Intelligent clinical teaching system and method

Abstract

The invention provides an intelligent clinical teaching system and method, which are used for responding to query statements of multiple rounds of questions and answers by an intelligent agent simulating a patient through a virtual patient, obtaining corresponding answer statements based on a personalized knowledge sub-graph corresponding to the virtual patient, and feeding back the answer statements to a learner, each round of questions and answers corresponding to an inquiry node of an inquiry path. The intelligent agent not only can support complex semantic generation and teaching reasoning and realize that each round of questions and answers accord with a medical logic relationship, but also can realize accurate control of multiple rounds of questions and answers on an inquiry path, so that the intelligent degree of clinical teaching is improved, and the teaching, research and development requirements of medical knowledge can be better met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to an intelligent clinical teaching system and method. BACKGROUND

[0002] Due to the complexity of medical knowledge system, the traditional clinical teaching mode mainly adopts the professor form between teachers and students, and such teaching form is limited by the lack of teachers and is difficult to meet the training of large-scale clinical medical talents.

[0003] With the development of computer technology, a virtual patient system emerges as the times require. The virtual patient simulates a patient and interacts with a learner to realize the consolidation learning of medical knowledge and the clinical simulation practice. However, the current virtual patient system generally has the problem of insufficient intelligence, for example: the virtual patient system has insufficient modeling ability of medical knowledge structure and is difficult to support complex semantic generation and teaching reasoning; the context adaptation and interaction control ability based on the language model is limited and is difficult to support multi-round teaching interaction in line with the inquiry path. SUMMARY

[0004] The present application proposes an intelligent clinical teaching system and method, which can solve one of the problems in the background art.

[0005] To achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0006] In a first aspect, a clinical teaching method is provided, the clinical teaching method comprising:

[0007] In response to a current query sentence of a learner, a pre-trained intelligent agent for simulating a patient by a virtual patient returns a current answer sentence for the current query sentence to the learner; and

[0008] In response to a next query sentence of a learner, the intelligent agent returns a next answer sentence for the next query sentence to the learner,

[0009] The intelligent agent is configured to perform semantic analysis on a query sentence, determine an inquiry node corresponding to a query element on an inquiry path, query an answer element corresponding to the query element from a personalized knowledge subgraph corresponding to the virtual patient, generate an answer sentence on the inquiry node based on the answer element, the query element and the answer element meet a medical logical relationship, and the personalized knowledge subgraph is constructed based on a medical knowledge graph.

[0010] Based on the above technical scheme, the intelligent agent for imitating the patient with the virtual patient responds to the query statement of the multi-round question and answer, obtains the corresponding answer statement based on the personalized knowledge subgraph corresponding to the virtual patient, and feeds back to the learner, and each round of question and answer corresponds to the inquiry node of the inquiry path. The intelligent agent can not only support complex semantic generation and teaching reasoning, implement each round of question and answer in accordance with the medical logical relationship, but also can realize accurate control of multi-round question and answer on the inquiry path, improve the intelligent degree of clinical teaching, and better meet the development requirements of medical knowledge teaching and research.

[0011] In a possible design of the first aspect, the intelligent agent comprises, which are sequentially arranged:

[0012] The input embedding module is configured to convert the query statement and the position encoding representing the inquiry node into a time-dependent block input representation.

[0013] The encoder module is configured to query the answer element from the corpus, and

[0014] The decoder module is configured to generate the answer statement based on the answer element, a confidence variable and / or a teaching control variable. The confidence variable is used to call a fuzzy answer as the answer statement when a confidence requirement is met. The teaching control variable includes a teaching goal, a tone style and / or an expression preference.

[0015] Based on the above technical scheme, through the design of the confidence variable, the fuzzy answer of the patient when answering uncertain questions can be realized, so that the virtual patient increases the design of the simulated answer and is closer to the real inquiry situation.

[0016] In a possible design of the first aspect, the encoder module comprises, which are sequentially arranged: a self-attention layer, a first normalization layer, a first feedforward network and a second normalization layer. The output of the input embedding module is connected to the input of the self-attention layer. The output of the input embedding module and the input of the first normalization layer, the output of the first normalization layer and the input of the second normalization layer are connected through a residual connection.

[0017] The decoder module comprises, which are sequentially arranged: a masking attention layer, a third normalization layer, a multi-head attention layer, a fourth normalization layer, a second feedforward network and a fifth normalization layer. The output of the input embedding module is connected to the input of the masking attention layer. The output of the second normalization layer is connected to the input of the multi-head attention layer. The output of the input embedding module and the input of the third normalization layer, the output of the third normalization layer and the input of the fourth normalization layer, the output of the fourth normalization layer and the input of the fifth normalization layer are connected through a residual connection.

[0018] In a possible design of the first aspect, the intelligent agent further includes:

[0019] a graph neural network configured to obtain at least one diagnosis result of a category disease based on the virtual patient knowledge subgraph, with the symptoms, the fuzzy membership degree, and the node confidence as inputs.

[0020] According to the technical solution, the fuzzy membership degree and the node confidence are introduced for double modeling, so that the fuzziness and subjective difference in describing the symptoms by different patients can be reflected, and the virtual patient is closer to the real consultation situation; and the situation of multiple diseases coexisting is supported.

[0021] In a possible design of the first aspect, the graph neural network includes, in sequence, an input layer, a feature fusion layer, a first activation function layer, a regularization layer, and a classification output layer, and the attention mechanism is added before the first activation function layer.

[0022] In a possible design of the first aspect, the clinical teaching method further includes:

[0023] The medical knowledge graph is pruned according to the relationship confidence and the graph structure entropy.

[0024] In a possible design of the first aspect, the intelligent agent further includes:

[0025] an emotion recognition module configured to generate a multi-distribution emotion vector based on the answer sentence;

[0026] a facial expression driving module configured to implement mapping from the emotion vector to facial expressions, with the output of the emotion recognition module as input; and

[0027] a posture behavior library selection module configured to determine a current action and an execution time thereof based on a context dialogue state and a performed action through temporal logic, with the output of the emotion recognition module, the symptoms, and the temporal encoding as input.

[0028] According to the technical solution, multi-modal consistency modeling of language-expression-action can be implemented, behavior personification is improved, and the virtual patient is closer to the real patient.

[0029] In a possible design of the first aspect, the emotion recognition module includes, in sequence, a language model and a second activation function layer,

[0030] the facial expression driving module includes, in sequence, a classification mapping layer and a linear layer,

[0031] the posture behavior library selection module includes a third activation function layer.

[0032] In a possible design of the first aspect, the intelligent agent triggers a feedback mechanism dynamically by using a constructed teaching interaction quality evaluation function, and influences the output of the answer statement, facial expression and posture behavior, the teaching interaction quality evaluation function including a question quality score, a strategy rationality score and a path integrity score.

[0033] In a second aspect, an intelligent clinical teaching system is provided, and the intelligent clinical teaching system includes:

[0034] an interaction module, and

[0035] a processing module configured to: in response to a current query statement of a learner, return a current answer statement corresponding to the current query statement to the learner by using a pre-trained intelligent agent that imitates a patient as a virtual patient; and in response to a next query statement of the learner, return a next answer statement corresponding to the next query statement to the learner by using the intelligent agent, the intelligent agent being configured to: perform semantic analysis on the query statement, determine an interrogation node corresponding to a query element on an interrogation path, query an answer element corresponding to the query element from a personalized knowledge subgraph corresponding to the virtual patient, and generate an answer statement on the interrogation node based on the answer element, the query element and the answer element satisfying a medical logic relationship, the personalized knowledge subgraph being constructed based on a medical knowledge graph.

