Method and system for improving expandability of large language model in low-AI health vegetarian population

Abstract

The invention belongs to the crossing field of artificial intelligence and medical information technology, and relates to a method and system for improving the expandability of a large language model in low-AI health vegetarian population. The method comprises the following steps: constructing a main chat robot, and obtaining dialogue log data of a user and the main chat robot; extracting an interaction obstacle type list according to the dialogue log data; constructing an interaction obstacle-learning instruction mapping form according to the interaction obstacle type list; generating video learning content for each learning instruction according to the interaction obstacle-learning instruction mapping form; and adaptively distributing the video learning content, and accurately putting the video learning content to a specified interface position of the target user. According to the method, the dialogue ability of a large model and a patient can be improved, and the expandability problem of low-AI healthy vegetarian people is solved.

Description

TECHNICAL FIELD

[0001] The application relates to the field of artificial intelligence and medical information technology, in particular to a method and system for improving the scalability of large language models (LLMs) in medical health consultation scenarios, especially aiming at the problem of user AI health literacy differences, forming a supplementary architecture of the main chat robot by accurately constructing learning function modules, and realizing intelligent health consultation services for low health literacy groups. BACKGROUND

[0002] The application potential of large language models (LLMs) in primary medical health consultation scenarios has attracted much attention, but their actual deployment faces the challenge of user interaction ability gap. The ability of individuals to effectively understand, evaluate and use artificial intelligence (AI) technology (such as large language models) in the medical health scene (i.e. AI health literacy) is uneven, which significantly reduces the ability of LLMs to generate clinically relevant recommendations and leads to potential risks of reliability and safety of AI tools. Among them, the quality of user dialogue is low, and low AI health literacy users often input vague descriptions, which leads the model to be unable to generate accurate health recommendations. Users may blindly accept the wrong suggestions of large language models or lack the ability to question wrong conclusions. Due to limited technology exposure, the AI health literacy of rural areas and the elderly is low, further widening the gap in the accessibility of medical services.

[0003] Current technologies use a one-way optimization paradigm, only improving the model side (such as medical fine-tuning, knowledge enhancement), and ignore the adaptability of user interaction ability, failing to solve the fundamental problem of insufficient user interaction ability. SUMMARY

[0004] The core method of the application focuses on the construction and optimization of primary medical health consultation models, based on the collaborative design of community users (patients, family members, community staff) and expert teams (model developers, clinical personnel and public health experts), to build chat robots with more universal dialogue interaction capabilities for medical health consultation, especially for the problem of low AI health literacy of patient groups that cannot be solved by the model alone. Precise construction of learning function modules forms a supplementary architecture of the main chat robot, thereby improving the dialogue ability of large models and patients and solving the scalability problem in low AI health literacy groups.

[0005] The technical solution adopted by the application is as follows: A method for improving the scalability of large language models in low AI health literacy groups, comprising the following steps: Constructing a main chat robot to obtain dialogue log data of users and the main chat robot; According to the dialogue log data, extracting a list of interaction barriers; Based on the list of interaction barrier types, construct an interaction barrier-learning instruction mapping form; Based on the interaction barrier-learning instruction mapping form, generate video learning content for each learning instruction; Adaptive distribution of video learning content, precisely targeting specific screen locations for the intended users.

[0006] Furthermore, the step of extracting a list of interaction obstacle types based on the dialogue log data includes: Perform data preprocessing, including cleaning the data, removing empty dialogue records or invalid interactions, filtering out duplicate or test-related non-user-generated content, and dividing the data into multiple batches by random sampling based on the total amount of data and computing resources; Semantic analysis methods were used to initially identify common problem patterns from dialogue log data. Combined with collected user feedback and experience data, model developers and domain experts collaborated to conduct a comprehensive analysis to provide a first version of the definition of interaction barrier types and a list of specific manifestations. Using machine learning methods, the first version of the obstacle type is input as the initial seed. Starting from the first batch of data, the dialogue logs are automatically labeled to generate obstacle labels, which are then manually checked by the collaborative team. Feedback is collected, and the obstacle type definition or machine learning model is adjusted based on the feedback. The machine labeling and manual checking process is repeated for subsequent batches, iterating until the last batch of data is completed, and a complete list of interaction obstacle types and their specific manifestations is output.

