Intelligent medical inquiry method, device and equipment and storage medium

Abstract

The invention provides an intelligent medical inquiry method, device and equipment and a computer readable storage medium. According to the method, user intentions are classified by calling a large language model based on interaction information from a user, so that recall of a disease probability and recall of an optimal inquiry point are realized by calling different agents based on another large language model under the condition of determining that the classification of the user intentions is medical inquiry. Therefore, corresponding disease probability distribution and further inquiry information are generated and displayed to the user in a chart form. According to the method, a user can quickly obtain the dynamically updated suspected disease probability distribution diagram, and the user experience is greatly improved. According to the method, tasks of the large language model are decomposed through multiple agents, each subtask can be optimized independently, the training efficiency and the training effect of the model are greatly improved, and through combination of models of different scales, calculation power configuration can be optimized, and the model training cost is reduced.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence, and more specifically, to an intelligent medical consultation method, apparatus, device, and storage medium. Background Technology

[0002] Human-computer dialogue interaction, as an important branch of computer science, is used to achieve natural language communication between humans and computers. Human-computer dialogue is a significant challenge in artificial intelligence (AI) technology, and with the rapid development of AI, research on human-computer dialogue is becoming increasingly popular. The main goal of human-computer dialogue interaction technology is to enable computers to understand and generate human language and engage in natural conversations with humans. This requires computers to possess multiple capabilities, including natural language processing, machine learning, and knowledge representation. In recent years, with breakthroughs in deep learning technology, human-computer dialogue interaction technology has made significant progress. Deep learning models can learn language patterns and semantic knowledge from massive amounts of text data, thereby improving computers' ability to understand and generate natural language.

[0003] Currently, human-computer dialogue interaction technology has become a core technology for various intelligent dialogue applications, significantly impacting the efficiency and convenience of human-computer interaction. For example, in the medical field, medical assistants simulating doctor consultations serve as a novel medical tool, realistically mimicking doctor consultations through human-computer dialogue interaction, thereby providing users with convenient and efficient medical services. Therefore, an effective human-computer dialogue interaction method is needed to achieve superior intelligent medical consultations. Summary of the Invention

[0004] Embodiments of this disclosure provide an intelligent medical consultation method, apparatus, device, and computer-readable storage medium.

[0005] This disclosure provides an intelligent medical consultation method applied to a human-computer dialogue application. The method includes: responding to first dialogue information input by a user in a human-computer dialogue, classifying the user's intent based on the first dialogue information using a first major language model; responding to the user's intent being classified as a medical consultation, invoking a guiding decision-making agent based on a second major language model to determine whether to continue the consultation; responding to the guiding decision-making agent's determination to continue the consultation, invoking a clinical probabilistic reasoning agent based on the second major language model to recall at least two suspected diseases and their probability distributions, and invoking a consultation point determination agent based on the second major language model to recall an optimal consultation point; based on the at least two suspected diseases and their probability distributions, and the optimal consultation point, invoking a consultation agent based on the second major language model to generate second dialogue information, and outputting the second dialogue information in the human-computer dialogue. The second dialogue information includes a suspected disease probability distribution map and first consultation information. The suspected disease probability distribution map presents the at least two suspected diseases and their probability distributions in a graphical form, and the first consultation information is used to collect user information required for disease diagnosis.

[0006] This disclosure provides an intelligent medical consultation device applied to a human-computer dialogue application. The device includes: an intent classification module configured to, in response to first dialogue information input by a user in a human-computer dialogue, classify the user's intent based on the first dialogue information using a first major language model; a guidance decision module configured to, in response to the user's intent being classified as a medical consultation, invoke a guidance decision agent based on a second major language model to determine whether to continue the consultation; and a continue consultation module configured to, in response to the guidance decision agent's determination to continue the consultation, invoke a clinical probabilistic reasoning agent based on the second major language model to recall at least two suspected diseases and the at least two suspected... The system calculates the probability distribution of suspected diseases and calls a consultation point based on a second major language model to determine the optimal consultation point for the intelligent agent. An information output module is configured to, based on the probability distribution of the at least two suspected diseases and the optimal consultation point, call a consultation intelligent agent based on the second major language model to generate second dialogue information and output the second dialogue information in the human-computer dialogue. The second dialogue information includes a probability distribution map of suspected diseases and first consultation information. The probability distribution map of suspected diseases presents the probability distribution of the at least two suspected diseases in the form of a chart. The first consultation information is used to collect user information required for disease diagnosis.

[0007] Embodiments of this disclosure provide an intelligent medical consultation device, including: one or more processors; and one or more memories, wherein the one or more memories store a computer-executable program, and when the processor executes the computer-executable program, it performs the intelligent medical consultation method as described above.

[0008] Embodiments of this disclosure provide a computer-readable storage medium having computer-executable instructions stored thereon, which, when executed by a processor, are used to implement the intelligent medical consultation method described above.

[0009] Embodiments of this disclosure provide a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform an intelligent medical consultation method according to embodiments of this disclosure.

[0010] The method provided in this disclosure is based on user interaction information received during human-computer dialogue. It classifies user intent by invoking a large language model. If the user intent is determined to be a medical consultation, it recalls disease probabilities and optimal consultation points by invoking different agents based on another large language model. This generates and displays the corresponding disease probability distribution and consultation information to the user in graphical form, along with user information needed for further diagnosis. The method provided in this disclosure enables users to quickly obtain dynamically updated probabilities of suspected diseases and achieve accurate disease diagnosis through multi-turn dialogue, significantly improving the user experience. Furthermore, decomposing the large language model's tasks using multiple agents facilitates individual optimization of each sub-task, greatly improving training efficiency and effectiveness. Combining models of different sizes also optimizes computational power allocation and reduces model training costs. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some exemplary embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0012] Figure 1A This is a schematic diagram illustrating a human-computer dialogue interaction scenario according to an embodiment of the present disclosure;

[0013] Figure 1BThis is a schematic diagram illustrating a dialog interface at a user terminal according to an embodiment of the present disclosure;

[0014] Figure 2A This is a flowchart illustrating an intelligent medical consultation method according to an embodiment of the present disclosure;

[0015] Figure 2B This is a schematic diagram illustrating an intelligent medical consultation method according to an embodiment of the present disclosure;

[0016] Figure 3 This is a schematic diagram illustrating an intelligent medical consultation method according to an embodiment of the present disclosure;

[0017] Figure 4 This is a schematic diagram illustrating the dialogue interface of a human-computer dialogue application according to an embodiment of the present disclosure;

[0018] Figure 5 This is a schematic diagram illustrating the updating of a suspected disease probability distribution map based on first dialogue information input by the user in a human-computer dialogue according to an embodiment of the present disclosure;

[0019] Figure 6A This is a schematic diagram illustrating the implementation process of an intelligent medical consultation method according to an embodiment of the present disclosure;

[0020] Figure 6B This is a schematic diagram illustrating another implementation process of the intelligent medical consultation method according to an embodiment of the present disclosure;

[0021] Figure 6C This is a schematic diagram illustrating the agent architecture of an intelligent medical consultation method according to an embodiment of the present disclosure;

[0022] Figure 7 This is a schematic diagram illustrating an intelligent medical consultation device according to an embodiment of the present disclosure;

[0023] Figure 8 A schematic diagram of an intelligent medical consultation device according to an embodiment of the present disclosure is shown; and

[0024] Figure 9 A schematic diagram of the architecture of an exemplary computing device according to an embodiment of the present disclosure is shown. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this disclosure more apparent, exemplary embodiments according to this disclosure will now be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this disclosure, and not all embodiments of this disclosure. It should be understood that this disclosure is not limited to the exemplary embodiments described herein.

[0026] In this specification and accompanying drawings, steps and elements that are substantially the same or similar are indicated by the same or similar reference numerals, and repeated descriptions of these steps and elements are omitted. Furthermore, in the description of this disclosure, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance or order.

[0027] In embodiments of this disclosure, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0028] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0029] To facilitate the description of this disclosure, the following concepts related to this disclosure are introduced.

[0030] The intelligent medical consultation method disclosed herein can be based on artificial intelligence (AI). Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. In other words, artificial intelligence is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. By studying the design principles and implementation methods of various intelligent machines, artificial intelligence enables the intelligent medical consultation method disclosed herein to dynamically update the probability distribution of categories associated with the scenario classification based on the user's input and the question in real time, and then display this information to the user.

[0031] The intelligent medical consultation method disclosed herein can be based on Natural Language Processing (NLP). NLP is an important field within computer science and artificial intelligence. It studies various theories and methods that enable effective communication between humans and computers using natural language. NLP is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field involves natural language, that is, the language people use in daily life, and thus it has a close relationship with linguistic research. NLP technologies typically include text processing, semantic understanding, machine translation, robot question answering, knowledge graphs, and other technologies.

[0032] The intelligent medical consultation method disclosed herein can be based on a Large Language Model (LLM). A LLM is a language model in the field of Natural Language Processing (NLP) trained on a large corpus of language data using an autoregressive approach to analyze and process language text, aiming to understand and generate human language. Trained on massive amounts of text data, LLMs can perform a wide range of tasks, including text summarization, translation, and sentiment analysis. A key characteristic of LLMs is their massive scale, potentially containing billions of parameters to help them learn complex patterns in language text data. Therefore, in the intelligent medical consultation method disclosed herein, the trained LLM can be applied to the processing of interactive information in human-computer dialogue to provide efficient and accurate dialogue services, thereby improving the user experience.

[0033] In summary, the solutions provided by the embodiments of this disclosure involve technologies such as artificial intelligence, natural language processing, and large language models. The embodiments of this disclosure will be further described below with reference to the accompanying drawings.