[0036] In a third aspect, an electronic device is provided, and the electronic device includes a processor and a memory coupled to the processor, the memory being configured to store a computer program, and the processor being configured to execute the computer program stored in the memory, so that the electronic device performs the clinical teaching method according to any one of possible implementation manners of the first aspect.

[0037] In a fourth aspect, a computer readable storage medium is provided, and the computer readable storage medium includes a computer program or instructions, and when the computer program or instructions are run on a computer, the computer program or instructions make the computer execute the clinical teaching method according to any one of possible implementation manners of the first aspect.

[0038] In a fifth aspect, a computer program product is provided, and the computer program product includes a computer program or instructions, and when the computer program or instructions are run on a computer, the computer program or instructions make the computer execute the clinical teaching method according to any one of possible implementation manners of the first aspect. BRIEF DESCRIPTION OF DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or related description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0040] Figure 1 is the main flowchart of the clinical teaching method provided by the first embodiment of the present application;

[0041] Figure 2 is one of the implementation manners of the agent structure provided by the first embodiment of the present application;

[0042] Figure 3 is the second implementation manner of the agent structure provided by the first embodiment of the present application based on the first implementation manner;

[0043] Figure 4 is the third implementation manner of the agent structure provided by the first embodiment of the present application based on the second implementation manner;

[0044] Figure 5 is the technical roadmap of the AI virtual patient system for the clinical teaching of the gastroenterology department provided by the second embodiment of the present application;

[0045] Figure 6 is the word cloud diagram of the digestive system disease provided by the second embodiment of the present application;

[0046] Figure 7 is the multi-level causal diagram reasoning architecture provided by the second embodiment of the present application;

[0047] Figure 8 is the neural network reasoning schematic diagram provided by the second embodiment of the present application;

[0048] Figure 9 is the knowledge adaptation and cutting mechanism schematic diagram driven by the teaching scene provided by the second embodiment of the present application;

[0049] Figure 10 is the clinical context generation technical roadmap provided by the second embodiment of the present application;

[0050] Figure 11 is the scene simulation technical roadmap schematic diagram provided by the second embodiment of the present application. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solutions and advantages of the present application more clear, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0052] It should be noted that although the functional modules are divided in the device schematic diagram, and the logical sequence is shown in the flowchart, in some cases, the steps shown or described can be performed in a manner different from the module division in the device or the sequence in the flowchart. The terms "first", "second", etc. in the specification and claims and the above-described drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0053] 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 application belongs. The terminology used herein is for the purpose of describing embodiments of the present application only and is not intended to limit the present application.

[0054] Embodiment one

[0055] As shown in Figure 1 The present embodiment provides a clinical teaching method, which comprises:

[0056] Step S101, in response to a current query sentence of a learner, a pre-trained intelligent agent for imitating a patient with a virtual patient returns a current answer sentence for the current query sentence to the learner; and

[0057] Step S102, in response to a next query sentence of a learner, the intelligent agent returns a next answer sentence for the next query sentence to the learner,

[0058] The intelligent agent is configured to perform semantic analysis on the query sentence, determine an interrogation node corresponding to a query element on an interrogation path, query a corresponding personalized knowledge subgraph of the virtual patient to obtain an answer element corresponding to the query element, generate an answer sentence on the interrogation node based on the answer element, the query element and the answer element meet a medical logical relationship, and the personalized knowledge subgraph is constructed based on a medical knowledge graph.

[0059] Specifically, the above-mentioned clinical teaching method is applicable to a scenario in which a medical knowledge learner interacts with an artificial intelligence (AI) system, the AI system imitates a patient, presents a virtual patient in the form of video and audio, and the learner learns medical knowledge and practices simulation through interaction with the virtual patient presented by the AI system.

[0060] The AI system can not only realize interaction with the learner, but also realize multi-round interaction on different interrogation nodes on the interrogation path.

[0061] The inquiry path can include multiple levels. For example, the inquiry path can include first-level nodes such as symptoms, examinations, diagnoses, treatments, and the like, and second-level nodes can be further set under the first-level nodes. For example, symptoms can be divided into distributed symptoms of different organs, examinations can be divided into examinations at different stages, and treatments can be divided into treatments in different aspects such as traditional Chinese medicine and Western medicine. In actual situations, selective adoption can be made.

[0062] The intelligent agent can be a neural network that can implement all functions, or a collection of different types of neural networks, including but not limited to large language models (LLM), graph neural networks, and multi-layer perceptrons (MLP).

[0063] On the one hand, in order to understand the learner's intention, the intelligent agent is trained to perform semantic analysis, i.e., semantic analysis of the learner's voice or text input, and on the other hand, in order to interact with the learner in a question-and-answer manner, the intelligent agent is trained to find specific answer sentences to respond to the learner's query sentences. This specificity is mainly reflected in two aspects: first, it is necessary to determine whether the learner's query sentence conforms to the arrangement of the inquiry path, and second, it is necessary to find a specific answer sentence to respond to the query sentence, i.e., the query and the answer conform to the scientific medical logic relationship.

[0064] The virtual patient corresponds to a personalized knowledge subgraph, for example, the virtual patient has specific symptom manifestations, specific examination results, specific diagnosis schemes, and specific treatment schemes. The personalized knowledge subgraph is constructed based on a medical knowledge graph and conforms to the scientific medical logic relationship.

[0065] Based on the above technical solutions, the intelligent agent for simulating patients with virtual patients responds to query sentences in multiple rounds of questioning, obtains corresponding answer sentences based on the personalized knowledge subgraph corresponding to the virtual patient, and feeds back to the learner. Each round of questioning corresponds to an inquiry node of the inquiry path, and the intelligent agent can not only support complex semantic generation and teaching reasoning and implement each round of questioning to conform to the medical logic relationship, but also can implement precise control of multiple rounds of questioning on the inquiry path, improve the intelligent degree of clinical teaching, and better meet the development requirements of medical knowledge teaching and research.

[0066] Further, as shown in Figure 2 The intelligent agent includes, in sequence:

[0067] An input embedding module for converting the query sentence and the position encoding representing the inquiry node into a time-dependent block input representation;

[0068] An encoder module for querying the answer elements from the personalized knowledge subgraph, and

[0069] A decoder module is configured to generate the answer statement based on the answer element, a confidence variable, and / or a teaching control variable. The confidence variable is configured to invoke a fuzzy answer as the answer statement when a confidence requirement is satisfied. The teaching control variable includes a teaching goal, a tone style, and / or an expression preference.

[0070] Specifically, the input embedding module can embed the patient text, case paragraph, test image, or contextual question, and add position encoding to form a multi-modal semantic sequence. The module constructs a time-dependent token input representation through an embedding layer and a position encoding layer as the input of the encoder.

[0071] The encoder module is configured to process the relevant medical paragraphs retrieved from the case corpus and encode the semantic content thereof. The encoder is composed of a stack of multiple layers of Transformer structure, each layer including a self-attention layer, a feed-forward network (FFN), a normalization layer (LayerNorm), and a residual connection. The encoder can extract semantic information and causal clues from the case paragraph to provide medical constraints and references for the decoder.

