has the ability to ask questions and answer questions, and can ask patients about the retrieved professional symptoms in popular natural language; second, it has the ability to use tools, and can use symptom-disease bidirectional probability networks and external medical knowledge bases. This ability helps reduce model hallucinations and provides a certain degree of interpretability; third, it has the ability to make autonomous decisions. Doctor Agent constantly thinks about whether to diagnose the patient or further ask about new symptoms to increase diagnostic confidence throughout the consultation process.

[0025] The Med-Team Agents module can generate comprehensive and detailed solutions when dealing with more complex consultation scenarios, using the collective wisdom of professionals from different disciplines, integrating insights from different professions, overcoming the limitations of consultations in a single field, and achieving accurate diagnosis. Specifically, in Med-Team Agents, a Manager Agent is designed to give it the role of an attending physician, and multiple Disease Agents are designed as disease experts. These Agents share a message queue to store the patient's consultation records:

[0026] A next =Med-Team Agents (prompt, memory, Manager Agent, Disease Agents), where Disease Agents can call external knowledge bases to enhance their professional capabilities in this field, and can respond to calls from Manager Agents and return suspicious symptoms based on historical diagnostic information. Manager Agents have some functions similar to Doctor Agents. S-D Obtain tool1 for ranking disease weights. The Manager Agent uses tool1 to obtain ranked_diseases according to the patient's symptoms, and decides whether a diagnosis can be made based on the disease weights. If a diagnosis can be made, the consultation ends; otherwise, the Manager Agent should communicate with other DiseaseAgents to obtain symptom inquiry suggestions. Specifically, the Manager Agent can compare the ranked_diseases of the current round. k and ranked_diseases of historical rounds k-1, obtain the list of diseases whose weights are increased_diseases, obtain the relevant symptoms from the corresponding Disease Agents, and ask the intersection symptoms with the highest probability to obtain the patient's feedback.

[0027] The present invention is applicable to various automatic diagnosis tasks, and the positive effects of adopting the present invention are:

[0028] This paper proposes a novel multi-agent cooperative diagnostic method for the field of medical automatic diagnosis, which solves the problem of lack of interpretability of traditional "black box" models. In addition, this method effectively simulates the real scene of single-department expert diagnosis and multi-department expert consultation, and based on the strategy of symptom-disease bidirectional probability network prompts, it can significantly improve the efficiency and accuracy of medical automatic diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a schematic diagram of the principle of an automatic diagnosis method of adaptive multi-agent cooperation oriented to explainability of the present invention;

[0030] Figure 2 This is a flow chart of an automatic diagnosis method for adaptive multi-agent cooperation oriented to explainability of the present invention. DETAILED DESCRIPTION

[0031] In order to enable those skilled in the art to better understand the solutions of the present invention and to make the above-mentioned objectives, technical solutions and advantages of the present invention more obvious and understandable, the present invention is further described in detail below in conjunction with embodiments and accompanying drawings of the embodiments.

[0032] See also Figure 1 , Figure 1The schematic diagram of the principle of an automatic diagnosis method for adaptive multi-agent cooperation for interpretability of the present invention is shown. The present invention consists of a pre-diagnosis module, a Doctor Agent module and a Med-Team Agents module. The pre-diagnosis module constructs a directed probability network from symptoms to diseases, calculates the disease weight by comparing the patient's explicit symptoms with the known condition, and a large discriminant model determines whether to enter the Doctor Agent module or the Med-TeamAgents module based on the result. The Doctor Agent module simulates a single expert diagnosis scenario, can combine the medical knowledge base and the two-way probability network, and call the external knowledge base and predefined tools according to the strategy, based on the patient's symptoms, to determine whether the diagnosis is confirmed, or further inquire about related symptoms to assist in diagnosis. The Med-Team Agents module simulates the consultation scenario of multiple disease experts, and constructs a Manager Agent and multiple Disease Agents, which simulate the attending physician and multiple disease experts respectively. The Manager Agent dynamically manages multiple Disease Agents, determines which DiseaseAgents to use to generate related symptoms based on the patient's symptom information, and finally makes a diagnosis.