[0034] Figure 1A This is a schematic diagram illustrating a human-computer dialogue interaction scenario according to an embodiment of the present disclosure. Figure 1B This is a schematic diagram illustrating a dialog interface at a user terminal according to an embodiment of the present disclosure.

[0035] like Figure 1AAs shown, a user can send user input (e.g., a request to obtain information from the server, or a question from the user to the server) to the server via a user terminal, so that the server can respond to the user input. Subsequently, the server can return the generated machine response information to the user terminal via the network. Optionally, the user terminal may specifically include smartphones, tablets, laptops, desktop computers, in-vehicle terminals, wearable devices, etc., but is not limited to these. The user terminal may also be a client with a browser or various applications (including system applications and third-party applications) installed. Optionally, the network may be an Internet of Things (IoT) based on the Internet and / or telecommunications networks, which can be a wired network or a wireless network. For example, it may be an electronic network capable of information exchange, such as a local area network (LAN), a metropolitan area network (MAN), a wide area network (WAN), or a cellular data communication network. The user terminal and the server can be directly or indirectly connected through wired or wireless communication methods, which is not limited herein. Optionally, the server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. Of course, it should be understood that, in addition to... Figure 1A Beyond the interactive scenarios shown, human-computer dialogue interaction can be implemented solely on the user terminal side, and this disclosure does not impose any restrictions on this. For example, a functional module for human-computer dialogue interaction can be deployed on the user terminal side, enabling the user terminal to independently implement human-computer dialogue interaction without relying on a server.

[0036] The development of LLM-based intelligent assistants is a significant research direction in the field of Natural Language Processing (NLP). LLM, as a deep learning-based NLP model, has wide applications, including language generation, text classification, and information retrieval. Currently, LLM-based intelligent assistant development schemes primarily involve inputting relevant knowledge during LLM training and training the model through autoregressive tasks to enable intelligent question-answering capabilities. In this approach, LLM is considered a general intelligent model capable of learning and simulating human language abilities, thereby achieving natural language understanding and generation. Therefore, LLM-based intelligent assistant development schemes have significant research and application value, especially for intelligent assistants in specific domains (such as legal, medical, and customer service), where they represent a very promising application solution.

[0037] In the construction of LLM-based intelligent assistants, human-computer dialogue can be categorized into three types based on function: task completion, question-and-answer, and chat. Task completion-type human-computer dialogue is primarily used to fulfill specific user needs, such as booking movie tickets, flight tickets, or playing music. These dialogues are characterized by clearly defined user needs, often requiring multiple rounds to resolve. As an example, in the medical field, medical assistants simulating doctor consultations serve as a novel medical tool, realistically mimicking doctor consultations through task completion-type human-computer dialogue, thereby providing users with convenient and efficient medical services.

[0038] The aforementioned medical assistant has become a highly anticipated technology in the medical field. It can be used to understand user queries and extract relevant medical information, store structured knowledge about diseases, symptoms, and treatment plans, and predict possible diseases based on the user's symptoms and medical history. For example... Figure 1B As shown, users can interact with medical assistants through text conversations. The medical assistant can conduct multiple rounds of consultations with users based on information extracted from the question-and-answer dialogues and combined with its stored structured knowledge, and then make a diagnosis based on the information provided by the user.

[0039] However, in the existing medical assistant interactions described above, users only answer the assistant's questions to receive a diagnosis. They cannot understand the reasons behind the assistant's questions, nor can they know the basis for the diagnosis. This can lead to users being unable to understand or trust the assistant's actions, thus affecting their ability to make informed medical decisions and significantly reducing the user experience. Furthermore, these medical assistants often fail to continuously update their disease prediction probability distribution based on real-time user feedback. This means that when users provide new information or symptoms, these medical assistants cannot adjust their disease probability predictions, potentially leading to inaccurate diagnoses and treatment recommendations.

[0040] Therefore, this disclosure aims to provide more accurate and personalized medical advice, in which possible diseases and their probabilities can be displayed to users in the form of charts. This will enable users to easily compare the probabilities of different diseases and make more informed medical decisions, and the disease probability distribution will be continuously updated based on users' real-time feedback to ensure that the medical assistant can continuously improve its diagnosis as users provide more information.

[0041] Based on this, this disclosure provides an intelligent medical consultation method. Based on user interaction information received during human-computer dialogue, the method categorizes user intent using a large language model. If the user intent is determined to be a medical consultation, different agents based on another large language model are invoked to recall disease probabilities and optimal consultation points. This generates and displays the corresponding disease probability distribution and consultation information to the user in graphical form, along with user information needed for further diagnosis. The method provided by the embodiments of this disclosure enables flexible and efficient human-computer dialogue interaction using different models of varying sizes. This allows users to quickly obtain dynamically updated probabilities of suspected diseases and achieve accurate disease diagnosis through multi-turn dialogue, thereby significantly improving the user experience.

[0042] Figure 2A This is a flowchart illustrating an intelligent medical consultation method 200 according to an embodiment of the present disclosure. Figure 2B This is a schematic diagram illustrating an intelligent medical consultation method according to an embodiment of the present disclosure. Figure 3 This is a schematic diagram illustrating an intelligent medical consultation method according to an embodiment of the present disclosure. Figure 4 This is a schematic diagram illustrating the dialogue interface of a human-computer dialogue application according to an embodiment of the present disclosure.

[0043] As described above, the intelligent medical consultation method disclosed herein can guide users to continuously provide information related to the task under the scenario classification through multi-turn dialogue interaction, thereby updating the probability distribution of the task and finally outputting the task result for user reference. Figure 3 As shown, for each round of dialogue interaction, the probability distribution of the task can be dynamically updated based on user feedback (e.g., the first dialogue information). Optionally, the intelligent medical consultation method of this disclosure can be applied to a human-computer dialogue application based on a large language model, which can be installed, for example, on a user terminal and perform data processing through a server. Optionally, the user can open a new dialogue interface in the human-computer dialogue application and initiate a human-computer dialogue by entering the information they wish to query.

[0044] like Figure 2A As shown, in step S201, in response to the first dialogue information input by the user in the human-computer dialogue, the user's intent can be classified based on the first dialogue information using a first large language model.

[0045] Optionally, at the beginning of each round of dialogue interaction, user intent can be identified based on the interaction information entered by the user in the dialogue interface of the human-computer dialogue application to determine the scenario classification of the current dialogue interaction, and corresponding subsequent operations can be performed in this round of dialogue interaction. That is, the domain and task scenario of the current human-computer conversation can be determined, such as a consultation scenario in the medical field or a general question-and-answer scenario.

[0046] Optionally, a pre-trained large language model (the first large language model) can be used to identify intent from the user's input interaction information, such as determining the scenario classification of the current dialogue interaction, and then performing subsequent processing corresponding to that scenario based on the interaction information in the current dialogue interaction. For example, Figure 4 An example human-computer dialogue is shown in the dialogue interface of a human-computer dialogue application. When a user enters the interactive information "What's wrong with my sudden headache?" in the dialogue interface, the first language model can be used to perform intent recognition on the input to determine the current human-computer dialogue scenario category, such as a medical consultation scenario. Then, based on the input, subsequent dialogue is conducted for the consultation scenario.

[0047] Optionally, the primary language model can be a large language model that has undergone supervised fine-tuning (SFT), which is customized based on data from the intent recognition task, thereby improving performance and effectiveness on the intent recognition task. Considering limited computing power, for the intent recognition task, a small-scale model can be used to obtain the primary language model for intent recognition through SFT, thereby reducing training costs while ensuring training effectiveness. After supervised fine-tuning, this primary language model can better perform intent recognition and scene classification based on user input interaction information.

[0048] Specifically, the generation process of the aforementioned first large language model involves several key steps. First, during the generation of the first large language model, a small-scale large language base model (such as baichuan7B, baichuan13B, llama2, etc.) needs to be acquired. Second, for the intent classification task, a certain amount of labeled training data can be collected to perform SFT on the large language model. That is, this training data contains user input information (interaction information) and also includes intent annotation information for the user input. Based on this, the SFT can be performed on the base model using this training data to ensure that the SFT-processed model can effectively classify and recognize the user's input intent. By using a small-scale large language model as the base model for intent recognition, not only can the requirements of the user intent classification task be met, but computational resources can also be saved.

[0049] In step S202, in response to the user's intent being classified as a medical consultation, a guiding decision-making agent based on the second major language model may be invoked to determine whether to continue the consultation.

[0050] like Figure 3 As shown, after obtaining the first dialogue information (310) input by the user, it can be determined that the input is intended for recognition (301). Furthermore, if it is determined that the current input is related to medical consultation, for example, if it is determined that the user's intention is classified as medical consultation, subsequent guidance decisions in the medical consultation scenario can be executed based on the user input.

[0051] Optionally, for each round of dialogue interaction, the execution of guided decisions may include determining whether to output further questions to the user (i.e., continue the consultation) or to complete the current disease diagnosis to output the disease diagnosis result (i.e., make a diagnosis).

[0052] As an example, if a dialogue system needs to gather more information to complete the current disease diagnosis, it can ask the user further questions. For instance, in a medical consultation scenario, if the dialogue system is diagnosing a user's medical condition, it may need to ask further questions about the user's symptoms, medical history, and lifestyle. Conversely, if the dialogue system has already gathered enough information to complete the current disease diagnosis, it can output the diagnosis result. For example, if the dialogue system is diagnosing a user's medical condition, it can make a diagnosis and output the diagnosis result.

[0053] On the one hand, when the collected information is insufficient, the dialogue system can make a guiding decision to continue the consultation, in order to request further information from the user by outputting new consultation information.

