Digital intelligent body medical education inquiry method based on clinical database

Abstract

The invention discloses a digital intelligence medical education inquiry method based on a clinical database. The method comprises the following steps: S1, constructing a clinical information database; s2, constructing a digital-intelligent body model and training the digital-intelligent body model to obtain a trained digital-intelligent body model; and S3, performing inquiry teaching on students by using the trained digital intelligent body model, evaluating inquiry questions proposed by the students, and generating an evaluation report. According to the method, the digital intelligent body model is constructed, the question-answer pair is used as the index, and the knowledge graph is introduced to optimize the answer strategy, so that the questions of the students can be more accurately understood, professional and detailed answers are given, the students are guided to think deeply, the clinical thinking and inquiry ability of the students are favorably cultivated, and the teaching quality is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical education, and specifically relates to a digital, intelligent, and physical medical education consultation method based on a clinical database. Background Art

[0002] Medical education currently relies primarily on real patients or standardized patients (SPs) for teaching. Real patients are commonly used in clinical teaching, and medical students can conduct interviews with them during actual diagnosis and treatment. Standardized patients are specially trained individuals who simulate the symptoms, signs, and medical history of specific diseases, providing medical students with standardized interview practice scenarios.

[0003] However, existing technologies still have some shortcomings. In terms of real patients, the conditions are diverse and uncontrollable, making it difficult to meet the needs of systematic and standardized teaching. The severity of the conditions and the combination of symptoms of different patients vary greatly, and it may not be possible to accurately present specific teaching content, which is not conducive to medical students systematically learning the typical diagnosis points of various diseases. For example, for some rare disease symptoms, it is difficult to encounter suitable real patients for teaching in daily clinical practice. In terms of standardized patients, the training cost is high, and a lot of manpower, material resources and time are required to train standardized patients to master the simulated manifestations of various diseases. Moreover, the number of standardized patients is limited, which makes it difficult to meet the frequent practice needs of many medical students. In addition, the performance of standardized patients is easily affected by factors such as emotions and physical condition, making it difficult to ensure the stability and consistency of the teaching process. The simulation of the same disease by different standardized patients at different times may be different, affecting students' accurate understanding and learning of disease characteristics. Summary of the Invention

[0004] The main purpose of this invention is to overcome the shortcomings and deficiencies of the existing technology and provide a digital-intelligence-based medical education and consultation method based on a clinical database. By constructing a digital-intelligence-based model and then using question-answer pairs as indexes, the knowledge graph is introduced to optimize the answer strategy. This method can more accurately understand students' questions and give professional and detailed answers. At the same time, it can guide students to think deeply, help cultivate students' clinical thinking and consultation skills, and improve teaching quality.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] In a first aspect, the present invention provides a digital, intelligent, and physical medical education and consultation method based on a clinical database, comprising the following steps:

[0007] S1. Constructing a clinical information database includes the following steps:

[0008] Collect various clinical information, clean the collected clinical information, classify and label the clinical information according to given classification standards, structure the labeled clinical data, and analyze it using STATA to build a clinical information database; the clinical information includes case records, basic medical knowledge, medical literature, and other clinical medical data;

[0009] S2. Build and train a digital-intelligence-body model to obtain the trained digital-intelligence-body model, including the following steps:

[0010] Construct a digital-intelligence-body model and pre-train it using basic medical knowledge to obtain a pre-trained digital-intelligence-body model. The pre-trained digital-intelligence-body model includes a first processing module, a second processing module, and a third processing module. For the same case, the first processing module is used to parse the preset medical questions to obtain key information of the questions. The second processing module is used to generate an answer strategy based on the key information of the questions. Relevant data is retrieved from the clinical information database based on the answer strategy to generate an answer. The answer is fitted with the medical question to obtain a question-answer pair. Using the question-answer pair as an index, the third processing module is used to introduce a knowledge graph to optimize the answer strategy and obtain the optimal answer strategy. Continue to train the digital-intelligence-body model for various cases to obtain a trained digital-intelligence-body model.

[0011] S3. Use the trained digital-intelligence-body model to conduct question-asking instruction for students, evaluate the questions raised by students, and generate an evaluation report. This includes the following steps:

[0012] The trained digital-intelligence model plays the role of a standardized patient. The teacher selects a case, and students ask the standardized patient questions based on the selected case in the interactive interface. The standardized patient obtains answers based on the optimal answer strategy and records the questions and answers. The recorded questions are evaluated according to preset evaluation indicators and an evaluation report is generated.

[0013] As a preferred technical solution, step S1 includes:

[0014] Use data cleaning tools and algorithms to identify collected clinical information and remove duplicate records and erroneous data to obtain cleaned data;

[0015] Classify the cleaned data according to medical standards and annotate key information, including symptoms, disease diagnosis, and treatment methods;

[0016] The labeled data are converted into a structured format, stored in a relational database, and analyzed using STATA to facilitate rapid retrieval and query to obtain a clinical information database.

