Case teaching plan generation method based on medical education droop class large model

Abstract

The invention relates to the crossing field of artificial intelligence and medical education, and provides a medical education droop class large model-based case teaching plan generation method. In order to solve the problems that traditional cases are high in reuse rate and typical cases are insufficient, multi-modal data fusion processing (clinical records / textbooks / expert experiences), domain adaptive model training (Tranform + LoRA parameter migration), constraint decoding generation (regular beam search + medical ontology verification), teaching target dynamic adaptation (Bloom vector fusion + reinforcement learning) and a three-dimensional quality evaluation system are adopted, and a three-dimensional quality evaluation system is established. And efficient generation and accurate optimization of the case teaching plan are realized. The method comprises the following five steps: 1) heterogeneous data fusion; 2) two-stage model training; 3) generating a structured instruction; 4) dynamically adjusting the teaching content; and 5) knowledge updating and stability guarantee. The system solves the problems of insufficient case diversity, low clinical fitting degree and other pain points, and improves the medical education quality.

Description

TECHNICAL FIELD

[0001] The application belongs to the cross field of artificial intelligence and medical education, and provides a case teaching plan generation method based on a medical education vertical large model. BACKGROUND

[0002] In the field of medical education, cases play a crucial role in the cultivation and education of medical students. Traditional case acquisition methods mainly rely on real cases accumulated by hospitals in long-term clinical practice. These real cases record the entire process of patients from the appearance of symptoms, diagnosis, to treatment, providing medical students with a direct view of real clinical scenarios and are an indispensable important material in the medical education system.

[0003] However, the current traditional cases have exposed many problems in the application process. According to the statistical data of Chinese Medical Education Journal in 2023, the reuse rate of teaching cases in China's top three hospitals is as high as 78%. This means that a large number of teaching activities rely on the repeated use of cases, which not only leads to a lack of freshness and timeliness of teaching content, making it difficult to meet the needs of constantly updated medical knowledge and clinical practice, but also may make students lack sufficient ability to respond to complex and variable situations in actual clinical practice. More seriously, the proportion of typical cases in existing teaching cases is less than 5%, which has a serious structural defect. Typical cases can clearly show the typical characteristics of diseases and diagnosis and treatment ideas, which are of great significance for students to master the nature of diseases and standardize the diagnosis and treatment process. The lack of typical cases makes it difficult for students to fully understand various diseases in the learning process, affecting the quality and effectiveness of teaching.

[0004] Under the background of the continuous development of computer technology, some computer-aided generation systems have emerged. However, most existing systems use finite state machine architecture, such as the generation method disclosed in CN112487194A based on a symptom-disease mapping table, which strictly limits the symptom combination space within the dimensions of the preset rule library. This results in a lack of flexibility and scalability of the system, making it difficult to cope with complex and diverse clinical symptoms and disease manifestations, and unable to generate diversified and personalized cases. In addition, although some simple case simulation software has begun to appear, they can generate some basic case scenarios through preset rules and limited data, to some extent, assisting teaching work. However, due to the limited data and simple rules, the generated case scenarios are often too idealized, with a large gap from real clinical situations, and cannot provide students with a learning experience close to real clinical environments.

[0005] In summary, whether it is a traditional case acquisition method or an existing computer-aided generation system, there are respective limitations, which are difficult to meet the needs of medical education for high-quality and diversified cases. Therefore, it is imminent to develop an innovative and efficient case generation method. SUMMARY

[0006] The present application aims to solve the problem of lack of diversity, pertinence and timeliness of teaching content caused by high case reuse rate and insufficient typical cases in existing medical education, and realizes efficient generation and precise optimization of case teaching plans through multi-modal medical data fusion processing, domain adaptive model training, constraint decoding generation mechanism, dynamic adaptation of teaching objectives and three-dimensional quality evaluation system.

[0007] To achieve the above purpose, the present application provides the following technical solutions:

[0008] The present application provides a case teaching plan generation method based on a medical education vertical category large model, comprising the following steps:

[0009] Step 1: Multi-modal medical data fusion processing:

[0010] Collect clinical case records, medical textbooks and expert experience data, after data cleaning and missing value processing, label disease types, symptom characteristics and teaching attributes according to ICD-11 disease classification standard and SNOMED CT symptom terminology system, and realize heterogeneous data fusion through graph attention network.

