Multi-round health dialogue method and system based on dynamic user state and real-time knowledge constraint

Abstract

The invention discloses a multi-round health dialogue method and system based on a dynamic user state and real-time knowledge constraint, relates to an artificial intelligence technology, and provides a scheme for solving the problems of weak interpretability and the like in the prior art. Constructing a medical health vector knowledge base and a corresponding vector retriever; based on distributed storage and SSL / TLS encryption, a user health archive library is constructed by using a cloud database technology; constructing a historical health state summary agent, a user health state updating agent, a system reply type decision agent and a multi-round health dialogue reply agent by using the medical health big language model; and constructing a multi-agent collaborative framework to realize multiple rounds of health conversations. The method has the advantages that the multi-agent collaborative framework based on the dynamic user health state and the real-time knowledge constraint is constructed by utilizing the medical health big language model, the behavior mode of doctors during inquiry is simulated, the multi-round health dialogue ability of the model is further improved, the interpretability and inquiry controllability are enhanced, and medical health knowledge is complemented.

Description

Technical Field

[0001] The present invention relates to artificial intelligence technology, and in particular to a multi-round health dialogue method and system with dynamic user status and real-time knowledge constraints, which is applied to the field of medical interaction. Background Art

[0002] At present, medical resources are in short supply, and there is an urgent need for an artificial intelligence method or system that can play the role of "family doctor" to achieve various medical and health interaction tasks including but not limited to medical and health knowledge consultation, multi-round medical and health consultation, etc., to meet the diverse medical and health needs of the general public. In recent years, the rise of large language models has brought new hope to the field of artificial intelligence and medical interaction technology. At present, there are mainly the following research directions in the field of artificial intelligence and medical interaction technology: training-based medical and health large language models, and agent-based medical and health large language models.

[0003] Among them, the training-based medical and health big language model can be divided into the medical and health big language model based on continued pre-training, the medical and health big language model based on fine-tuning, the medical and health big language model based on reinforcement learning, and the medical and health big language model based on hybrid training. These training-based medical and health big language models are usually used to inject medical knowledge, or align their response style with the needs of the doctor-patient communication scenario. However, these methods are not only often criticized for not having interpretability and good controllability of consultation, but also limited by limited data. Therefore, although many researchers have paid attention to tasks such as medical and health knowledge consultation and multi-round medical and health consultation, there is still no good solution.

[0004] Agent-based medical and health big language models refer to the implementation of medical and health big language models as intelligent agents, and their flexible application in different medical and health tasks. This type of method generally does not require training, and the focus and implementation methods are also complex and diverse. Currently, more methods focus on multi-department joint diagnosis, which improves the interpretability and comprehensiveness of the diagnosis to a certain extent. However, this method sacrifices the model's response speed, and it is no longer possible to perform tasks other than diagnosis; a small number of methods pay attention to other medical and health interaction tasks, or doctor assistant tasks, etc.; only a few methods pay attention to medical and health interaction tasks such as consultation and knowledge consultation at the same time.

[0005] In summary, the current large medical and health language models have shortcomings such as insufficient medical and health knowledge, weak interpretability, and weak controllability of medical consultations, and there is no suitable solution yet. Summary of the invention

[0006] The purpose of the present invention is to provide a multi-round healthy dialogue method and system with dynamic user status and real-time knowledge constraints to solve the problems existing in the above-mentioned prior art.

[0007] The multi-round healthy dialogue method with dynamic user status and real-time knowledge constraints in the present invention comprises the following steps:

[0008] S1. Build a medical health vector knowledge base and a corresponding vector retrieval engine;

[0009] S2. Based on distributed storage and SSL / TLS encryption, use cloud database technology to build a user health archive;

[0010] S3. Using the medical and health language model, we build a historical health status summary agent, a user health status update agent, a system response type decision agent, and a multi-round health dialogue response agent;

[0011] S4. Build a multi-agent collaborative framework to realize multi-round healthy dialogue.