[0017] As a preferred technical solution, the pre-training using basic medical knowledge is specifically as follows:

[0018] The digital-agent model is pre-trained using basic medical knowledge based on a semi-supervised learning algorithm. The semi-supervised learning algorithm includes: learning the model based on the labeled basic medical knowledge, and distributing the unlabeled basic medical knowledge on both sides of the hyperplane according to the objective function; the objective function is specifically:

[0019]

[0020] Among them, w represents the normal vector of the hyperplane, C1 represents the penalty parameter of the labeled sample, C2 represents the penalty parameter of the unlabeled sample, ξ i represents the slack variable of the labeled sample, represents the slack variable of unlabeled samples, l represents the number of labeled samples, and u represents the number of unlabeled samples.

[0021] As a preferred technical solution, step S2 includes:

[0022] Fine-tuning phase: Define the state space S, action space A, and reward function R. The state space S includes the current case data, the student's previous medical questions, and the digital agent model's previous responses. The action space A includes the set of possible responses given by the digital agent model. The reward function R is used to measure the quality of each response. Using reinforcement learning, optimize the first module's ability to parse medical questions and extract key information.

[0023] The reinforcement learning-based optimization of the digital-intelligence model's ability to parse medical questions and extract key information from them is as follows: when the digital-intelligence model gives accurate answers and guides students to ask further medical questions, a positive reward is given; if the answer is inaccurate or fails to effectively guide students, a negative reward is given, until the long-term accumulated reward is maximized.

[0024] As a preferred technical solution, the method of retrieving relevant data from a clinical information database and generating an answer based on an answer strategy includes: continuing to train the digital-intelligence-body model using machine learning, locating clinical data labeled corresponding to the medical question from the clinical information database based on the answer strategy, integrating the corresponding labeled clinical data, and outputting an answer based on basic medical knowledge; the basic medical knowledge includes language patterns and semantic relationships in medical texts;

[0025] When outputting answers based on basic medical knowledge, if there are detailed records in the case, the answer will be given directly. If not, the second processing module will give reasonable speculative answers based on common conditions of similar cases and other clinical medical knowledge, and prompt students to ask further questions about relevant details.

[0026] As a preferred technical solution, the method of locating the clinical data marked corresponding to the medical questions includes: using the search conditions as a guide, querying the matching information of the clinical data marked corresponding to the medical questions, and screening and sorting the matching information according to the learned patterns and weights, giving priority to obtaining matching information with high relevance and reliability to the current medical questions, and the matching information includes question-answer pairs, taboo information and related case information; at the same time, according to the current case, the taboo information is eliminated.

[0027] As a preferred technical solution, the method of using the question-answer pair as an index and introducing the knowledge graph to optimize the answer strategy using the third processing module includes the following steps:

[0028] Further collect medical knowledge to build a medical knowledge graph. Medical knowledge includes disease knowledge, symptom knowledge, and drug knowledge.

[0029] The knowledge graph is integrated with the second processing module to provide supplementary auxiliary information and reasoning paths when retrieving information in the clinical information database to generate an optimized answer strategy. The auxiliary information includes the relationship between drugs and diseases and the applicability of different drugs. The reasoning path includes a way to embed the auxiliary information into the question-answer pair.

[0030] As a preferred technical solution, when constructing a knowledge graph, key knowledge information is extracted from medical knowledge as nodes, and the relationship between the key information is used as lines; the key knowledge information includes disease names, symptom manifestations and drug names, and the relationship between the key knowledge information includes the relationship between disease names and symptom manifestations, the relationship between drug names and applicable disease names, and the relationship between drug names and applicable symptom manifestations.

[0031] As a preferred technical solution, the fusion of the knowledge graph and the second processing module includes fusion at the data level, fusion of the model training process, and fusion of the reasoning and answering process;

[0032] The data-level fusion specifically includes: using question-answer pairs as references, using knowledge graph embedding technology to map entities and relationships related to question-answer pairs in the knowledge graph into a low-dimensional vector space, obtaining vectors of medical knowledge, and integrating the vectors of medical knowledge into question-answer pairs;

[0033] The fusion of the model training process is specifically as follows: auxiliary information in the knowledge graph is input as additional features, and the auxiliary information is embedded in the question-answer pair;

[0034] The fusion of the reasoning and answering processes is specifically as follows: based on the key information of the question as an index, relevant nodes and connections are searched in the knowledge graph, and answers are generated in combination with the data in the clinical information database; when generating answers, the nodes and connections corresponding to the key information of the question are obtained, and the basic medical knowledge related to the nodes in the clinical information database is retrieved, and the answers are organized and output according to the basic medical knowledge related to the nodes.

[0035] As a preferred technical solution, step S3 includes case selection and loading, consultation interaction, and teaching evaluation:

[0036] Case selection and loading: The teacher selects a case based on teaching needs, loads the selected case data into the digital-agent model, and initializes the state of the digital-agent so that it simulates the standardized patient corresponding to the case.