[0033] See also Figure 2 , Figure 2 This is a flowchart of an explainable adaptive multi-agent cooperative automatic diagnosis method for a light consultation task in the present invention, and the specific steps include:

[0034] S1. Construct a two-way probability matrix of symptoms and diseases based on the light medical consultation dataset; use D = [d1, d2, ..., d m ] to represent the disease set, where m is the number of diseases in the disease set; S = [s1, s2, ..., s n ] to represent the symptom set, where n represents the number of symptoms in the symptom set; according to the distribution of the data set, the probability of suffering from each disease can be obtained Assume that the patient has d i In the case of disease, j The probability of symptoms P(s j |d i ); From this, a disease-to-symptom matrix can be constructed:

[0035]

[0036] Among them, M S-D ∈R m×n ; In addition, given the symptom status information, the conditional probability of each disease can be obtained as follows:

[0037]

[0038] in

[0039] P(s j )=∑ d∈D P(s j |d i )*P(d i ),

[0040] P(d i |s j ) indicates that the patient has s j Symptoms, disease i The probability of constructing the matrix M from symptoms to diseases S-D :

[0041]

[0042] Among them, M S-D ∈R n×m ;

[0043] S2. In the pre-diagnosis stage, the patient's symptoms are obtained and the symptom pool is updated. A disease weight list is generated based on the known symptoms, and a large discriminant model is used to route according to the disease weight list to distinguish between single-expert diagnosis and multi-expert consultation.

[0044] S21. Use the pre-designed prompt template to guide LLM to accurately extract symptom entities from the patient's complaint, and then represent it as an n-dimensional vector S o , where if S o The element value in is 1, which means the symptom exists, 0 means the symptom is unknown, and -1 means the symptom is denied;

[0045] The specific contents of prompt are as follows:

[0046] “You are currently an outstanding AI medical expert. Your task is to extract key symptom information based on the patient’s description, and follow the following rules when extracting:

[0047] 1) Please extract key symptom information that is useful for diagnosis and treatment;

[0048] 2) Please fill in the "{}" part according to the following format

[0049] {'Symptom 1': 'Yes', 'Symptom 2': 'No'}

[0050] 3) The output can be loaded via Python's json.loads() function

[0051] ## Output:

[0052] {}";

[0053] S22. Using the constructed symptom-disease probability network M S-D Calculate the disease weight list corresponding to the patient's existing symptoms; then use the product matrix method based on MS-D to obtain the weight of each disease:

[0054] D p =S o ×M S-D ,

[0055] Where D p is an m-dimensional vector, in which the element value represents the weight corresponding to each disease. The higher the weight, the more similar the patient's symptoms are to the disease.

[0056] S23. Design prompts to guide LLM to analyze the complexity of the patient's illness and classify the patient's subsequent diagnosis into single-expert diagnosis or multi-expert consultation.

[0057] The specific contents of prompt are as follows:

[0058] "You are currently an outstanding AI medical expert. Your task is to analyze the complexity of the patient's illness from the disease weight list and decide whether to use a single-specialty expert diagnosis or a multi-specialty expert diagnosis plan. When making the judgment, follow the following rules:

[0059] 1) If the weight value of a certain disease is prominent, the complexity is low. If there is more than one disease with a high weight value, the patient's condition is more complicated;

[0060] 2) For patients with complex conditions, multiple specialists should be consulted, while for patients with simpler conditions, they should be referred to a single specialist to improve efficiency;

[0061] 3) Please choose one from [Single subject expert, Multi-subject expert] as your output.

