Personalized learning-oriented medical digital teaching material intelligent construction system and method

Abstract

The invention belongs to the field of medical education, and particularly relates to a medical digital textbook intelligent construction system and method for personalized learning, and the system comprises a medical knowledge graph construction and management module which is used for constructing and storing a medical knowledge graph; the learner modeling and analyzing module is used for establishing and updating a learner cognitive model, and the cognitive model is associated with entity nodes in the medical knowledge graph through node unique identifiers; the learning module is also used for outputting a mastery degree value of the target learning content through the learner cognitive model according to the cognitive state information of the learner; the personalized content engine module is used for dynamically selecting and combining content atoms based on the knowledge sub-graph and the mastery degree value corresponding to the target learning content in combination with a preset teaching rule, and generating a personalized teaching material page sequence; and the interaction and presentation module is used for rendering the page sequence into a visual user interface. The problems that an existing medical digital teaching material is fixed and rigid in content organization and poor in knowledge association characterization capacity are solved.

Description

Technical Field

[0001] This invention belongs to the field of medical education, and in particular relates to an intelligent construction system and method for digital medical textbooks oriented towards personalized learning. Background Technology

[0002] Currently, the digital teaching materials widely used in medical education (such as e-books, PDFs, and online course pages) are essentially digital replicas of paper textbooks. Their content organization is fixed, linear, and uniform, which cannot adapt to the significant differences among learners, such as medical students, interns, and specialist trainees in terms of knowledge background, learning stage, cognitive ability, and personal learning goals.

[0003] Most existing so-called adaptive learning platforms adopt a simple knowledge point-question bank mapping model. They determine a learner's mastery of a certain knowledge tag based on test results and recommend pre-set, complete learning modules or questions accordingly. This approach has obvious drawbacks: 1. The adaptive unit is the entire chapter, section, or set of questions, and it cannot make fine-tuning adjustments to the specific content within a chapter.

[0004] 2. The medical knowledge system contains complex and rigorous logical connections (such as the connections between etiology, pathology, clinical manifestations, diagnosis, and treatment), while a simple tagging system cannot represent this deep network relationship, resulting in personalized recommendations lacking disciplinary logical support.

[0005] 3. Failed to continuously model the changes in learners' knowledge mastery over time and due to the transfer of knowledge relevance; personalized strategies were singular and static. Summary of the Invention

[0006] This invention provides an intelligent construction system and method for medical digital textbooks oriented towards personalized learning, which solves the problems of fixed and rigid content organization and weak knowledge association representation ability in existing medical digital textbooks.

[0007] This invention provides a basic solution: an intelligent construction system for digital medical textbooks oriented towards personalized learning, comprising: The Medical Knowledge Graph Construction and Management Module is used to construct and store medical knowledge graphs represented in a graph structure. The learner modeling and analysis module is used to establish and update learner cognitive models, which are associated with entity nodes in the medical knowledge graph through unique node identifiers; it is also used to output the mastery value of the target learning content based on the learner's cognitive state information through the learner cognitive model. The personalized content engine module dynamically selects and combines content atoms based on the knowledge subgraph and mastery value corresponding to the target learning content, combined with preset teaching rules, to generate a personalized textbook page sequence. The interaction and presentation module is used to render the page sequence generated by the personalized content engine module into a visual user interface.

[0008] Preferably, it also includes a teaching content atom library construction module, which parses the original electronic versions of medical textbooks, clinical diagnosis and treatment guidelines, and standardized medical terminology systems into multiple content atoms based on the experience of experts; it is also used to associate content atoms with entity nodes of the medical knowledge graph; each content atom includes content carrier data, presentation and interaction definitions, semantic and teaching metadata.

[0009] More preferably, the medical knowledge graph construction and management module includes a sub-module for original textbook data acquisition and preprocessing, a sub-module for joint extraction of entity relationships, a sub-module for knowledge fusion and conflict resolution, and a sub-module for graph storage and query interface; The original textbook data acquisition and preprocessing submodule is used to preprocess the original textbook data, which includes electronic versions of original medical textbooks, clinical diagnosis and treatment guidelines, and standardized medical terminology systems. The entity relationship extraction submodule is used to automatically identify medical knowledge entities and semantic relationships between entities from preprocessed standardized text, forming preliminary knowledge triples. The knowledge fusion and conflict resolution submodule is used to integrate the extraction results and construct a knowledge graph. The graph storage and query interface submodule is used to persistently store the fused, standardized knowledge graph data.