[0007] Furthermore, the step of constructing an interaction obstacle-learning instruction mapping form based on the list of interaction obstacle types includes: Based on the list of interaction obstacle types, we gain a deep understanding of how each obstacle affects the user experience and learning process, set clear learning objectives for each obstacle type, and formulate multiple possible learning instructions based on the set learning objectives to form a set of potential learning instructions. The set of potential learning instructions is temporarily integrated into the main chatbot for testing and feedback iteration. During the testing process, when a certain type of interaction obstacle is detected in the user interaction, a potential learning instruction text format is automatically and randomly selected and presented to the tester. After receiving the instruction and learning the content, the tester continues to talk to the chatbot. After each round of testing, the effectiveness index of each potential learning instruction is collected, and the effectiveness score is calculated based on the weighted average. Based on the effectiveness score, the content of the learning instruction is added or deleted. For deleted instructions, the expert team designs new instruction variants based on community user experience and feedback, adds them to the set of potential learning instructions, and repeats the testing, feedback collection, and decision-making steps until the termination condition is met, outputting an interaction barrier-learning instruction mapping form.

[0008] Furthermore, the step of generating video learning content for each learning instruction based on the interaction barrier-learning instruction mapping form includes: Develop a basic requirements form based on the input learning instructions. and actual needs form Output the initial requirements form ; In the initial requirement form Based on this foundation, each learning instruction is analyzed, and a conversational and easy-to-understand video narration / explanation script is generated based on the analysis results. Natural language generation technology is used to ensure that the script conforms to [the required language]. and Based on the script, multiple modules are integrated to output preliminary video learning content; Initial video learning content is integrated into the main chatbot. When a specific type of obstacle is detected in real time during user interaction, the corresponding video content is automatically triggered and presented to the test user. The user watches the video while interacting with the chatbot and then continues to interact. User feedback is collected after each round of testing, and the forms are updated based on the user feedback to meet actual needs. This generates an updated requirements form. ; Once the termination condition is met, the final video learning content for each learning instruction is output. The video learning content has multimodal characteristics, including virtual avatars, audio explanations, and synchronized subtitles, to ensure easy understanding and access.

[0009] Furthermore, the adaptive distribution of video learning content employs the following distribution rules: Rule 1, Passive Distribution Before User Consultation: When a new user completes registration or an existing user successfully logs in, text and video learning content will be automatically displayed on the user's homepage and consultation startup page. Users can choose to watch based on their personal interests and needs. Rule 2, proactive distribution before user consultation: When a new user registration or a user's first consultation intent after logging in is detected, the learning content most closely matching the user's characteristics is automatically pushed to their homepage or consultation launch page by combining the user's historical status information and historical information. Rule 3, Real-time Distribution During User Consultations: During a user's consultation session, if a specific interaction obstacle is detected in real time, the rule engine will be immediately triggered. Based on the specific obstacle type detected, the rule engine will match the generated corresponding video learning content and accurately push the video to the user in the current session through instant message push, pop-up prompts, or sidebar recommendations.

[0010] Furthermore, the above method also includes a feedback loop step, which periodically monitors user interactions and triggers content updates when new obstacles or significant changes are detected, in order to ensure dynamic adaptability.

[0011] A system for improving the scalability of large language models in populations with low AI health literacy, comprising: The main chatbot building module is used to build the main chatbot; The interaction obstacle log data acquisition module is used to acquire the dialogue log data between the user and the main chatbot; The interaction obstacle type and feature extraction module is used to extract a list of interaction obstacle types based on dialogue log data; The interaction barrier and learning instruction building module is used to build an interaction barrier-learning instruction mapping form based on the list of interaction barrier types; The multimodal learning instruction content generation module is used to generate video learning content for each learning instruction based on the interaction barrier-learning instruction mapping form; The adaptive distribution module is used to adaptively distribute video learning content and accurately deliver it to the specified interface location of the target user.

[0012] Furthermore, the system also includes a feedback loop module for periodically monitoring user interactions and triggering content updates when new obstacles or significant changes are detected, in order to ensure dynamic adaptability.

[0013] This invention addresses the AI ​​health literacy gap in existing intelligent health consultation systems and user interactions by proposing a new paradigm of collaborative optimization between users and the model, moving away from the traditional approach of unidirectional model optimization. Specifically, this is embodied in a dynamic and adaptive systematic solution based on interactive data and user collaborative design. The technical solution of this invention revolves around a method and system for improving the scalability of large language models among populations with low AI health literacy. Specific innovations and beneficial effects are as follows: 1. A dynamic content generation method based on the deconstruction of user interaction barriers. Existing technologies typically directly augment models with medical scenario knowledge, while this invention uses a method that reverse-engineers interaction barriers from user-model interaction failure cases (dialogue logs). Through a human-computer collaborative annotation model, it systematically identifies, defines, and quantifies various barriers (e.g., dialogue initiation failure, risk misjudgment, etc.) that affect the effectiveness of users with low health literacy.