[0072] The decoder module is responsible for generating the target patient answer according to the system instruction, the teaching control variable, and the contextual information. The decoder is also composed of a stack of multiple layers of Transformer modules, each layer including a masked attention layer, a cross-attention layer, a feed-forward network, and a normalization layer. The masked attention mechanism is used to ensure the autoregressive mechanism of the generated sequence, and the cross-attention mechanism aligns the retrieved semantic representation in the encoder with the current decoding state to achieve semantic enhancement and control guidance. Finally, the decoder outputs a patient answer content that meets the clinical context, medical causal rationality, and teaching goal.

[0073] The uncertainty and teaching control mechanism participate in the construction and generation of prompts through teaching (intention) control variables and confidence functions (confidence variables). The system designs a confidence evaluation function based on the entropy value of the language model to dynamically adjust the fuzziness and expression style of the answer content. At the same time, by calculating the dynamic confidence threshold based on individual variables (such as age, education level, cognitive state), the system can control the generation of fuzzy expressions such as "I can't remember" and "I think so". The teaching intention control vector is embedded in the prompt to guide the virtual patient's tone (anxious, calm), expression style (verbose, concise), and other personality parameters, thereby improving the teaching adaptability and interaction authenticity of the system.

[0074] The model structure can also access a text and image combined test report generation module to realize a personalized patient reply with images and text. The system synchronously labels semantic sources of each sentence when outputting the sentence, including metadata information such as case ID, knowledge path, and atlas node, to ensure traceability and explainability of the generated content, and support closed-loop verification and evaluation of medical teaching.

[0075] Based on the above technical solution, through the design of the confidence variable, the fuzzy answer of the patient when answering uncertain questions can be realized, so that the virtual patient increases the design of the simulated answer, and is closer to the real diagnosis situation.

[0076] Further, as shown in Figure 3 The intelligent agent further includes:

[0077] A graph neural network is configured to obtain at least one category disease diagnosis result based on the virtual patient knowledge subgraph, with the symptoms, fuzzy membership degree, and node confidence as inputs.

[0078] Specifically, the graph neural network includes an input layer, a feature fusion layer, a first activation function SotfMax layer, a regularization (Dropout) layer, and a classification output (MLP) layer arranged in sequence, and an attention mechanism (att) is added before the first activation function layer.

[0079] The knowledge graph block is configured to construct a unified medical knowledge representation system that integrates structured graphs and unstructured data. Based on the word cloud of digestive system diseases, high-frequency co-occurring entities are extracted through word frequency statistics and context distribution analysis. The term variants are normalized to standard medical entities through synonym mapping to construct a term alignment table. A four-level disease classification ontology framework (digestive internal medicine / gastrointestinal surgery / pediatric gastroenterology / infectious diseases) is designed to construct an initial concept graph containing multi-dimensional nodes such as disease entities, symptom manifestations, diagnostic indicators, and treatment plans. The patient data is input to generate a specific node graph corresponding to the patient's symptoms, i.e., the above-mentioned personalized knowledge subgraph.

[0080] The uncertainty modeling block includes fuzzy-confidence dual modeling, which can effectively reflect the uncertainty sources of interactive information in the virtual patient system. In view of the term ambiguity, implicit causality, and cross-dependence of multi-level and multi-granularity information such as anatomical structure, physiological mechanism, pathogenesis, and test indicators in digestive system diseases, an uncertainty representation and causal reasoning mechanism is introduced to enhance the model's ability to express and understand complex clinical knowledge. Through input of the virtual patient's description of symptoms, the activation function (Sigmoid function) is used to output the confidence (Conf), reflecting the fuzziness and subjective differences in describing symptoms by different users.

[0081] The neural network inference block includes a feature fusion layer, an activation function (SoftMax) layer, a regularization (Dropout) layer, and a classification output (MLP) layer, and an attention mechanism (Attention) is introduced. The input patient symptoms, fuzzy membership degree, and confidence are subjected to the feature fusion layer to output initial nodes. Then, through the attention mechanism, all nodes receive information from their neighbor nodes and update their own representation. Then, after the activation function layer and the classification output layer, the disease nodes are classified and predicted, and the probability that it is the target diagnosis is output. The diagnosis output result is a multi-class distribution, supporting the situation of multiple diseases coexisting.

[0082] Based on the above technical solution, by introducing fuzzy membership degree and node confidence double modeling, the fuzziness and subjective difference existing when different patients describe symptoms can be reflected, which is closer to the real consultation situation; and the situation of multiple diseases coexisting is supported.

[0083] Further, the clinical teaching method further includes:

[0084] According to the relationship confidence and the graph structure entropy, the medical knowledge graph is pruned, so that the core subgraph can be extracted according to the teaching needs, irrelevant information is pruned, a targeted knowledge closed loop is constructed, and diagnosis thinking training and teaching design are supported.

[0085] Further, as shown in Figure 4 The intelligent agent further includes:

[0086] An emotion recognition module for generating a multi-distribution emotion vector based on the answer statement;

[0087] A facial expression driving module taking the output of the emotion recognition module as input for realizing mapping from the emotion vector to facial expressions; and

[0088] A posture behavior library selection module taking the output of the emotion recognition module, symptoms, and time sequence encoding as input for determining the current action and its execution time based on the context dialogue state and the executed action through time sequence logic.

[0089] The emotion recognition module generates a multi-distribution emotion vector based on the dialogue content generated by the language model through the combination of the RoBERTa module (or BERT module) and the activation function (SoftMax) module.

[0090] The facial expression driving module includes a classification output (MLP) layer and a linear layer. After the input emotion vector is subjected to the mapping function from emotion to facial action unit, a plurality of facial action units are obtained. These facial action units are subjected to a linear combination layer to obtain the comprehensive facial expression of the virtual character, realizing real-time dynamic expression rendering of the virtual character.

[0091] The posture behavior library selection module inputs the emotion vector, symptom information and time sequence code into an activation function (SoftMax) module, determines the timing and duration of action execution based on the context of the dialogue state and the state of the actions performed by the virtual patient through time sequence logic. Finally, the rendering layer is completed by a three-dimensional skeleton binding system or a virtual human driving engine.

[0092] The situational feedback guiding mechanism evaluates the interactive behavior based on the quality of the student's questions, the rationality of the strategy and the completeness of the path. The output score S of the teaching interactive quality evaluation function dynamically triggers the feedback mechanism of the virtual patient, affects the expression output and action selection of the virtual patient, and generates the next round of dialogue of the virtual patient.

[0093] Based on the above technical solutions, multi-modal consistency modeling of language-expression-action can be realized, behavior personification is improved, and the virtual patient is closer to the real patient.

[0094] The embodiment also provides an intelligent clinical teaching system, which comprises:

[0095] An interactive module, and

[0096] A processing module, which returns a current answer statement for a current query statement of a learner to the learner in response to the current query statement by a pre-trained intelligent agent for imitating a patient by a virtual patient, and returns a next answer statement for a next query statement of the learner to the learner in response to the next query statement, the intelligent agent being configured to perform semantic analysis on the query statement, determine an inquiry node corresponding to a query element on an inquiry path, query an answer element corresponding to the query element from a personalized knowledge subgraph corresponding to the virtual patient, generate an answer statement on the inquiry node based on the answer element, the query element and the answer element meeting a medical logical relationship, and the personalized knowledge subgraph being constructed based on a medical knowledge graph.