[0037] Interactive Diagnosis: Students ask standardized patients questions about their selected cases in an interactive interface, which is used to input questions and display and broadcast the output of the digital-intelligence model. The digital-intelligence model analyzes the student's questions in real time, combines the loaded case data with the training model, generates answers, and displays them on the interactive interface. After generating the answers, the digital-intelligence model outputs and displays prompts to guide students in asking questions.

[0038] Teaching evaluation: Evaluate the recorded medical questions according to the preset evaluation indicators and generate an evaluation report; the evaluation indicators include the accuracy of the questions, the comprehensiveness of the questions, the logic of the questions and the ability to judge the condition;

[0039] The accuracy of the questions is specifically to check whether the questions raised by students are accurate in relation to the information of the current case;

[0040] The comprehensiveness of the questions is specifically analyzed as follows: whether the questions raised by students cover multiple aspects of case-related information;

[0041] The logic of the questions is specifically to determine whether the order in which students ask questions conforms to medical logic;

[0042] The ability to judge the condition is specifically: based on the students' integration and analysis of case information during the consultation process, whether they can preliminarily infer the possible direction of the disease. If the students can propose a reasonable disease hypothesis based on the information obtained, then the ability to judge the condition is strong; otherwise, the ability is weak.

[0043] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0044] (1) This invention combines natural language processing and knowledge graph technology, and the digital intelligence body can understand students' questions more accurately and give professional and detailed answers. At the same time, it guides students to think deeply, helps to cultivate students' clinical thinking and questioning ability, and improves teaching quality.

[0045] (2) Compared to the high human, material, and time costs required to train a large number of standardized patients, the one-time investment in building a digital-intelligence system is relatively low, and it can be reused over a long period of time, greatly reducing the cost of teaching consultations. Furthermore, the digital-intelligence system answers questions based on preset rules and training models, unaffected by factors such as emotions and physical condition. This ensures the stability of the teaching process and the consistency of answers, facilitating standardized learning and training for students.

[0046] (3) The present invention designs an interactive interface that can meet the medical students' needs for convenient consultation. After the consultation, the student's consultation performance can be evaluated based on the interaction record between the digital intelligence body and the student. The evaluation results are presented to teachers and students in the form of a report, helping students understand their own learning situation and providing a reference for teachers to adjust their teaching strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0048] Figure 1 This is a flowchart of a digital-intelligence-physical medical education and consultation method based on a clinical database according to an embodiment of the present invention;

[0049] Figure 2 This is a schematic diagram of the structure of the digital-intelligence-body model according to an embodiment of the present invention;

[0050] Figure 3 Schematic diagram of a digital agent acting as a standardized patient according to an embodiment of the present invention.

[0051] Numbering description: 1-first processing module; 2-second processing module; 3-third processing module. DETAILED DESCRIPTION

[0052] In order to enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0053] References to "embodiments" in this application mean that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described in this application may be combined with other embodiments.

[0054] See also Figure 1 This embodiment provides a digital-intelligence-physical medical education and consultation method based on a clinical database, comprising the following steps:

[0055] S1. Build a clinical information database.

[0056] First, digital data collection. We obtained the center's electronic medical record data, covering both outpatient and inpatient records. We also collected supplementary data, including medical research literature and clinical guidelines. This step aims to comprehensively gather a wide range of clinical information, providing a data foundation for the subsequent construction of a rich digital-intelligence-body model.

[0057] Next, data cleaning is performed. Data cleaning tools and algorithms, such as regular expression matching and statistical analysis, are used to process the collected data. Duplicate records and erroneous data, such as unreasonable values ​​and incorrect date formats, are identified and removed to ensure data accuracy and consistency.

[0058] Next, data is classified and labeled. Data is categorized according to medical standards, such as the ICD classification criteria and the Medical Subject Headings (MeSH). Key information, such as symptoms, disease diagnosis, and treatment methods, is also labeled to provide accurate data support for subsequent digital-intelligence model training.

[0059] Finally, data structuring is performed. The cleaned, classified, and labeled data is converted into a structured format, stored in a relational database (such as MySQL), and analyzed using STATA to facilitate rapid retrieval and query, making it easier for the digital intelligence model to access data.

[0060] S2. Build and train a digital-intelligence-body model to obtain the trained digital-intelligence-body model.

[0061] In step S2, the digital intelligence model is constructed. In this embodiment, the model includes a natural language processing module 1, a machine learning module 2, and a knowledge graph module 3. Figure 2 and Figure 3 shown.

[0062] To enable the model to understand students' questions, the natural language processing module 1 is first trained. Natural language processing models (such as BERT and GPT) are pre-trained using a large amount of medical text data, including clinical records, medical textbooks, and research papers. Subsequently, the model is fine-tuned using annotated clinical data for the consultation teaching scenario, ensuring that it can accurately understand and process medical students' questions.