[0014] 2. Collaborative Design and Testing Feedback Mechanism. This invention does not involve experts unilaterally developing teaching materials; instead, it constructs a collaborative design and testing feedback mechanism between experts and users. The learning content initially set by the expert team (e.g., how to describe symptoms) is immediately integrated into the main system, and its effectiveness is tested by real users (especially representatives with low health literacy) in simulated consultation scenarios. By collecting effectiveness scores (e.g., improved comprehension, satisfaction) and setting clear iteration termination conditions (e.g., voting consensus), the most efficient and practical learning instructions are ultimately selected, thereby greatly improving the accuracy of interventions and user acceptance.

[0015] 3. Adaptive Generation Strategy for Learning Content. In terms of learning content generation, this invention introduces an adaptive generation strategy based on a dual-requirement form. Basic requirements are defined by experts to ensure the scientific rigor and standardization of the content; user needs are defined by users and community staff and designed as dynamically updatable templates to incorporate personalized elements such as cultural adaptability and emotional preferences. Based on these two forms, the adaptive generation strategy automatically generates multimodal videos combining virtual avatars, conversational scripts, and synchronized subtitles. This is then iteratively optimized through the aforementioned testing feedback loop, ensuring that the content's presentation highly matches the cognitive habits of users with low health literacy.

[0016] 4. Context-Aware Adaptive Distribution Strategy Based on a Rule Engine. This invention employs a multi-rule distribution strategy to ensure that learning content is pushed to the right people at the right time, in the right format. Passive / active distribution before consultation is primarily used for new user onboarding or predictive recommendations based on historical records. Real-time distribution during consultation is the core innovation of this invention. The system can identify specific interaction obstacles (e.g., uncontrolled conversation duration) in real time within the dialogue flow and immediately trigger the push of corresponding micro-learning videos (e.g., how to control conversation duration). This learn-as-you-go, scenario-based teaching model seamlessly embeds the learning process into the consultation task, thereby improving the timeliness and relevance of intervention.

[0017] 5. Feedback Loop Module for Sustainable Optimization. The system of this invention is not a static tool built once, but rather an ecosystem with self-evolving capabilities. Its feedback loop module periodically monitors interaction data and employs an event-driven strategy: the update loop is triggered only when a new obstacle or significant change is detected, ensuring the system can continuously adapt to changes in user behavior and avoiding unnecessary consumption of computational resources. Attached Figure Description

[0018] Figure 1 This is a flowchart outlining the steps involved in improving the scalability of large language models among populations with low AI health literacy.

[0019] Figure 2This is a modular diagram of a system designed to improve the scalability of large language models among populations with low AI health literacy. Detailed Implementation

[0020] 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 specific embodiments and accompanying drawings.

[0021] This invention proposes a method to improve the scalability of large language models in populations with low AI health literacy. This method is a dynamic generation approach for accompanying E-learning modules based on interaction barrier deconstruction. Addressing the challenges in AI health consultant interactions, it achieves the accurate construction of supplementary pre-consultation learning tools through four stages: interaction barrier type and feature extraction, interaction barrier and learning instruction construction, multimodal learning instruction content generation, and adaptive distribution. This method forms a complementary architecture with the main chatbot system. The steps of this method are as follows: Figure 1 As shown below, the specific implementation process of each stage and step is explained.

[0022] Phase 0: Construction of the main chatbot and acquisition of interaction obstacle log data.

[0023] 0.1 Based on the existing technical architecture, a preliminary prototype of a medical and health consultation chatbot will be built, possessing basic question-and-answer, intent recognition, and dialogue management capabilities. An extensible interface will be implemented to support the integration of subsequent functional modules and log recording. A secure online interaction platform will be built, supporting online annotation of dialogue text.

[0024] 0.2 Community users (especially patients or their families with low health literacy) interact with existing robot models to obtain dialogue log data and provide feedback on their user experience.

[0025] Phase 1: Types and features of interaction barriers.

[0026] Input: The dialogue log data stream obtained in Phase 1.

[0027] Processing procedure: 1.1 Data Preprocessing: Clean the data, remove empty dialogue records or invalid interactions, and filter out duplicate or test-related non-user-generated content. Based on the total amount of data and computing resources, divide the data into multiple batches through random sampling.

[0028] 1.2.1 Common problem patterns are initially identified from the entire dialogue log data using semantic analysis methods (e.g., information extraction, entity recognition, semantic classification). Simultaneously, combined with user feedback data collected in Phase 0, a comprehensive analysis is conducted collaboratively by model developers and domain experts (e.g., clinicians, public health experts) to provide a first-version definition of interaction barrier types and a list of specific manifestations of these barriers. Specific manifestations refer to the specific content included in each interaction barrier type. Examples include users repeatedly asking the same question, misunderstanding medical terminology, inability to accurately describe symptoms, deviating from the core message in multiple rounds of dialogue, failing to point out key errors and provide correct information, and misunderstanding system suggestions.