[0012] The multi-round healthy dialogue system with dynamic user status and real-time knowledge constraints in the present invention uses the method to complete the multi-round healthy dialogue task.

[0013] The multi-round health dialogue method and system with dynamic user status and real-time knowledge constraints described in the present invention have the advantage of constructing a multi-agent collaborative framework based on dynamic user health status and real-time knowledge constraints by utilizing the medical health big language model, simulating the behavior pattern of doctors during consultation, further improving the model's multi-round health dialogue capability, enhancing the explainability and controllability of consultation, and at the same time supplementing the lack of medical health knowledge. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flowchart of the multi-round health dialogue method described in the present invention.

[0015] Figure 2 It is a structural diagram of the historical health status summary intelligent agent described in the present invention.

[0016] Figure 3 It is a structural diagram of the user health status updating agent described in the present invention.

[0017] Figure 4 It is a structural diagram of the system response type decision-making agent described in the present invention.

[0018] Figure 5 It is a structural diagram of the multi-round healthy dialogue replying intelligent agent described in the present invention. DETAILED DESCRIPTION

[0019] like Figures 1 to 5 As shown, the multi-round healthy dialogue method with dynamic user status and real-time knowledge constraints in the present invention:

[0020] S1 builds a medical health vector knowledge base and a corresponding vector retriever.

[0021] It includes the following sub-steps:

[0022] S11. Collect several drug instruction documents and several disease introduction documents in the medical field, and divide all collected documents into several text blocks according to subtitles.

[0023] Each drug instruction document contains the drug's shape, pharmacology and toxicology, indications, usage and dosage, adverse reactions, contraindications, interactions, specifications, precautions, etc. Each disease introduction document contains the department to which it belongs, disease definition, cause, symptoms, diagnosis, drug treatment, daily care, etc.

[0024] S12. Using a dense vector search tool library and a text embedding model, all text blocks are constructed into a searchable medical health vector knowledge base. The dense vector search tool library is FAISS.

[0025] The text embedding model is trained by a neural network and has the feature of encoding a text block of arbitrary length into an embedding vector of fixed dimension.

[0026] S13. Construct a vector retriever using the text embedding model and the similarity retrieval function in the dense vector retrieval tool library.

[0027] The vector retriever has an input x query , an output y kn . Enter x query Is the query text input, which belongs to the text type. Output y kn It is the retrieved relevant knowledge and also belongs to the text type.

[0028] The input and output relationship of the vector retriever Retriever is:

[0029] y kn =Retriever(x query )=Cvt tl2t (SimSearch(Embedding(x query ),10));

[0030] Among them, Embedding(·) is the text embedding model, Retriever(·) is the vector retriever, and SimSearch(·) is the similarity retrieval function. The SimSearch function receives two inputs, the query vector and the top-k, and returns the top-k text blocks that are most similar to the query vector in the form of a list. Cvt tl2t (·) is a conversion function that converts a text list into text in a specified format.

[0031] S2. Use cloud database technology to build a user health record library, and ensure patient information security through distributed storage and SSL / TLS encryption technology.

[0032] In the user health archive library, each user's user health archive is stored separately as a JSON object type and indexed using a user's unique identification code. The user's unique identification code is a string of several digits, generated using the classic table ID self-increment method. When the user successfully registers, the unique identification code is automatically generated and distributed according to the internal parameters. Once generated, it is uniquely bound to the user and cannot be modified later.

[0033] The user health file contains two key-value pairs: user basic information and user's full historical health status data. The user basic information contains three key-value pairs: user's name, date of birth, and gender. The values ​​of the three key-value pairs of name, date of birth, and gender are filled in by the user when the user registers, and are all stored as string types. The user's full historical health status data contains several key-value pairs, each of which uses the date as the key and the user's health status on the corresponding date as the value. The user's health status on the corresponding date is composed of five key-value pairs: chief complaint, current medical history, past history, personal history, and family history.

