Doctor-patient communication training method and system based on large model dialogue engine
By constructing a training dataset for doctor-patient communication and combining it with instruction fine-tuning and reinforcement learning to train a large-scale model dialogue engine, the problems of insufficient multi-scenario coverage and real-time evaluation in existing doctor-patient communication training technologies have been solved. This has enabled personalized feedback and standardized demonstrations, thereby improving the communication skills of medical staff.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-11
- Publication Date
- 2026-03-13
AI Technical Summary
Existing doctor-patient communication training methods are insufficient to cover a variety of medical scenarios, lack real-time feedback and objective evaluation, and cannot provide accurate personalized improvement suggestions and standardized dialogue demonstrations.
We constructed a training dataset for doctor-patient communication, and used a combination of instruction fine-tuning and reinforcement learning to train a large-scale dialogue engine, obtaining the trainer's standard dialogue performance and providing personalized feedback reports.
It enables accurate identification of communication risks in multiple scenarios, provides personalized improvement suggestions and standardized dialogue demonstrations, and enhances the communication skills and effectiveness of medical staff.
Smart Images
Figure CN121662425A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent dialogue technology, and in particular to a method and system for training doctor-patient communication based on a large-model dialogue engine. Background Technology
[0002] Doctor-patient communication, as an indispensable part of medical services, directly affects patient trust, treatment outcomes, and the probability of medical disputes; its importance is self-evident. In the healthcare industry, effective communication is not only the cornerstone of harmonious doctor-patient relationships but also key to improving medical quality and patient satisfaction. However, with the increasing complexity of medical scenarios and the diversification of patient needs, how to achieve accurate and effective communication in different situations has become a major challenge that healthcare professionals urgently need to address.
[0003] Currently, although many medical institutions and training programs attempt to improve doctor-patient communication skills through role-playing or case-based teaching, these methods often have limitations. They mostly rely on fixed scripts or limited simulated scenarios, making it difficult to cover the various unexpected situations and emotional fluctuations that may occur in real medical environments. More importantly, these methods lack real-time feedback on dynamic changes during the communication process, making it difficult to help medical staff quickly adjust their strategies and cope with complex patient emotions and needs in actual interactions.
[0004] The core technological challenge in doctor-patient communication training lies in constructing a standardized training system covering multiple medical scenarios and achieving accurate identification and tiered management of communication risks. Traditional training methods rely on role-playing and theoretical learning, failing to provide standardized patient response patterns and hindering objective quantitative assessment of trainees' communication performance. In complex medical environments, healthcare professionals face highly uncertain and specialized communication scenarios, with significant differences in the types and levels of communication risks across different scenarios. This necessitates a training system capable of accurately identifying and providing early warnings of potential communication risks. Furthermore, assessing healthcare professionals' communication skills involves multiple dimensions, including professionalism, empathy, risk management, and efficiency. Existing technologies lack a comprehensive evaluation index system and real-time dynamic assessment capabilities, failing to provide healthcare professionals with precise, personalized improvement suggestions and standardized dialogue demonstrations. Summary of the Invention
[0005] The purpose of this invention is to provide a training method and system for doctor-patient communication based on a large model dialogue engine, which provides medical staff with accurate personalized improvement suggestions and standardized dialogue demonstrations in doctor-patient communication in multiple scenarios.
[0006] To achieve the above objectives, the present invention provides the following solution:
[0007] This invention provides a training method for doctor-patient communication based on a large-model dialogue engine, comprising:
[0008] Construct a doctor-patient communication training dataset, wherein the training dataset includes doctor-patient dialogue samples in multiple scenarios, communication risk labels, and standard response templates;
[0009] The basic large model is trained based on the training dataset to obtain a dedicated dialogue engine for doctor-patient communication. The training of the basic large model includes a method that combines instruction fine-tuning and reinforcement learning.
[0010] The training scenario and trainee identity information are input into the dedicated dialogue engine for doctor-patient communication to obtain the trainee's standard dialogue performance.
[0011] The evaluation results are obtained based on the trainee's standard dialogue performance and actual dialogue voice data;
[0012] Based on the assessment results, obtain a personalized feedback report.
[0013] Optionally, constructing a doctor-patient communication training dataset includes:
[0014] By collecting raw data from dialogues in multiple scenarios, we obtained doctor-patient dialogue samples in multiple scenarios. The raw data includes scenarios such as routine consultation, medical condition disclosure, informed consent, and dispute mediation.