[0054] Optionally, the intelligent medical consultation method of this disclosure can determine the disease diagnosis result corresponding to the user's input information by utilizing various intelligent agents based on a large language model. These intelligent agents can progressively refine the disease diagnosis information used to determine the diagnosis result based on each user's interactive input, and then concretely display the disease diagnosis information to the user. Optionally, these intelligent agents can be constructed based on a second large language model. According to embodiments of this disclosure, the first large language model can be different from the second large language model, and the first large language model is smaller in scale than the second large language model. Optionally, the first large language model used for intent recognition can be smaller in scale than the second large language model used to construct various intelligent agents. Specifically, the first large language model can be a small-scale small model base model, while the second large language model can be a large-scale medical large language model, which can be obtained by SFT training on SFT data constructed from medical knowledge based on a large-scale large model base model. By combining models of different sizes in this way, computational power configuration can be optimized, model training costs can be reduced, and computational resources can be saved.

[0055] Optionally, the aforementioned guidance decision-making can be executed using a guidance decision-making agent based on a second major language model. That is, after determining that the current dialogue interaction scenario is a medical consultation scenario based on the acquired initial dialogue information, the guidance decision-making agent can be directly invoked to execute the guidance decision. For example, the guidance decision-making agent itself can have the ability to execute guidance decisions, simulating a doctor's actual consultation to make decisions about the next action.

[0056] As described above, the aforementioned guided decision-making agent can be based on a large-scale base model (i.e., a second large language model). Therefore, for guided decision-making tasks, a certain amount of labeled training data can be collected to perform SFT on the large base model. This training data can be constructed based on medical knowledge, enabling the SFT-processed large base model (i.e., the second large language model) to possess structured data and professional knowledge in the medical field. Medical knowledge plays a crucial role in constructing the second large language model. SFT data constructed using medical knowledge is used to train the base model using SFT, resulting in the second large language model. The structured data and domain-specific knowledge allow the second large language model to perform better in the medical field. Based on the second large language model, prompting engineering (PE) can be used to train it to generate responses that conform to the guided decision-making task, thereby creating a guided decision-making agent. Specifically, a series of prompts related to guided decision-making can be designed. These prompts provide relevant context and guidance to the model, enabling the second large language model to generate responses to the guided decision-making task, i.e., determining whether to continue the consultation or make a diagnosis, thus creating a guided decision-making agent.

[0057] like Figure 3 As shown, if the guiding decision-making agent determines to continue the consultation, further consultation operations can be performed, including recalling the probability distribution of suspected diseases as shown in 303, and recalling the optimal consultation point as shown in 304, to display the current disease diagnosis information to the user and output further consultation information. Optionally, the intelligent medical consultation method of this disclosure can determine the disease diagnosis result through multiple rounds of human-computer dialogue, wherein each round of human-computer dialogue, except for the last round, can include the user's response information to the consultation information output to them in the previous round (e.g., as the first dialogue information 310), and the further consultation information output to the user (and the updated probability distribution map of suspected diseases, i.e., the second dialogue information 320).

[0058] In step S203, in response to the guidance decision agent's determination to continue the consultation, a clinical probabilistic reasoning agent based on the second largest language model can be invoked to recall at least two suspected diseases and the probability distribution of the at least two suspected diseases, and a consultation point based on the second largest language model can be invoked to determine the agent to recall the optimal consultation point.

[0059] As described above, similar to the guided decision-making agent, the aforementioned clinical probabilistic reasoning agent and consultation point determination agent can be based on a large-scale large model base model (i.e., a second large language model). Therefore, a certain scale of labeled training data can be collected to perform SFT on the large model base model. This training data can be constructed based on medical knowledge, enabling the SFT-processed large model base model (i.e., the second large language model) to possess structured data and professional knowledge in the medical field. Medical knowledge plays a crucial role in constructing the second large language model. SFT data constructed using medical knowledge is used to train the base model using SFT, thereby obtaining the second large language model. Through structured data and professional domain knowledge, the second large language model can perform better in the medical field. Based on the same second large language model, PE can be used to train the second large language model for clinical probabilistic reasoning tasks and consultation point determination tasks respectively, generating responses that conform to the clinical probabilistic reasoning tasks and consultation point determination tasks, thereby creating clinical probabilistic reasoning agents and consultation point determination agents. Specifically, for the clinical probabilistic reasoning task, a series of prompts related to clinical probabilistic reasoning can be designed. These prompts can provide the model with relevant context and guidance, enabling the second language model to generate a response for the clinical probabilistic reasoning task, namely, recalling at least two suspected diseases and their probability distributions, thereby creating a clinical probabilistic reasoning agent. Similarly, for the consultation point determination task, a series of prompts related to consultation point determination can be designed. These prompts can provide the model with relevant context and guidance, enabling the second language model to generate a response for the consultation point determination task, namely, recalling the optimal consultation point, thereby creating a consultation point determination agent.

[0060] According to embodiments of this disclosure, the clinical probabilistic reasoning agent may include: a disease recall agent and a disease probabilistic reasoning agent, wherein the disease recall agent may be configured to recall at least two suspected diseases related to the human-computer dialogue based on the human-computer dialogue including the first dialogue information, and the disease probabilistic reasoning agent may be configured to recall the probability of each of the at least two suspected diseases being diagnosed based on the human-computer dialogue.

[0061] Optionally, based on each user's interaction with the consultation information in the dialogue interface, a clinical probabilistic reasoning agent can update the probability distribution map of suspected diseases displayed to the user based on each input. Updating the probability distribution map can include introducing new categories, deleting previous categories, and changing the probabilities of each category. For example, in a consultation scenario, the dialogue system (or application) can adjust the disease probability distribution based on the user's responses. This adjustment can include adjusting the probabilities of previously inferred diseases, introducing new diseases, or deleting previous diseases.

[0062] For example, after acquiring new interaction information from user input, a clinical probabilistic reasoning agent can be used to refine and update the previous prediction of the suspected disease probability distribution (i.e., the probability distribution of diagnosis for each of at least two suspected diseases related to the human-computer dialogue) based on the new interaction information, i.e., recall the latest suspected disease probability distribution (303). Optionally, a disease recall agent can be invoked to recall suspected diseases related to the current human-computer dialogue, and a disease probabilistic reasoning agent can be invoked to determine the probability of diagnosis for each of the recalled suspected diseases, thereby generating an updated suspected disease probability distribution.

[0063] Optionally, the updated probability distribution of suspected diseases can be output to the user in the form of a chart in the dialog interface, i.e., outputting a probability distribution map of suspected diseases.

[0064] Figure 5 This is a schematic diagram illustrating the updating of a suspected disease probability distribution map based on first dialogue information input by the user in a human-computer dialogue according to an embodiment of the present disclosure. Figure 5 As shown, in each round of dialogue interaction, the user is shown an updated probability distribution map of suspected diseases. In response to the question "Do you have symptoms of nausea, vomiting, abdominal pain or fever?" in the previous round of dialogue interaction, the user inputs the new interaction information "nausea". Therefore, in the current human-computer dialogue, the probability distribution map of suspected diseases is updated based on the new interaction information. For example, the probability of "food poisoning" increases by 10%, while the probability of "bacterial diarrhea" decreases by 3%.

[0065] Therefore, through the dynamic updating and display of the probability distribution of suspected diseases, users can quickly obtain key information related to their own category, while also learning about the possibilities of other categories. This allows users to have a more comprehensive understanding of their current situation, enabling them to make better decisions.

[0066] After recalling the latest probability distribution of suspected diseases, further consultation is needed to determine consultation information for the next round of dialogue. Optionally, if the guiding decision is to continue consultation, multiple consultation points can be recalled by calling a consultation point determination agent based on the second largest language model, and the optimal consultation point can be selected from them.

[0067] According to embodiments of this disclosure, the consultation point determination agent may include a consultation point recall agent and a consultation point screening agent, wherein the consultation point recall agent may be configured to recall at least two consultation points based on the at least two suspected diseases, and the consultation point screening agent may be configured to select the optimal consultation point from the at least two consultation points based on the at least two suspected diseases and the human-computer dialogue.

[0068] Optionally, the consultation point recall agent can recall multiple consultation points related to the context of the current human-computer dialogue based on its prior knowledge. These consultation points can be used to collect more information to help diagnose the current disease. In this case, it is necessary to recall all possible consultation points and exclude as many category hypotheses as possible from the list of suspected diseases.

[0069] Optionally, a consultation point screening agent can be used to select the optimal consultation point from a pool of recalled consultation points. For example, the consultation point screening agent can exclude used consultation points based on the human-computer dialogue record and select the optimal consultation point based on a priority ranking of multiple consultation points. This priority ranking can be associated with one or more of the following: the relevance of the consultation point to the current human-computer dialogue context and related categories, the consultation point's ability to distinguish different categories, and the consultation point's ability to gather more information, etc. Of course, the above method of selecting the optimal consultation point based on a priority ranking of multiple consultation points is merely an example and not a limitation in this disclosure.

[0070] As an example, in a medical consultation scenario, if the current human-computer dialogue is as follows:

[0071] Patient: I have a headache.

[0072] Doctor: When did the headache start?

[0073] Patient: It started last night.