[0011] Step 2: Domain adaptive model training:

[0012] Select a pre-training model with Transformer architecture, use a two-stage strategy of pre-training with general medical data and fine-tuning with education vertical data, use low-rank adaptive matrix LoRA for efficient parameter transfer, and optimize the model with cross-entropy loss with weight decay;

[0013] Step 3: Constraint decoding case generation:

[0014] Parse the structured instructions generated by the user's input disease type and complexity requirements, generate a preliminary case teaching plan through a rule-constrained beam search algorithm, and perform entity verification based on medical ontology;

[0015] Step 4: Dynamic adaptation of teaching objectives:

[0016] After vectorizing the Bloom teaching objectives, fuse the case features through a learnable gating network, dynamically adjust the depth and length distribution of the teaching content, and optimize the teaching applicability combined with reinforcement learning.

[0017] Step 5: Three-dimensional quality evaluation update:

[0018] We construct a medical accuracy discriminator and a teaching applicability assessment model, continuously optimize the system through a dynamic knowledge update mechanism, and ensure the stability of knowledge transfer with KL divergence constraints.

[0019] In the above scheme, step 1 includes the following steps:

[0020] Step 1.1: Collaborate with hospitals through web crawlers to obtain clinical case records, medical images, medical literature, and expert experience data;

[0021] Step 1.2: Use data cleaning algorithms and tools to remove duplicate content, erroneous information and irrelevant symbols, and use mean imputation, model prediction imputation or deletion strategies to handle missing values ​​to ensure the integrity and accuracy of the data;

[0022] Step 1.3: Annotate the data based on medical ontology, which includes the ICD-11 disease classification standard and the SNOMED CT symptom terminology system. The annotation content includes disease type, symptom characteristics and teaching attributes.

[0023] Step 1.4: Construct a heterogeneous data graph structure: Use clinical records (EMR), medical imaging reports, and evidence-based medicine literature as nodes in the heterogeneous graph, and establish cross-modal association edges through entity links;

[0024] Step 1.5: Perform graph attention network computation:

[0025] a) Node representation update:

[0026]

[0027] b) The formula for calculating the attention coefficient of a graph attention network is:

[0028] α ij =softmax(LeakyReLU(a T [Wh i ||Wh j ]))in W represents the attention weight of node j to node i in layer l. (l) This represents the trainable weight matrix of the l-th layer. This represents the feature vector of neighbor node j at layer l. Let denot be the set of neighboring nodes of node i, σ denote the non-linear activation function, a be a learnable parameter vector, and h be the set of neighboring nodes of node i. i and h j The feature vectors of node i and node j are respectively, [Wh i |Wh j ] indicates that Wh i and Wh i The concatenated vector.

[0029] In the above scheme, step 2 includes the following steps:

[0030] Step 2.1: Select a pre-trained model based on the Transformer architecture and initialize the model through random initialization or pre-trained parameter transfer;

[0031] Step 2.2: Perform multi-stage training:

[0032] Pre-training on large-scale general medical data to learn common medical features;

[0033] In the medical education vertical data, a low-rank adaptation matrix (LoRA) is used for efficient parameter fine-tuning. The fine-tuning process is expressed as follows:

[0034] ΔW=A·B T in

[0035] Original weights The domain knowledge transfer is achieved by training only low-rank adaptation matrices A and B under the frozen conditions.

[0036] Step 2.3: Optimize the model using a weighted decaying cross-entropy loss function:

[0037]

[0038] Where y i The true label of the i-th sample, p θ (y i |x i The model, under parameter θ, given input x i The predicted label is y i The probability, ||·|| F The Frobenius norm of the matrix, λ = 0.01, controls the regularization strength.

[0039] In the above scheme, step S3 includes the following steps:

[0040] Step 3.1: Parse the user's input request for generating medical records, and extract the disease type, complexity, and special requirements using natural language processing technology;

[0041] Step 3.2: Generate structured instructions based on the parsing results. The instructions include disease type, complexity level, complication requirements, and case element templates.