[0063] Specifically, during the pre-training phase, the model learns basic knowledge such as language patterns and semantic relationships in medical texts. A semi-supervised learning algorithm is used. Traditional machine learning algorithms rely on large amounts of labeled data during training, but labeled data is expensive to obtain. The introduction of semi-supervised learning algorithms can effectively alleviate this problem. Taking clinical data as an example, the amount of unlabeled data far exceeds the amount of labeled data. Algorithms like the Semi-supervised Support Vector Machine (S3VM) can learn using a small number of labeled samples and a large number of unlabeled samples.

[0064] Assume that the labeled sample set is L = {(x1, y1), (x2, y2), ..., (x l ,y l )}, the unlabeled sample set is U={x l+1 ,x l+2 ,…,x l+u The S3VM algorithm constructs a hyperplane that not only correctly classifies labeled samples but also distributes unlabeled samples as close to both sides of the hyperplane as possible. The digital-intelligence model is pre-trained using basic medical knowledge based on a semi-supervised learning algorithm. The semi-supervised learning algorithm involves learning the model based on labeled basic medical knowledge and distributing the unlabeled basic medical knowledge on both sides of the hyperplane according to an objective function. The objective function is specifically:

[0065]

[0066] Where w· represents the normal vector of the hyperplane, C1 represents the penalty parameter of the labeled sample, C2 represents the penalty parameter of the unlabeled sample, ξ i represents the slack variable of the labeled sample, represents the slack variable of unlabeled samples, l represents the number of labeled samples, and u represents the number of unlabeled samples.

[0067] After completing pre-training, it is possible to fully utilize the information in unlabeled data, improve the model's ability to understand and process complex clinical data, and thus improve the accuracy and generalization of answers.

[0068] During the consultation interaction, clinical information is selected and categorized according to its type, and then the consultation interaction is conducted based on the corresponding clinical information type. Taking "case" as the training basis, when training the digital-intelligence model, this embodiment inputs relevant question-answer pairs for the same case. After multiple rounds, the optimal answer strategy for that case is obtained. Then, training is performed case by case, ultimately obtaining a trained digital-intelligence model.

[0069] Specifically, the training steps for the same case include the training of the natural language processing module 1, the machine learning module 2, and the knowledge graph module 3, as follows:

[0070] Fine-tuning phase of Natural Language Processing Module 1: The digital agent can draw on reinforcement learning to optimize its response strategy during its interactions with students. This approach considers the digital agent's response process as a Markov decision process (MDP). Define a state space S, which contains information such as the current case data, the student's question history, and the digital agent's previous responses; an action space A, which represents the set of possible responses the digital agent can give; and a reward function R, which measures the quality of each action (response).

[0071] What needs to be explained is that the quality of each action (answer) depends on whether it meets the standards of clinical thinking logic. For example, fewer questions can obtain more core information. Guidance refers to giving suggestions when students are asked how to ask questions, such as: "You can pay attention to whether I have lost weight recently." When the digital intelligence body gives an accurate answer and guides students to think deeply, it will be given a positive reward; if the answer is inaccurate or fails to effectively guide students, it will be given a negative reward. The goal of the digital intelligence body is to maximize long-term cumulative rewards. where is a discount factor that balances the importance of current rewards and future rewards.

[0072] Next, the machine learning module 2 needs to learn through continuous trial and error, and the digital intelligence model gradually finds the optimal answer strategy to improve the quality of interaction with students. The machine learning module 2 generates an answer strategy based on the key information of the question, and retrieves relevant data from the clinical information database based on the answer strategy to generate an answer. The answer and the medical question are fitted to obtain a question-answer pair. Among them, this embodiment will select a suitable machine learning algorithm, such as a decision tree, support vector machine, neural network, etc., to construct the machine learning module 2 of the digital intelligence for answering questions. Using the question-answer pairs in the clinical data as training samples, the training model enables it to retrieve relevant information from the clinical database based on the input questions and generate reasonable answers. During the training process, the parameters and structure of the algorithm are continuously adjusted to improve the generalization ability of the model and the accuracy of the answers. For example, for the neural network algorithm, parameters such as the number of hidden layers and the number of neurons can be adjusted. Through multiple iterative training, the model can better fit the relationship between questions and answers in the clinical data.

[0073] It should be emphasized that the specific implementation process of retrieving relevant information from clinical databases is achieved through three key aspects: data processing foundation, model collaborative retrieval, and answer generation optimization.

[0074] (1) Retrieval based on a solid data processing foundation: When constructing the clinical database, a series of rigorous steps including data collection, cleaning, classification and labeling, and structured processing were carried out. Data from multiple channels, including the center's outpatient and inpatient medical records, medical research literature, and clinical guidelines, were collected and cleaned to remove duplicate and erroneous data. The data was classified and labeled using ICD classification standards and medical subject headings, and finally stored in a structured format in a relational database. This standardized processing makes the data highly orderly and usable. For example, when a student asks "What disease might a patient have with a cough and low fever?", the digital intelligence body can quickly locate case data in the database containing "cough" and "low fever" symptoms based on symptom labeling, providing accurate materials for subsequent analysis.