[0034] S3. Using the medical and health big language model, we construct four agents, namely, the historical health status summary agent, the user health status update agent, the system response type decision agent, and the multi-round health dialogue response agent.

[0035] It includes the following sub-steps:

[0036] S31. Construct a historical health status summary agent.

[0037] The historical health status summary agent SumAgent(·) is as follows Figure 2 As shown, there is an input x sum , an output y sum Output y sum It is a summary of historical health status. It is a JSON object that contains five key-value pairs: chief complaint, current medical history, past medical history, personal history, and family history. sum It is the user's full history health status data, corresponding to the user's full history health status data in step S2.

[0038] The input-output relational expression of the historical health status summary agent SumAgent(·) is:

[0039] y sum =SumAgent(x sum )=Std hs(MedLLM([Sys sum ;Cvt sum2t (x sum );Usr sum ]));

[0040] Among them, MedLLM(·) is a medical and health language model, which has the characteristic of returning text-type output given a text-type input. Sys sum It is the system prompt word of the historical health status summary agent, which includes three parts: description of the historical health status summary task, explanation of the meaning of each part of the health status, and a small number of examples of the historical health status summary task. sum It is the user prompt word of the historical health status summary agent, specifically refers to the generation format requirements of the historical health status summary task. [Sys sum ;Cvt sum2t (x sum );Usr sum ] means concatenating the three parts of text. hs (·) is the health status normalization function, which normalizes the content generated by MedLLM and converts it into the same sum Same format. sum2t (·) is the sum A conversion function that converts a JSON object of the same structure into plain text.

[0041] S32. Build a user health status update agent.

[0042] The user health status update agent UpdAgent(·) is as follows Figure 3 As shown, there are three input and An output y upd Output y upd The latest user health status is a JSON object containing five key-value pairs: chief complaint, current medical history, past medical history, personal history, and family history. is the user's health status for the day, also a JSON object, and is consistent with the output y upd Have the same key-value pair structure. Input It is a conversation history list, which contains a series of JSON object type elements, each of which contains two key-value pairs: role and content. It is the current user input, which is of text type.

[0043] The input and output relationship of the user health status update agent UpdAgent(·) is:

[0044]

[0045] Among them, Sys upd It is the system prompt word for the user health status update agent, which includes the description of the user health status update task, the explanation of the meaning of each part of the health status, and a small number of examples of the user health status update task. upd It is the user prompt word of the user health status update agent, specifically refers to the generation format requirements of the user health status update task. Indicates concatenating five parts of text. hs2t (·) is to be A conversion function that converts a JSON object of the same structure into plain text. ch2t (·) is to be A conversion function that converts a JSON object of the same structure into plain text.

[0046] S33. Construct a system response type decision-making agent.

[0047] The system responds to the type decision agent DcsAgent(·) as Figure 4 As shown, there are four input and An output y dcs Output y dcs Is the system reply type, belongs to the text type, and is one of the 17 reply types. The 17 reply types are: "Explanation", "Inquiry", "Diagnosis", "Suggestion", "Explain and ask", "Inquiry and explain", "Diagnosis and explain", "Explain and diagnose", "Suggestion and explain", "Explanation and suggest", "Diagnosis and ask", "Inquiry and diagnose", "Suggestion and ask", "Inquiry and suggest", "Diagnosis and suggest", "Suggestion and diagnose", "Polite reply". Input It is a series of related knowledge retrieved, which belongs to text type. Input is the latest user health status, a JSON object, and is consistent with the input Have the same key-value pair structure. Input is a list of conversation history, with the input have the same structure. Input It is the current user input, which is of text type.

[0048] The input and output relationship of the system reply type decision agent DcsAgent(·) is:

[0049]

[0050] Among them, Sys dcsIt is the system prompt word for the system response type decision agent, which includes three parts: description of the system response type decision task, explanation of the meaning of each response type, and a small number of examples of the system response type decision task. dcs It is the user prompt word of the system response type decision agent, specifically refers to the generation format requirements of the system response type decision task. Indicates concatenating six parts of text. hs2t (·) Function and Cvt ch2t (·) The function is exactly the same as the two functions with the same name in sub-step S32.