[0042] Step 3.3: Input the instructions into the trained model and generate preliminary case study plans through a constraint decoding algorithm. The decoding process satisfies the following:

[0043]

[0044] where ρ final (ω t ) represents the final adjusted probability of generating the word ω at time step t, t is the entity verification function, ω t represents the specific word generated at time step t, ω represents the word sequence generated, and ρ θ represents the probability of generating the word ω based on the model parameters θ under the condition of the word sequence before the given time step t, E t t represents the set of medical entity constraints to be satisfied at time t, ω <t represents the word sequence before the given time step t;

[0045] Step 3.4: Use the medical entity verification module to perform consistency verification on the generated text according to the ICD-11 disease classification standard and the SNOMED CT symptom terminology system.

[0046] In the above solution, the step S4 includes the following steps:

[0047] Step 4.1: Convert the Bloom's teaching objective classification into a 128-dimensional embedding vector E obj , and construct a multi-scale prompt template including a basic template and an extended template;

[0048] Step 4.2: Dynamically fuse the target vector and the case features through a learnable gating network:

[0049] g = σ(W · [E obj ; E case )

[0050] E obj is the teaching objective embedding vector (128-dimensional), E case is the case feature vector, W is the trainable weight matrix, [·; ·] represents the vector concatenation operation, and the output gating value g dynamically adjusts the depth and duration allocation of the teaching content;

[0051] Step 4.3: Call the GQE model to evaluate the teaching applicability: <00​​​​​​​​​​

[0055] In the above scheme, step 5 includes the following steps:

[0056] Step 5.1: Construct a medical accuracy discriminator:

[0057] An adversarial training network was constructed based on the PubMed database to detect pathological inconsistencies in generated cases.

[0058] Step 5.2: Conduct an assessment of the teaching suitability of the curriculum.

[0059] The overall score is calculated using the GQE model:

[0060] Score=α·CLINICAL+β·EDU+γ·COHESION(α+β+γ=1)

[0061] Where CLINICAL is the medical accuracy score, EDU is the teaching objective matching degree, and COHESION is the logical coherence.

[0062] Step 5.3: Implement dynamic knowledge updates:

[0063] When updating the authoritative knowledge base (UpToDate), the model parameters are adjusted using an incremental learning framework, with the following constraints:

[0064]

[0065] in Denotes divergence, This represents the output probability distribution of the model before the parameter update. This represents the output probability distribution of the model after parameter updates, where ∈ = 0.1 is a preset threshold;

[0066] Step 5.4: Iteratively optimize the generation quality using a reinforcement learning framework, with the evaluation score as the reward signal.

[0067] Because the present invention employs the above-mentioned technical means, it has the following beneficial effects:

[0068] 1. This invention addresses the problem of insufficient case diversity in existing medical education systems through multimodal medical data fusion processing technology (step 1), and ensures the timeliness of case information through a dynamic knowledge update mechanism (step 5.3). Specifically, it achieves precise alignment of heterogeneous data representations by constructing a heterogeneous data graph structure covering clinical records, medical imaging reports, and evidence-based medicine literature (step 1.4), combined with graph attention network computation (step 1.5). This technique automatically identifies key information in different modalities of data through an attention mechanism, such as focusing on lesion areas when processing medical imaging reports and extracting diagnostic-related indicators when analyzing clinical records. This cross-modal data fusion approach breaks the limitations of traditional single data sources, expanding the data dimension of case generation from a single clinical record to a multimodal knowledge system, significantly improving the richness and clinical relevance of case content. Simultaneously, through medical ontology annotation (step 1.3) and a dynamic knowledge update mechanism (step 5.3), the system can connect to authoritative medical databases in real time, ensuring that the pathological logic of generated cases is consistent with the latest treatment guidelines, effectively solving the problems of high reuse rate and lagging knowledge updates in traditional teaching cases.

[0069] 2. This invention addresses the issues of low parameter transfer efficiency and high computational cost in large-scale models within medical education scenarios through domain-adaptive model training technology (step 2), achieving efficient parameter transfer via a low-rank adaptation matrix (LoRA). Specifically, in the Transformer architecture pre-trained model, a two-stage strategy (steps 2.1-2.3) is employed to achieve collaborative learning of general medical knowledge and educational vertical features: first, pre-training is performed on large-scale general medical data to acquire basic medical capabilities; then, the original weights are frozen, and only the low-rank adaptation matrix (ΔW = A·B) is fine-tuned. T Domain knowledge transfer is performed using a method where r=8. This technique reduces the number of parameters in traditional fine-tuning methods to 1 / 8 of the original size (r=8). Simultaneously, it constrains the model optimization direction through a weighted decay cross-entropy loss function (step 2.3), significantly reducing memory usage (by 70%) and training time (by 3 times) while maintaining the integrity of medical knowledge during parameter updates. This efficient transfer strategy avoids the computational resource waste caused by excessive parameters in traditional methods and ensures the model continuously absorbs the latest medical findings through a dynamic knowledge update mechanism (step 5.3), providing a scalable solution for large-scale model applications in medical education scenarios.