[0062] ##The following are the disease weights that need to be judged:

[0063] Disease weight:{disease_list}

[0064] Output: ";

[0065] S3. Use a single agent (Doctor Agent) to further inquire and confirm symptoms. Use prompts to guide the agent to observe the patient's known symptoms, and then use tools to assist in making decisions: continue to ask the patient about related symptoms and repeat this process; or diagnose a disease when conditions are met and terminate the cycle;

[0066] The specific content of DoctorAgent's prompt is as follows:

[0067] “You are a medical expert, your name is Doctor. Your task is to use the tool to continuously retrieve and ask the patient if they have other symptoms based on the known symptoms provided by the patient. Only one symptom is asked in each round.

[0068] ##The specific process includes:

[0069] Step 1: Accept the symptoms given by the patient and add them to the symptom dictionary. Symptoms are key-value pairs of two strings, such as 'egg drop stool':'unknown'.

[0070] Step 2: Use tool1 to retrieve possible diseases and their weights based on the patient's symptom dictionary.

[0071] Step 3: Use tool tool2 to retrieve several symptoms and select one from them.

[0072] Step 4: Ask the patient whether he has the symptom, refer to the inquiry format: Do you have XX symptoms?

[0073] ##Notice:

[0074] 1. You need to strictly follow the above process until you think you can diagnose a disease for the patient. Refer to the diagnosis format: You may have XX disease.

[0075] 2. Ask the patient only about one symptom at a time, and do not reveal the disease when asking about the symptoms.

[0076] 3. Symptom type has three character strings, "unknown", "true", and "false". It is prohibited to interpret the patient's reply "unknown" as False or "false".

[0077] 4. The patient's answer has three options: true, false, and unknown. If the patient answers "true", it means that the patient has the symptom; if the answer is "false", it means that the patient does not have the symptom; if the answer is "unknown", it means that the patient does not know whether the symptom exists.

[0078] S31. Assign the role of specialist to LLM and enhance the medical professional ability of LLM by combining with external medical knowledge base, making it a Doctor Agent. Bind a M-based S-D Get the disease weight ranking tool1 and a M-based D-S Tool2 to get the list of related symptoms;

[0079] The specific contents of Agent Tool are as follows:

[0080] 1.Agent tool tool1 definition content

[0081] "Based on known symptoms, retrieve several diseases and probabilities"

[0082] 2.Agent tool tool2 definition content

[0083] “By specifying a disease and known symptoms, several other unrepresented symptoms are retrieved for the model to choose from.”

[0084] S32, Doctor Agent obtains the patient's symptoms and is prompted to call tool1 from the tool pool to obtain the disease weight ranking ranked_diseases according to the symptoms that have been revealed;

[0085] S33, Doctor Agent makes a decision based on the disease weight result, whether a final diagnosis can be made. If a diagnosis can be made, the doctor agent exits the consultation; otherwise, tool2 is called to obtain the symptom ranking corresponding to the disease with the highest weight and the top 20% symptom list symptom_list is returned as the tool call result;

[0086] symptom_list=Top 20％ (tool2(Max(ranked_diseases)))

[0087] S34. Doctor Agent makes the best symptom selection based on its own knowledge and the supplement of external medical knowledge base, rewrites the selected symptoms into user-friendly natural sentence inquiries, sends them to the patient, and waits for the patient's response.

[0088] S4. Use Med-Team Agents to handle complex medical consultation scenarios. In this mode, ManagerAgent is responsible for managing the calls of multiple Disease Agents and performing symptom inquiry or diagnosis based on the returned results and the patient's known symptom information.

[0089] The specific contents of prompt are as follows:

[0090] 1.Manager Agent prompt:

[0091] "You are an attending physician. Your task is to use the tool to obtain a list of disease weights based on the known symptoms provided by the patient, and to determine whether the patient can be diagnosed. During this process, you can work with other disease specialists to obtain accurate symptom information and ask the patient questions.

[0092] ##The specific process includes:

[0093] Step 1: Accept the symptoms given by the patient and add them to the symptom dictionary. Symptoms are key-value pairs of two strings, such as 'egg drop stool':'unknown'.

[0094] Step 2: Use tool1 to retrieve possible diseases and their weights based on the patient's symptom dictionary.