[0015] The doctor-patient dialogue samples are classified into risk tags using preset labeling rules to obtain the communication risk tags;
[0016] Based on the doctor-patient dialogue samples and the corresponding communication risk tags, obtain the standard response template for each scenario.
[0017] Optionally, training the base large model based on the training dataset includes:
[0018] Extract key semantic information from doctor-patient dialogue samples, and output the initial dialogue content based on the extracted key semantic information from the basic large model.
[0019] If the initial dialogue content output by the basic large model during training does not match the semantics of the target scene, then the optimized large model can be obtained by adjusting the training parameters.
[0020] The optimized large model is then adjusted using reinforcement learning methods.
[0021] Optionally, adjusting the optimized large model using reinforcement learning methods includes:
[0022] The communication risk labels and standard response templates are obtained from the training dataset, and it is determined whether the response content of the basic model in the communication scenario deviates from the professional standard.
[0023] If the model deviates from the professional standard, the optimized large model is adjusted based on the updated training dataset by updating the dialogue samples in the training dataset.
[0024] Optionally, the evaluation results obtained based on the trainee's standard dialogue performance and actual dialogue voice data include:
[0025] The actual dialogue voice data is cleaned to obtain cleaned actual voice data.
[0026] The trainees' standard dialogue performance and the cleaned actual voice data were evaluated separately to obtain evaluation indicators for standard dialogue and actual dialogue. The evaluation indicators include professional indicators and risk control indicators.
[0027] The evaluation metrics of the standard dialogue and the actual dialogue are compared to obtain the evaluation results.
[0028] Optionally, cleaning the actual dialogue voice data includes: performing a speech-to-text operation on the actual dialogue voice data, and cleaning the actual dialogue voice data using regular expressions.
[0029] Optional evaluation metrics for obtaining standard dialogues include:
[0030] Natural language processing techniques are used to segment and semantically analyze the text data of the standard dialogue to obtain keywords;
[0031] If the keyword matches a preset professional indicator system, then the keyword will be classified as a professional indicator.
[0032] Extract warning words or tone markers related to risk control from the standard dialogue, and classify the number of words with warning words or tone markers related to risk control as risk control indicators.
[0033] This invention also provides a doctor-patient communication training system based on a large-model dialogue engine, comprising:
[0034] The dataset construction module is used to construct a doctor-patient communication training dataset, wherein the training dataset includes multi-scenario doctor-patient dialogue samples, communication risk labels, and standard response templates;
[0035] The model building module is used to train the basic large model based on the training dataset to obtain a dedicated dialogue engine for doctor-patient communication. The training of the basic large model includes a method that combines instruction fine-tuning and reinforcement learning.
[0036] The dialogue generation module is used to obtain the training scenario and trainee identity information and input them into the dedicated dialogue engine for doctor-patient communication to obtain the trainee's standard dialogue performance.
[0037] The evaluation module is used to obtain evaluation results based on the trainee's standard dialogue performance and actual dialogue voice data;
[0038] The feedback module is used to obtain personalized feedback reports based on the evaluation results.
[0039] The beneficial effects of this invention are as follows: By constructing a doctor-patient dialogue dataset covering multiple scenarios such as routine consultations, patient information disclosure, informed consent, and dispute mediation, and integrating communication risk tags and standard response templates, this invention addresses the core issues of scenario complexity and risk diversity in doctor-patient communication. This invention optimizes the model's performance in terms of professionalism and scenario adaptability through a combination of fine-tuning and reinforcement learning on the basic language model, generating a dedicated dialogue engine. Based on a multi-dimensional indicator system of professionalism and risk control, this invention generates personalized feedback reports in real time, accurately analyzes shortcomings, and provides optimization suggestions and standard demonstrations, ultimately achieving a systematic improvement in doctor-patient communication skills and significantly enhancing trainees' coping abilities and communication effectiveness in complex scenarios. Attached Figure Description
[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0041] Figure 1 This invention provides a method for training doctor-patient communication based on a large-model dialogue engine. Detailed Implementation
[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0043] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0044] like Figure 1 As shown, this embodiment provides a method for training doctor-patient communication based on a large model dialogue engine, including:
[0045] Construct a doctor-patient communication training dataset, wherein the training dataset includes doctor-patient dialogue samples in multiple scenarios, communication risk labels, and standard response templates;
[0046] The basic large model is trained based on the training dataset to obtain a dedicated dialogue engine for doctor-patient communication. The training of the basic large model includes a method that combines instruction fine-tuning and reinforcement learning.