[0075] (2) Rely on multi-model collaboration to achieve accurate retrieval: The natural language processing model conducts in-depth analysis of students' questions and extracts key information. For example, when encountering the question "What medications are suitable for patients with hypertension?", the model can accurately identify key concepts such as "hypertension" and "drug treatment" and convert them into retrieval conditions that can be recognized by the database. The machine learning algorithm uses these retrieval conditions as a guide to find matching question-answer pairs or related case information from the clinical database. The algorithm will filter and sort the retrieved data based on the patterns and weights learned in previous training, and give priority to obtaining information that is highly relevant to the current question and has strong reliability. For example, by training and mastering the answer patterns of hypertension treatment drugs under different expressions, relevant data can be found more efficiently and accurately. The knowledge graph plays an auxiliary reasoning role in this process. It integrates a rich medical knowledge system. When the machine learning algorithm retrieves information, the knowledge graph can provide supplementary information and reasoning paths. For the problem of hypertension drugs, the knowledge graph can clarify the relationship between drugs and diseases, the applicability of different drugs, etc., helping the algorithm to understand the problem more comprehensively, optimize the search strategy, and ensure that the retrieved information is more complete and accurate.

[0076] (3) Generate reasonable answers through comprehensive analysis: The machine learning algorithm integrates the information retrieved from the database and combines it with the machine learning model trained by itself to generate preliminary answers. For example, for questions about hypertension medications, the algorithm will refer to answers to similar questions and organize the language to form a preliminary answer based on the specific medication information retrieved.

[0077] When outputting answers based on basic medical knowledge, if there are detailed records in the case, the answer will be given directly. If not, the machine learning module 2 will give reasonable speculative answers based on common conditions of similar cases and other clinical medical knowledge, and prompt students to ask further questions about relevant details.

[0078] The digital-intelligence model uses the loaded case data to adjust and optimize the initial responses. If the patient in the case has special medical conditions or medication contraindications, the digital-intelligence model will specifically address these in its responses, making them more relevant to the situation. For example, students will be informed, "Generally, patients with hypertension can use XX medication, but given that this patient also suffers from XX disease, caution and close monitoring are required when using it." Before outputting the response, the digital-intelligence model also performs logical and linguistic checks on the generated content to ensure that the response is coherent, organized, and in accordance with medical professional expression standards, ensuring clear understanding for students.

[0079] Although the above steps can obtain answer strategies and answers based on the questions, the answers at the current stage are simple and superficial. Therefore, in order to improve the accuracy and richness of the answers, this embodiment uses the question-answer pair as an index and uses the knowledge graph module 3 to introduce the knowledge graph to optimize the answer strategy and obtain the optimal answer strategy.

[0080] The construction and integration of the knowledge graph are as follows: Medical knowledge such as disease knowledge, symptom knowledge, and drug knowledge is collected to construct a medical knowledge graph. The knowledge graph is integrated with natural language processing models and machine learning algorithms to enable the digital agent to reason based on the knowledge graph and thus answer complex medical questions. In the process of building a digital agent based on a clinical database for medical consultation teaching, integrating the knowledge graph with natural language processing models and machine learning algorithms is the key to enabling the digital agent to efficiently simulate standardized patients and accurately answer medical questions. This can be specifically reflected in three aspects: data, model training, and reasoning and answering:

[0081] (1) Data-level fusion. The knowledge graph contains a wealth of medical knowledge, such as disease, symptoms, and drug knowledge, and stores concepts and relationships in a structured form. Medical text data used in natural language processing model training, such as clinical records, textbooks, and papers, provide material for entity and relationship extraction for the knowledge graph. When constructing the knowledge graph, key knowledge information, such as disease names, symptoms, and therapeutic drugs, as entities, as well as causal and correlation relationships between them, are extracted from these medical texts. The extracted information is integrated into the knowledge graph to enrich its content. Sample data for machine learning algorithm training, that is, question-answer pairs in clinical data, can also be associated with the knowledge graph. For example, for a question about disease treatment drugs, the knowledge graph can provide the corresponding relationship between drugs and diseases. These relationships can be used as a reference for machine learning algorithm training, allowing the algorithm to learn the connection between the question and the relevant information in the knowledge graph, thereby improving the accuracy of the model's answer to the question. The fusion of data and knowledge graph organically embeds the knowledge graph into the algorithm. In order to better integrate the knowledge graph into the digital intelligence model, the knowledge graph embedding technology is used to map the entities and relationships in the knowledge graph into a low-dimensional vector space. Using the TransE model, for each triple (h, r, t) in the knowledge graph (where h is the head entity, r is the relation, and t is the tail entity), the model assumes that h + r ≈ t. The vector representations of entities and relations are learned by minimizing the following loss function:

[0082]

[0083] Here, S is the set of positive sample triplets in the knowledge graph, S′ is the set of negative samples generated by randomly replacing the head or tail entity in the positive samples, d is a function that measures vector distance (such as Euclidean distance), γ is a boundary value greater than 0, and [x] + = max(0, x). This way, the knowledge in the knowledge graph is represented in vector form, making it easier to integrate with natural language processing models and machine learning algorithms.