[0097] The content of the above intelligent clinical teaching system is similar to the clinical teaching method, which will not be repeated here.

[0098] Embodiment two

[0099] This embodiment revolves around the overall technical path of "knowledge reasoning—semantic generation—behavioral simulation," focusing on four core challenges faced by AI virtual patients in medical teaching scenarios: the accuracy of knowledge expression, the rationality of language interaction, the realism of behavior generation, and the usability of system deployment. To this end, this embodiment conducts research in the following three aspects: (i) Addressing the complex heterogeneity of medical knowledge and clinical data, a medical ontology system oriented towards the digestive system is proposed. Combining entity recognition and semantic relation extraction technologies, a structured, multi-layered medical knowledge graph is constructed, and a graph reasoning chain mining and graph-text fusion mechanism is introduced to support knowledge enhancement and semantic generation of the language model; (ii) Addressing semantic preservation in multi-turn question answering within clinical contexts. Figure 5 This project addresses the overall technical roadmap and style control issues of an AI virtual patient system for clinical teaching in gastroenterology. It proposes designing a teaching goal-driven prompt control strategy, constructing a multi-turn context maintenance and context tracking mechanism, and introducing methods for fine-tuning patient language style and tone transfer to enhance the language model's interactive adaptability in medical contexts. (iii) To address the multimodal behavior generation problem of anthropomorphic virtual patients, it proposes constructing a conditional embedding mechanism based on features such as age, gender, and emotion, designing a cross-modal alignment and unified generation network between language, action, and expression, and introducing an emotion-driven speech and facial expression synthesis module to achieve realistic and natural interactive behavior output. Through the collaborative design of the above three technical paths, the project aims to establish an AI virtual patient system with knowledge schedulable, language controllable, and behaviorally believable characteristics, providing a reusable and scalable intelligent interactive platform for clinical teaching. The overall roadmap is as follows: Figure 5 As shown.

[0100] The proposed technical routes include technical routes ①, ② and ③, specifically:

[0101] Technical Approach ① Uncertainty Modeling and Reasoning Enhanced by Hybrid Knowledge Graph and Clinical Data

[0102] Technical approach ①.1 Heterogeneous uncertainty modeling: Integrating structured knowledge and unstructured clinical data to solve heterogeneous alignment and expression differences, and constructing a unified medical knowledge representation system.

[0103] In the process of intelligent modeling of digestive system diseases, we face the challenge of highly heterogeneous structured and unstructured data sources, such as... Figure 6As shown, the knowledge hierarchy of digestive system diseases is complex, with numerous entity types, necessitating the establishment of a unified and integrated representation system to support downstream tasks. Structured data typically originates from existing medical knowledge graphs, the International Classification of Diseases (ICD) coding system, and guidelines, possessing a clear hierarchical structure and conceptual boundaries. Unstructured data, such as electronic medical records (EMR), discharge summaries, surgical reports, image descriptions, and educational texts, often contains a large amount of free text, ambiguous descriptions, and context-dependent medical entities. This semantic difference and inconsistency in data representation severely restricts the feasibility of cross-source knowledge fusion and unified modeling. To address these issues, this technical approach proposes a heterogeneous uncertainty modeling method for digestive system disease knowledge, aiming to construct a unified medical knowledge representation system that integrates structured graphs and unstructured data. This method includes the following key steps:

[0104] (i) Constructing a unified conceptual framework based on medical ontology—Combining existing medical knowledge ontology, a structured conceptual hierarchy for digestive system diseases is established, covering four primary categories: gastroenterology diseases, gastrointestinal surgery diseases, pediatric digestive diseases, and digestive system infectious diseases. Each primary category is further refined into multi-level sub-disease structures, constructing a conceptual map containing multi-dimensional nodes such as disease entities, symptoms, diagnostic indicators, and treatment plans;

[0105] (ii) Integrating word cloud frequency and context distribution to achieve semantic enhancement of terms—Based on a word cloud constructed from gastroenterological diseases, and combined with word frequency statistics and weighting mechanisms, high-frequency co-occurring entities and important pathological concepts in medical texts are identified, such as “ulcer,” “acute abdomen,” and “rotavirus.” Valuable information fragments are extracted from unstructured corpora, and the extraction results are mapped back to the ontology concept framework to solve the problem of diverse terminology (such as the unified processing of “Helicobacter pylori infection” and “HP infection”), and to achieve heterogeneous synonym alignment and entity normalization.

[0106] Technical approach ①.2, Uncertainty representation and causal reasoning mechanism: Introduce confidence and fuzzy attributes to model uncertainty, and combine probabilistic graphs and neural reasoning to realize dynamic context-aware causal reasoning.

[0107] To address the issues of terminological ambiguity, implicit causality, and cross-dependencies inherent in multi-level, multi-granular information such as anatomical structures, physiological mechanisms, pathogenesis pathways, and laboratory indicators in digestive system diseases, traditional general ontology or static relational networks are insufficient for accurate modeling and reasoning. Therefore, it is necessary to introduce uncertainty representation and causal reasoning mechanisms to enhance the model's ability to express and understand complex clinical knowledge. The overall architecture is as follows: Figure 7 As shown, where D represents the disease and S represents the symptom, the specific steps include:

[0108] (i) Employing fuzzy logic to express the degree of fuzziness in symptoms. For example, subjective judgments such as "mild cough" and "moderate fever" can be modeled using fuzzy membership functions to reflect the fuzziness and subjective differences among users when describing symptoms. Simultaneously, confidence is introduced as a measure of the reliability of user input. Confidence considers not only the objective intensity of the symptoms themselves but also the credibility of the user's subjective judgment, modeled using a 3-1 formula:

[0109] Conf(s i )=σ(w T x i +b) (3-1)

[0110] Where σ is the Sigmoid function, x i For the feature representation of symptom input, w T Let b represent the learnable weight vector, b represent the bias term, and s represent the bias term. i The input symptoms are represented by 'Conf', and the confidence level is 'Confidence'. This fuzzy-confidence dual modeling mechanism can effectively reflect the sources of uncertainty in interactive information in a virtual patient system.

[0111] (ii) Based on medical knowledge graphs and data-driven structure learning, a multi-level causal graph is constructed. Nodes in the graph represent symptoms and diseases, respectively, and edges represent possible causal relationships, such as "persistent high fever" → "bacterial infection". This causal graph not only preserves the prior probability P(d) in traditional Bayesian networks, but also... j |s i ), where d j The system represents potential diseases and supports structural learning and dynamic edge weight updates, enabling it to adapt to the heterogeneous performance of different patient groups based on new data. Furthermore, we assign a dynamic weight to each edge in the graph, combined with a confidence score, allowing for greater causal influence on high-confidence symptoms during inference, thereby improving the accuracy and stability of diagnostic results.

[0112] (iii) Graph Neural Network Inference: Through graph neural network inference mechanisms, the system acquires the ability to dynamically perceive context and perform high-order causal modeling. Specific details are as follows... Figure 8 As shown.