[0051] S34. Build a multi-round healthy dialogue response agent.

[0052] The multi-round health dialogue reply agent ResAgent(·) is as follows Figure 5 As shown, there are five input and An output y res Output y res Is the final system response, which is of text type. Input is the system reply type, belonging to the text type, and the output y dcs have the same characteristics. Input It is a series of related knowledge retrieved, which belongs to text type. Input is the latest user health status, which is a JSON object and is consistent with the input Have the same key-value pair structure. Input is a list of conversation history, with the input have the same structure. Input It is the current user input, which is of text type.

[0053] The relational expression of the input and output of the multi-round healthy dialogue reply agent ResAgent is:

[0054]

[0055] Among them, Sys res (·) is the system prompt word function of the multi-round healthy dialogue reply agent, As a condition, the system prompt word pool is used to filter and obtain the system prompt word of the specified reply type. There are 17 different system prompt words in the system prompt word pool, corresponding to 17 different reply types, where each system prompt word contains two parts, namely, a description of the multi-round healthy dialogue reply task of the specified reply type and a small number of examples of the multi-round healthy dialogue reply task of the specified reply type. resIt is the user prompt word of the multi-round healthy dialogue reply agent, specifically refers to the generation format requirements of the multi-round healthy dialogue reply task.

[0056] Indicates concatenating six parts of text. hs2t (·) Function and Cvt ch2t (·) The function is exactly the same as the two functions with the same name in sub-step S32.

[0057] S4. Build a multi-agent collaborative framework to realize multi-round healthy dialogue.

[0058] It includes the following sub-steps:

[0059] S41. According to the current date and the unique identification code of the current user, search the user health archive to see if there is a user health status of the current user on the current date. If it exists, assign it to the historical user health status. If it does not exist, use the historical health status summary agent to summarize the user's full historical health status data of the current user in the user health archive, generate a historical health status summary, and upload it to the user health archive as the user health status of the current user on the current date. Specifically, a new key-value pair is created in the user's full historical health status data of the current user, with the current date as the key and the historical health status summary value as the historical health status summary, and then assigned to the historical user health status.

[0060] S42. Use the user health status to update the intelligent agent, taking the conversation history list, current user input, and historical user health status as input to generate the latest user health status.

[0061] S43. Using function Cvt hs2t (·) Convert the latest user health status into plain text, and use the vector retriever to query with the obtained plain text as the query text input to obtain a series of related knowledge.

[0062] S44. Utilize the system response type decision agent to generate the system response type using the obtained series of relevant knowledge, the latest user health status, the conversation history list, and the current user input as input.

[0063] S45. Utilize multiple rounds of health dialogue reply agents to generate the final system response using the obtained system response type, a series of relevant knowledge, the latest user health status, the dialogue history list, and the current user input as input.

[0064] S46. When the user continues the conversation, repeat sub-steps S41 to S45 to achieve multiple rounds of healthy conversations.

[0065] The multi-round healthy dialogue system with dynamic user status and real-time knowledge constraints in the present invention uses the method to complete the multi-round healthy dialogue task.

[0066] The technical solution proposed in the present invention not only makes up for the lack of knowledge in the current large medical and health model, but also focuses on improving the interpretability of the model and the controllability of the consultation. Specifically, it can actively ask questions at the right time and obtain relevant information based on the user's health status. It can also be explained by the timing of the consultation and the content of the consultation. In addition, the present invention also solves the problem that the current medical dialogue method is prone to errors or omissions in the memory of the user's health status in long contexts or across dates by constructing a user health archive, while avoiding exceeding the context length limit. A stronger ability to understand and remember the user's health status and a superior ability for multiple rounds of dialogue can be obtained.