[0084] (2) Integration of the model training process. The natural language processing model learns the language patterns and semantic relationships of medical texts in the pre-training stage, and then uses the annotated clinical data to fine-tune it to make it more suitable for the consultation teaching scenario. During the fine-tuning process, the information of the knowledge graph is introduced. For example, when the model processes a medical question, the model optimizes its understanding and representation of the question with the help of the concepts and relationships of diseases, symptoms, etc. related to the question in the knowledge graph. When the machine learning algorithm constructs the answer model of the digital intelligence, the knowledge graph provides auxiliary information. Taking the neural network algorithm as an example, during the training process, the relevant knowledge in the knowledge graph is input into the network as an additional feature, or the training samples are screened and enhanced according to the knowledge graph, so that the model learns more comprehensive and accurate knowledge and improves the rationality of the generated answers. Assuming that the original feature vector received by the input layer neuron is x, and the related feature vector obtained from the knowledge graph embedding is y, the two are spliced ​​into a new input vector z = [x; y]. During the training process, the weight W and bias b of the neural network are adjusted to minimize the loss function J(W, b), such as the mean square error loss function:

[0085]

[0086] Among them, n is the number of training samples, y i is the true label, is the model prediction value. In this way, the model can fully utilize the structured knowledge in the knowledge graph during training, improving its ability to understand and answer medical questions.

[0087] (3) Integration of reasoning and answering processes. When a medical student asks a question, the natural language processing model first analyzes the question and extracts key information. This key information is used as an index to search for relevant knowledge nodes and connections in the knowledge graph. For example, when a student asks "What are the common complications of diabetes?", the natural language processing model extracts key information such as "diabetes" and "common complications," finds the "diabetes" node and its relationship with "complications" in the knowledge graph, and obtains information such as "diabetic nephropathy" and "diabetic retinopathy." The machine learning algorithm generates the answer content based on the results of the natural language processing model analysis and the information obtained from the knowledge graph, combined with the data in the clinical database and the training model. If the knowledge graph shows that "diabetic nephropathy" is closely related to "long-term hyperglycemia," the machine learning algorithm will incorporate this relationship into the answer when generating the answer, making the answer more detailed and accurate, such as "Common complications of diabetes include diabetic nephropathy, and long-term hyperglycemia is one of the important factors leading to diabetic nephropathy."

[0088] S3. Use the trained digital-intelligence-body model to conduct question-asking instruction for students, evaluate the questions raised by students, and generate an evaluation report. This includes the following steps:

[0089] The trained digital-intelligence model plays the role of a standardized patient. The teacher selects a case, and students ask the standardized patient questions based on the selected case in the interactive interface. The standardized patient obtains answers based on the optimal answer strategy and records the questions and answers. The recorded questions are evaluated according to preset evaluation indicators and an evaluation report is generated.

[0090] Specifically, when teaching medical consultations, teachers select appropriate case data from the clinical database based on the teaching objectives and students' learning progress. The selected case data is loaded into the digital agent model, and the state of the digital agent is initialized so that it simulates the corresponding standardized patient. When selecting cases, teachers can choose representative and moderately difficult cases based on factors such as the course content and students' knowledge level. This ensures that students can effectively learn and master relevant medical consultation techniques and knowledge through interaction with the digital agent.

[0091] Medical students conduct medical consultations with the digital agent through a specially designed interactive interface. This interface features a question input box, an area for displaying the digital agent's answers, and case information prompts. It also supports voice input and output. Students can enter text questions using the keyboard or use voice input to convert their speech into text and send it to the digital agent. The digital agent analyzes the student's question in real time, combining it with loaded case data and trained models to generate answers that are displayed on the interactive interface. Furthermore, the digital agent can proactively guide students to further inquire based on the student's question, such as, "What else would you like to know about my condition?" When analyzing the question, the digital agent first uses the natural language processing module 1 to parse the question and extract key information. It then combines the case data with the answer model generated by the machine learning algorithm to retrieve relevant information from the clinical database and finally organize the language to generate a reasonable answer. For example, if a student asks, "What has the patient's diet been like recently?" the digital agent searches the case data for relevant records. If detailed records are available, it provides a direct answer. If not, the digital agent can provide a reasonable, speculative answer based on common features of similar cases and medical knowledge, and prompt the student to inquire for further details.

[0092] After the consultation, the system evaluates the student's performance based on the interaction records between the digital intelligence body and the student. Evaluation indicators include the accuracy, comprehensiveness, and logic of the questions, as well as the student's ability to judge the condition. The evaluation results are presented to teachers and students in the form of a report to help students understand their own learning situation and provide reference for teachers to adjust teaching strategies. The specific evaluation process is as follows:

[0093] (1) Question accuracy assessment: Check whether the questions asked by students are accurate in targeting key symptoms, medical history, and other information of the case. For example, if the main symptoms of the patient in the case are cough and fever, the student asks "Do you have symptoms of cough and fever?" The accuracy of the question is high; if the question is unrelated to the case, such as "How many people are there in your family?", the accuracy is low.