[0070] 3. This invention solves the problems of frequent logical contradictions and inconsistencies in medical entities in traditional case generation systems by using a constraint-based decoding case generation technology (step 3). It achieves precise control over the generated content through a rule-constrained beam search algorithm (step 3.3) and a medical ontology verification mechanism (step 3.4). Specifically, after generating structured instructions by parsing user input requirements (step 3.1), the model introduces an entity verification function during the decoding process. The generated text is subject to mandatory constraints to meet the requirements of the ICD-11 disease classification standard and the SNOMED CT symptom terminology system. This technique employs a dual verification mechanism: firstly, it verifies the medical rationality of the generated terms in real time during the decoding stage (e.g., symptom-disease correlation); secondly, it checks the consistency of annotations (e.g., the matching degree between symptom intensity and disease type) through an ontology reasoning engine during the post-processing stage. This multi-level verification system effectively solves the problem of idealized generated content caused by the limited dimensions of the rule base in traditional systems, enabling case texts to maintain clinical authenticity while meeting the dual requirements of logical coherence and medical standardization in teaching scenarios.

[0071] 4. This invention addresses the lack of targeted content design in traditional case generation systems through dynamic adaptation technology for learning objectives (step 4). It achieves precise matching of teaching content through Bloom's learning objective embedding vectors (step 4.1) and a learnable gating network (step 4.2). Specifically, after transforming the Bloom's six-level learning objectives (memorization, understanding, application, etc.) into 128-dimensional embedding vectors, a gating network (g = σ(W·[E)) is used... obj E case This technology dynamically integrates target vectors and case features, automatically adjusting content depth and time allocation according to the teaching level. It supports parameterized adjustment through multi-scale prompt templates (basic and extended templates). For example, when generating memory-based cases, it emphasizes disease definitions and typical symptoms, while increasing the complexity of differential diagnosis and treatment plans when generating application-based cases. Simultaneously, it combines a reinforcement learning framework (PPO algorithm) to perform closed-loop optimization of the generated content (step 4.4), continuously improving the output of the teaching applicability assessment model (Score = α.CLINICAL + β·EDU + γ·COHESION). This dynamic adaptation mechanism overcomes the limitations of traditional fixed template generation methods, enabling case content to accurately match the learning objectives of different teaching stages, significantly enhancing personalized teaching capabilities in medical education scenarios.

[0072] 5. This invention addresses the inadequacy of the quality control system in traditional case generation systems through a three-dimensional quality assessment and update technology (step 5). It constructs a multi-dimensional quality assurance system using a medical accuracy discriminator (step 5.1) and a teaching applicability assessment model (step 5.2). Specifically, the medical accuracy discriminator detects pathological contradictions in generated cases (such as logical conflicts between symptoms and disease diagnosis) through an adversarial training network, while the teaching applicability assessment model quantifies the comprehensive scores of medical accuracy, target matching, and logical coherence using the GQE framework. This technique ensures knowledge transfer stability through KL divergence constraints (step 5.3). When accessing authoritative knowledge bases such as UpToDate, the incremental learning framework only adjusts the low-rank adaptation matrix parameters (ΔW = A·B). T This avoids disturbing the original model structure. Simultaneously, a reinforcement learning framework (step 5.4) is used to iteratively optimize the model by using the evaluation score as a reward signal, forming a closed-loop system of "generation-evaluation-optimization". This dynamic quality control system effectively solves the problem of model performance fluctuations caused by knowledge updates in traditional methods, enabling the generated cases to continuously adapt to changing teaching needs while maintaining medical professionalism, providing a reliable, high-quality case generation solution for the medical education field.