[0010] More preferably, the learner modeling and analysis module includes a multi-source data acquisition submodule and a knowledge tracing model submodule. The multi-source data acquisition submodule generates learner cognitive state information based on interaction log data; The knowledge tracking model submodule, based on the Transformer architecture deep knowledge tracking model, takes the temporal cognitive state feature sequence as input, encodes it through a multi-head self-attention mechanism that integrates knowledge graph structural information, generates a comprehensive state vector, and finally maps it to a mastery probability vector that corresponds one-to-one with the nodes of the medical knowledge graph.

[0011] More preferably, the deep knowledge tracing model includes an input layer, a knowledge graph-enhanced multi-head self-attention encoding layer, a state convergence layer, and an output layer. The input layer is used to receive a sequence of cognitive state features; The knowledge graph-enhanced multi-head self-attention encoding layer is composed of multiple identical encoder blocks stacked together, each encoder block containing a multi-head self-attention sublayer and a feedforward neural network sublayer; The state convergence layer, after multiple layers of encoding, uses the vector representation of the last time step of the corresponding sequence output by the last encoder block as the comprehensive state vector of the learner's current overall knowledge state. The output layer is a fully connected layer with the number of neurons equal to the total number of core entity nodes defined in the medical knowledge graph. It is used to input the comprehensive state vector obtained in the previous step into this layer. After linear transformation, the output value is mapped to the interval [0, 1] using the Sigmoid activation function to obtain the mastery probability vector. More preferably, the graph structure modulation attention mechanism specifically involves calculating attention weights in the self-attention sub-layer based on the dot product similarity between the query vector and the key vector, combined with a prior bias matrix based on the medical knowledge graph. The specific calculation steps are as follows: 1) Calculate the scaling dot product attention score; 2) Construct the graph adjacency bias matrix This matrix has the same dimensions as the attention score matrix, and its elements... The value is determined by the strength of the relationship between the two knowledge point entities corresponding to the current sequence positions i and j in the knowledge graph. The relationship strength can be pre-calculated and assigned based on the edge type and path distance in the graph. 3) The graph bias matrix is ​​added to the calculation of the attention score to form the modulated attention weight.

[0012] More preferably, the total loss function of the deep knowledge tracing model includes the main loss and the graph regularization loss; The expression for the main loss is:

[0013] In the formula, M is the number of samples in a training batch. It is the true boundary result of the m-th sample; It is the predicted probability of the interaction result corresponding to the model; The expression for the graph regularization loss is:

[0014] In the formula, E is the set of all directed edges in the medical knowledge graph, representing entity pairs with predefined semantic relationships; are model pairs of edges, and These are the mastery predictions output by the head entity i and the tail entity j in the edge (i, j) of the model, respectively. This is the weight coefficient of the edge, which is set according to the importance or confidence level of the relation type.

[0015] More preferably, the personalized content engine module includes a teaching rule reasoning and triggering submodule, a content atom retrieval and adaptation submodule, and a teaching path planning and sequence generation submodule. The teaching rule reasoning and triggering submodule is used to generate a sequence of teaching action instructions by querying and matching the condition-action rules in the pre-set teaching rule base, based on the knowledge subgraph and mastery value corresponding to the current learning objective and the symbolic reasoning method based on the production rule system. The content atom filtering and adaptation submodule is used to filter out a candidate set of content atoms from the library based on the sequence of teaching action instructions and the set of knowledge points covered by the knowledge subgraph by querying the metadata index of the content atom library and calculating the adaptation degree. The teaching path planning and sequence generation submodule generates a structured textbook page sequence based on the topological structure of the knowledge subgraph and the candidate set of content atoms, using a graph theory-based path planning algorithm.

[0016] More preferably, the interaction and presentation module includes a rendering engine submodule, an interaction manager submodule, and a log collection and reporting submodule; The rendering engine submodule is used to parse the structured textbook page sequence and call the front-end component library and templates to dynamically generate and assemble the visual user interface displayed on the application. The interaction manager submodule is used to bind preset event listeners to interactive elements in the interface when the rendering engine generates the user interface, and to define the immediate front-end response logic after the interactive event is triggered. The log collection and reporting submodule is used to listen to all user behavior events on the interface, record and temporarily store these interaction log data in a structured format in real time, and send them asynchronously to the backend learner modeling and analysis module via the network.

[0017] Another basic solution provided by this invention: an intelligent construction method for medical digital teaching materials oriented towards personalized learning, applied to the above system, including: S1: Constructing a medical knowledge graph; S2: In response to a learner starting to learn, initialize or load the learner's cognitive model; S3: Based on the learner's personalized textbook page sequence, call the learner's cognitive model, calculate the mastery value corresponding to the learner's target learning content, and query the knowledge subgraph corresponding to the learner's target learning content; S4: Based on preset teaching rules, combined with knowledge subgraphs and mastery values, dynamically select and combine content atoms to generate a personalized textbook page sequence; S5: Generates front-end pages based on personalized textbook page sequences for learners to interact with, and records learners' interactive behaviors.