[0074] The new first dialogue information obtained from the user input is "started last night". Therefore, the suspected diseases related to the current human-computer dialogue recalled by the disease recall agent are:

[0075] ["Nervous system disorders", "Migraine", "Tension headache", "Intracranial hypertension", "Intracranial tumor", "Intracranial hemorrhage", "Intracranial infection", "Hypertension"]

[0076] In other words, based on the above human-computer dialogue, it is currently determined that the user may be suffering from one of the eight diseases mentioned above. For each disease, a disease probability inference agent can assign a probability value to determine the probability of the user having each disease, thereby generating the following suspected disease probability distribution:

[0077] ["Other neurological disorders, 0.1"]

[0078] "Migraine, 0.3"

[0079] "Tension headache, 0.3"

[0080] "Other increases in intracranial pressure, 0.05"

[0081] "Intracranial tumor (risk), 0.05"

[0082] "Intracranial hemorrhage (acute) (critical), 0.05"

[0083] Intracranial infection (acute) (critical), 0.05"

[0084] "High blood pressure, 0.1"

[0085] Therefore, based on the above probability distribution of suspected diseases, multiple consultation points can be recalled using the consultation point recall agent as follows:

[0086] [1. Timing of headache (List of ruled-out suspected diseases - migraine, tension headache; List of confirmed suspected diseases - traumatic brain injury (acute), intracranial infection (acute) (critical))]

[0087] 2. Severity of headache (List of ruled-out suspected diseases - tension headache; List of confirmed suspected diseases - brain tumor (critical), other increased intracranial pressure)

[0088] 3. Accompanying symptoms (List of excluded suspected diseases - other neurological diseases; List of confirmed suspected diseases - other cerebrovascular diseases, intracranial infection (acute) (critical)) ...]

[0090] For multiple consultation points identified during the recall, a consultation point filtering agent can be used to select the optimal consultation point, for example:

[0091] ["Timing of headache" (List of suspected diseases ruled out - migraine, tension headache; List of suspected diseases confirmed - traumatic brain injury (acute), intracranial infection (acute) (critical))).

[0092] Optionally, based on the selected optimal consultation point, consultation information corresponding to the optimal consultation point (e.g., "When did the headache start?") can be generated by calling a consultation agent based on the second largest language model, and the consultation information can be output to the user to execute the next round of dialogue interaction.

[0093] In step S204, based on the at least two suspected diseases and their probability distributions, as well as the optimal consultation point, a consultation agent based on the second language model can be invoked to generate second dialogue information, and the second dialogue information can be output in the human-computer dialogue. The second dialogue information may include a probability distribution map of suspected diseases and first consultation information. The probability distribution map of suspected diseases presents the at least two suspected diseases and their probability distributions in the form of a chart. The first consultation information is used to collect user information required for disease diagnosis.

[0094] Optionally, if the guiding decision is to continue the consultation, by recalling the probability distribution of suspected diseases and the optimal consultation point, the consultation agent based on the second largest language model can be invoked to generate and output the second dialogue information. The second dialogue information may include the probability distribution of suspected diseases (i.e., the probability distribution map of suspected diseases) displayed to the user in the form of a chart, as well as further consultation information (i.e., the first consultation information).

[0095] According to embodiments of this disclosure, the consultation intelligence agent may include: a disease probability distribution display intelligence agent and a consultation point display intelligence agent, wherein the disease probability distribution display intelligence agent may be configured to generate and display the probability distribution map of the suspected diseases to the user based on the at least two suspected diseases and the probability distribution of the at least two suspected diseases, and the consultation point display intelligence agent may be configured to generate and display the first consultation information to the user based on the optimal consultation point.

[0096] Optionally, a disease probability distribution display agent can be invoked to generate and display a suspected disease probability distribution map to the user based on the aforementioned suspected disease probability distribution.

[0097] Optionally, the first consultation information can be generated based on the recalled optimal consultation point by calling the consultation point display agent, and then displayed to the user.

[0098] like Figure 2BAs shown, after obtaining the user's initial dialogue information "What causes a sudden headache?", the current human-computer dialogue scenario can be classified as a "medical consultation scenario." This means the task of the current human-computer dialogue is to diagnose a disease based on the user's input. Therefore, based on this interaction information, the diseases corresponding to the user's input can be determined, i.e., the diseases the user may have, and the probability of the user having each of these diseases can be determined. For example, based on the interaction information "What causes a sudden headache?", it can be determined that the user may currently have three diseases: "ruptured cerebral blood vessel," "migraine," and "tension headache," with the probabilities of the three diseases decreasing in that order. Optionally, for Figure 2B The illustrated consultation scenario, based on the user's initial input, can output a second dialogue, which includes a probability distribution chart of suspected diseases. It can also simulate a doctor's consultation, outputting the initial consultation information needed to collect user information for disease diagnosis, in order to further understand the user's disease-related symptoms or conditions. As an example, based on the initial determination of the disease probability distribution, further questions can be asked the user, such as the initial consultation information, "Where is your headache located? Are there any accompanying symptoms such as nausea, vomiting, or blurred vision?", to further collect user information.

[0099] Optionally, in the intelligent medical consultation method disclosed herein, the latest probability distribution map of suspected diseases can be displayed to the user in the form of a chart, allowing the user to intuitively understand the diseases they may currently have and their probabilities. Furthermore, the intelligent medical consultation method disclosed herein can also employ other methods to display the probability distribution map of suspected diseases to the user.

[0100] According to embodiments of this disclosure, the disease recall agent and the consultation point recall agent can be obtained through reinforcement learning based on a second language model.

[0101] Optionally, the disease recall agent can be created based on the second large language model through the prompting engineering described above. It can recall various diseases related to the user's input information based on the second large language model, which possesses strong medical knowledge. Furthermore, considering that the large language model may suffer from hallucination problems, leading to inaccurate disease output, and that disease recall requires high data accuracy, some inaccurate negative samples generated during the prompting engineering can be used as partial training data to reinforce the second large language model. This allows the trained disease recall agent to accurately recall various diseases related to the user's input.

[0102] Alternatively, considering that the accuracy of data is also relatively high for consultation point recall, a consultation point recall agent can be created in a similar way to the disease recall agent. This would allow the trained consultation point recall agent to accurately recall multiple consultation points related to the current human-computer dialogue context based on its prior knowledge, in order to collect more information to help diagnose the current disease.

[0103] For example, the disease recall agent and the consultation point recall agent can be Dialogue Policy Optimization (DPO) models obtained by reinforcement learning on the second largest language model. By adopting models that have undergone reinforcement learning, such as the DPO model, the illusion problem of the large language model can be effectively avoided.

[0104] According to embodiments of this disclosure, the guiding decision agent, the disease probability reasoning agent, the consultation point screening agent, the disease probability distribution display agent, and the consultation point display agent can be obtained through prompting engineering based on a second language model.

[0105] Alternatively, using prompting engineering to create agents from large language models can involve designing a series of prompts so that the model can generate specific types of responses. By providing the model with relevant context and guidance, it can be trained to generate responses that meet specific needs, thereby creating various types of agents.

[0106] As an example, for the aforementioned disease probability inference agent, probabilistic continuation can be obtained by instructing the model to perform probability continuation. Specifically, contextual information can be passed to the model in plain text (e.g., split into dialogue text between the patient and the doctor), providing output samples for the model. Then, the model is instructed to perform continuation based on the new dialogue text to output the probability distribution of the suspected disease in the format of the output samples.

[0107] For example, the prompt text used for the second largest language model could be as follows: [

[0109] Me: The doctor-patient dialogue is as follows:

[0110] Patient: My baby is 5 months and 23 days old. Is a weight of 8kg normal? Please list five most likely diseases and their probability levels.

[0111] Your: Normal, 0.8; Malnutrition, 0.05; Growth hormone deficiency, 0.05; Hereditary disease, 0.03; Digestive system disease, 0.02; Endocrine disease, 0.02.

[0112] Me: The doctor-patient dialogue is as follows:

[0113] Patient: I have a burning sensation in my stomach that doesn't subside after eating. Will the anti-tuberculosis medication worsen it? Please list ten most likely diseases and their probability levels.

[0114] you: ]

[0116] Therefore, for the example of the prompt text above, the second language model can produce the following output: [

[0118] Gastritis or gastric ulcer, 0.1; hyperacidity, 0.08; drug side effects, 0.06; gastrointestinal infection, 0.05; gastric cancer, 0.04; gastroesophageal reflux disease, 0.03; stress response, 0.03; chronic gastritis, 0.02; gastric ulcer, 0.02; gastrointestinal ulcer, 0.01; gastrointestinal inflammation, 0.01. ]

[0120] As mentioned above, through prompting engineering, the disease probability reasoning agent can output the corresponding probability distribution of suspected diseases based on the current human-computer dialogue.

[0121] According to embodiments of this disclosure, the consultation point display agent can also be configured to generate and display explanatory information about the first consultation information to the user based on the doctor's thought process, explaining to the user the reason for outputting the first consultation information. Optionally, the consultation point display agent can explain to the user the reason for outputting the consultation information and / or the probability distribution map of suspected diseases through explanatory language, such as... Figure 4 and Figure 5 The text above the probability distribution chart of suspected diseases is shown in the image.

[0122] As an example, for a consultation point display agent, this can be achieved by instructing the model to continue writing. Specifically, context information can be combined into a single-round message and passed to the model as plain text to instruct the model to continue writing.