[0094] (2) Comprehensiveness of questions: Analyze whether students’ questions cover multiple aspects related to the case, such as symptoms, medical history, family history, lifestyle habits, etc. If students only ask questions about symptoms and do not cover other important aspects, the comprehensiveness is insufficient.

[0095] (3) Question logic assessment: Determine whether the order in which students ask questions conforms to medical logic. For example, asking about general symptoms first and then gradually asking about specific details and related causes is logical; if the questions are jumpy and confusing, then the logic is poor.

[0096] (4) Assessment of the ability to judge the condition: Based on the integration and analysis of case information during the consultation process, determine whether the student can preliminarily infer the possible direction of the disease. If the student can propose a reasonable disease hypothesis based on the information obtained, the ability to judge the condition is strong; otherwise, the ability is weak.

[0097] Based on the evaluation results of the above indicators, a detailed evaluation report is generated. The report not only includes the scores of each indicator, but also provides specific suggestions and improvement directions to help students improve their medical consultation skills in a targeted manner.

[0098] It should be noted that, for the sake of convenience, the aforementioned method embodiments are all expressed as a series of action combinations, but those skilled in the art should know that the present invention is not limited to the described order of actions, because according to the present invention, certain steps can be performed in other orders or simultaneously.

[0099] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0100] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A digital, intelligent, and physical medical education and consultation method based on a clinical database, characterized by: The steps include: S1. Constructing a clinical information database includes the following steps: Collect various clinical information, clean the collected clinical information, classify and label the clinical information according to given classification standards, structure the labeled clinical data, and analyze it using STATA to build a clinical information database; the clinical information includes case records, basic medical knowledge, medical literature, and other clinical medical data; S2. Build and train a digital-intelligence-body model to obtain the trained digital-intelligence-body model, including the following steps: Construct a digital-intelligence-body model and pre-train it using basic medical knowledge to obtain a pre-trained digital-intelligence-body model. The pre-trained digital-intelligence-body model includes a first processing module, a second processing module, and a third processing module. For the same case, the first processing module is used to parse the preset medical questions to obtain key information of the questions. The second processing module is used to generate an answer strategy based on the key information of the questions. Relevant data is retrieved from the clinical information database based on the answer strategy to generate an answer. The answer is fitted with the medical question to obtain a question-answer pair. Using the question-answer pair as an index, the third processing module is used to introduce a knowledge graph to optimize the answer strategy and obtain the optimal answer strategy. Continue to train the digital-intelligence-body model for various cases to obtain a trained digital-intelligence-body model. S3. Use the trained digital-intelligence-body model to conduct question-asking instruction for students, evaluate the questions raised by students, and generate an evaluation report. This includes the following steps: The trained digital-intelligence model plays the role of a standardized patient. The teacher selects a case, and students ask the standardized patient questions based on the selected case in the interactive interface. The standardized patient obtains answers based on the optimal answer strategy and records the questions and answers. The recorded questions are evaluated according to preset evaluation indicators and an evaluation report is generated.

2. The digital intelligence and physical medical education and consultation method based on clinical database according to claim 1 is characterized in that: The step S1 comprises: Use data cleaning tools and algorithms to identify collected clinical information and remove duplicate records and erroneous data to obtain cleaned data; Classify the cleaned data according to medical standards and annotate key information, including symptoms, disease diagnosis, and treatment methods; The annotated data are converted into a structured format, stored in a relational database, and analyzed using STATA to facilitate rapid retrieval and query to obtain a clinical information database.

3. The digital intelligence and physical medical education and consultation method based on clinical database according to claim 1 is characterized in that: The pre-training using basic medical knowledge is specifically as follows: The digital-agent model is pre-trained using basic medical knowledge based on a semi-supervised learning algorithm. The semi-supervised learning algorithm includes: learning the model based on the labeled basic medical knowledge, and distributing the unlabeled basic medical knowledge on both sides of the hyperplane according to the objective function; the objective function is specifically: ； in, represents the normal vector of the hyperplane, C1 represents the penalty parameter of the labeled sample, and C2 represents the penalty parameter of the unlabeled sample. represents the slack variable of the labeled sample, represents the slack variable for unlabeled samples, l represents the number of labeled samples, u Indicates the number of unlabeled samples.

4. The digital intelligence and physical medical education and consultation method based on clinical database according to claim 1 is characterized in that: The step S2 includes: Fine-tuning phase: defining the state space , action space and the reward function The state space , including current case data, students’ historical questions and historical answers of the digital agent model, the action space Including the set of possible answers given by the digital agent model, the reward function Used to measure the quality of each answer; based on reinforcement learning, it optimizes the first module's ability to analyze the questions and extract key information; The reinforcement learning-based optimization of the digital-intelligence model's ability to parse medical questions and extract key information from them is as follows: when the digital-intelligence model gives accurate answers and guides students to ask further medical questions, a positive reward is given; if the answer is inaccurate or fails to effectively guide students, a negative reward is given, until the long-term accumulated reward is maximized.