[0073] 6. This invention addresses the lack of case diversity in existing medical education systems through multi-source data fusion technology (covering six types of data including clinical records, imaging reports, and literature). It ensures the timeliness of case information through a dynamic knowledge update mechanism (connecting to authoritative knowledge bases such as UpToDate), enhances the relevance of content design through teaching objective embedding technology (supporting the generation of 6-level Bloom classification), guarantees professionalism through a three-dimensional quality assessment system (medical accuracy ≥95%, teaching applicability improved by 40%), and significantly reduces model training costs through low-rank adaptation technology (reducing memory usage by 70% and improving fine-tuning efficiency by 3 times). The synergistic effect of these technologies achieves comprehensiveness, timeliness, accuracy, and efficiency in medical education content, providing a systematic solution for clinical teaching and knowledge updating. Attached Figure Description

[0074] Figure 1 This is a simplified flowchart of the invention.

[0075] Figure 2 This is a software speed measurement output diagram of one embodiment. Detailed Implementation

[0076] The embodiments of the present invention will be described in detail below. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.

[0077] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without these specific details.

[0078] The purpose of this invention is to address the pain points in the current process of generating medical case teaching materials, such as low efficiency, lack of targeted and personalized content, and difficulty in meeting the needs of different levels of medical education. By using a technology based on a large-scale model of medical education vertical categories, it aims to achieve more efficient, accurate, and practically teaching-compliant case teaching material generation.

[0079] To achieve the above objectives, the present invention employs the following technical means:

[0080] A case generation method based on a large-scale model of medical education includes the following steps:

[0081] Step 1: Multimodal medical data fusion processing

[0082] During the data collection phase, we extensively explored multi-source medical data, including, but not limited to, a vast amount of medical literature, detailed clinical case records, and authoritative medical textbooks. Utilizing web crawling technology, we precisely extracted cutting-edge medical literature from authoritative medical databases such as PubMed and Wanfang Medical Network, as well as well-known medical journal websites. We actively collaborated with major hospitals to obtain authentic and abundant clinical case records. Simultaneously, we collected publicly available electronic versions of medical textbooks, comprehensively expanding our data sources.

[0083] In the data cleaning stage, professional data cleaning algorithms and tools are used to carefully identify and remove duplicate content from the data to avoid redundant information interfering with subsequent analysis; erroneous information is accurately identified and removed to ensure data reliability; irrelevant symbols and formatting marks are removed to purify the data format; and for missing values ​​in the data, strategies such as mean imputation, model prediction imputation, or deletion when the missing value ratio is too high are adopted according to the data characteristics and statistical methods to ensure the integrity and accuracy of the data, laying a solid foundation for subsequent processing.

[0084] The data annotation process utilizes professional medical annotation tools, organizing medical experts and professional annotators to meticulously annotate key information in the data, such as disease names, symptoms, diagnostic results, and treatment methods, based on internationally recognized medical standards and terminology systems, such as the ICD-11 disease classification standard and the SNOMED CT symptom terminology system. This provides structured and standardized data for subsequent model training.

[0085] Cross-modal data fusion engine (Step 1): In the medical field, data sources are wide-ranging and diverse in form, posing a significant challenge to the effective integration and utilization of data. To address this issue, we constructed a multimodal corpus that comprehensively covers clinical records (EMR), medical imaging reports, and evidence-based medicine literature. Clinical records (EMR) contain detailed patient medical history, diagnostic processes, treatment plans, and other information, forming the foundation of medical data; medical imaging reports present the patient's physical condition in image form, such as X-rays, CT scans, and MRI images, containing rich pathological information; evidence-based medicine literature gathers the latest findings in global medical research and clinical practice experience, providing a scientific basis for medical decision-making.

[0086] To enable these different modalities of data to work together, we adopted graph attention network technology.

[0087] Graph attention network computation:

[0088] Node representation update formula:

[0089]

[0090] Attention coefficient calculation:

[0091] α ij =softmax(LeakyReLU(a T [Wh i |Wh j ]))in W represents the attention weight of node j to node i in layer l. (l) This represents the trainable weight matrix of layer 1. This represents the feature vector of neighbor node j at layer l. Let represent the set of neighboring nodes of node i, σ represent the nonlinear activation function, and α be a learnable parameter vector. Graph attention networks can automatically identify and focus on key information in data based on the characteristics of different modalities through an attention mechanism. For example, when processing medical image reports, it can focus on lesion areas in the image; when processing clinical records, it can focus on analyzing key indicators related to disease diagnosis and treatment. In this way, it achieves precise alignment of heterogeneous data representations, breaks down barriers between data, and enables multimodal data to complement each other, providing comprehensive and high-quality data support for subsequent model training and case generation, greatly improving the model's ability to understand and analyze complex medical scenarios.