[0018] The principles and advantages of this invention are as follows: 1. A medical knowledge graph is constructed using a graph structure, which breaks down medical knowledge into knowledge point entity nodes and defines the logical connections between knowledge points as the connection edges between nodes, forming a hierarchical and logically rigorous knowledge network. This breaks through the limitations of the traditional tag system, accurately represents the complex logical connections within medical knowledge, provides solid disciplinary logic support for personalized content recommendations, and avoids problems such as logical breaks or lack of disciplinary relevance in recommended content.

[0019] 2. By establishing a mapping relationship between learner cognitive models and knowledge graph entity nodes through unique node identifiers, the mastery values ​​of each knowledge point are dynamically updated based on learners' cognitive states and time-series data. This captures the migration and change patterns of knowledge mastery, enabling dynamic modeling of learners' cognitive states. This breaks through the limitations of existing static label-based mastery assessments and can accurately reflect learners' cognitive changes over time and the influence of knowledge associations, providing a real-time and accurate basis for personalized strategy adjustments.

[0020] 3. The content of medical textbooks is broken down into the smallest granular content atoms. Based on the logical structure of the target knowledge subgraph and the learner's knowledge mastery level, combined with preset teaching rules, the content atoms are dynamically selected and combined to generate a non-fixed personalized textbook page sequence. This breaks the fixed linear organization of traditional digital medical textbooks and realizes refined and personalized content reorganization at the knowledge point level. It adapts to the differences in knowledge background, learning stage and cognitive ability of different groups such as medical students, interns and specialist trainees, and solves the defects of coarse-grained recommendation units in existing adaptive platforms. Attached Figure Description

[0021] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0022] The following detailed description illustrates the specific implementation method: The specific implementation process is as follows: Example 1 See Figure 1 A smart system for constructing digital medical textbooks for personalized learning includes: The Medical Knowledge Graph Construction and Management module is used to construct and store medical knowledge graphs represented in a graph structure.

[0023] The medical knowledge graph construction and management module includes a sub-module for original textbook data acquisition and preprocessing, a sub-module for entity relation extraction, a sub-module for knowledge fusion and conflict resolution, and a sub-module for graph storage and query interface. The original textbook data acquisition and preprocessing submodule is used to acquire and preprocess original textbook data, which includes electronic versions of original medical textbooks, clinical diagnosis and treatment guidelines, and standardized medical terminology systems. During data collection, raw textbook data is obtained from online resources via web crawlers or API calls. Based on the data format, appropriate parsers are used to extract content from the raw textbook data. This includes using regular expressions and dedicated parsing libraries to extract plain text and metadata from formats such as PDF and HTML. Specifically, for structured / semi-structured data, such as JSON and XML, corresponding parsing libraries are used; for unstructured text data, such as HTML, PDF, and Word, regular expressions, specific format parsers, OCR engines, and natural language processing models are used. This yields readable, continuous text strings and related structured fields. Finally, the obtained structured data is cleaned, structured, and standardized. This includes removing noise unrelated to medical entities, dividing the text into sentence or paragraph units, initially mapping the terminology in the text to the standard medical terminology system, and converting the text into a numerical feature representation that can be processed by machine learning models.

[0024] The entity relationship extraction submodule is used to automatically identify medical knowledge entities and semantic relationships between entities from preprocessed standardized text, forming preliminary knowledge triples. Specifically, the preprocessed standardized text is jointly extracted using a deep learning method based on a pre-trained medical language model. The internal encoder of the medical language model first encodes the context, outputting a context-related vector representation for each lexical unit. Head extraction is divided into two parallel or cascaded subtasks: the entity recognition subtask performs sequence labeling, identifying and classifying medical entities in the text, such as diseases, symptoms, drugs, and examination methods; the relation classification subtask classifies the identified entity pairs, determining whether there are predefined semantic relation types between them, such as "cause of disease," "clinical manifestation," and "treatment method." The model is trained end-to-end, learning to directly output "head entity-relation-tail entity" triples from the original text. Finally, a structured extraction result is output, including the identified medical entities and their types, as well as relation triples between entities, providing raw materials for subsequent knowledge fusion.