[0047] The training scenario and trainee identity information are input into the dedicated dialogue engine for doctor-patient communication to obtain the trainee's standard dialogue performance.
[0048] The evaluation results are obtained based on the trainee's standard dialogue performance and actual dialogue voice data;
[0049] Based on the assessment results, obtain a personalized feedback report.
[0050] Furthermore, the construction of the doctor-patient communication training dataset includes:
[0051] By collecting raw data from dialogues in multiple scenarios, we obtained doctor-patient dialogue samples in multiple scenarios. The raw data includes scenarios such as routine consultation, medical condition disclosure, informed consent, and dispute mediation.
[0052] The doctor-patient dialogue samples are classified into risk tags using preset labeling rules to obtain the communication risk tags;
[0053] Based on the doctor-patient dialogue samples and the corresponding communication risk tags, obtain the standard response template for each scenario.
[0054] Specifically, when constructing the basic dataset, diverse dialogue corpora can be collected by simulating doctor-patient communication scenarios. For example, in a typical consultation scenario, a patient asks a doctor about treatment options for a cold, and the doctor explains in detail the precautions for medication and possible side effects. This dialogue is recorded in the corpus, covering everyday communication content. In a scenario where a doctor informs a patient of an early lung shadow on a test result, the doctor's tone must be cautious and professional to avoid causing panic. These dialogues provide rich foundational data for subsequent annotation.
[0055] When classifying risk labels, based on preset rules, dialogue content can be categorized into three risk levels: low, medium, and high, and further subdivided into types such as emotional conflict and misunderstanding risk. For example, in a dispute mediation dialogue, if a patient is emotionally agitated due to poor treatment results, and the doctor's words show slight impatience, this dialogue might be labeled as high-risk, categorized as emotional conflict. This labeling method helps to accurately identify potential problems, laying the foundation for subsequent analysis. When obtaining standard template content from the labeled dataset, response frameworks can be matched according to scenario classification. In informed consent scenarios, if a doctor needs to inform about surgical risks, the standard template might include preoperative preparation, a description of the surgical risk probability (e.g., 5% to 10%), and postoperative recovery suggestions. Template matching ensures that the doctor's communication is standardized and comprehensive, reducing the risk of misunderstanding.
[0056] Furthermore, training the base large model based on the aforementioned training dataset includes:
[0057] Extract key semantic information from doctor-patient dialogue samples, and output the initial dialogue content based on the extracted key semantic information from the basic large model.
[0058] If the initial dialogue content output by the basic large model during training does not match the semantics of the target scene, then the optimized large model can be obtained by adjusting the training parameters.
[0059] The optimized large model is then adjusted using reinforcement learning methods.
[0060] Furthermore, by incorporating reinforcement learning methods, the optimized large model is adjusted in the following ways:
[0061] The communication risk labels and standard response templates are obtained from the training dataset, and it is determined whether the response content of the basic model in the communication scenario deviates from the professional standard.
[0062] If the model deviates from the professional standard, the optimized large model is adjusted based on the updated training dataset by updating the dialogue samples in the training dataset.
[0063] Specifically, the dataset is classified and grouped according to different communication scenarios to obtain a set of scenario dialogues. For these classified dialogue sets, the basic language model is initially adapted using pre-defined rules. Key semantic information is extracted from the dialogue sets to determine the model's semantic understanding framework in different scenarios. Based on the extracted key semantic information, fine-tuning training is implemented to adjust the output logic of the basic language model. If the model output does not match the target scenario semantics during fine-tuning, the training parameters are adjusted to obtain an optimized semantic adaptation model. Using the optimized semantic adaptation model, reinforcement learning methods are combined to further adjust the model's dialogue generation capabilities. Feedback information is obtained from the training dataset to determine the accuracy of the model's responses in specific communication scenarios. Based on the feedback information, the professionalism of the model's responses in doctor-patient communication is iteratively addressed. If the feedback information shows that the response content deviates from professional standards, the sample content in the training dataset is updated to determine an improved dialogue generation strategy. The improved dialogue generation strategy is then used to finally integrate the dedicated dialogue engine. Validation data is obtained from multiple communication scenarios to determine the engine's adaptability and stability in different scenarios. The dedicated dialogue engine is continuously optimized and adjusted based on the validation data. To address any shortcomings identified during validation, supplementary training samples are obtained to arrive at the final doctor-patient communication dialogue engine.