[0113] The process mainly consists of three steps: First, input encoding and representation construction. Each node (symptom or disease) in the causal graph needs to have its state vector h initialized. i (0) This is to facilitate the subsequent propagation of graph information. We fuse multi-source information from nodes, including: original symptom / disease embeddings x i Fuzzy membership degree μ i Confidence score Conf(s) i Therefore, the initial node representation is constructed as follows:

[0114] h i (0) =Fuse(x i ,μ i ,Conf(s i (3-2)

[0115] Here, `fuse` represents the fusion feature function. Secondly, a graph neural network propagation mechanism is introduced: in each layer of the GNN, a node receives information from its neighboring nodes and updates its own representation. Since causal relationships have directionality and varying strength, we incorporate an attention mechanism weighted by causal weights and confidence into the message passing mechanism. In each round of propagation, the information received by node i from its neighboring nodes j∈N(i) is represented as:

[0116] m ij =α ij ·Wh j (l) (3-3)

[0117] Where α ij The attention weights are based on causality and confidence, where W represents the learnable message transformation matrix and l represents the information propagation in the l-th round of the graph. Their calculation method is as follows:

[0118] α ij =SoftMax(f att (h i (l) ,h j (l) ,P ij )·Conf(s j (3-4)

[0119] Among them, P ij The causal strength of edge (i,j) in the graph can be derived from expert knowledge or structural learning; Conf(s j ) represents the confidence level of symptom node j, controlling its influence during the propagation process; f att It is a learnable attention function, such as a bidirectional attention mechanism or dot product attention. The node state is updated as follows:

[0120] h j (l) =σ(W s h i (l) +∑ j∈N(i) m ij (3-5)

[0121] Where σ is a nonlinear activation function, W sThis involves node self-updating weights. This process can be repeated L times, allowing the node to gradually perceive a broader causal context. Finally, disease node aggregation and diagnostic prediction: after multiple layers of information propagation, each disease node d... k The representation of This integrates information from all potential symptom nodes and their relationships. We classify and predict disease nodes, outputting the probability that each node represents the target diagnosis:

[0122]

[0123] Among them, W o represents the learnable parameters of the output layer, b represents the prior probability compensation for the disease category, and the diagnostic output is a multi-class distribution, supporting scenarios where multiple diseases coexist. The system can return the top N high-probability candidates, along with diagnostic explanations (such as symptom weights, attention scores, etc.).

[0124] Technical approach ①.3, Teaching scenario-oriented knowledge adaptation and trimming mechanism: Extract core sub-graphs based on teaching needs, trim irrelevant information, construct targeted knowledge loops, and support diagnostic thinking training and instructional design.

[0125] This mechanism starts from "teaching objectives" and constructs a knowledge subgraph trimming process based on the perception of teaching scenarios. The core steps are as follows:

[0126] (i) Modeling of teaching objectives: Formalizing course objectives or teaching tasks (e.g., "training in the differential diagnosis of upper gastrointestinal bleeding") into query conditions, such as... Figure 9 As shown, clearly define the key disease nodes, symptom nodes, and diagnostic pathway types required;

[0127] (ii) Subgraph Extraction: Based on entities and their semantic relationships in the knowledge graph, and combined with graph traversal or graph attention mechanisms (such as GAT), select subgraph structures closely related to the target. For example: retain only symptoms (such as melena, hematemesis), causes (such as gastric ulcer, esophageal varices), and examination items related to "upper gastrointestinal bleeding." Filter out irrelevant diseases (such as respiratory diseases) and their related information;

[0128] (iii) Irrelevant Information Pruning and Edge Denoising: To achieve efficient focusing of the teaching subgraph, this step dynamically prunes low-relevance nodes and edge paths using both relation confidence and graph structure entropy as indicators. The specific implementation consists of the following three steps.

[0129] First, relation confidence assessment: A confidence scoring function is constructed based on the semantic association strength between entities in the knowledge graph (such as co-occurrence frequency and causal weight). For any edge (i,j), its confidence C... ij Calculated using the following formula:

[0130]

[0131] Wherein P(d) j │s i ) for symptoms i With disease d j The causal probability, f ij Let α be the entity co-occurrence frequency, α be the balance coefficient, and max(f) represent the maximum co-occurrence frequency of all edges within the same evaluation range. Below the threshold θ... C Edges are considered low-confidence paths and included in the pruning candidate set.

[0132] Secondly, graph structure entropy quantization: The structure entropy H of node i is calculated based on the degree of disorder in the local topology of the subgraph. i :

[0133]

[0134] Among them, w ij W represents the edge weight. i Let be the weighted degree of node i. High-entropy nodes (H i >θ H ) represents information redundancy or semantic divergence, θ H This represents the entropy threshold. Nodes exceeding this value are considered high-entropy nodes and should be pruned to increase the density of the subgraph.

[0135] Finally, dynamic pruning and teaching loop reinforcement: Based on teaching objectives (e.g., "differential diagnosis of upper gastrointestinal bleeding"), a two-way attention mechanism (GAT) is used to reorder the importance of nodes. High-weighted connected paths to core disease nodes (e.g., "gastric ulcer" and "esophageal varices") are retained, while isolated nodes and low-weight branches are pruned.

[0136] (iv). Construction of closed-loop knowledge path: Construct a closed-loop knowledge chain from symptoms → differential diagnosis → examination items → treatment suggestions to strengthen students' causal and decision-making thinking training;

[0137] (v) Adaptation and System Deployment: Generate structured subgraphs and visualize them, embedding them into virtual patient systems or online teaching platforms to provide controllable knowledge resources for instructional design.

[0138] Technical Approach ②: Retrieval Enhancement-Driven Controllable and Credible Clinical Context Generation

[0139] Technical Approach ②.1: Uncertainty Control Strategy: This involves introducing confidence variables (such as sampling thresholds and probability distributions) to simulate patients' "fuzzy memory" and "uncertain expression."

[0140] In the virtual patient system, to simulate the "fuzzy memory" and "uncertain expression" of real patients in clinical practice, we designed an uncertainty control strategy. This strategy uses the token probability P(W) of the language model. t |W <t Based on this, a confidence evaluation function is introduced:

[0141]

[0142] Among them, W t W represents the current word generated by the language model at time step t. <t This represents all words generated up to time step t, where t represents the time step in the generated sequence. The uncertainty of the generated content is quantified by calculating Shannon entropy. We set a dynamic threshold to determine whether the answer uses a vague expression style; for example, when Confidence(x) < θ, the system calls a vague template such as "I don't remember clearly" or "I think so," generating an answer with uncertainty. Simultaneously, Top-k sampling and Top-p strategies are introduced, combined with temperature T to control the diversity of generated responses (high T simulates hesitation, low T simulates certainty). Furthermore, we encode patient individual attributes (such as age, educational background, and mood) as variables and incorporate them into the confidence function, through:

[0143] θ i =f(a i )=θ0+CognitiveRisk(a i (3-10)

[0144] Where θ0 represents the basic confidence threshold, a i Let f represent the patient's individual attribute vector, f represent the dynamic threshold calculation function, and θ represent the value of the vector. i The final confidence threshold represents the individualized outcome for each patient, while CognitiveRisk represents the cognitive risk function. The system dynamically adjusts the expression confidence levels of different patients to better reflect the language characteristics of specific patient groups.