[0025] The knowledge fusion and conflict resolution submodule is used to integrate the extraction results and construct a knowledge graph. Specifically, the extracted results are first aligned to entity identity, and the similarity between entities described from different sources is calculated, including similarity based on name strings, similarity based on attributes, and graph embedding similarity based on contextual relationships. When the overall similarity exceeds a preset threshold, they are identified as the same entity and merged. Then, for the same entity with the same attribute or relationship described differently, an evidence-based decision algorithm is executed. Specifically, a confidence weight is assigned to each data source, and a timeliness decay factor is assigned to each piece of information. Through weighted voting or a probability model, the information with the highest overall confidence is adopted as the standard value, and other conflicting information is marked as alternatives or historical versions, while recording their source and level of evidence. Finally, a standardized medical knowledge graph data model that has been disambiguated, deduplicated, and labeled with version and source is output.

[0026] The graph storage and query interface submodule is used to persistently store the fused standardized knowledge graph data. Specifically, the standardized medical atlas data model is imported into a dedicated graph database for storage. In the graph database, knowledge entities are stored as nodes, semantic relationships are stored as edges, and entity and relationship attributes are stored as key-value pairs. Furthermore, this submodule provides a set of application programming interfaces (APIs) by encapsulating the graph database's native query language. These interfaces support various query modes, including exact queries based on entity ID or name, conditional queries based on entity type and attributes, graph traversal queries based on relationship paths, and multi-hop inference queries. All query operations are executed through an optimized graph query engine to ensure millisecond-level response times even with massive amounts of medical knowledge data.

[0027] The teaching content atom library construction module, based on expert experience, parses original electronic versions of medical textbooks, clinical diagnosis and treatment guidelines, and standardized medical terminology systems into multiple content atoms; it is also used to associate content atoms with entity nodes of the medical knowledge graph; each content atom includes content carrier data, presentation and interaction definitions, semantic and teaching metadata.

[0028] Specifically, based on expert analysis of the original materials at the pedagogical level, an atom decomposition strategy and metadata scheme are planned. The content carrier data is the core data that carries specific teaching information, and its forms include, but are not limited to: plain text, rich text (including formatting), image data, audio and video streams, executable interactive scripts, parameterized templates, etc. Presentation and interaction are defined to describe how the atom is rendered on the front-end interface, and to define the interactive behaviors that learners can perform with it and their corresponding system response logic. Semantics and teaching metadata are used to describe and retrieve the tags of the atom, including: associated knowledge points, unique identifiers of one or more medical knowledge graph entity nodes explained by the atom; teaching functions, such as "concept definition", "principle explanation", "positive example demonstration", "negative example analysis", "standard procedure", "clinical decision-making exercise", "formative assessment", etc.; cognitive attributes, estimated cognitive load level, suggested prerequisite knowledge, and target audience level (e.g., undergraduate / resident physician).

[0029] The learner modeling and analysis module is used to establish and update learner cognitive models, which are associated with entity nodes in the medical knowledge graph through unique node identifiers; it is also used to output the mastery value of the target learning content based on the learner's cognitive state information through the learner cognitive model. The learner modeling and analysis module includes a multi-source data acquisition submodule and a knowledge tracing model submodule. The multi-source data acquisition submodule generates learner cognitive state information based on interaction log data. The interaction log data includes answer record data (correctness / incorrectness, reaction time), content consumption data (video viewing progress, page dwell time, completion rate), external evaluation data (standardized test scores, skills assessment scores), direct feedback data (self-assessment of "mastery / lack of mastery" of knowledge points or content, difficulty rating), active annotation and note-taking data (highlighted text, annotation content, note frequency and location), search and query behavior data (search keywords within the system, query frequency, query context), metacognitive and regulatory behavior data (marking "key points / difficult points", customizing learning plans, adjusting learning goals), and interaction detail data (mouse hover trajectory, page scrolling mode, quick jumps and replay sequences between different content blocks).