[0123] For example, the prompt text used for the second largest language model could be as follows: [

[0125] {'role':'user','content':'[{"role":"system","content":"You are a large language model.\\nYou are a kind and friendly general practitioner.\\nI will talk to you as a patient. Please combine your knowledge of common diseases in the patient's age group: Each time, you should list the decimal probability of the disease or human body system that the patient's symptoms may indicate, and explain the disease or human body system that your next question as a doctor will target and your questioning approach. Finally, you should ask only one question to help with the diagnosis.\\n\\nYour thinking is:\\n1. First, if the human body system from which the patient's symptoms originate is unclear, ask questions to confirm which human body system the patient's symptoms indicate.\\n2. Next, after clarifying the source of the patient's symptoms, use questions to determine whether it is an acute or chronic problem, such as asking when the symptoms appeared.\\n3. Then, because of the patient's subjective description Not entirely reliable; ask questions to gather reliable information from sources other than the patient, such as previous diagnoses by other doctors and the effects of medication after treatment. 4. Secondly, prioritize ruling out dangerous and serious diseases. 5. Gradually eliminate possible diseases to arrive at a final diagnosis. You are the doctor; do not ask the patient if they wish to go to the hospital, undergo examinations, or take measures to improve their condition. You need to strictly follow the above thought process when asking questions. Unless the patient's symptoms are directly related to genes / allergies, do not ask about family medical history / allergy history. If the patient mentions multiple symptoms, ask to clarify the primary and accompanying symptoms. When listing the probability of a disease, if the disease is dangerous or serious, add "(serious)" after the disease name. Reply format: <probabilities>List the patient's normal probability, the probability of possible diseases or corresponding body systems, and the overall probability of other possible diseases in comma-separated (CSV) format. Format: "Normal, Normal Probability\\\nDisease Name, Corresponding Probability\\\nOther Diseases, Probability of Other Conditions". <explanation>Directly, concisely, and specifically using only two declarative sentences, explain the disease or body system that needs to be addressed when asking the next question based on the patient's answer, and the specific line of reasoning behind the question, such as confirming a high-probability disease or ruling out a low-probability serious disease. <question>Only one question can be asked at a time. It should be concise (subject omitted), conversational, and easy to understand. Ensure only one question is asked at a time. "},{"role":"user","content":"After using the Qingrun Adjustable Nasal Cleanser to clean my nose, my nose is becoming increasingly uncomfortable, and I'm also experiencing tearing. Can I still use it?"},{"role":"assistant","content":" <probabilities>Rhinitis, 0.3; Sinusitis, 0.2; Allergy, 0.15; Dry nose, 0.1; Other, 0.25 ]

[0127] Therefore, for the example of the prompt text above, the second language model can generate a probability distribution "{"role":"assistant","content":" <probabilities>The explanatory text for the section "Rhinitis, 0.3; Sinusitis, 0.2; Allergy, 0.15; Dry Nasal Syndrome, 0.1; Other, 0.25" can be further developed. This explanatory text can be generated by automatically simulating a doctor's consultation logic using a large language model guided by prompting engineering based on the Doctor's Thought Chain (COT). The Doctor's Thought Chain is a thought process or mode followed by doctors when diagnosing and treating diseases. This chain can include steps such as the formation, observation, testing, and verification of diagnostic hypotheses to help doctors quickly and accurately diagnose patients' conditions. Therefore, through the aforementioned prompting engineering, the consultation point display agent can systematically analyze and reason based on the patient's symptoms, signs, and medical history using the medical diagnostic thought chain to determine the most likely diagnosis and formulate corresponding treatment plans. This allows for a better understanding of the patient's condition, improving diagnostic accuracy and treatment effectiveness.

[0128] Through the processing described above, in this round of dialogue interaction, in response to the guiding decision-making agent's determination to continue the consultation, further consultation information can be output, and the next round of dialogue interaction can be carried out based on the consultation information.

[0129] According to embodiments of this disclosure, the intelligent medical consultation method of this disclosure may further include: responding to a response from the user to the first consultation information input in the human-computer dialogue, and receiving the response as the first dialogue information. Optionally, the input for the next round of dialogue interaction may be a response from the user to the first consultation information output to them in the current round of dialogue interaction. That is, the response can be used as the first dialogue information in the next round of dialogue interaction to iteratively execute the operations described in steps S201-S204 above, thereby realistically simulating a doctor's consultation by executing multiple rounds of dialogue interaction.

[0130] On the other hand, such as Figure 3 As shown, when there is sufficient information, the guiding decision agent can determine to make a diagnosis, and in this case, it can provide the user with the diagnosis result information (330).

[0131] According to embodiments of this disclosure, in response to the determination of the given diagnostic result by the guiding decision-making agent, diagnostic result information can be generated based on the probability distribution of the at least two suspected diseases using a second language model.

[0132] Optionally, for the final round of dialogue interaction, instead of outputting the dialogue information including the updated probability distribution map of suspected diseases and further consultation information, the diagnosis result information of the disease can be output. That is, if the guiding decision-making agent can determine to output the diagnosis result information based on the latest first dialogue information input by the user, the diagnosis result information can be output in the dialogue interface instead of outputting the new probability distribution map of suspected diseases and further consultation information.

[0133] According to embodiments of this disclosure, the diagnostic result information may include a predicted result of the disease. The diagnostic result information may also include one or more of the following: treatment level, referral information for the disease, treatment recommendations, and reassurance information.

[0134] Optionally, if it is determined that a disease diagnosis can be made based on the current dialogue information, a final disease diagnosis result can be output to the user, which may include, for example, information indicating the disease category.

[0135] Optionally, diagnostic results can be generated using the second largest language model based on the recall probability distribution of suspected diseases. As an example, the system can output the user's judgment of the category to which their query belongs; for instance, if the system is diagnosing the user's medical condition, it could output the diagnosis "You may have migraines," such as... Figure 4 As shown.

[0136] Optionally, in addition to the disease diagnosis results, the dialogue system can also provide other indicative information about the disease. For example, the system can provide information on the current disease's treatment level, symptoms of migraines, treatment options and / or preventative measures, and reassurance information for the user. For instance, the current disease treatment level provided to the user can include the three levels described above, providing information on the severity of the disease for reference. For example, treatment level 1 might indicate a high probability of a diagnosis and recommend immediate medical attention; treatment level 2 might indicate that a certain disease cannot be ruled out and regular follow-up visits are necessary; and treatment level 3 might indicate mild symptoms, uncertainty about whether it is a disease, and recommend home observation. Furthermore, the diagnosis results can also present guidance information about the current disease, such as guidance on the disease, including its causes, symptoms, diagnostic methods, treatment options, and prognosis, to help patients better understand their health problems so they can make more informed decisions and take appropriate actions to manage and treat the disease. For example, treatment suggestions can be provided to users within the diagnostic results information. These suggestions can be tailored to the specific circumstances of the illness. Common treatment suggestions may include medication, surgery, physical therapy, and psychotherapy. Specifically, for example, for migraines, treatment suggestions may include avoiding triggering factors, medication, and relaxation techniques. Furthermore, reassuring information can be provided to users within the diagnostic results information, such as words of emotional support, encouragement, and comfort, aimed at helping patients maintain a calm and optimistic attitude when facing illness. Further, the dialogue system can also invoke relevant neural network models to provide suggestions on the best next steps. For example, if the dialogue system diagnoses a user with the flu, the model may suggest that the user seek medical attention or rest and drink plenty of water.

[0137] By providing a final diagnosis when a disease can be diagnosed, a clear answer and useful information can be given to the user, which can help the user understand their situation and take appropriate action.

[0138] According to embodiments of this disclosure, in response to the user intent being classified as question-and-answer, a second language model is used to generate answer information for the first dialogue information, which is used to answer the questions involved in the first dialogue information.

[0139] Optionally, if the scenario classification of this round of dialogue interaction is determined to be unrelated to a medical consultation scenario, such as when the scenario classification of this round of dialogue interaction is determined to be a medical question-and-answer scenario, the system can output answer information to the user's inquiry, that is, conduct human-computer dialogue interaction in the form of question-and-answer replies. For example, if the user's input information in this round of dialogue interaction is detected as "Can ibuprofen relieve headaches?", the system can output a reply to the user regarding this information, such as "Yes, ibuprofen is an over-the-counter drug, usually used to relieve mild to moderate pain, such as headaches, joint pain, or toothaches."

[0140] According to embodiments of this disclosure, the first dialogue information may include one or more of text information, diagnostic result images, and voice information. Optionally, the first dialogue information in this disclosure may be, in addition to text information, a user's diagnostic result image, such as a laboratory report or X-ray film, as well as voice information. Therefore, the aforementioned large language model used to process this information can have multimodal capabilities.

[0141] According to embodiments of this disclosure, in response to the user intent being classified as report interpretation, a second language model is used to generate interpretation information for the first dialogue information, used to interpret the medical diagnostic result report involved in the first dialogue information. Optionally, if the user's intent may be to request interpretation of the medical diagnostic result report, it may be necessary to utilize a second language model to generate interpretation information for the medical diagnostic results provided in the first dialogue information. In this case, the first dialogue information may contain information about the patient's symptoms, medical test results, and diagnosis. By utilizing a second language model, this information can be input and the model can be asked to generate an interpretation of this information. The model can generate an interpretation report based on its training data and language understanding capabilities to explain the meaning of the medical diagnostic results, possible diagnoses, and recommendations to the user.

[0142] As described above, the intelligent medical consultation method of this disclosure can determine the disease diagnosis result corresponding to the user's input information using various intelligent agents based on a large language model. These intelligent agents can progressively refine the disease diagnosis information used to determine the disease diagnosis result based on the user's interaction information with each input, and concretely display the disease diagnosis information, such as the aforementioned suspected disease probability distribution map, to the user. Optionally, the aforementioned suspected disease probability distribution map can include predictions of the possible category of the current human-computer dialogue and its associated probability. By extracting feedback information from multiple rounds of dialogue, the suspected disease probability distribution map can be progressively refined to present a more accurate probability distribution map to the user.

[0143] Optionally, for each round of dialogue interaction after the first round, if it is detected that the user's input information deviates from the current scenario category, such as inputting information related to legal consultation in a medical consultation scenario, or if the user's input information is not a valid response to the previous medical consultation information, such as inputting a response information that is unrelated to the previous medical consultation information but related to the medical consultation scenario, then the suspected disease probability distribution map will not be updated for this round of dialogue interaction.