[0095] Step 3: Extract the list of diseases with increased weights and communicate with relevant disease experts to obtain the corresponding symptom list.

[0096] Step 4: Select a symptom from the intersection of the symptom lists corresponding to each disease and ask the patient. Refer to the inquiry format: Do you have XX symptom?

[0097] ##Notice:

[0098] The disease experts you can communicate with include [pneumonia experts, pediatric hand, foot and mouth disease experts, allergic rhinitis experts, pediatric diarrhea experts, upper respiratory tract infection experts]

[0099] 2. You need to strictly follow the above process until you think you can diagnose a disease for the patient. Refer to the diagnosis format: You may have XX disease.

[0100] 3. Ask the patient only about one symptom at a time, and do not reveal the disease when asking about the symptoms.

[0101] 4. Symptom type has three character strings, "unknown", "true", and "false". It is prohibited to interpret the patient's reply "unknown" as False or "false".

[0102] 5. The patient's answer can be in three situations: true, false, and unknown. If the patient answers "true", it means that the patient has the symptom; if the answer is "false", it means that the patient does not have the symptom; if the answer is "unknown", it means that the patient does not know whether the symptom exists.

[0103] 2.Disease Agent prompt:

[0104] "You are a professional pneumonia expert. Your task is to respond to the call from the Manager Agent and call the tool to retrieve other pneumonia symptoms based on the known conversation records."

[0105] S41. Assign the attending physician role to the LLM and bind an M-based S-DObtain the disease weight ranking tool1 to make it a Manager Agent. Initialize the Disease Agents corresponding to all diseases, combine the external medical knowledge base to increase the medical professional capabilities of the Agent, and bind the M-based D-S Tool2 to get the list of related symptoms;

[0106] The specific contents of Agent Tool are as follows:

[0107] 1.Agent tool tool1 definition content

[0108] "Based on known symptoms, retrieve several diseases and probabilities"

[0109] 2.Agent tool tool2 definition content

[0110] "By specifying a disease and known symptoms, several other unrepresented symptoms are retrieved for the model to select.";

[0111] S42, Manager Agent obtains the patient's symptoms and is prompted to call tool1 from the tool pool to obtain the disease weight ranking ranked_diseases according to the manifested symptoms;

[0112] S43, Manager Agent compares the ranked_diseases of the current round k and ranked_diseases in previous rounds k-1 , get the list of diseases whose weights are increased_diseases;

[0113] S44, Manager Agent asks the Disease Agent corresponding to the disease in increased_diseases, and Disease Agent calls tool2 to obtain the symptom list symptom_list sorted by probability, and selects from it based on its own knowledge and external medical knowledge base. symptoms as a response;

[0114] S45. After obtaining relevant symptoms from corresponding Disease Agents, the Manager Agent selects the symptom with the highest probability in the intersection and asks the patient.

[0115] In the above, the specific embodiments of the present invention are described with reference to the accompanying drawings. However, those skilled in the art will appreciate that various changes and substitutions may be made to the specific embodiments of the present invention without departing from the spirit and scope of the present invention. These changes and substitutions are all within the scope defined by the claims of the present invention.