[0030] Specifically, cognitive state information includes mastery characteristics, engagement characteristics, and ability benchmark characteristics. Mastery characteristics include the correct answer rate for knowledge point k within a time window t; the average response time for knowledge point k within a time window t; and the standard deviation or coefficient of variation of the response time. Engagement characteristics include the proportion of effective dwell time on content atom a (related to knowledge point k) to the total time; the completion rate of video content; and the total duration and frequency of continuous learning sessions. Ability benchmark characteristics include standardized scores mapped from external standardized test scores to knowledge graph nodes; quantitative values ​​of self-assessment feedback, such as converting "mastery / not mastery" into binary labels; confusion and attention characteristics; metacognition and strategy characteristics; and cognitive negative... Features of cognitive load and fluency, and features of confusion and attention include the frequency of text related to knowledge point k being highlighted or annotated, the frequency of search queries related to knowledge point k, and the total duration and trajectory density of mouse hover over the content area explaining knowledge point k; features of metacognition and strategy include the set of knowledge points actively marked as "difficult" by learners, the number of times the learning plan is customized and the consistency of execution, and the pattern of review behavior, such as the frequency of backtracking to a specific theoretical chapter after answering a question incorrectly; features of cognitive load and fluency include abrupt changes in page scrolling speed, and the distribution of time intervals for interactive events, such as alternation between dense and rapid clicks and long pauses, which may reflect cognitive struggles, and changes in behavioral patterns from smooth linear browsing to non-linear frequent jumps, etc. Finally, all cognitive state information is combined into a time-evolving, multi-dimensional feature vector sequence as input to the learner's cognitive model, used to estimate and predict the learner's dynamic mastery of each knowledge node. The knowledge tracing model submodule, based on a deep knowledge tracing model with a Transformer architecture, takes the temporal cognitive state feature sequence as input and encodes it through a multi-head self-attention mechanism that integrates knowledge graph structural information, generating a comprehensive state vector, which is ultimately mapped to a mastery probability vector corresponding one-to-one with each node in the medical knowledge graph. The deep knowledge tracing model is linked to the medical knowledge graph in the following ways: 1) Each neuron in the output layer of the deep knowledge tracing model corresponds to a specific entity node in the knowledge graph through a unique identifier for the entity node; 2) During model training, the optimization objective includes a regularization loss term based on the strength of relationships between knowledge graph nodes, which constrains the model's output on associated nodes; during state updates, the graph-enhanced attention mechanism is modulated by the adjacency relationships of the knowledge graph.

[0031] Specifically, the deep knowledge tracing model includes an input layer, a knowledge graph-enhanced multi-head self-attention encoding layer, a state convergence layer, and an output layer. The input layer is used to receive a sequence of cognitive state features; The knowledge graph-enhanced multi-head self-attention encoding layer is composed of multiple identical encoder blocks stacked together, each encoder block containing a multi-head self-attention sublayer and a feedforward neural network sublayer; Specifically, the graph structure modulation attention mechanism involves calculating attention weights in the self-attention sub-layer based on the dot product similarity between the query vector and the key vector, combined with a prior bias matrix based on the medical knowledge graph. The specific steps are as follows: 1) Calculate the scaling dot product attention score; 2) Construct the graph adjacency bias matrix This matrix has the same dimensions as the attention score matrix, and its elements... The value is determined by the strength of the relationship between the two knowledge point entities corresponding to the current sequence positions i and j in the knowledge graph. The relationship strength can be pre-calculated and assigned based on the edge type and path distance in the graph. 3) The spectral bias matrix is ​​added to the calculation of the attention score to form the modulated attention weights; The feedforward neural network sublayer performs a nonlinear transformation on the output of the self-attention sublayer, enhancing the model's expressive power.

[0032] The state convergence layer, after multiple layers of encoding, uses the vector representation of the last time step of the corresponding sequence output by the last encoder block as a comprehensive state vector of the learner's current overall knowledge state. This vector compresses all the effective information and knowledge structure associations of the historical interaction sequence.

[0033] The output layer is a fully connected layer with the number of neurons equal to the total number N of core entity nodes defined in the medical knowledge graph. It is used to input the comprehensive state vector obtained in the previous step into this layer. After linear transformation, the output value is mapped to the interval [0, 1] using the Sigmoid activation function. Finally, an N-dimensional mastery probability vector is obtained. The i-th component of this vector represents the learner's current mastery probability of the i-th entity node in the knowledge graph, as predicted by the model.

[0034] The total loss function of a deep knowledge tracing model includes the main loss and the graph regularization loss; The expression for the main loss is:

[0035] In the formula, M is the number of samples in a training batch. It is the true boundary result of the m-th sample; It is the predicted probability of the interaction result corresponding to the model; The expression for the graph regularization loss is:

[0036] In the formula, E is the set of all directed edges in the medical knowledge graph, representing entity pairs with predefined semantic relationships; are model pairs of edges, and These are the mastery predictions output by the head entity i and the tail entity j in the edge (i, j) of the model, respectively. This is the weight coefficient of the edge, which is set according to the importance or confidence level of the relation type; The deep knowledge tracking model employs a two-stage training strategy. First, it is pre-trained on a large dataset of medical learner interactions, enabling the model to learn the mapping from general interaction sequences to knowledge states. After deployment, online learning or periodic incremental learning is used to fine-tune the model parameters using new data generated by the system, allowing it to continuously adapt to changes in data distribution. During the training phase, an adaptive moment estimation optimizer is used for parameter optimization. Learning rate warm-up and cosine annealing strategies are employed during training to stabilize the model and improve its final performance. To prevent overfitting, dropout techniques are extensively used in the encoder and feedforward layers.