[0144] Optionally, for the above-mentioned consultation scenario, category predictions of different granularities (e.g., including disease categories and specific diseases) can be generated to cover all possible illnesses of the user, while taking into account both physiological (normal) and pathological factors. Furthermore, based on the user's disease category and the severity of current symptoms, further prompts can be provided to the user, such as tags like "acute / critical," to help the user obtain key information more intuitively.

[0145] The following will refer to Figures 6A-6C An example implementation of the intelligent medical consultation method disclosed herein is provided.

[0146] Figure 6A This is a schematic diagram illustrating the implementation process of an intelligent medical consultation method according to an embodiment of the present disclosure. Figure 6B This is a schematic diagram illustrating another implementation process of the intelligent medical consultation method according to an embodiment of the present disclosure. Figure 6C This is a schematic diagram illustrating the agent architecture of an intelligent medical consultation method according to an embodiment of the present disclosure.

[0147] like Figure 6A As shown, for each user input in a dialogue interaction (first dialogue information), the first language model can be used to classify the scene through intent recognition.

[0148] If the current human-computer dialogue scenario is determined to be a medical consultation scenario, the guidance decision-making agent based on the second largest language model can be invoked to execute the above guidance decision.

[0149] Optionally, the guidance decision-making agent can perform the above-mentioned guidance decision by classifying the level of medical visits. For example, the guidance decision-making agent may have the following characteristics: Figure 6C The multi-agent architecture shown can be a combination of an LLM and rules, with the goal of completing its own task (i.e., a sub-goal). Hierarchical relationships can exist between agents; for example, a pediatrician could be a subordinate agent of the head of a differential diagnosis expert group, or a subordinate agent of the head of the inquiry department. Each subordinate agent can serve as a resource that a superior agent can access, and each agent can choose to perform tasks by itself or by accessing its resources. Therefore, in a consultation scenario, the user's disease diagnosis can be completed through the collaborative work of various agents between the user and the application acting as their medical assistant. Thus, in the classification of consultation levels, the voting results of various agents regarding consultation levels (e.g., ...) can be combined. Figure 6A The conclusions of the medical service level classification section (such as "continue asking questions" or "diagnosis-home observation") determine the medical service level classification result. This classification result can correspond to the results of the aforementioned guided decision-making. For example, a medical service level of 0 can represent "continue asking questions," while a medical service level of 1 can determine the category, i.e., output a diagnosis result. Furthermore, medical service levels can be further subdivided based on the clarity of the diagnosis or the severity of the disease (because different diseases may correspond to different medical certifications) to distinguish multiple medical service levels. For example, in the case of dividing into three medical service levels, medical service level 1 can indicate a high probability of diagnosing a certain disease, recommending immediate medical attention; medical service level 2 can indicate that a certain disease cannot be ruled out, requiring regular follow-up visits; and medical service level 3 can indicate mild symptoms, uncertainty about whether it is a certain disease, recommending home observation, etc. Of course, the above medical service level classification method is only used as an example and not a limitation in this disclosure.

[0150] Optionally, if the guiding decision-making agent determines to continue the consultation based on the level of medical visits, a clinical probabilistic reasoning agent based on the second largest language model can be invoked to perform subsequent recall of the disease list, probability assignment, and display of the disease probability list. For example... Figure 6A As shown, when the guiding decision-making agent utilizes the multi-agent architecture described above to make a guiding decision on continuing consultation (continuing to ask questions) by combining the voting results of various agents regarding the level of medical attention, the disease recall agent and the disease probability inference agent in the clinical probability inference agent can be used to recall the disease list and perform probability assignment, respectively. Specifically, the clinical probability inference agent can have the aforementioned multi-agent architecture to perform voting processing, so as to synthesize the probability voting results of various agents to generate a comprehensive probability distribution of suspected diseases. Furthermore, based on the probability assignment results, the disease probability distribution display agent can generate and display a probability distribution map of the suspected diseases to the user, such as... Figure 6A As shown.

[0151] Optionally, if the guiding decision-making agent determines to continue the consultation based on the level of medical care, such as Figure 6A As shown, the intelligent agent can also utilize consultation points to perform the recall of consultation point lists and selection of optimal consultation points as described above for consultation scenarios. Based on the selected optimal consultation point, it can output new consultation information to the user. Furthermore, it can explain to the user the reasons for outputting this consultation information and the list of suspected disease probabilities through interpretive language, such as... Figure 4 and Figure 5 The text above the list of category probabilities is shown in the table. When the guiding decision-making agent determines to make a diagnosis (i.e., provide a diagnosis result), it can output the diagnosis result information to the user.

[0152] Optionally, such as Figure 6A As shown, for the consultation scenario, the clinical probabilistic reasoning agent and consultation point determination agent disclosed herein can have a multi-agent architecture similar to the guidance decision agent to perform voting processing, so as to make a comprehensive and accurate judgment by integrating the structured knowledge of each agent.

[0153] Figure 6B Another implementation process of the intelligent medical consultation method disclosed herein is presented. In this method, a unified, trained large-scale language model is used for each processing part, for the same human-computer dialogue and user input.

[0154] As described above, the intelligent medical consultation method of this disclosure determines the scenario classification of the human-computer dialogue upon receiving the initial input interaction information from the user. Based on the initial input interaction information, it generates and displays a list of category probabilities and associated questions related to the scenario classification to the user in a graphical form. Furthermore, upon receiving subsequent input interaction information from the user regarding the associated questions, it updates the category probability list based on this interaction information and displays it to the user, enabling the user to quickly access the dynamically updated category probability list. Through multi-turn dialogue, it achieves accurate classification of the human-computer dialogue, thereby significantly improving the user experience. Specifically, for consultation scenarios, the intelligent medical consultation method of this disclosure can provide a medical assistant simulating a doctor's consultation based on large-scale language model underlying technology. This medical assistant can predict diseases based on the symptoms presented by the user and display multiple related diseases and their probabilities graphically, facilitating the user's quick access to key information and significantly improving the user experience. It can also highly simulate a doctor's consultation scenario, guiding the user to continuously provide disease-related information through multi-turn dialogue, updating the probability distribution of related diseases, and ultimately outputting a disease diagnosis result for the user's reference.

[0155] As described above, the intelligent medical consultation method of this disclosure is based on the interaction information received from the user during human-computer dialogue. It classifies the user's intent by invoking a large language model. If the user's intent is determined to be a medical consultation, it invokes different agents based on another large language model to recall disease probabilities and optimal consultation points. This generates and displays the corresponding disease probability distribution and consultation information to the user in graphical form, along with user information needed for further disease diagnosis. The method provided by the embodiments of this disclosure enables flexible and efficient human-computer dialogue interaction using different models of varying sizes. This allows users to quickly obtain dynamically updated probability distribution maps of suspected diseases and achieve accurate disease diagnosis through multi-turn dialogue, thereby significantly improving the user experience. Specifically, for consultation scenarios, the intelligent medical consultation method disclosed herein can provide a medical assistant that simulates a doctor's consultation based on the underlying technology of a large language model. This medical assistant can predict diseases based on the symptoms proposed by the user and display multiple related diseases and their probabilities in a graphical form, making it easy for users to quickly obtain key information and greatly improving the user experience. It can also highly simulate a doctor's consultation scenario, guide users to continuously provide feedback on disease-related information through multiple rounds of dialogue, update the probability distribution of related diseases, and finally output disease diagnosis results for the user's reference.

[0156] Figure 7 This is a schematic diagram illustrating an intelligent medical consultation device 700 according to an embodiment of the present disclosure.

[0157] According to embodiments of this disclosure, the intelligent medical consultation device 700 may include an intent classification module 701, a guidance decision module 702, a continued consultation module 703, and an information output module 704.

[0158] The intent classification module 701 can be configured to, in response to first dialogue information input by the user in a human-computer dialogue, classify the user's intent based on the first dialogue information using a first language model. Optionally, the intent classification module 701 can perform the operations described above with reference to step S201.

[0159] Optionally, at the beginning of each round of dialogue interaction, the user's intent can be identified based on the interaction information entered by the user in the dialogue interface of the human-computer dialogue application to determine the scenario classification of the current dialogue interaction, and to execute the corresponding subsequent operations in this round of dialogue interaction. That is, the domain and task scenario of the current human-computer dialogue can be determined, such as a consultation scenario or question-and-answer scenario in the medical field, or a legal consultation scenario in the legal field.

[0160] Optionally, a pre-trained large language model (the first large language model) can be used to identify intent from the user's input interaction information, such as determining the scenario classification of the current dialogue interaction, and then performing subsequent processing corresponding to that scenario based on the interaction information in the current dialogue interaction. For example, Figure 4 An example human-computer dialogue is shown in the dialogue interface of a human-computer dialogue application. When a user enters the interactive information "What's wrong with my sudden headache?" in the dialogue interface, the first language model can be used to perform intent recognition on the input to determine the current human-computer dialogue scenario category, such as a medical consultation scenario. Then, based on the input, subsequent dialogue is conducted for the consultation scenario.

[0161] Optionally, the primary language model can be a supervised, fine-tuned language model that has been customized based on data from the intent recognition task, thereby improving its performance and effectiveness on the intent recognition task. With supervised fine-tuning, this primary language model can better perform intent recognition and scene classification based on user input interaction information.

[0162] The guidance decision module 702 can be configured to, in response to the user's intent to classify the inquiry as a medical consultation, invoke a guidance decision agent based on a second major language model to determine whether to continue the consultation. Optionally, the guidance decision module 702 can perform the operations described above with reference to step S202.