Claims

1. An adaptive multi-agent collaborative automatic diagnosis method for interpretability, characterized by The following steps are involved: S1. Construct a two-way probability matrix of symptoms and diseases based on the data set, including the following two parts: Probability matrix M for locating diseases from symptoms S-D , each element of which represents the patient's performance j Symptoms of disease i The probability of i |s j ) represents; the probability matrix M of locating symptoms from the disease D-S , each element of which represents a patient suffering from a specific disease d i Symptoms j The probability of j |d i )express; S2. In the pre-diagnosis stage, the probability matrix MS based on symptoms to diseases -D Extract symptom information from the patient's complaint, calculate the disease weight list, and pass it to the discriminant model for routing based on the results, distinguishing between single-expert diagnosis and multi-expert consultation; S3, using the intelligent agent model Doctor Agent to simulate the scenario of single expert diagnosis; The Doctor Agent intelligent agent model combines the medical knowledge base and the bidirectional probability matrix, and calls the external knowledge base and predefined tools according to the strategy. It can determine whether the diagnosis is confirmed based on the symptoms that the patient has shown, or further inquire about related symptoms to assist in diagnosis; S4. Use the multi-intelligent agent model Med-Team Agents to simulate the scenario of multi-expert consultation; Med-Team Agents builds a Manager Agent and multiple Disease Agents to simulate the attending physician and multiple disease experts respectively; DiseaseAgents have professional domain knowledge, can respond to Manager Agent's inquiries, and return suspicious symptoms based on historical conversation information; Manager Agent calls external knowledge bases and predefined tools according to strategies, determines whether the patient is diagnosed based on the symptoms he has already manifested, or asks Disease Agents to obtain symptom prompts and make symptom choices to assist in diagnosis.

2. The automatic diagnosis method for adaptive multi-agent cooperation for interpretability according to claim 1, characterized in that: The step S2 specifically includes: S21, using prompt templates to guide the large language model to extract symptom entities from patient complaints and convert them into vector representations; S22, based on M S-D The weight of each disease is obtained by using the product matrix method; S23. Use prompt templates to guide the large language model to analyze the complexity of the patient's disease and classify the patient's subsequent diagnosis into single-expert diagnosis or multi-expert consultation situations.

3. The method of automatic diagnosis based on adaptive multi-agent cooperation for interpretability according to claim 1, characterized in that: The step S3 specifically includes: S31, initialize Doctor Agent, set the specialist role for it and use the external medical knowledge base to enhance its professional capabilities; bind Doctor Agent based on M S-D Obtain the Agent tool tool1 for disease weight ranking. Tool1 can transform known symptoms into n-dimensional state vectors and compare them with M S-D Calculate the disease weight list by multiplying the product; bind the Doctor Agent based on M D-S The Agent tool tool2 that obtains the list of related symptoms can retrieve M D-S And combined with the knowledge base, obtain a probability-ranked list of symptoms related to the specified disease; S32, using the prompt template to guide the Doctor Agent to obtain the patient's symptoms, and calling tool1 from the tool pool to calculate the disease weight ranking according to the symptoms that have been revealed; S33, using the prompt template to guide the Doctor Agent to make a decision based on the disease weight result, whether a final diagnosis can be made. If a diagnosis can be made, the consultation will be exited; otherwise, tool2 will be called to obtain the symptom ranking corresponding to the disease with the highest weight and the top 20% of the symptoms will be returned as the tool call result; S34. Use the prompt template to guide the Doctor Agent to make the best symptom selection based on its own knowledge combined with the external medical knowledge base, and rewrite the selected symptoms into user-friendly natural sentence inquiries, wait for the patient's response, and then return to execute S2.

4. The method of automatic diagnosis based on adaptive multi-agent cooperation for interpretability according to claim 1, characterized in that: The step S4 specifically includes: S41, initialize Med-Team Agents, set Manager Agent as the attending physician and multiple DiseaseAgents as disease experts; bind tool1 to Manager Agent, and bind tool2 to Disease Agents; S42, using the prompt template to guide the Manager Agent to obtain the patient's symptoms, and calling tool1 from the tool pool to calculate a disease weight ranking list based on the manifested symptoms, recorded as ranked_diseases; S43, Manager Agent compares the ranked_diseases of the current round k and ranked_diseases in previous rounds k-1 , get the list of diseases with increased weights; S44, the Manager Agent inquires the Disease Agent corresponding to the disease in the disease list, and the Disease Agent calls tool2 to obtain a list of suspected symptoms and selects 50% of the symptoms as responses based on its own knowledge and the external medical knowledge base; S45. After obtaining relevant symptoms from corresponding Disease Agents, the Manager Agent selects the symptom with the highest probability in the intersection and asks the patient.

Images