[0029] 1.2.2 Based on the first version of obstacle types, immediately begin the machine annotation process. Using machine learning methods (such as a BERT-based text classifier), input the first version of obstacle types as the initial seed, automatically annotate the dialogue logs starting from the first data batch, generating obstacle labels, which are then manually reviewed by a collaborative team. After manual review, feedback is collected (including annotation accuracy, obstacle type applicability, etc.), and the obstacle type definition or machine learning model is adjusted based on the feedback. For subsequent batches, repeat the machine annotation and manual review process. Iterate until the last data batch is completed.

[0030] Output: Let the set of interaction obstacle types be... ,in It is the number of obstacle types, each , n = 1…N, represent a specific type of obstacle. Output a complete list of interactive obstacle types I and the specific manifestations of the interactive obstacles.

[0031] Phase 2: Interaction Barriers and Learning Instruction Construction

[0032] Input: A complete list of interaction obstacle types and their specific manifestations from the output of Phase 1.

[0033] Processing procedure: 1) Formulating potential learning instructions: A team of experts (model developers, clinicians, and public health professionals) develops these instructions based on a list of barrier types. This involves gaining a deep understanding of how each obstacle affects the user experience and learning process, including analyzing the specific difficulties or misunderstandings users may encounter when facing that obstacle. For each type of obstacle... Set clear learning objectives, which are the knowledge or skills users should acquire after completing the learning instructions. Based on these objectives, design specific learning instructions, drafting multiple possible instructions to form a potential set of learning instructions. This set is denoted as: , where K It is for obstacles The number of potential instructions per , is a learning instruction.

[0034] 2) Iterative testing and feedback: This involves adapting the set of potential learning instructions. It was temporarily integrated into the main chatbot system, and a collaborative design team (patients, families, community workers, model developers, clinicians, and public health experts) was invited to test it. During the testing, the system detected obstacles in user interactions. (Based on real-time dialogue analysis) Automatically and randomly select a command The instructions are presented to the testers in text format. After receiving the instructions, the testers learn the content and then continue their conversation with the chatbot. At the end of each round of testing, a post-test questionnaire is used to collect data on each potential instruction. The effectiveness metrics include improvements in user comprehension, enhanced interaction fluency, and satisfaction ratings, with an effectiveness score calculated based on a weighted average. Based on the effectiveness score, the expert team adds or removes learning instructions. If the effectiveness score... If the value exceeds a set threshold and the expert panel reaches a consensus through voting, the instruction is retained; otherwise, it is deleted. For deleted instructions, the expert team designs new instruction variants based on the experiences and feedback of community users (patients, families, and community staff) and adds them to the potential set of learning instructions. The testing, feedback collection, and decision-making steps are repeated until the termination condition is met (the collaborative group reaches a consensus through voting or discussion, and the new iteration no longer significantly improves the instruction's effectiveness). Community user experience and feedback include effectiveness metrics and user experience data from other unstructured text records.

[0035] Output: Interaction Barrier-Learning Instruction Mapping Form: A structured list or database containing the final learning instruction tool corresponding to each interaction barrier type. The output is in plain text format.

[0036] Phase 3: Generation of multimodal learning instruction content.

[0037] Input: Stage 2 output interaction barrier - learning instruction mapping form.

[0038] Processing procedure: 1) Requirements Form Construction: Based on the input learning instructions, the collaborative design team will develop two requirements forms: Basic required forms The process is proposed by a team of experts based on learning instructions, including defining general video generation standards, such as video length (e.g., 30-60 seconds), language style (conversational, easy to understand), core learning objectives, key information points, visual elements (e.g., virtual avatars, animated icons), audio requirements (e.g., speech rate, tone), and synchronized subtitles.

[0039] Actual need form The form, initially submitted by users and community staff, considers factors such as cultural compatibility, emotional preferences, and learning preferences. The collaborative team, based on interview and test data drafts, structured them into a modifiable template that allows for dynamic adjustment and updates based on feedback in subsequent iterations.

[0040] Output initial requirements form This is a composite form containing all the basic and practical requirements, stored in the form of a structured database or configuration file.

[0041] 2) Content generation: In the form request... Based on this, each learning instruction is analyzed to extract core learning objectives, key information, and style points, as well as... The system analyzes user needs and preferences (such as those from community feedback). Based on the analysis results, it generates a conversational and easy-to-understand video narration / explanation script. Natural language generation technology is used to ensure the script meets user needs and preferences. and Based on the script, multiple modules (such as virtual human rendering, text-to-speech audio generation, and subtitle synchronization tools) are integrated to output preliminary video learning content. .