[0067] For those skilled in the art, various other corresponding changes and deformations can be made according to the technical solutions and concepts described above, and all of these changes and deformations should fall within the protection scope of the claims of the present invention.

Claims

1. A multi-round healthy dialogue method with dynamic user status and real-time knowledge constraints, characterized by: The following steps are involved: S1. Build a medical health vector knowledge base and a corresponding vector retrieval engine; S2. Based on distributed storage and SSL / TLS encryption, use cloud database technology to build a user health archive; S3. Using the medical and health language model, we build a historical health status summary agent, a user health status update agent, a system response type decision agent, and a multi-round health dialogue response agent; S4. Build a multi-agent collaborative framework to realize multi-round healthy dialogue.

2. According to claim 1, the multi-round healthy dialogue method with dynamic user status and real-time knowledge constraints is characterized by: The step S1 comprises the following sub-steps: S11. Collect several drug instruction documents and several disease introduction documents, and divide all the collected documents into several text blocks according to subtitles; S12. Use the dense vector retrieval tool library and text embedding model to construct all text blocks into a searchable medical health vector knowledge base; The text embedding model is trained by a neural network and is used to encode text blocks of arbitrary length into an embedding vector of fixed dimension; S13. Construct a vector retriever using the text embedding model and the similarity retrieval function in the dense vector retrieval tool library; the vector retriever has an input x query , an output y kn ; Input x query is the query text input, which belongs to the text type; output y kn It is the retrieved relevant knowledge, which also belongs to the text type; y kn =Retriever(x query )=Cvt tl2t (SimSearch(Embedding(x query ),10)); Among them, Embedding(·) is the text embedding model, Retriever(·) is the vector retriever, SimSearch(·) is the similarity retrieval function; SimSearch function receives two inputs, the query vector and top-k, and returns the top-k text blocks that are most similar to the query vector in the form of a list; Cvt tl2t (·) is a conversion function that converts a text list into text in a specified format.

3. According to claim 2, the multi-round healthy dialogue method with dynamic user status and real-time knowledge constraints is characterized by: In the sub-step S11, the drug instructions document includes the shape, pharmacology and toxicology, indications, usage and dosage, adverse reactions, contraindications, interactions, specifications, and precautions of the drug; the disease introduction document includes the department to which it belongs, disease definition, cause, symptoms, diagnosis, drug treatment, and daily care; and the dense vector search tool library is FAISS.

4. The multi-round healthy dialogue method with dynamic user status and real-time knowledge constraints according to claim 3 is characterized in that: In step S2, the user health record of each user in the user health record library is stored separately as a JSON object type and indexed using a user unique identification code; the user unique identification code is a string of several digits, generated using the classic table ID self-increment method, and when the user successfully registers, the unique identification code is automatically generated according to the internal parameters and distributed and bound; The user health record contains two key-value pairs: user basic information and user's full historical health status data; the user basic information contains three key-value pairs: user's name, date of birth, and gender; the user's full historical health status data contains several key-value pairs, each key-value pair uses the date as the key and the user's health status on the corresponding date as the value; the user's health status on the corresponding date is composed of five key-value pairs: chief complaint, current medical history, past medical history, personal history, and family history.

5. The multi-round healthy dialogue method with dynamic user status and real-time knowledge constraints according to claim 4 is characterized in that: In step S3, the historical health status summary agent is represented as SumAgent(·); there is an input x sum , an output y sum ; Output y sum It is a summary of historical health status, including five key-value pairs: chief complaint, current medical history, past medical history, personal history, and family history; input x sum The user's full historical health status data in step S2; The input and output satisfy the relationship: y sum N SumAgent ( x sum )mStd hs (MedLLM([Sys sum Cvt sum2t (x sum )Usr sum ])) Among them, MedLLM(·) is a medical and health language model, which has the feature of returning text-type output given a text-type input; Sys sum It is the system prompt word of the historical health status summary agent; Usr sum It is the user prompt word of the historical health status summary agent; [Sys sum ;Cvt sum2t (x sum );Usr sum ] means concatenating the three parts of text; Std hs (·) is the health status normalization function, which normalizes and converts the content generated by MedLLM(·) into the same sim Same format; Cvt sum2t (·) is the sum A conversion function that converts a JSON object of the same structure into plain text.