[0064] Large-scale models based on the Transformer architecture, such as the GPT series and BERT, are selected and optimized and adjusted according to the training needs of doctor-patient communication. For example, a module specifically for handling medical domain knowledge is added to the model structure to better understand and process medical terminology, disease knowledge, and other content. Collected and processed doctor-patient communication data is used as training data, and a combination of supervised learning, unsupervised learning, and reinforcement learning is employed to train the large-scale model. In the supervised learning phase, labeled data is used to train the model, enabling it to learn correct doctor-patient communication patterns and language expressions. In the unsupervised learning phase, the model automatically learns the statistical patterns and semantic understanding of language from a large amount of unlabeled doctor-patient communication data. In the reinforcement learning phase, a reward mechanism is set up, rewarding or penalizing the model based on the performance of the responses generated by the model in actual doctor-patient communication scenarios (such as patient satisfaction, communication effectiveness evaluation, etc.), guiding the model to continuously optimize its response strategy and improve the quality and effectiveness of its responses. During training, various techniques were employed to optimize the model, such as adjusting the learning rate, using regularization to prevent overfitting, and using multimodal data fusion (e.g., combining speech and image information) to enhance the model's ability to understand complex information, thereby improving model performance and generalization ability. Regarding the details of model parameter tuning, the learning rate was set to 0.001 at the beginning of training and adjusted at a decay rate of 0.1 every 10 training epochs. L2 regularization was used with a regularization coefficient of 0.0001 to prevent overfitting. The AdamW optimizer was selected as the training algorithm, with β1 set to 0.9, β2 set to 0.999, and epsilon set to 1e-8 to improve training stability and convergence speed.
[0065] This embodiment incorporates a task awareness module into the multi-head attention calculation. Task labels (such as "Children's Doctor-Patient Communication - Creative Script Generation") are input, and a small classifier outputs weight adjustment coefficients, which are then weighted and fused with the original attention weights. The formula is:
[0066] ;
[0067] in Adjust the weight matrix specifically for each task.
[0068] The system designs a variety of training scenarios for doctor-patient communication, covering common diseases (such as the common cold, hypertension, and diabetes), rare diseases (such as Huntington's disease and cystic fibrosis), complex conditions (such as multiple organ failure and late-stage malignant tumors), and special patient groups (such as children, Alzheimer's patients, and patients with mental illnesses). Each scenario includes detailed background information, the patient's symptoms, psychological state, and possible questions and requests. Medical personnel first select a training scenario and then communicate with a large-scale dialogue engine that acts as the patient. The large-scale model generates realistic patient responses based on the scenario settings and the medical personnel's input, including verbal expressions and emotional reactions. Based on the patient's responses, the medical personnel use learned communication skills, such as listening, questioning, explaining, and reassuring. After the communication session, the system evaluates the medical personnel's performance based on preset assessment indicators (such as the accuracy, completeness, and effectiveness of the communication content, the effect on patient emotional reassurance, and the application of communication skills), and provides a detailed evaluation report pointing out strengths and weaknesses. Medical staff can reflect on and summarize based on the assessment report, and then choose the same or different scenarios to train again, continuously improving their communication skills.
[0069] It's important to note that the dataset may include dialogue samples from routine consultations where patients describe their symptoms, as well as contextual content from doctors explaining test results when informing patients of their medical conditions. Assuming the dataset contains 5000 dialogue records, with routine consultations accounting for 40%, medical condition explanations for 30%, and other scenarios such as informed consent and dispute mediation each accounting for 15%, this initial screening ensures a balanced sample size for each scenario, preventing model bias during training. When classifying and grouping the dataset, dialogues in routine consultation scenarios can be further subdivided into initial consultations and follow-up consultations. Initial consultation dialogues may involve more symptom descriptions and medical history inquiries, while follow-up consultations focus on feedback on treatment effectiveness. After classification, assuming 1000 initial consultation dialogues and 800 follow-up consultation dialogues, this refined grouping helps the model more accurately adapt to different communication needs.