[0042] 3) Testing and Feedback: [This will be done / will ... Temporarily integrated into the main chatbot system. When the system detects specific obstacle types during user interaction in real time, it automatically triggers the presentation of corresponding video content to test users. The collaborative design team and test users (especially representatives with low health literacy) were invited to conduct interaction tests. Users watched videos while conversing with the chatbot and then continued interacting. After each round of testing, user feedback was collected through a post-test questionnaire. Based on the feedback, the collaborative team discussed and updated the forms to meet actual needs. (e.g., adjusting language style). Create an updated requirements form. The process terminates when user satisfaction scores reach a predetermined threshold (e.g., ≥80%) and the collaborative team reaches a consensus through voting (that the new iteration will no longer significantly improve the effect). If the termination condition is met, the loop exits, and the final content is generated. If not, return to step 2), regenerate the video content using the requirement form, and repeat the test.

[0043] Output: The final video learning content for each learning instruction is stored in a video file format (such as MP4) along with metadata (such as the corresponding obstacle type and generation parameters). The video content is multimodal, including virtual avatars, audio narration, and synchronized subtitles, to ensure ease of understanding and access.

[0044] Phase 4: Adaptive distribution based on users' health literacy needs, dynamically adjusting the output of user learning content.

[0045] Inputs: Text output from Phase 2 (interaction barrier-learning instruction mapping form) and video learning content output from Phase 3; user historical interaction data; current / real-time user status information.

[0046] Processing procedure: System settings distribution rule engine Rule 1 (Passive Distribution Before User Consultation) When the system detects that a new user has completed registration or an existing user has successfully logged in, it will automatically display text and video learning content on the user's homepage and consultation startup page. Users can choose to watch based on their personal interests and needs.

[0047] Rule 2 (Proactive Distribution Before User Inquiry) When the system detects a new user registration or a user's first inquiry intent after logging in, it automatically pushes the most relevant common learning content to the user's homepage or inquiry launch page by combining the user's historical status information and historical information.

[0048] Rule 3 (Real-time Distribution During User Consultations): During a user's consultation session, log analysis from Phase 1 identifies specific interaction obstacles in real time (e.g., detecting 'lack of proactive clarification of diagnostic questions' or 'out-of-control conversation duration' in consecutive dialogue rounds), immediately triggering the rule engine. Based on the identified obstacle type, the engine matches the corresponding video generated in Phase 3 and accurately pushes the short video to the user in the current session via instant message push, pop-up notifications, or sidebar recommendations.

[0049] Output: Precisely deliver the selected learning content to the specified interface location of the target user.

[0050] Building upon stages 0 to 4 described above, this invention further includes a feedback loop step or a feedback loop module. The feedback loop module is designed to periodically monitor user interaction, triggering content updates only when new obstacles or significant changes are detected, ensuring the dynamic adaptability of the learning tool. A significant change refers to a difference in the proportion of obstacles between the current and previous periods exceeding a system-preset threshold (or passing a significance test). The system automatically collects and preprocesses new dialogue log data periodically (e.g., monthly) and proceeds to stage 1. If a new obstacle or significant change is detected, stage 2 is triggered; otherwise, data is simply recorded and the system waits for the next monitoring cycle without proceeding to subsequent stages.

[0051] Example: Phase 0: Construction of the main chatbot and acquisition of interaction obstacle log data 0.1 Chatbot Prototype Construction: Based on the existing AI architecture, the development team fine-tuned the model to build a prototype healthcare consultation chatbot. This prototype possesses basic question-and-answer, intent recognition, and dialogue management capabilities, and integrates scalable API interfaces for logging and module integration. The platform is deployed on a secure cloud server and supports real-time dialogue and text annotation functions.

[0052] 0.2 Data Collection: Invite community users (especially those with low health literacy, such as the elderly or family members of patients with chronic diseases) to interact with the robot. Through the online platform, users participate in consultation sessions, generating dialogue log data (including text dialogues, timestamps, and user feedback). Simultaneously, collect users' subjective experience feedback (e.g., through questionnaires) to initially accumulate raw data streams.

[0053] Phase 1: Types and features of interaction barriers.

[0054] Input: The dialogue log data stream obtained in Phase 0 (including text and feedback).

[0055] Processing procedure: 1.1 Data preprocessing: Automated scripts clean the data, remove invalid records (such as empty dialogues or test content), and divide the data into five batches (100 records per batch) through random sampling.