[0163] Optionally, for each round of dialogue interaction, the execution of guided decisions may include determining whether to output further questions to the user (i.e., continue the consultation) or to complete the current disease diagnosis to output the disease diagnosis result (i.e., make a diagnosis).

[0164] As an example, if a dialogue system needs to gather more information to complete the current disease diagnosis, it can ask the user further questions. For instance, in a medical consultation scenario, if the dialogue system is diagnosing a user's medical condition, it may need to ask further questions about the user's symptoms, medical history, and lifestyle. Conversely, if the dialogue system has already gathered enough information to complete the current disease diagnosis, it can output the diagnosis result. For example, if the dialogue system is diagnosing a user's medical condition, it can make a diagnosis and output the diagnosis result.

[0165] On the one hand, when the collected information is insufficient, the dialogue system can make a guiding decision to continue the consultation, in order to request further information from the user by outputting new consultation information.

[0166] Optionally, the intelligent medical consultation device of this disclosure can determine the disease diagnosis result corresponding to the user's input information using various intelligent agents based on a large language model. These intelligent agents can progressively refine the disease diagnosis information used to determine the diagnosis result based on each user's interactive input, and then concretely display the disease diagnosis information to the user. Optionally, these intelligent agents can be constructed based on a second large language model. According to embodiments of this disclosure, the first large language model may be different from the second large language model, and the first large language model is smaller in scale than the second large language model. Optionally, the first large language model used for intent recognition may be smaller in scale than the second large language model used to construct the various intelligent agents.

[0167] Optionally, the aforementioned guidance decision-making can be executed using a guidance decision-making agent based on a second major language model. That is, after determining that the current dialogue interaction scenario is a medical consultation scenario based on the acquired initial dialogue information, the guidance decision-making agent can be directly invoked to execute the guidance decision. For example, the guidance decision-making agent itself can have the ability to execute guidance decisions, simulating a doctor's actual consultation to make decisions about the next action.

[0168] The continued consultation module 703 can be configured to, in response to the guiding decision agent's determination to continue consultation, invoke a clinical probabilistic reasoning agent based on a second major language model to recall at least two suspected diseases and their probability distributions, and invoke a consultation point based on the second major language model to determine the optimal consultation point for the agent to recall. Optionally, the continued consultation module 703 can perform the operations described above with reference to step S203.

[0169] Optionally, based on each user's interaction with the consultation information in the dialogue interface, a clinical probabilistic reasoning agent can update the probability distribution map of suspected diseases displayed to the user based on each input. Updating the probability distribution map can include introducing new categories, deleting previous categories, and changing the probabilities of each category. For example, in a consultation scenario, the dialogue system (or application) can adjust the disease probability distribution based on the user's responses. This adjustment can include adjusting the probabilities of previously inferred diseases, introducing new diseases, or deleting previous diseases.

[0170] For example, after acquiring new interaction information from the user, a clinical probabilistic reasoning agent can be used to refine and update the previous predictions of the suspected disease probability distribution (i.e., the probability distribution of diagnosis for each of at least two suspected diseases related to the human-computer dialogue) based on the new interaction information, i.e., recalling the latest suspected disease probability distribution. Optionally, a disease recall agent can be invoked to recall suspected diseases related to the current human-computer dialogue, and a disease probabilistic reasoning agent can be invoked to determine the probability of diagnosis for each of the recalled suspected diseases, thereby generating an updated suspected disease probability distribution.

[0171] After recalling the latest probability distribution of suspected diseases, further consultation is needed to determine consultation information for the next round of dialogue. Optionally, if the guiding decision is to continue consultation, multiple consultation points can be recalled by calling a consultation point determination agent based on the second largest language model, and the optimal consultation point can be selected from them.

[0172] Optionally, the consultation point recall agent can recall multiple consultation points related to the context of the current human-computer dialogue based on its prior knowledge. These consultation points can be used to collect more information to help diagnose the current disease. In this case, it is necessary to recall all possible consultation points and exclude as many category hypotheses as possible from the list of suspected diseases.

[0173] Optionally, based on the selected optimal consultation point, consultation information corresponding to the optimal consultation point (e.g., "When did the headache start?") can be generated by calling a consultation agent based on the second largest language model, and the consultation information can be output to the user to execute the next round of dialogue interaction.

[0174] The information output module 704 can be configured to, based on the at least two suspected diseases and their probability distributions, and the optimal consultation point, invoke a consultation agent based on a second language model to generate second dialogue information, and output the second dialogue information in the human-computer dialogue. The second dialogue information includes a suspected disease probability distribution map and first consultation information. The suspected disease probability distribution map presents the at least two suspected diseases and their probability distributions in a graphical form. The first consultation information is used to collect user information required for disease diagnosis. Optionally, the information output module 704 can perform the operations described above with reference to step S204.

[0175] Optionally, if the guiding decision is to continue the consultation, by recalling the probability distribution of suspected diseases and the optimal consultation point, the consultation agent based on the second largest language model can be invoked to generate and output the second dialogue information. The second dialogue information may include the probability distribution of suspected diseases (i.e., the probability distribution map of suspected diseases) displayed to the user in the form of a chart, as well as further consultation information (i.e., the first consultation information).

[0176] Optionally, a disease probability distribution display agent can be invoked to generate and display a suspected disease probability distribution map to the user based on the aforementioned suspected disease probability distribution.

[0177] Optionally, the first consultation information can be generated based on the recalled optimal consultation point by calling the consultation point display agent, and then displayed to the user.

[0178] Optionally, the intelligent medical consultation device disclosed herein can display the latest probability distribution map of suspected diseases to users in the form of a chart, allowing users to intuitively understand the diseases they may currently have and their probabilities. Furthermore, the intelligent medical consultation device disclosed herein can also employ other methods to display the probability distribution map of suspected diseases to users.

[0179] In addition, the intelligent medical consultation device disclosed herein may also include modules for implementing other steps in the above-described intelligent medical consultation method.

[0180] For example, on the other hand, given sufficient information, the guiding decision-making agent can determine to make a diagnosis, and in this case, can provide the user with the diagnostic result information. For the final round of dialogue interaction, instead of outputting the dialogue information including the updated suspected disease probability distribution map and further consultation information, the disease diagnosis result information can be output. That is, if the guiding decision-making agent can determine to output diagnostic result information based on the latest first dialogue information input by the user, the diagnostic result information can be output in the dialogue interface without outputting the new suspected disease probability distribution map and further consultation information.

[0181] Optionally, if a disease diagnosis can be made based on the current dialogue information, a final disease diagnosis result can be output to the user. This diagnosis result may include, for example, information indicating the disease category. Optionally, the diagnosis result information can be generated using a second language model based on the probability distribution of recalled suspected diseases. As an example, the dialogue system's judgment on the category to which the user's query belongs can be output to the user; for example, if the dialogue system is diagnosing the user's medical condition, it can output the diagnosis result "You may have migraines."

[0182] Optionally, in addition to the disease diagnosis results, the dialogue system can also provide other indicative information about the disease. For example, the system can provide information on the current disease's treatment level, symptoms of migraines, treatment options and / or preventative measures, and reassurance information for the user. For instance, the current disease treatment level provided to the user can include the three levels described above, providing information on the severity of the disease for reference. For example, treatment level 1 might indicate a high probability of a diagnosis and recommend immediate medical attention; treatment level 2 might indicate that a certain disease cannot be ruled out and regular follow-up visits are necessary; and treatment level 3 might indicate mild symptoms, uncertainty about whether it is a disease, and recommend home observation. Furthermore, the diagnosis results can also present guidance information about the current disease, such as guidance on the disease, including its causes, symptoms, diagnostic methods, treatment options, and prognosis, to help patients better understand their health problems so they can make more informed decisions and take appropriate actions to manage and treat the disease. For example, treatment suggestions can be provided to users within the diagnostic results information. These suggestions can be tailored to the specific circumstances of the illness. Common treatment suggestions may include medication, surgery, physical therapy, and psychotherapy. Specifically, for example, for migraines, treatment suggestions may include avoiding triggering factors, medication, and relaxation techniques. Furthermore, reassuring information can be provided to users within the diagnostic results information, such as words of emotional support, encouragement, and comfort, aimed at helping patients maintain a calm and optimistic attitude when facing illness. Further, the dialogue system can also invoke relevant neural network models to provide suggestions on the best next steps. For example, if the dialogue system diagnoses a user with the flu, the model may suggest that the user seek medical attention or rest and drink plenty of water.

[0183] By providing a final diagnosis when a disease can be diagnosed, a clear answer and useful information can be given to the user, which can help the user understand their situation and take appropriate action.

[0184] Optionally, if the scenario classification of this round of dialogue interaction is determined to be unrelated to a medical consultation scenario, such as when the scenario classification of this round of dialogue interaction is determined to be a medical question-and-answer scenario, the system can output answer information to the user's inquiry, that is, conduct human-computer dialogue interaction in the form of question-and-answer replies. For example, if the user's input information in this round of dialogue interaction is detected as "Can ibuprofen relieve headaches?", the system can output a reply to the user regarding this information, such as "Yes, ibuprofen is an over-the-counter drug, usually used to relieve mild to moderate pain, such as headaches, joint pain, or toothaches."

[0185] Optionally, if the user's intent might be to request an interpretation of a medical diagnostic report, a second language model might be needed to generate an interpretation of the medical diagnostic results provided in the first dialogue. In this case, the first dialogue might contain information about the patient's symptoms, medical test results, and diagnosis. By utilizing a second language model, this information can be input, and the model can be asked to generate an interpretation of this information. Based on its training data and language understanding capabilities, the model can generate an interpretation report to explain the meaning of the medical diagnostic results, possible diagnoses, and recommendations to the user.