[0037] The personalized content engine module dynamically selects and combines content atoms based on the knowledge subgraph and mastery value corresponding to the target learning content, combined with preset teaching rules, to generate a personalized textbook page sequence. The personalized content engine module includes a teaching rule reasoning and triggering submodule, a content atom retrieval and adaptation submodule, and a teaching path planning and sequence generation submodule. The teaching rule reasoning and triggering submodule is used to generate a sequence of teaching action instructions by querying and matching the condition-action rules in the pre-set teaching rule base, based on the knowledge subgraph and mastery value corresponding to the current learning objective and the symbolic reasoning method based on the production rule system. The pre-built teaching rule base includes the following: Cognitive state response rules are used to execute different preset actions based on the learner's current mastery level of a specific knowledge point. For example, when the mastery level of a target knowledge point is determined to be lower than the preset basic mastery threshold, and the knowledge point type is marked as a core concept, firstly, a content atom marked as a core concept in detail is forcibly inserted. This atom usually contains the most basic and core definitions and principles. Secondly, the system will delay the presentation of complex application cases related to this core concept to ensure that learners fully understand the basics before encountering advanced applications, avoiding cognitive overload. When the mastery level of a target knowledge point is higher than the preset proficiency threshold, the default teaching content will be replaced with a knowledge graph relationship view that can show the position of the knowledge point in the overall knowledge network, or a review of the latest cutting-edge developments in the field. At the same time, advanced clinical decision-making questions are inserted to challenge and deepen the learner's comprehensive application and clinical reasoning abilities. Knowledge structure dependency rules are used to ensure the logical coherence and cognitive rationality of teaching content, and to force the system to follow the inherent logic of the discipline, which is to start with the basics and then apply them, and to address the root cause and then treat the problem. For example, when planning a learning path, the system will continuously monitor the mastery of all prerequisite knowledge points that the current target knowledge point depends on. If the mastery of any prerequisite knowledge point is found to be lower than the understanding threshold, the system will automatically insert a prerequisite knowledge review module at the beginning of the current chapter. Learner profile adaptation rules are used to adjust the depth, breadth, and focus of content based on the learner's identity as displayed in their user information. For example, when the learner is identified as an undergraduate clinical medicine student, the system will set the textbook description as the dominant content version, that is, it will prioritize the authoritative and systematic expositions in classic textbooks as the main content framework. At the same time, the focus of the teaching content will be placed on in-depth explanations of pathophysiological mechanisms to help undergraduates build a solid theoretical foundation. The evidence conflict decision rule is used to intelligently decide the content presentation strategy based on the teaching context when the same knowledge point in the knowledge graph has multiple sources and versions of descriptions (such as textbook theories, the latest clinical guidelines, and different academic viewpoints). For example, when the system detects differences in recommendations between different authoritative guidelines for the same clinical problem, such as the inconsistent recommendations of Guideline A and Guideline B, the system will also use a comparative description method, juxtaposing the recommendations of different guidelines and inserting a structured comparison table of the applicable populations and conditions of different guidelines. This helps learners clearly understand the scope and preconditions of various recommendations, enabling them to make the most reasonable choice based on the specific situation in future clinical practice.

[0038] The content atom filtering and adaptation submodule is used to filter out a candidate set of content atoms from the library based on the sequence of teaching action instructions and the set of knowledge points covered by the knowledge subgraph by querying the metadata index of the content atom library and calculating the adaptation degree. The teaching path planning and sequence generation submodule generates a structured textbook page sequence based on the topological structure of the knowledge subgraph and the candidate set of content atoms, using a graph theory-based path planning algorithm.

[0039] The interaction and presentation module is used to render the page sequence generated by the personalized content engine module into a visual user interface, and also to collect learners' interaction log data and send it to the learner modeling and analysis module.

[0040] The interaction and presentation module includes a rendering engine submodule, an interaction manager submodule, and a log collection and reporting submodule. The rendering engine submodule is used to parse the structured textbook page sequence and call the front-end component library and templates to dynamically generate and assemble the final visual user interface that can be displayed on a web browser or mobile application. Specifically, the rendering engine submodule contains an interpreter that identifies layout instructions (such as "place content atom A in the left sidebar"), style instructions (such as "highlight"), and resource links (such as image URLs and video sources) in the page description, and converts the instructions into specific HTML, CSS, and JavaScript code, and instantiates the corresponding interactive components (such as video players, collapsible panels, and interactive charts), and finally outputs a complete and interactive web page view. The interaction manager submodule is used to bind preset event listeners to interactive elements (such as buttons, question selection items, video control bars, and note highlighting tools) in the interface when the rendering engine generates the user interface, and to define the immediate front-end response logic after the interactive event is triggered, so as to provide smooth interactive feedback. Specifically, based on the interaction definition metadata defined in the content atom, front-end logic is configured for each interaction point. For example, when a learner clicks on a multiple-choice option, the module will immediately process the event, determine whether it is correct or incorrect, and provide immediate feedback on the interface (such as displaying a checkmark or cross and explaining the reason) without waiting for a server response, thus ensuring the immediacy of the interaction. The log collection and reporting submodule is used to listen to all user behavior events on the interface (bound by the interaction manager), record and temporarily store these interaction log data in a structured format in real time, and send them asynchronously to the backend learner modeling and analysis module via the network. The interaction log data includes: Event types (e.g., click, answer submission, video pause); Related objects (such as content atom ID, knowledge point ID, question ID); Event details (such as selected answer, video timestamp, and notes text); Timestamp (precise client time); Interactive log data is encapsulated in a standard format (such as JSON) and sent to the server's log receiving interface via HTTP request or WebSocket connection.