[0145] Technical Route ②.2, Retrieval Enhanced Medical Semantic Generation Mechanism: Using semantic retrieval of relevant real cases to guide the model to generate "patient responses" to ensure medical causal consistency and traceability.

[0146] To achieve highly credible voice and video generation of virtual patients in medical teaching scenarios and enhance the medical consistency and causal plausibility of their speech content, this project proposes a retrieval-enhanced medical semantic generation mechanism. This mechanism employs a core strategy of "semantic-driven retrieval + structured knowledge guidance + traceable generation," such as... Figure 10As shown, the system focuses on addressing the medical logic bias problem in language model generation. In its workflow, the system first uses a retrieval agent to perform semantic retrieval and structural extraction of gastroenterology-related diseases, constructing a patient knowledge graph subgraph oriented towards the current question-and-answer intent. Through medical entity recognition and graph modeling, it extracts etiology, symptoms, examinations, diagnoses, and treatments from real or synthetic cases.

[0147] Five key path nodes. To further improve the robustness of semantic matching, the system constructs a medical semantic vector library D = {d1, d2, ..., d...}. n Each case record is transformed into a vector representation using medical pre-trained models such as BioBERT:

[0148] h i =Encoder(d i (3-11)

[0149] The data is then stored in a vector database. Here, Encoder represents encoding. When a student asks a question or enters an interaction context with a virtual patient, the system first semantically encodes the query statement q, generates a query vector, and compares its similarity with a medical semantic vector database (e.g., using cosine similarity).

[0150]

[0151] Here, sim represents cosine similarity. The top-K most relevant case paragraphs are retrieved. These paragraphs often contain clear causal paths, such as "Helicobacter pylori infection → chronic gastritis → upper abdominal pain," possessing good clinical teaching value. Subsequently, the retrieval results are concatenated as external knowledge fragments into the language model's prompt. Through retrieval-enhanced generation (RAG) techniques such as RAGFusion, the generative model possesses large-scale language modeling capabilities while using real medical history as semantic anchors for answer generation. To ensure that the generated process not only has accurate medical content but also good causal coherence and pedagogical controllability, the KG query generation agent transforms natural language questions into structured queries to further invoke multi-hop medical logical paths from the knowledge graph. This approach constructs a collaborative mechanism of "semantic retrieval + graph structure reasoning," enabling the generated content to have a complete logical support chain. In the generation phase, the abstract agent first constructs a patient expression logic template based on retrieval and graph paths, such as "I've been experiencing stomach pain and nausea lately. The doctor said it might be gastritis and scheduled a gastroscopy for me." Next, the rewritten agent optimizes and anthropomorphizes the template language according to the teaching objectives, enhancing naturalness and interactive realism while maintaining causal consistency. When the patient's expression involves specific "examination" or "test" nodes, the system further introduces a large-model-driven image-text test report generation mechanism. By using the examination type and contextual pathological information as input, the image-text multimodal generation model outputs the corresponding test report card. The report includes standardized text descriptions (e.g., "Gastroscopy revealed hyperemia and edema of the gastric antral mucosa, suggestive of chronic superficial gastritis"), and simultaneously generates or retrieves highly similar medical images (e.g., images from gastroscopy, CT scans, and lab report screenshots), binding them for output. The image-text combined test report is presented to students as supplementary interactive content, enhancing clinical immersion and judgment training. Each test report is assigned a unique ID and a corresponding knowledge path identifier, ensuring clear content origins and traceable results, supporting subsequent teaching analysis. Finally, in the output phase, the system uses a metadata annotation mechanism to bind each "patient response" with its source case ID, knowledge chain path, graph subgraph ID, and image / text report ID, ensuring the generated content is interpretable and traceable. Teachers or students can visually trace the knowledge source of the statements, achieving a closed-loop teaching experience "from semantics to evidence."

[0152] Technical Route ②.3, Teaching Intent Control Mechanism: Based on teaching objectives and application scenarios, dynamically adjust the anthropomorphic characteristics (such as emotions and tone) of the virtual patient.

[0153] The mechanism takes "teaching-driven personality regulation" as its core concept and introduces a multi-dimensional control vector c = [g1, g2, g3, ...] to control the behavior of virtual patients. It encodes key attributes such as teaching objectives (e.g., "recognizing symptoms" and "ruling out misdiagnosis"), tone style (e.g., "anxious" and "calm"), and expression preferences (e.g., "talkative" and "cautious"). g represents the key attribute.

[0154] Let the system-level instruction be I, the teaching control vector be c (encoding teaching objectives, tone style, and expression preferences), the patient's individual attributes be P (such as age, education level, etc.), and the dialogue context be context. Through their respective embedding mappings, we obtain:

[0155] Instruction = E I (I),C=E C (c), Patient=E P (P), context = E Context (context)

[0156] Among them, E I E C E P E Context This represents the corresponding learnable embedding map.

[0157] The generated control vectors can be embedded into the language model's generation prompt or directly input into the multimodal control interface.

[0158] Prompt final =[Instruction,C,Patient,context] (3-13)

[0159] In this system, Instruction represents system-level instructions, C represents teaching objectives and style parameters, Patient represents patient-specific attributes, and Context represents the context of the current dialogue. During the generation phase, the system dynamically adjusts the patient's response strategies based on the student's questioning style and the teaching stage. For example, if the current teaching objective is "to improve students' ability to identify irrelevant information," the system will introduce some irrelevant but distracting background information, which the patient will then express verbatim. When the objective changes to "to assess students' diagnostic accuracy," the patient will instead provide a clear, concise, and focused description of their medical history to assess whether the student has the ability to quickly extract key information. Regarding emotion and expression modeling, the system uses an emotional dialogue model framework to dynamically generate emotionally charged token sequences based on the teaching context, such as phrases like "I'm a little worried you might have misdiagnosed me" or "Could you hurry up? I'm really uncomfortable." Semantic control information is also simultaneously transmitted to the speech synthesis and expression-driven modules to achieve a consistent mapping of "sound, speech, emotion, and expression," enhancing the human-likeness and immersion of the dialogue.

[0160] Technical Approach ③ Context-Driven Multimodal Virtual Human Behavior Simulation and Situational Feedback

[0161] Technical Route ③.1, Semantic-Driven Emotion-Face Mapping: Based on the emotional features output by the language model, this project drives the synthesis of virtual patient facial expression units to enhance the immersion and clinical realism of the virtual patient system. It constructs a multimodal semantic-behavior mapping mechanism, comprehensively simulating the non-verbal interaction behaviors of patients in real clinical settings, from semantic-driven expression generation and action invocation to contextual feedback based on teaching quality. In the semantic-driven emotion-face mapping part, based on the response content generated by the language model (such as ChatGPT, T5, BioGPT), we extract the emotion vector e = [e1, e2, ..., e] using an emotion recognition module (which can use RoBERTa / BERT combined with Softmax). n ] indicates the distribution of emotions such as anxiety, anger, and confusion:

[0162] e=EmotionClassifier(Utterance) (3-14)

[0163] EmotionClassifier represents the distribution of emotions, and utterance represents the utterance, referring to the content of the interactive response.

[0164] The specific implementation of the emotion recognition module involves feeding the utterance into a text encoder (such as RoBERTa / BERT) to obtain a sentence vector representation h, which is then passed through a linear layer and SoftMax to obtain the probability distribution e of K emotion classes:

[0165] e = SoftMax(W e h+b e )

[0166] Where, the vector representation is h = TextEnc(utterance), W e and b e These are learnable parameters.