[0056] 1.2 Obstacle Identification and Iteration: Semantic analysis tools (such as BERT-based classifiers) were used to initially identify problem patterns from logs, and combined with analysis of user feedback by experts (clinical doctors, public health experts), the first version of interaction disorder types were defined, including "conversation initiation failure point", "ignoring error message correction", "uncontrolled conversation duration" and "patient lack of initiative in guiding the conversation".

[0057] Machine annotation begins: Using the first batch of obstacle types as a seed, data batches are automatically annotated to generate obstacle labels (e.g., classifying dialogue fragments into specific obstacles). A collaborative team (experts and developers) conducts manual reviews and provides feedback. The iterative process continues until the last batch of data annotation and manual review is completed.

[0058] Output: Final set of interaction obstacle types

[0059] Phase 2: For each barrier type, a collaborative team (including clinical staff, developers, and user representatives) will work on each barrier type. The impact on user experience is analyzed, and learning objectives are set. After testing and feedback iteration, n learning strategies are obtained, forming a learning instruction mapping form, in the format of {[obstacle type 1: learning strategy 1],[obstacle type 2: learning strategy 2],…,[obstacle type n: learning strategy n]}.

[0060] Phase 3: Define general standards, such as video length 30-60 seconds, conversational language, virtual avatar, synchronized subtitles, and speech speed suitable for the elderly. Based on user preferences (such as cultural compatibility and emotional preferences), set up an updatable requirement form template. Parse each instruction and generate a narration script. Integrate multi-module tools (such as virtual avatar rendering and TTS audio generation) to output an initial video (V). Integrate the video into the robot and test user feedback after viewing. Update based on feedback (such as adjusting the language style) and regenerate the video. After the iteration reaches the termination condition, finally output a short educational video no longer than 60 seconds, including virtual narrator explanation, clear voice, synchronized subtitles, and content highly consistent with the input educational instructions.

[0061] Phase 4: For different user groups, the selected learning content will be precisely delivered to the designated interface location of users with low health literacy according to one of the following rules.

[0062] Rule 1 (Passive Distribution Before Consultation): When users register or log in, learning content (text and video) is displayed on the homepage, and users can choose for themselves.

[0063] Rule 2 (Proactive Distribution Before Consultation): Based on the user's history (such as previous consultation records), push matching content (such as videos of common obstacles) to the launch page.

[0064] Rule 3 (Real-time Distribution During Consultation): During a real-time conversation, if the system detects an obstacle (such as "conversation duration out of control"), it will immediately trigger the corresponding video and push it through a pop-up window or sidebar.

[0065] Feedback loop module: For example, after three months of operation, log analysis shows the addition of "information overload obstacle", triggering a complete update loop to ensure that the learning tool adapts dynamically.

[0066] To verify the effectiveness of this invention in individuals with low AI health literacy, a simulation test was designed to verify the real-world performance of the E-learning module. The experiment included 300 patients with low AI health literacy, randomly divided into an experimental group (n=150) and a control group (n=150). The control group used a basic chatbot without the methods described in this invention for health consultations. The experimental group used a chatbot integrating the complete set of technologies described in this invention (including dynamically generated and adaptively distributed E-learning modules) for consultations. After the consultations, the differences between the two groups on the following key indicators were quantitatively compared using professional assessment tools and analysis of the dialogue logs.

[0067] This experiment found that the method of the present invention can improve patients' AI health literacy by 20.6%; improve the completeness of human-computer health consultation dialogue information by 12.8%; improve the listening ability of the chat model by 7.5%; and improve the intelligent doctor's attention to patients by 12.6%. Details are shown in Table 1.

[0068] Table 1 Evaluation indicators Specific meaning Promotion amplitude Effect analysis Patient AI health literacy The ability of users to effectively understand, evaluate and use AI for health consultation 20.6% Directly proves that the invention effectively improves the core ability of target users, providing key evidence for solving the root problem. Dialogue information integrity The keyness, completeness and structuredness of health information provided by users 12.8% Indicates that users have learned how to describe their conditions more clearly and comprehensively, thereby laying a better data foundation for the model to generate accurate recommendations. Model listening ability The ability of the AI model to accurately understand the user's intention and capture key information in the dialogue 7.5% As the quality of user input information is higher (more complete and clear), the model's understanding ability is also improved accordingly. Model's attention to patients The relevance, care and ability to actively guide the dialogue of AI responses 12.6% Higher-quality interactions enable the model's deeper reasoning and empathy capabilities, making its responses more targeted and the user experience better. Currently, top-tier medical AI models have demonstrated performance approaching or even surpassing that of human experts on closed test sets. However, in practical deployments, especially in direct-to-patient (B2C) consultation scenarios, their effectiveness is significantly reduced. This is not due to insufficient knowledge or reasoning ability of the models themselves, but rather the inability to achieve high-quality interaction due to differences in user capabilities, forming the so-called "last mile problem." This invention accurately identifies and aims to solve this last mile problem. It acknowledges that handing powerful AI models to users lacking usage skills is inefficient and even dangerous. Therefore, this invention removes a key barrier to the implementation of intelligent medical consultation by improving users' AI health literacy, ensuring that the model's capabilities can be fully and correctly utilized.