[0092] Step 2: Domain Adaptive Pre-training and Instruction Fine-tuning

[0093] When selecting and initializing the model, factors such as the characteristics of medical data, model performance, and scalability are comprehensively considered. A suitable large language model architecture for the medical field is chosen, such as a Transformer-based pre-trained model. The model parameters are initialized by selecting appropriate initialization methods, such as random initialization or transfer initialization based on pre-trained model parameters, to determine the initial state of the model and provide a good start for subsequent training.

[0094] First, the model is pre-trained on large-scale general medical data, allowing it to fully learn the common features of medical knowledge and build a solid foundation of medical knowledge. Then, it is fine-tuned on specific medical education data, enabling the model to deeply understand the needs and characteristics of medical education scenarios. In this process, the following methods are employed:

[0095] The LoRA (Low-Rank Adaptation) parameter-efficient fine-tuning method cleverly injects educational domain features into a basic medical model and sets an adaptation matrix with a rank r = 8, effectively achieving domain knowledge transfer and significantly improving the model's performance in the medical education field. During the efficient parameter fine-tuning process, the weight update amount can be expressed as:

[0096] ΔW=A·B T Where A∈R d×r , B∈R k×r It is a low-rank fitting matrix with rank r = 8 and original weights W ∈ R. d×k By freezing the original parameters and training only the adaptation matrix, domain knowledge transfer can be achieved.

[0097] During training, preprocessed data is input into the model in carefully designed batches and sequences. The difference between the model's output and the true label is calculated using the cross-entropy loss function, employing weighted decay cross-entropy loss.

[0098]

[0099] where y i is the true label of the i-th sample, p θ (y i |x i ) is the probability that the model predicts the label y given the input x under the parameter θ, ||·|| i is the Frobenius norm of the matrix, and λ = 0.01 controls the regularization strength. Then, with the help of the backpropagation algorithm, the parameters of the model are efficiently adjusted to continuously optimize the performance of the model so that it can accurately capture the complex features and patterns in medical data.

[0100] Step 3: Case generation based on constrained decoding

[0101] In the requirement analysis stage, using text parsing techniques in natural language processing, the case generation requirements input by the user are deeply analyzed. By means of lexical analysis, syntactic analysis, and semantic understanding, key information such as disease type, complexity, and special requirements are accurately extracted to provide an accurate basis for subsequent instruction generation.

[0102] In the instruction generation link, according to the parsed requirements, combined with the input format and requirements of the model, clear, definite, and model-understandable instructions are generated. For example, when the user's requirement is to generate a pneumonia case of medium complexity and include complication information, the instruction "Generate a pneumonia case of medium complexity, including complication information. The case content should cover the patient's basic information, symptoms, diagnosis process, treatment plan, etc." is generated.

[0103] In the case generation and optimization process, the instruction is input into the trained model, and the model generates case text based on the learned knowledge. To ensure that the generated case text meets medical professional standards and actual requirements, a beam search algorithm based on rule constraints is adopted:

[0104]

[0105] where ρ final (ω t ) represents the final adjusted probability of generating the word ω at time step t, t is the entity verification function, ω is the specific word generated at time step t, ω represents the generated word sequence, ρ t represents the probability of generating the word ω based on the model parameters θ under the condition of the word sequence (ω<t) before the given time step t, θ and E t is the probability of generating the word ω t ​​​​This represents the set of medical entity constraints that must be satisfied at time t. A medical entity verification module is also included, which strictly adheres to the ICD-11 disease classification standard and the SNOMED CT symptom terminology system to verify and optimize the generated text, ensuring the accuracy, standardization, and professionalism of the case records.

[0106] 4. Dynamic adaptation of teaching objectives:

[0107] Innovative teaching guidance and generation framework:

[0108] Target Vectorization: Bloom's learning objectives classification is transformed into 128-dimensional embedding vectors. Multi-Scale Cue Templates: Basic Template: Contains a fixed structure of [disease type, complexity level, teaching duration]. Extended Template: Supports parameterized adjustment of [key symptom weights, differential diagnosis depth, complication probability]. Dynamic Attention Mechanism: Through a learnable gating network g = σ(W·[E obj E case The target vector and case features are fused. Among them, E... obj Embed a vector (128-dimensional) to represent the teaching objectives, E case Let W be the case feature vector, W be the trainable weight matrix, and [·;·] represent vector concatenation operations. The output gate value g dynamically adjusts the depth and duration of the teaching content.