[0070] During fine-tuning training, if the model's output is found to be mismatched with the target scenario—for example, generating overly rigid terminology in a medical explanation scenario—training parameters can be adjusted to include more samples with more colloquial expressions. Assuming that after adjustment, the model's semantic matching accuracy in testing improves from 70% to 85%, this indicates that fine-tuning effectively enhances the relevance of the communication. When adjusting dialogue generation capabilities using reinforcement learning methods, doctor-patient dialogues can be simulated, allowing the model to dynamically optimize its responses based on patient feedback. For instance, in a dispute mediation scenario where the patient is dissatisfied with the doctor's explanation, the model needs to learn to adjust its tone and increase empathetic expressions, such as "I understand how you feel," thereby improving the acceptability of the dialogue.
[0071] Furthermore, based on the trainee's standard dialogue performance and actual dialogue voice data, the evaluation results include:
[0072] The actual dialogue voice data is cleaned to obtain cleaned actual voice data.
[0073] The trainees' standard dialogue performance and the cleaned actual voice data were evaluated separately to obtain evaluation indicators for standard dialogue and actual dialogue. The evaluation indicators include professional indicators and risk control indicators.
[0074] The evaluation metrics of the standard dialogue and the actual dialogue are compared to obtain the evaluation results.
[0075] Furthermore, cleaning the actual dialogue voice data includes: performing a speech-to-text operation on the actual dialogue voice data, and cleaning the actual dialogue voice data using regular expressions.
[0076] Furthermore, the evaluation metrics for obtaining standard dialogues include:
[0077] Natural language processing techniques are used to segment and semantically analyze the text data of the standard dialogue to obtain keywords;
[0078] If the keyword matches a preset professional indicator system, then the keyword will be classified as a professional indicator.
[0079] Extract warning words or tone markers related to risk control from the standard dialogue, and classify the number of words with warning words or tone markers related to risk control as risk control indicators.
[0080] Specifically, when acquiring dialogue data from training subjects, the system can collect voice or text information in real-time from doctor-patient communication scenarios. For example, if a doctor is discussing a patient's condition, the system records their conversation and converts it into text for preprocessing. This process filters out irrelevant background noise and repetitive expressions, resulting in a clear dialogue record that lays the foundation for subsequent analysis.
[0081] For the initially compiled dialogue content, natural language processing (NLP) technology is used for word segmentation and semantic analysis. For example, statements from doctors asking patients about their medical history are broken down into keywords such as "medical history," "duration," and "symptoms," and their semantic meaning is analyzed. If these keywords match preset professional indicator systems such as "accuracy" or empathy indicator systems such as "compassion," they are categorized accordingly. This categorization helps to accurately assess whether the doctor's expression meets professional requirements. In the feature extraction stage, warning words or tone markers related to risk control are extracted from the dialogue. For example, if a doctor mentions "potential deterioration" or has a hurried tone in the dialogue, and the system identifies 3 warning words, exceeding the preset threshold of 2, it is judged as having a risk tendency, and the risk level is recorded as "medium." This allows for the timely detection of potential communication problems, reminding trainees to adjust their expression.
[0082] By acquiring dynamic evaluation data from the system, user performance shortcomings and related evaluation results are extracted and organized into structured information. Based on the extracted performance shortcomings and evaluation results, a pre-established analysis model is used to determine the user's specific deficiencies, resulting in detailed analysis results. If the analysis results show that a certain performance is below a preset threshold, corresponding improvement directions are generated for that deficiency, and specific suggestions are determined. By matching the improvement directions with a preset scenario library, corresponding scenarios that match the user's context are obtained, resulting in suitable scenario information. Based on the matched corresponding scenarios, a standard dialogue template database is accessed to extract suitable dialogue content, determining standard dialogue example content. A personalized adjustment mechanism is employed, combining the user's evaluation results and suggestions, to generate the final personalized feedback report.
[0083] This embodiment also provides a doctor-patient communication training system based on a large model dialogue engine, including:
[0084] The dataset construction module is used to construct a doctor-patient communication training dataset, wherein the training dataset includes multi-scenario doctor-patient dialogue samples, communication risk labels, and standard response templates;
[0085] The model building module is used to train the basic large model based on the training dataset to obtain a dedicated dialogue engine for doctor-patient communication. The training of the basic large model includes a method that combines instruction fine-tuning and reinforcement learning.
[0086] The dialogue generation module is used to obtain the training scenario and trainee identity information and input them into the dedicated dialogue engine for doctor-patient communication to obtain the trainee's standard dialogue performance.
[0087] The evaluation module is used to obtain evaluation results based on the trainee's standard dialogue performance and actual dialogue voice data;
[0088] The feedback module is used to obtain personalized feedback reports based on the evaluation results.