[0069] Good human-computer interaction design emphasizes system usability, learnability, and accessibility. Traditional model optimization focuses only on the system's performance, neglecting user learnability. Adult learning theory states that adult learning is characterized by practicality, autonomy, experiential relevance, and immediacy. Adults are most willing to learn knowledge that can immediately solve their current practical problems. This invention is not a simple technical patchwork, but a deep integration of human-computer interaction design and adult learning theory. Its technical solutions (e.g., real-time distribution strategies in consultations) embody the principles of immediacy and practicality, providing precise guidance when users encounter specific obstacles, promoting efficient learning methods.

[0070] The following example illustrates the use of the method and solution of the present invention in a specific application.

[0071] Example: Chronic disease management consultation for the elderly User characteristics: A 65-year-old diabetic patient with low AI health literacy.

[0072] Interaction process: 1) Consultation initiated: The user enters "I am not feeling well" (vague description); 2) Obstacle recognition: The system detected an obstacle called "Dialogue Initiation Failure Point"; 3) Real-time intervention: Trigger rule 3 and push out a micro-video on "How to clearly describe symptoms"; 4) Learning outcome: After watching, the user re-entered "My blood sugar control has been poor recently, and I often feel dizzy"; 5) Quality Improvement: The model provides accurate suggestions based on clear descriptions; 6) System optimization: Through feedback loops, similar cases are continuously collected to optimize the learning content generation strategy.

[0073] Another embodiment of the present invention provides a system for improving the scalability of large language models in populations with low AI health literacy, such as... Figure 2 As shown, it includes: The main chatbot building module is used to build the main chatbot; The interaction obstacle log data acquisition module is used to acquire the dialogue log data between the user and the main chatbot; The interaction obstacle type and feature extraction module is used to extract a list of interaction obstacle types based on dialogue log data; The interaction barrier and learning instruction building module is used to build an interaction barrier-learning instruction mapping form based on the list of interaction barrier types; The multimodal learning instruction content generation module is used to generate video learning content for each learning instruction based on the interaction barrier-learning instruction mapping form; The adaptive distribution module is used to adaptively distribute video learning content and accurately deliver it to the specified interface location of the target user.

[0074] The above division of modules is merely illustrative. In practical applications, the functions described above can be assigned to different functional modules as needed to complete all or part of the functions described in the aforementioned method. The specific working process of each module can be found in the corresponding process in the aforementioned method embodiments, and will not be repeated here.

[0075] Another embodiment of the present invention provides a computer device (computer, server, smartphone, etc.) including a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing steps of the method of the present invention.

[0076] Another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, disk, optical disk) that stores a computer program, which, when executed by a computer, implements the steps of the method of the present invention.

[0077] Another embodiment of the present invention provides a computer program product, the computer program product including a computer program, which, when executed by a computer, implements the steps of the method of the present invention.

[0078] The specific embodiments of the present invention disclosed above are intended to help understand the content of the present invention and to implement it accordingly. Those skilled in the art will understand that various substitutions, changes, and modifications are possible without departing from the spirit and scope of the present invention. The present invention should not be limited to the content disclosed in the embodiments of this specification; the scope of protection of the present invention is defined by the claims.

Claims

1. A method for improving the scalability of large language models in populations with low AI health literacy, characterized in that, Includes the following steps: Build a main chatbot and obtain the dialogue log data between the user and the main chatbot; Extract a list of interaction barrier types based on the dialogue log data; Based on the list of interaction barrier types, construct an interaction barrier-learning instruction mapping form; Based on the interaction barrier-learning instruction mapping form, generate video learning content for each learning instruction; Adaptive distribution of video learning content, precisely targeting specific screen locations for the intended users.

2. The method according to claim 1, characterized in that, The step of extracting a list of interaction obstacle types based on dialogue log data includes: Perform data preprocessing, including cleaning the data, removing empty dialogue records or invalid interactions, filtering out duplicate or test-related non-user-generated content, and dividing the data into multiple batches by random sampling based on the total amount of data and computing resources; Semantic analysis methods were used to initially identify common problem patterns from dialogue log data. Combined with collected user feedback and experience data, model developers and domain experts collaborated to conduct a comprehensive analysis to provide an initial version of the definition of interaction barrier types and a list of specific manifestations. Using machine learning methods, the first version of the obstacle type is input as the initial seed. Starting from the first batch of data, the dialogue logs are automatically labeled to generate obstacle labels, which are then manually checked by the collaborative team. Feedback is collected, and the obstacle type definition or machine learning model is adjusted based on the feedback. The machine labeling and manual checking process is repeated for subsequent batches, iterating until the last batch of data is completed, and a complete list of interaction obstacle types and their specific manifestations is output.