6. The multi-round healthy dialogue method with dynamic user status and real-time knowledge constraints according to claim 5 is characterized in that: In step S3, the user health status update agent is represented as UpdAgent(·); it has three inputs: and An output y ipd ; Output y upd is the latest user health status; input is the user's health status on that day, and the output y upd Have the same key-value pair structure; input is a list of conversation history; input It is the current user input, which is of text type; The input and output satisfy the relationship: Among them, Sys upd It is the system prompt word for the user health status update agent; Usr upd is the user prompt word for the user health status update agent; Indicates concatenating five parts of text; Cvt hs2t (·) is to be Conversion function that converts a JSON object of the same structure into plain text; Cvt ch2t (·) is to be A conversion function that converts a JSON object of the same structure into plain text.

7. The multi-round healthy dialogue method with dynamic user status and real-time knowledge constraints according to claim 6 is characterized in that: In step S3, the system response type decision agent is represented as DcsAgent(·); there are four inputs and An output y dcs ; Output y dcs It is a system reply type, belonging to the text type; input is the retrieved relevant knowledge, which belongs to the text type; input is the latest user health status, a JSON object, and is consistent with the input Have the same key-value pair structure; input is a list of conversation history, with the input has the same structure; input It is the current user input, which is of text type; The input and output satisfy the relationship: Among them, Sys dcs is the system prompt word for the system response type decision agent; Usr dcs is the user prompt word for the system response type decision agent; Indicates concatenating six parts of text.

8. The multi-round healthy dialogue method with dynamic user status and real-time knowledge constraints according to claim 7 is characterized in that: In step S3, the multi-round health dialogue reply agent is represented as ResAgent(·); there are five inputs and An output y res ; Output y res is the final system response, which is of text type; input is the system reply type, belonging to the text type, and the output y dcs have the same characteristics; input is the retrieved relevant knowledge, which belongs to the text type; input is the latest user health status, which is a JSON object and is consistent with the input Have the same key-value pair structure; input is a list of conversation history, with the input has the same structure; input It is the current user input, which is of text type; The input and output satisfy the relationship: Among them, Sys res (·) is the system prompt word function of the multi-round healthy dialogue reply agent; Usr res It is the user prompt word for the multi-round healthy dialogue reply agent; Indicates concatenating six parts of text.

9. The multi-round healthy dialogue method with dynamic user status and real-time knowledge constraints according to claim 8 is characterized in that: The step S4 comprises the following sub-steps: S41. According to the current date and the unique identification code of the current user, whether there is a user health status of the current user on the current date in the user health archive, if it exists, it is assigned to the historical user health status, if not, the historical health status summary agent is used to summarize the user's full historical health status data of the current user in the user health archive, generate a historical health status summary, and upload it to the user health archive as the user health status of the current user on the current date; S42. Update the agent using the user health status, using the conversation history list, the current user input, and the historical user health status as input to generate the latest user health status; S43. Convert the latest user health status into plain text, and use the vector retriever to query the obtained plain text as the query text input to obtain relevant knowledge; S44. Using the system response type decision agent, the system response type is generated by taking the obtained relevant knowledge, the latest user health status, the conversation history list, and the current user input as input; S45. Using multiple rounds of health dialogue reply agent, the obtained system reply type, relevant knowledge, the latest user health status, the dialogue history list, and the current user input are used as input to generate the final system reply; S46. When the user continues the conversation, repeat sub-steps S41 to S45 to achieve multiple rounds of healthy conversations.

10. A multi-round healthy dialogue system with dynamic user status and real-time knowledge constraints, characterized by: The multi-round health dialogue task is completed using the multi-round health dialogue method as described in any one of claims 1-9.

Images