[0089] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for training doctor-patient communication based on a large-model dialogue engine, characterized in that, include: Construct a doctor-patient communication training dataset, wherein the training dataset includes doctor-patient dialogue samples in multiple scenarios, communication risk labels, and standard response templates; The basic large model is trained based on the training dataset to obtain a dedicated dialogue engine for doctor-patient communication. The training of the basic large model includes a method that combines instruction fine-tuning and reinforcement learning. The training scenario and trainee identity information are input into the dedicated dialogue engine for doctor-patient communication to obtain the trainee's standard dialogue performance. The evaluation results are obtained based on the trainee's standard dialogue performance and actual dialogue voice data; Based on the assessment results, obtain a personalized feedback report.
2. The doctor-patient communication training method based on a large-model dialogue engine according to claim 1, characterized in that, The training dataset for doctor-patient communication includes: By collecting raw data from dialogues in multiple scenarios, we obtained doctor-patient dialogue samples in multiple scenarios. The raw data includes scenarios such as routine consultation, medical condition disclosure, informed consent, and dispute mediation. The doctor-patient dialogue samples are classified into risk tags using preset labeling rules to obtain the communication risk tags; Based on the doctor-patient dialogue samples and the corresponding communication risk tags, obtain the standard response template for each scenario.
3. The doctor-patient communication training method based on a large-model dialogue engine according to claim 1, characterized in that, Training the base large model based on the aforementioned training dataset includes: Extract key semantic information from doctor-patient dialogue samples, and output the initial dialogue content based on the basic model described by the extracted key semantic information. If the initial dialogue content output by the basic large model during training does not match the semantics of the target scene, then the optimized large model can be obtained by adjusting the training parameters. The optimized large model is then adjusted using reinforcement learning methods.
4. The doctor-patient communication training method based on a large model dialogue engine according to claim 3, characterized in that, The optimization of the large model is adjusted using reinforcement learning methods, including: Obtain the communication risk labels and standard response templates from the training dataset, and determine whether the response content of the basic model in the communication scenario deviates from the professional standard. If the model deviates from the professional standard, the optimized large model is adjusted based on the updated training dataset by updating the dialogue samples in the training dataset.
5. The doctor-patient communication training method based on a large-model dialogue engine according to claim 1, characterized in that, The evaluation results, based on the trainee's standard dialogue performance and actual dialogue voice data, include: The actual dialogue voice data is cleaned to obtain cleaned actual voice data. The trainees' standard dialogue performance and the cleaned actual voice data were evaluated separately to obtain evaluation indicators for standard dialogue and actual dialogue. The evaluation indicators include professional indicators and risk control indicators. The evaluation metrics of the standard dialogue and the actual dialogue are compared to obtain the evaluation results.
6. The doctor-patient communication training method based on a large model dialogue engine according to claim 5, characterized in that, Cleaning the actual dialogue voice data includes: performing a speech-to-text operation on the actual dialogue voice data, and cleaning the actual dialogue voice data using regular expressions.
7. The doctor-patient communication training method based on a large model dialogue engine according to claim 5, characterized in that, Evaluation metrics for obtaining standard dialogues include: Natural language processing techniques are used to segment and semantically analyze the text data of the standard dialogue to obtain keywords; If the keyword matches a preset professional indicator system, then the keyword will be classified as a professional indicator. Extract warning words or tone markers related to risk control from the standard dialogue, and classify the number of words with warning words or tone markers related to risk control as risk control indicators.
8. A doctor-patient communication training system based on a large model dialogue engine, used to implement the doctor-patient communication training method based on a large model dialogue engine as described in any one of claims 1-7, characterized in that, include: The dataset construction module is used to construct a doctor-patient communication training dataset, wherein the training dataset includes multi-scenario doctor-patient dialogue samples, communication risk labels, and standard response templates; The model building module is used to train the basic large model based on the training dataset to obtain a dedicated dialogue engine for doctor-patient communication. The training of the basic large model includes a method that combines instruction fine-tuning and reinforcement learning. The dialogue generation module is used to obtain the training scenario and trainee identity information and input them into the dedicated dialogue engine for doctor-patient communication to obtain the trainee's standard dialogue performance. The evaluation module is used to obtain evaluation results based on the trainee's standard dialogue performance and actual dialogue voice data; The feedback module is used to obtain personalized feedback reports based on the evaluation results.