[0041] Example 2 A method for intelligently constructing digital medical textbooks for personalized learning, applied to the aforementioned system, includes: S1: Constructing a medical knowledge graph; S2: In response to a learner starting to learn, initialize or load the learner's cognitive model; S3: Based on the learner's personalized textbook page sequence, call the learner's cognitive model, calculate the mastery value corresponding to the learner's target learning content, and query the knowledge subgraph corresponding to the learner's target learning content; S4: Based on preset teaching rules, combined with knowledge subgraphs and mastery values, dynamically select and combine content atoms to generate a personalized textbook page sequence; S5: Generates front-end pages based on personalized textbook page sequences for learners to interact with, and records learners' interactive behaviors.

[0042] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A smart construction system for digital medical textbooks oriented towards personalized learning, characterized in that, include: The Medical Knowledge Graph Construction and Management Module is used to construct and store medical knowledge graphs represented in a graph structure. The learner modeling and analysis module is used to establish and update learner cognitive models, which are associated with entity nodes in the medical knowledge graph through unique node identifiers; it is also used to output the mastery value of the target learning content based on the learner's cognitive state information through the learner cognitive model. The personalized content engine module dynamically selects and combines content atoms based on the knowledge subgraph and mastery value corresponding to the target learning content, combined with preset teaching rules, to generate a personalized textbook page sequence. The interaction and presentation module is used to render the page sequence generated by the personalized content engine module into a visual user interface.

2. The intelligent construction system and method for medical digital teaching materials oriented towards personalized learning according to claim 1, characterized in that: It also includes a module for constructing a teaching content atom library, which, based on the experience of experts, parses the original electronic versions of medical textbooks, clinical diagnosis and treatment guidelines, and standardized medical terminology systems into multiple content atoms; it is also used to associate content atoms with entity nodes of the medical knowledge graph; each content atom includes content carrier data, presentation and interaction definitions, semantic and teaching metadata.

3. The intelligent construction system and method for medical digital teaching materials oriented towards personalized learning according to claim 2, characterized in that: The medical knowledge graph construction and management module includes a sub-module for original textbook data acquisition and preprocessing, a sub-module for entity relation extraction, a sub-module for knowledge fusion and conflict resolution, and a sub-module for graph storage and query interface. The original textbook data acquisition and preprocessing submodule is used to acquire and preprocess original textbook data, which includes electronic versions of original medical textbooks, clinical diagnosis and treatment guidelines, and standardized medical terminology systems. The entity relationship extraction submodule is used to automatically identify medical knowledge entities and semantic relationships between entities from preprocessed standardized text, forming preliminary knowledge triples. The knowledge fusion and conflict resolution submodule is used to integrate the extraction results and construct a knowledge graph. The graph storage and query interface submodule is used to persistently store the fused, standardized knowledge graph data.

4. The intelligent construction system and method for medical digital teaching materials oriented towards personalized learning according to claim 2, characterized in that: The learner modeling and analysis module includes a multi-source data acquisition submodule and a knowledge tracing model submodule. The multi-source data acquisition submodule generates learner cognitive state information based on interaction log data; The knowledge tracking model submodule, based on the Transformer architecture deep knowledge tracking model, takes the temporal cognitive state feature sequence as input, encodes it through a multi-head self-attention mechanism that integrates knowledge graph structural information, generates a comprehensive state vector, and finally maps it to a mastery probability vector that corresponds one-to-one with the nodes of the medical knowledge graph.