[0186] According to another aspect of this disclosure, an intelligent medical consultation device is also provided. Figure 8 A schematic diagram of an intelligent medical consultation device 2000 according to an embodiment of the present disclosure is shown.

[0187] like Figure 8 As shown, the intelligent medical consultation device 2000 may include one or more processors 2010 and one or more memories 2020. The memories 2020 store computer-readable code, which, when executed by the one or more processors 2010, can perform the intelligent medical consultation method as described above.

[0188] The processor in the embodiments of this disclosure can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor, and can be based on an x86 architecture or an ARM architecture.

[0189] In general, the various exemplary embodiments of this disclosure can be implemented in hardware or dedicated circuitry, software, firmware, logic, or any combination thereof. Some aspects can be implemented in hardware, while others can be implemented in firmware or software that can be executed by a controller, microprocessor, or other computing device. When aspects of embodiments of this disclosure are illustrated or described as block diagrams, flowcharts, or using some other graphical representation, it will be understood that the blocks, apparatuses, systems, techniques, or methods described herein can be implemented as non-limiting examples in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0190] For example, the method or apparatus according to embodiments of this disclosure can also be used by means of Figure 9 The architecture of the computing device 3000 shown is used for implementation. For example... Figure 9 As shown, the computing device 3000 may include a bus 3010, one or more CPUs 3020, a read-only memory (ROM) 3030, a random access memory (RAM) 3040, a communication port 3050 connected to a network, an input / output component 3060, a hard disk 3070, etc. The storage devices in the computing device 3000, such as the ROM 3030 or the hard disk 3070, may store various data or files used for processing and / or communication of the intelligent medical consultation method provided in this disclosure, as well as program instructions executed by the CPU. The computing device 3000 may also include a user interface 3080. Of course, Figure 9 The architecture shown is merely exemplary and can be omitted as needed when implementing different devices. Figure 9 One or more components in the computing device shown.

[0191] According to another aspect of this disclosure, a computer-readable storage medium is also provided. The computer storage medium stores computer-readable instructions. When the computer-readable instructions are executed by a processor, the intelligent medical consultation method according to embodiments of this disclosure, as described with reference to the above-drawn figures, can be performed. The computer-readable storage medium in the embodiments of this disclosure may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory may be random access memory (RAM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct memory bus random access memory (DR RAM). It should be noted that the memory used in the methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0192] Embodiments of this disclosure also provide a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform an intelligent medical consultation method according to embodiments of this disclosure.

[0193] Embodiments of this disclosure provide an intelligent medical consultation method, apparatus, device, and computer-readable storage medium.

[0194] The method provided in the disclosed embodiments is based on user interaction information received during human-computer dialogue. It classifies user intent by invoking a large language model. If the user intent is determined to be a medical consultation, it recalls disease probabilities and optimal consultation points by invoking different agents based on another large language model. This generates and displays the corresponding disease probability distribution and consultation information to the user in graphical form, along with user information needed for further diagnosis. The method provided in this disclosure enables users to quickly obtain dynamically updated probabilities of suspected diseases and achieve accurate disease diagnosis through multi-turn dialogue, significantly improving the user experience. Furthermore, decomposing the large language model task using multiple agents facilitates individual optimization of each sub-task, greatly improving training efficiency and effectiveness. Combining models of different sizes also optimizes computational power allocation and reduces model training costs.

[0195] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing at least one executable instruction for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0196] In general, the various exemplary embodiments of this disclosure can be implemented in hardware or dedicated circuitry, software, firmware, logic, or any combination thereof. Some aspects can be implemented in hardware, while others can be implemented in firmware or software that can be executed by a controller, microprocessor, or other computing device. When aspects of embodiments of this disclosure are illustrated or described as block diagrams, flowcharts, or using some other graphical representation, it will be understood that the blocks, apparatuses, systems, techniques, or methods described herein can be implemented as non-limiting examples in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0197] The exemplary embodiments of this disclosure described in detail above are merely illustrative and not restrictive. Those skilled in the art will understand that various modifications and combinations can be made to these embodiments or their features without departing from the principles and spirit of this disclosure, and such modifications should fall within the scope of this disclosure.< / probabilities> < / probabilities> < / question> < / explanation> < / probabilities>

Claims

1. An intelligent medical consultation method, applied to a human-computer dialogue application, the method comprising: In response to the first dialogue information input by the user in the human-computer dialogue, the user's intent is classified based on the first dialogue information using the first language model; In response to the user's intent being classified as a medical consultation, a guiding decision-making agent based on the second largest language model is invoked to determine whether to continue the consultation. In response to the guidance decision-making agent's determination to continue the consultation, a clinical probabilistic reasoning agent based on the second largest language model is invoked to recall at least two suspected diseases and the probability distribution of the at least two suspected diseases, and a consultation point based on the second largest language model is invoked to determine the optimal consultation point for the agent to recall. Based on the at least two suspected diseases and their probability distributions, as well as the optimal consultation point, a consultation agent based on a second language model is invoked to generate second dialogue information, and the second dialogue information is output in the human-computer dialogue. The second dialogue information includes a probability distribution map of suspected diseases and first consultation information. The probability distribution map of suspected diseases presents the at least two suspected diseases and their probability distributions in the form of a chart. The first consultation information is used to collect user information required for disease diagnosis.

2. The method as described in claim 1, wherein, The first large language model is different from the second large language model, and the first large language model is smaller in scale than the second large language model.

3. The method as described in claim 1, wherein, The clinical probabilistic reasoning agent includes a disease recall agent and a disease probabilistic reasoning agent, wherein the disease recall agent is configured to recall at least two suspected diseases related to the human-computer dialogue based on the human-computer dialogue including the first dialogue information, and the disease probabilistic reasoning agent is configured to recall the probability of each of the at least two suspected diseases being diagnosed based on the human-computer dialogue.

4. The method of claim 1, wherein, The consultation point determination agent includes a consultation point recall agent and a consultation point screening agent, wherein the consultation point recall agent is configured to recall at least two consultation points based on the at least two suspected diseases, and the consultation point screening agent is configured to select the optimal consultation point from the at least two consultation points based on the at least two suspected diseases and the human-computer dialogue.

5. The method of claim 1, wherein, The consultation intelligence agent includes a disease probability distribution display intelligence agent and a consultation point display intelligence agent, wherein the disease probability distribution display intelligence agent is configured to generate and display the probability distribution map of the suspected diseases to the user based on the at least two suspected diseases and the probability distribution of the at least two suspected diseases, and the consultation point display intelligence agent is configured to generate and display the first consultation information to the user based on the optimal consultation point.

6. The method according to any one of claims 3-5, wherein, The disease recall agent and the consultation point recall agent are obtained through reinforcement learning based on the second language model; The guided decision-making agent, the disease probability reasoning agent, the consultation point screening agent, the disease probability distribution display agent, and the consultation point display agent are obtained through prompting engineering based on the second language model.

7. The method of claim 5, wherein, The consultation point determination agent is also configured to generate explanatory information for the optimal consultation point based on the doctor's thought process, and the consultation point display agent is also configured to display the explanatory information for the optimal consultation point to the user, the explanatory information being used to explain the reason for selecting the optimal consultation point.

8. The method of claim 1, further comprising: In response to the user's reply to the first consultation information input in the human-computer dialogue, the reply information is received as the first dialogue information.

9. The method of claim 1, wherein, In response to the determination of the given diagnostic result by the guiding decision-making agent, diagnostic result information is generated based on the probability distribution of the at least two suspected diseases using a second major language model.

10. The method of claim 9, wherein, The diagnostic results information includes disease predictions; The diagnostic results information also includes one or more of the following: The information includes the level of medical care, guidance information for the described disease, treatment suggestions, and reassurance information.

11. The method of claim 1, wherein, In response to the user intent being categorized as question-and-answer, a second language model is used to generate answer information for the first dialogue information, which is used to answer the questions involved in the first dialogue information.

12. The method of claim 1, wherein, In response to the user intent being categorized as report interpretation, a second language model is used to generate interpretation information for the first dialogue information, which is used to interpret the medical diagnostic result report involved in the first dialogue information.

13. The method of claim 1, wherein, The first dialogue information includes one or more of the following: text information, diagnostic result images, and voice information.

14. An intelligent medical consultation device, applied to a human-computer dialogue application, the device comprising: The intent classification module is configured to respond to the first dialogue information input by the user in the human-computer dialogue, and to classify the user's intent based on the first dialogue information using the first language model. The guidance decision module is configured to respond to the user's intent by classifying it as a medical consultation and to invoke a guidance decision agent based on the second largest language model to determine whether to continue the consultation. The continued consultation module is configured to respond to the guidance decision agent's determination to continue consultation by calling a clinical probabilistic reasoning agent based on the second largest language model to recall at least two suspected diseases and the probability distribution of the at least two suspected diseases, and calling a consultation point based on the second largest language model to determine the agent to recall the optimal consultation point. The information output module is configured to, based on the at least two suspected diseases and their probability distributions, and the optimal consultation point, invoke a consultation agent based on a second language model to generate second dialogue information, and output the second dialogue information in the human-computer dialogue. The second dialogue information includes a probability distribution map of suspected diseases and first consultation information. The probability distribution map of suspected diseases presents the at least two suspected diseases and their probability distributions in the form of a chart. The first consultation information is used to collect user information required for disease diagnosis.

15. An intelligent medical consultation device, comprising: One or more processors; as well as One or more memories storing a computer-executable program, which, when executed by the processor, performs the method of any one of claims 1-13.

16. A computer-readable storage medium having stored thereon computer-executable instructions, which, when executed by a processor, are used to implement the method of any one of claims 1-13.

Images