[0109] Generative quality assessment module for fusion discriminator networks

[0110] Three-dimensional quality assessment system: Medical accuracy discriminator: An adversarial training network built based on PubMed literature to detect physiological and pathological contradictions. Teaching applicability assessment: Using the GQE (Generated Quality Estimator) model, the conformity formula is as follows:

[0111] Score=α·CLINICAL+β·EDU+y.COHESION(α+β+γ=1)

[0112] Iterative optimization mechanism: A closed-loop process of generation-evaluation-optimization is achieved through a reinforcement learning framework (PPO algorithm).

[0113] Enhanced Adaptability of Teaching (Step 4): The goal of medical education is to cultivate medical professionals with solid theoretical knowledge and rich practical experience. Different teaching stages and objectives require different case studies. To meet diverse teaching needs, we integrated Bloom's Taxonomy of Learning Objectives into the case generation system. Bloom's Taxonomy categorizes learning objectives into six levels: memorization, comprehension, application, analysis, evaluation, and creation. We focus on generating memorization-based, comprehension-based, and application-based cases.

[0114] For memory-based case studies, the focus is primarily on helping students solidify their foundational knowledge, such as disease definitions, symptoms, and diagnostic criteria. The generated cases include typical symptoms and diagnostic processes for common diseases, allowing students to deepen their understanding of basic knowledge through case study. Comprehension-based case studies emphasize cultivating students' understanding of disease pathological mechanisms. These cases include detailed disease development processes, pathological changes, and treatment principles, guiding students to think deeply about the essence of diseases. Application-based case studies aim to develop students' ability to apply their knowledge to solve practical clinical problems. These cases typically present complex clinical scenarios, requiring students to formulate reasonable diagnostic and treatment plans based on patient symptoms, examination results, and other information. In this way, the case generation system can generate highly targeted cases according to different teaching objectives and students' learning stages, effectively improving the quality and effectiveness of medical teaching and helping medical students comprehensively enhance their professional abilities.

[0115] 5. Three-dimensional quality assessment update:

[0116] Dynamic Knowledge Update Mechanism (Step 2): Medical knowledge updates rapidly. To ensure the system keeps pace with medical advancements, we designed an innovative incremental learning framework. The incremental learning process constrains the parameter update magnitude:

[0117]

[0118] in Denotes divergence, This represents the output probability distribution of the model before the parameter update. This represents the output probability distribution of the model after parameter updates, with ∈ = 0.1 as a preset threshold to ensure the stability of knowledge updates. This framework supports the system in real-time online access to updated content from authoritative knowledge bases. Taking UpToDate as an example, it is a globally renowned clinical decision support tool containing a vast amount of medical knowledge and the latest clinical research findings. Through connections with authoritative knowledge bases such as UpToDate, our system can promptly obtain the latest disease diagnostic standards, treatment methods, drug information, etc.

[0119] In integrating new knowledge into the existing model, we introduced a KL divergence control strategy. KL divergence is an indicator that measures the difference between two probability distributions. By calculating the KL divergence of the parameter distributions before and after the model update, we can precisely control the adjustment range of the model parameters. When new knowledge is added, the model gradually learns and adapts to the new knowledge while maintaining the existing knowledge, avoiding overfitting or performance degradation caused by rapid knowledge updates. This dynamic knowledge update mechanism ensures that the knowledge the system relies on is always up-to-date, while guaranteeing the stability of model performance and providing reliable knowledge support for case generation.

[0120] In summary, the present invention has the following effects:

[0121] Deep fusion of multi-source data:

[0122] By extensively and systematically collecting multi-source medical data, the system organically integrates diverse knowledge resources in the medical field. Medical literature contains cutting-edge research findings and information on rare cases, while clinical case records visually present real-world medical scenarios. This fusion effectively broadens the data dimensions for case generation, making the generated cases richer in content, closer to clinical practice, and significantly enhancing their application value and practicality.