5. The intelligent construction system and method for medical digital teaching materials oriented towards personalized learning according to claim 4, characterized in that: The deep knowledge tracing model includes an input layer, a knowledge graph-enhanced multi-head self-attention encoding layer, a state convergence layer, and an output layer. The input layer is used to receive a sequence of cognitive state features; The knowledge graph-enhanced multi-head self-attention encoding layer is composed of multiple identical encoder blocks stacked together, each encoder block containing a multi-head self-attention sublayer and a feedforward neural network sublayer; The state convergence layer, after multiple layers of encoding, uses the vector representation of the last time step of the corresponding sequence output by the last encoder block as the comprehensive state vector of the learner's current overall knowledge state. The output layer is a fully connected layer with the number of neurons equal to the total number of core entity nodes defined in the medical knowledge graph. It is used to input the comprehensive state vector obtained in the previous step into this layer. After linear transformation, the output value is mapped to the interval [0, 1] using the Sigmoid activation function to obtain the mastery probability vector.

6. The intelligent construction system and method for medical digital teaching materials oriented towards personalized learning according to claim 4 or 5, characterized in that: Specifically, the graph structure modulation attention mechanism calculates attention weights in the self-attention sub-layer based on the dot product similarity between the query vector and the key vector, combined with the prior bias matrix based on the medical knowledge graph. The specific calculation steps are as follows: 1) Calculate the scaling dot product attention score; 2) Construct the graph adjacency bias matrix This matrix has the same dimensions as the attention score matrix, and its elements... The value is determined by the strength of the relationship between the two knowledge point entities corresponding to the current sequence positions i and j in the knowledge graph. The relationship strength can be pre-calculated and assigned based on the edge type and path distance in the graph. 3) The graph bias matrix is ​​added to the calculation of the attention score to form the modulated attention weight.

7. The intelligent construction system and method for medical digital teaching materials oriented towards personalized learning according to claim 4 or 5, characterized in that: The total loss function of a deep knowledge tracing model includes the main loss and the graph regularization loss; The expression for the main loss is: In the formula, M is the number of samples in a training batch. It is the true boundary result of the m-th sample; It is the predicted probability of the interaction result corresponding to the model; The expression for the graph regularization loss is: In the formula, E is the set of all directed edges in the medical knowledge graph, representing entity pairs with a pre-defined semantic relationship; These are the opposite edges of the model. and These are the mastery predictions output by the head entity i and the tail entity j in the edge (i, j) of the model, respectively. This is the weight coefficient of the edge, which is set according to the importance or confidence level of the relation type.

8. The intelligent construction system and method for medical digital teaching materials oriented towards personalized learning according to claim 2, characterized in that: The personalized content engine module includes a teaching rule reasoning and triggering submodule, a content atom retrieval and adaptation submodule, and a teaching path planning and sequence generation submodule. The teaching rule reasoning and triggering submodule is used to generate a sequence of teaching action instructions by querying and matching the condition-action rules in the pre-set teaching rule base, based on the knowledge subgraph and mastery value corresponding to the current learning objective and the symbolic reasoning method based on the production rule system. The content atom filtering and adaptation submodule is used to filter out a candidate set of content atoms from the library based on the sequence of teaching action instructions and the set of knowledge points covered by the knowledge subgraph by querying the metadata index of the content atom library and calculating the adaptation degree. The teaching path planning and sequence generation submodule generates a structured textbook page sequence based on the topological structure of the knowledge subgraph and the candidate set of content atoms, using a graph theory-based path planning algorithm.

9. The intelligent construction system and method for medical digital teaching materials oriented towards personalized learning according to claim 2, characterized in that: The interaction and presentation module includes a rendering engine submodule, an interaction manager submodule, and a log collection and reporting submodule. The rendering engine submodule is used to parse the structured textbook page sequence and call the front-end component library and templates to dynamically generate and assemble the visual user interface displayed on the application. The interaction manager submodule is used to bind preset event listeners to interactive elements in the interface when the rendering engine generates the user interface, and to define the immediate front-end response logic after the interactive event is triggered. The log collection and reporting submodule is used to listen to all user behavior events on the interface, record and temporarily store interaction log data in a structured format in real time, and send it asynchronously to the backend learner modeling and analysis module via the network.

10. A method for intelligently constructing digital medical textbooks for personalized learning, characterized in that, The system applied to any one of claims 1 to 9 comprises: S1: Constructing a medical knowledge graph; S2: In response to a learner starting to learn, initialize or load the learner's cognitive model; S3: Based on the learner's personalized textbook page sequence, call the learner's cognitive model, calculate the mastery value corresponding to the learner's target learning content, and query the knowledge subgraph corresponding to the learner's target learning content; S4: Based on preset teaching rules, combined with knowledge subgraphs and mastery values, dynamically select and combine content atoms to generate a personalized textbook page sequence; S5: Generates front-end pages based on personalized textbook page sequences for learners to interact with, and records learners' interactive behaviors.

Images