3. The method according to claim 1, characterized in that, The step of constructing an interaction barrier-learning instruction mapping form based on the list of interaction barrier types includes: Based on the list of interaction obstacle types, we gain a deep understanding of how each obstacle affects the user experience and learning process, set clear learning objectives for each obstacle type, and formulate multiple possible learning instructions based on the set learning objectives to form a set of potential learning instructions. The set of potential learning instructions is temporarily integrated into the main chatbot for testing and feedback iteration. During the testing process, when a certain type of interaction obstacle is detected in the user interaction, a potential learning instruction text format is automatically and randomly selected and presented to the tester. After receiving the instruction and learning the content, the tester continues to talk to the chatbot. After each round of testing, the effectiveness index of each potential learning instruction is collected, and the effectiveness score is calculated based on the weighted average. Based on the effectiveness score, the content of the learning instruction is added or deleted. For deleted instructions, the expert team designs new instruction variants based on community user experience and feedback, adds them to the set of potential learning instructions, and repeats the testing, feedback collection, and decision-making steps until the termination condition is met, outputting an interaction barrier-learning instruction mapping form.

4. The method according to claim 1, characterized in that, The step of generating video learning content for each learning instruction based on the interaction barrier-learning instruction mapping form includes: Develop a basic requirements form based on the input learning instructions. and actual needs form Output the initial requirements form ; In the initial requirement form Based on this foundation, each learning instruction is analyzed, and a conversational and easy-to-understand video narration / explanation script is generated based on the analysis results. Natural language generation technology is used to ensure that the script conforms to [the required language]. and Based on the script, multiple modules are integrated to output preliminary video learning content; Initial video learning content is integrated into the main chatbot. When a specific type of obstacle is detected in real time during user interaction, the corresponding video content is automatically triggered and presented to the test user. The user watches the video while interacting with the chatbot and then continues to interact. User feedback is collected after each round of testing, and the forms are updated based on the user feedback to meet actual needs. This generates an updated requirements form. ; Once the termination condition is met, the final video learning content for each learning instruction is output. The video learning content has multimodal characteristics, including virtual avatars, audio explanations, and synchronized subtitles, to ensure easy understanding and access.

5. The method according to claim 1, characterized in that, The adaptive distribution of video learning content employs the following distribution rules: Rule 1, Passive Distribution Before User Consultation: When a new user completes registration or an existing user successfully logs in, text and video learning content will be automatically displayed on the user's homepage and consultation startup page. Users can choose to watch based on their personal interests and needs. Rule 2, proactive distribution before user consultation: When a new user registration or a user's first consultation intent after logging in is detected, the learning content most closely matching the user's characteristics is automatically pushed to their homepage or consultation launch page by combining the user's historical status information and historical information. Rule 3, Real-time Distribution During User Consultations: During a user's consultation session, if a specific interaction obstacle is detected in real time, the rule engine will be immediately triggered. Based on the specific obstacle type detected, the rule engine will match the generated corresponding video learning content and accurately push the video to the user in the current session through instant message push, pop-up prompts, or sidebar recommendations.

6. The method according to claim 1, characterized in that, It also includes a feedback loop step, which periodically monitors user interactions and triggers content updates when new obstacles or significant changes are detected to ensure dynamic adaptability.

7. A system for improving the scalability of large language models in populations with low AI health literacy, characterized in that, include: The main chatbot building module is used to build the main chatbot; The interaction obstacle log data acquisition module is used to acquire the dialogue log data between the user and the main chatbot; The interaction obstacle type and feature extraction module is used to extract a list of interaction obstacle types based on dialogue log data; The interaction barrier and learning instruction building module is used to build an interaction barrier-learning instruction mapping form based on the list of interaction barrier types; The multimodal learning instruction content generation module is used to generate video learning content for each learning instruction based on the interaction barrier-learning instruction mapping form; The adaptive distribution module is used to adaptively distribute video learning content and accurately deliver it to the specified interface location of the target user.

8. The system according to claim 7, characterized in that, It also includes a feedback loop module for periodically monitoring user interactions and triggering content updates when new obstacles or significant changes are detected, to ensure dynamic adaptability.

9. A computer device, characterized in that, It includes a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the method of any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a computer, implements the method according to any one of claims 1 to 6.

Images