[0167] Next, the emotion vectors are mapped to the facial expression units of the virtual character, driving specific muscle movements (e.g., AU1: inner eyebrow lift, AU12: corners of the mouth turn up):

[0168] AU i =f map (e j ),f map (e j )=σ(W AU e j +b AU Expression = Σ j α i ·AU i (3-15)

[0169] Among them, f map The function representing the mapping from emotions to facial action units, e j Let W represent the j-th emotion distribution, σ represent the activation function, and W represent the activating function. AU and b AU Let α be the mapping parameter. i This represents the muscle action weighting coefficient, and Expression represents facial expression. By utilizing this, real-time dynamic expression rendering is achieved, ensuring that the patient's semantic state is consistent with their facial expression, thus improving the naturalness of the interaction.

[0170] Technical Route ③.2, Posture Behavior Library and Dynamic Motion Planning: Establish a library of classic consultation interaction actions (such as fingertips, head stroking, etc.) and dynamically call them according to semantics and context.

[0171] In the posture behavior library and dynamic action planning module, this project pre-defines a classic clinical action library (such as "holding one's abdomen," "rubbing one's head," "frowning," "pointing to a pain point," "restlessness," etc.) and assigns semantic labels to each action (such as "abdominal pain" → holding one's abdomen, "dizziness" → rubbing one's head). The action selection strategy is controlled by the following multi-factor decision function:

[0172] P(α k |s t ,c t =SoftMax(W α [sympton; Emotion; Historty] t (3-16)

[0173] Where s t c represents the current semantic state of the symptoms. t For the context of the dialogue state, α k W represents the k-th action in the preset action library. α The Sympton represents the learnable parameters of the action decision layer. t Extracted key symptom information; Emotion indicates the shift from c t Extracted patient emotional state, Historty t This represents the timing code of the action executed in the most recent t steps. The timing and duration of the action execution are determined by the timing planning module. The rendering layer is handled by a 3D skeletal rigging system or a virtual human driving engine. Figure 11 This paper showcases the overall technical approach to scenario simulation. The method first defines a neutral basic template, then gradually introduces anthropomorphic features to construct diverse and realistic virtual patient avatars. These anthropomorphic features include gender, age group, education level, emotional state, and tone of voice. The example in the image is a male patient, aged 18-39, illiterate, currently experiencing anxiety, and asking the doctor a question in a complaining tone. In this specific scenario, the system automatically generates corresponding verbal expressions and body movements based on the set personality traits and the corresponding description of the patient's condition. For example, the patient might verbally express "My head has been hurting a lot," accompanied by painful gestures such as covering their head with their hands and furrowing their brow, enhancing the doctor's clinical perception. This technical solution effectively improves the realism of the interaction between virtual patients and doctors, helping to build a teaching and training environment that is closer to clinical practice. It supports doctors in practicing their ability to handle diverse patient behavior patterns in simulated scenarios, and is particularly suitable for teaching objectives such as consultation strategy development and communication skills training.

[0174] Technical Route ③.3, Posture Behavior Library and Dynamic Action Planning: Contextual Feedback Guidance Mechanism: Based on the quality of student questions, the rationality of strategies, and the completeness of the path, we will evaluate interactive behavior and drive feedback speech and actions. Regarding the contextual feedback guidance mechanism, we will construct a multi-dimensional interactive quality assessment module for teaching. This module evaluates student performance through the following dimensions: (i) Question quality score (e.g., whether it focuses on key medical history); (ii) Strategy rationality (e.g., whether it proceeds according to clinical logic); (iii) Path completeness (e.g., whether it covers core elements such as chief complaint, present medical history, and physical signs). We define the teaching interaction assessment function:

[0175] S=λ1Q question +λ2Q logic +λ3Q coverage (3-17)

[0176] Among them, Q questionIndicates the quality rating of the question, Q logic Indicating the rationality of the strategy, Q coverage λ represents path integrity, and λ represents weight.

[0177] The specific implementation of question quality scoring:

[0178]

[0179] Among them, E target E represents the set of key entities that should be covered in this case. asked Let S represent the set of entities extracted from student questions. core S represents the core slot set. asked This represents the set of slots that the student has reached, Redundancy represents semantic redundancy, α represents the weight coefficient, and [0,1] represents normalization.

[0180] Specific implementation of strategy rationality:

[0181]

[0182] Where, π asked This represents the student's consultation path (categorizing each round of questions into a standardized node sequence), π canon This represents the standard path set for instruction (which can be given according to department / course scenario), where D represents the sequence edit distance. max This represents the theoretical maximum distance.

[0183] The specific implementation of path integrity:

[0184]

[0185] Where β = {chief complaint, present illness, past medical history, physical signs, examination / diagnosis}, β can be expanded to a finer granularity if necessary, and b is covered means that element b is covered.

[0186] The score S will dynamically trigger the virtual patient's feedback mechanism. For example, when the student repeatedly asks ineffective questions, the patient will show signs of "frowning and hesitant tone," and generate prompts such as "Doctor, you just asked..., I feel much better." When the student completes a high-quality consultation, the system will play a smiling emoticon and positive feedback language, such as "Thank you, I feel much better."

[0187] This embodiment has the following innovative features:

[0188] Innovation Point ① Constructing a structured medical knowledge graph for clinical reasoning, supporting the logical consistency and content verifiability of language generation.

[0189] Technological Innovation: This project focuses on typical diseases of the digestive system, integrating medical ontology construction, structured medical record extraction, and graph reasoning chain mining technologies to construct a hierarchical medical knowledge graph covering the five elements of "cause, symptoms, examination, diagnosis, and treatment." It also embeds language generation processes to enhance the model's medical logical constraints and interpretability on semantic output.

[0190] Pedagogical value: With the help of semantic generation mechanism supported by knowledge graph, students can clearly understand the causal path between disease description and diagnostic suggestions, which helps to cultivate their disease understanding ability and clinical logical reasoning ability, and breaks through the teaching limitations of existing systems that "respond reasonably but have unclear causes and effects".

[0191] Innovation Point ②: Constructing a teaching objective-driven semantic generation and context adaptation mechanism to enhance the professionalism and controllability of language interaction.

[0192] Technological Innovation: This project proposes a semantic control framework of "teaching objectives - Prompt template - knowledge enhancement generation". Combining structured knowledge graphs and patient corpus fine-tuning methods, it constructs a multi-turn dialogue generation engine with the ability to maintain context, semantic consistency and tone adjustment, so as to achieve flexible adaptation of language content between professional expression and anthropomorphic style.

[0193] Pedagogical Value: Enhances students' adaptability and language expression training in multi-round consultation and communication simulations. By accurately recreating consultation scenarios for patients in different roles (such as anxious patients, patients who refuse treatment, etc.), it achieves a closed-loop interaction of "task-driven—language generation—semantic feedback," helping students train their language organization skills and communication strategy coping abilities under real clinical logic.

[0194] Innovation Point ③: Proposes a multimodal consistency generation mechanism driven by anthropomorphic features to achieve collaborative expression of language and behavior in virtual patients.