5. The digital intelligence and physical medical education and consultation method based on clinical database according to claim 1 is characterized in that: Retrieving relevant data from the clinical information database and generating answers based on the answer strategy includes: continuing to train the digital-intelligence model using machine learning, locating clinical data labeled accordingly to the medical question from the clinical information database based on the answer strategy, integrating the corresponding labeled clinical data, and outputting answers based on basic medical knowledge; the basic medical knowledge includes language patterns and semantic relationships in medical texts; When outputting answers based on basic medical knowledge, if there are detailed records in the case, the answer will be given directly. If not, the second processing module will give reasonable speculative answers based on common conditions of similar cases and other clinical medical knowledge, and prompt students to ask further questions about relevant details.

6. The digital intelligence and physical medical education and consultation method based on clinical database according to claim 5 is characterized in that: The method for locating clinical data labeled corresponding to the medical questions includes: using search conditions as a guide, querying matching information of clinical data labeled corresponding to the medical questions, and screening and sorting the matching information according to learned patterns and weights, giving priority to obtaining matching information with high relevance and reliability to the current medical questions, the matching information including question-answer pairs, contraindication information and related case information; at the same time, according to the current case, the contraindication information is eliminated.

7. The digital intelligence and physical medical education and consultation method based on clinical database according to claim 1 is characterized in that: The method of using the question-answer pair as an index and introducing the knowledge graph to optimize the answer strategy using the third processing module includes the following steps: Further collect medical knowledge to build a medical knowledge graph. Medical knowledge includes disease knowledge, symptom knowledge, and drug knowledge. The knowledge graph is integrated with the second processing module to provide supplementary auxiliary information and reasoning paths when retrieving information in the clinical information database to generate an optimized answer strategy. The auxiliary information includes the relationship between drugs and diseases and the applicability of different drugs. The reasoning path includes a way to embed the auxiliary information into the question-answer pair.

8. The digital intelligence and physical medical education and consultation method based on clinical database according to claim 7 is characterized in that: When constructing a knowledge graph, key knowledge information is extracted from medical knowledge as nodes, and the relationships between the key information are used as lines; the key knowledge information includes disease names, symptom manifestations and drug names, and the relationships between the key knowledge information include the relationship between disease names and symptom manifestations, the relationship between drug names and applicable disease names, and the relationship between drug names and applicable symptom manifestations.

9. The digital intelligence and physical medical education and consultation method based on clinical database according to claim 8 is characterized in that: The fusion of the knowledge graph and the second processing module includes fusion at the data level, fusion of the model training process, and fusion of the reasoning and answering process; The data-level fusion specifically includes: using question-answer pairs as references, using knowledge graph embedding technology to map entities and relationships related to question-answer pairs in the knowledge graph into a low-dimensional vector space, obtaining vectors of medical knowledge, and integrating the vectors of medical knowledge into question-answer pairs; The fusion of the model training process is specifically as follows: auxiliary information in the knowledge graph is input as additional features, and the auxiliary information is embedded in the question-answer pair; The fusion of the reasoning and answering processes is specifically as follows: based on the key information of the question as an index, relevant nodes and connections are searched in the knowledge graph, and answers are generated in combination with the data in the clinical information database; when generating answers, the nodes and connections corresponding to the key information of the question are obtained, and the basic medical knowledge related to the nodes in the clinical information database is retrieved, and the answers are organized and output according to the basic medical knowledge related to the nodes.

10. The digital intelligence and physical medical education and consultation method based on clinical database according to claim 1 is characterized in that: Step S3 includes case selection and loading, consultation interaction, and teaching evaluation: Case selection and loading: The teacher selects a case based on teaching needs, loads the selected case data into the digital-agent model, and initializes the state of the digital-agent so that it simulates the standardized patient corresponding to the case. Interactive consultation: Students ask standardized patients questions about selected cases in an interactive interface, where the interface is used to input questions and display and broadcast the output of the digital-intelligence model; The digital-intelligence model analyzes students' questions in real time, combines the loaded case data with the training model, generates answers, and displays them on the interactive interface. After generating the answers, the digital-intelligence model outputs and displays prompts to guide students in asking questions. Teaching evaluation: Evaluate the recorded medical questions according to the preset evaluation indicators and generate an evaluation report; the evaluation indicators include the accuracy, comprehensiveness, logic and ability to judge the condition of the disease; The accuracy of the questions is specifically to check whether the questions raised by students are accurate in relation to the information of the current case; The comprehensiveness of the questions is specifically analyzed as follows: whether the questions raised by students cover multiple aspects of case-related information; The logic of the questions is specifically to determine whether the order in which students ask questions conforms to medical logic; The ability to judge the condition is specifically: based on the students' integration and analysis of case information during the consultation process, whether they can preliminarily infer the possible direction of the disease. If the students can propose a reasonable disease hypothesis based on the information obtained, then the ability to judge the condition is strong; otherwise, the ability is weak.

Images