[0195] Technological Innovation: This project designs a multimodal behavior generation network based on individual variables such as age, gender, and emotion. This network jointly drives semantic generation, speech synthesis, facial expression, and body movement modules, constructing an integrated output system of language, speech, expression, and movement. Furthermore, it introduces a cross-modal alignment mechanism and saliency control point design to enhance the realism and immersion of the virtual patient.

[0196] Pedagogical value: The system can dynamically generate personalized patient performances with emotional fluctuations, facial expressions, and tone changes, enabling medical students to gain a more realistic training experience in nonverbal cue recognition and empathetic communication, and to carry out high-quality clinical communication skills training without relying on real patients.

[0197] This project comprehensively breaks through the technical bottlenecks of current AI virtual patient systems in knowledge modeling, language control, and behavioral expression through three core technical approaches. It also deeply integrates with the "knowledge-communication-judgment" capability chain of medical teaching to build a new generation of intelligent clinical teaching platform that is oriented towards real-world scenarios and has teaching effectiveness and promotional value.

[0198] Specific application examples

[0199] For example, a 38-year-old male sales manager complained of "recurrent burning pain in the upper abdomen for 3 months, which has worsened in the past week." The system first uses the knowledge graph of technical route ① to automatically associate typical features of "peptic ulcer" (such as postprandial pain, nocturnal pain, and NSAIDs medication history), and at the same time combines the patient's occupational characteristics (frequent business trips, irregular eating habits) and unstructured information such as self-reported "symptoms worsen when under stress" to construct a personalized disease model.

[0200] When medical students begin taking patient consultations, the semantic generation system of technical route ② dynamically adjusts the patient's response style according to the teaching stage. In the initial consultation, the patient will respond with a typical clinical description: "The pain usually occurs 1-2 hours after meals, and eating something can relieve it." As the consultation progresses, when students ask about specific medication history, the system will simulate the patient's memory ambiguity: "I think I took some painkillers... I can't remember the specific names, but I bought them at the pharmacy when I had a headache."

[0201] During the physical examination, the multimodal system of technical route ③ will simultaneously present the patient's nonverbal expressions: for example, the patient expresses themselves verbally, accompanied by painful postures such as covering their head with their hands and furrowing their brows, enhancing the doctor's clinical perception; when the student considers differential diagnoses, the patient will provide corresponding feedback based on the clinical rationality assessed by the system—if the student considers serious diagnoses such as "gastric cancer" too early, the patient will show obvious anxiety; if the student inquires about smoking and drinking history according to the standard, more detailed supplementary medical history will be obtained.

[0202] Throughout the consultation process, the system records the student's clinical thinking process in real time and generates a structured assessment report at the end, which includes: question quality score, strategy rationality and path integrity dimensions.

[0203] This application also provides an electronic device, including: a processor, and a memory coupled to the processor, the memory being used to store a computer program; the processor being used to execute the computer program stored in the memory, so that the electronic device performs the method as described in any of the above embodiments.

[0204] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A clinical teaching method, characterized in that, The clinical teaching methods include: In response to the learner's current query, a pre-trained agent, designed to mimic a patient using a virtual patient, returns a current answer to the learner for the current query; and In response to the learner's next query, the agent returns the next answer statement to the learner for the next query. The intelligent agent is used to: perform semantic analysis on the query statement, determine the consultation node on the consultation path corresponding to the query element, query the personalized knowledge subgraph corresponding to the virtual patient to obtain the answer element corresponding to the query element, generate the answer statement on the consultation node based on the answer element, wherein the query element and the answer element conform to the medical logical relationship, and the personalized knowledge subgraph is constructed based on the medical knowledge graph.

2. The clinical teaching method as described in claim 1, characterized in that, The intelligent agent includes the following components configured sequentially: An input embedding module is used to convert the query statement and the location code representing the consultation node into a time-related block input representation; The encoder module is used to query the answer elements from the personalized knowledge subgraph, and The decoder module is used to generate the response statement based on the response elements, confidence variables, and / or instructional control variables. The confidence variables are used to call a fuzzy response as the response statement when the confidence requirement is met. The instructional control variables include: instructional objectives, tone style, and / or expression preferences.

3. The clinical teaching method as described in claim 2, characterized in that, The encoder module includes, in sequence, a self-attention layer, a first normalization layer, a first feedforward network, and a second normalization layer. The output of the input embedding module is connected to the input of the self-attention layer. The output of the input embedding module is connected to the input of the first normalization layer, and the output of the first normalization layer is connected to the input of the second normalization layer via residual connections. The decoder module includes, in sequence: a masking attention layer, a third normalization layer, a multi-head attention layer, a fourth normalization layer, a second feedforward network, and a fifth normalization layer. The output of the input embedding module is connected to the input of the masking attention layer, the output of the second normalization layer is connected to the input of the multi-head attention layer, and the output of the input embedding module is connected to the input of the third normalization layer, the output of the third normalization layer is connected to the input of the fourth normalization layer, and the output of the fourth normalization layer is connected to the input of the fifth normalization layer via residual connections.

4. The clinical teaching method as described in claim 2, characterized in that, The intelligent agent also includes: A graph neural network is used to obtain disease diagnosis results for at least one category based on the virtual patient knowledge subgraph, with symptoms, fuzzy membership degree and node confidence as input.

5. The clinical teaching method as described in claim 4, characterized in that, The graph neural network comprises, in sequence: an input layer, a feature fusion layer, a first activation function layer, a regularization layer, and a classification output layer, with an attention mechanism added before the first activation function layer.

6. The clinical teaching method as described in claim 4, characterized in that, The clinical teaching methods also include: The medical knowledge graph is trimmed based on relation confidence and graph structure entropy.

7. The clinical teaching method as described in claim 2, characterized in that, The intelligent agent also includes: An emotion recognition module is used to generate multi-distributed emotion vectors based on the response statement; A facial expression driving module, taking the output of the emotion recognition module as input, is used to map the emotion vector to facial expressions; and The posture behavior library selection module takes the output of the emotion recognition module, symptoms and timing codes as input, and is used to determine the current action and its execution time based on the context dialogue state and the actions already performed through timing logic.

8. The clinical teaching method as described in claim 7, characterized in that, The emotion recognition module includes, in sequence, a language model and a second activation function layer. The facial expression driving module includes, in sequence, a classification mapping layer and a linear layer. The pose behavior library selection module includes a third activation function layer.

9. The clinical teaching method as described in claim 8, characterized in that, The intelligent agent dynamically triggers a feedback mechanism through a constructed teaching interaction quality evaluation function, influencing the output of answer statements, facial expressions, and gestures. The teaching interaction quality evaluation function includes: question quality score, strategy rationality score, and path integrity score.

10. An intelligent clinical teaching system, characterized in that, The intelligent clinical teaching system includes: Interactive modules, and The processing module, in response to the learner's current query, a pre-trained agent, used to mimic a patient in a virtual patient mode, returns a current answer statement to the learner for the current query; and in response to the learner's next query, the agent returns a next answer statement to the learner for the next query. The agent is configured to: perform semantic analysis on the query, determine the consultation node corresponding to the query element on the consultation path, retrieve the answer element corresponding to the query element from the personalized knowledge subgraph corresponding to the virtual patient, and generate the answer statement at the consultation node based on the answer element. The query element and the answer element conform to a medical logical relationship, and the personalized knowledge subgraph is constructed based on a medical knowledge graph.

Images