[0123] Model self-learning and dynamic optimization:

[0124] Leveraging the advanced training mechanism of large-scale models, these models possess the ability to continuously learn new knowledge and absorb new clinical experience. As a constant stream of new data is incorporated into the training system, the models can adaptively update parameters and optimize algorithms, thereby ensuring that the generated cases always align with the latest developments in the medical field and providing timely teaching materials for medical education.

[0125] Highly personalized customization:

[0126] This invention can accurately generate case studies based on diverse user needs, fully meeting the requirements of different levels and types of medical teaching scenarios. For medical beginners, it can generate basic case studies with concise structures and key points to help them solidify their medical foundation; for advanced medical training, it can generate complex cases with intricate diagnostic and treatment processes to enhance learners' ability to handle complex clinical problems.

[0127] Data annotation methods based on medical ontology

[0128] A multi-dimensional annotation framework is constructed using an internationally standardized terminology system: Disease classification annotation: A three-level annotation system (major category - subcategory - specific disease) is established based on the ICD-11 coding system; Symptom annotation: Fine-grained annotation is performed using the SNOMED CT symptom terminology tree (including dimensions such as symptom intensity and duration); Teaching attribute annotation: A triplet annotation template containing [knowledge level, cognitive dimension, teaching duration] is established; Dynamic verification mechanism: An ontology reasoning engine automatically detects annotation consistency (such as symptom-disease correlation verification).

[0129] Multimodal data fusion technology:

[0130] This innovative approach constructs a multimodal corpus, integrating various data modalities such as clinical records, medical imaging reports, and evidence-based medicine literature. Utilizing graph attention network technology, it automatically focuses on key information based on the characteristics of different data modalities, achieving precise alignment of heterogeneous data representations. This breaks down data barriers, providing comprehensive and high-quality data for model training and case generation, significantly enhancing the model's understanding and analysis capabilities in complex medical scenarios. Traditional methods often rely on only a single type of data, making it difficult to acquire comprehensive medical knowledge; however, this technology achieves deep fusion of multi-source data, significantly broadening the data dimensions.

[0131] Dynamic knowledge update mechanism:

[0132] A unique incremental learning framework is designed, which constrains the parameter update magnitude and utilizes a KL divergence control strategy to ensure stable model performance when incorporating new knowledge. It can also access authoritative knowledge bases (such as...) in real time online.

[0133] UpToDate updates content in a timely manner, providing information on disease diagnostic criteria and treatment methods, ensuring that case generation aligns with the latest medical developments, and solving the problem of lagging knowledge updates in traditional methods and existing systems.

[0134] Data annotation system based on international standards:

[0135] A multi-dimensional annotation framework is constructed using an internationally standardized terminology system. Disease classification is based on the ICD-11 coding system to establish a three-level annotation system. Symptom annotation uses the SNOMED CT symptom terminology tree for fine-grained annotation and also includes teaching attribute annotation. A dynamic verification mechanism is also included, which automatically detects annotation consistency through an ontology inference engine to ensure the accuracy and standardization of data annotation, providing a high-quality data foundation for model training.

[0136] Hints for the engineering architecture of course objective embedding vectors:

[0137] We transform Bloom's classification of learning objectives into 128-dimensional embedded vectors to construct an innovative framework for generating instructional guidance. We design multi-scale prompt templates; the basic template contains key instructional parameters, while extended templates support more parameterized adjustments. Utilizing a dynamic attention mechanism, we fuse target vectors and case features through a learnable gating network to accurately generate cases based on different learning objectives, enhancing the relevance of instruction—a capability not found in traditional and existing technologies.

[0138] Generative quality evaluation module for fusion discriminator networks:

[0139] A three-dimensional quality assessment system was constructed. A medical accuracy discriminator, based on PubMed literature, was used to build an adversarial training network to detect physiological and pathological inconsistencies. The teaching applicability assessment used the GQE model. A reinforcement learning framework (PPO algorithm) was employed to achieve a closed-loop process of generation, evaluation, and optimization, continuously improving the quality of generated cases and ensuring that cases meet medical professional standards and teaching requirements.

Claims

1. A method for generating case study teaching plans based on a large-scale model of medical education verticals, characterized in that, Includes the following steps: Step 1: Multimodal medical data fusion processing: Clinical case records, medical textbooks, and expert experience data were collected. After data cleaning and missing value processing, disease types, symptom characteristics, and teaching attributes were labeled according to the ICD-11 disease classification standard and the SNOMED CT symptom terminology system. Heterogeneous data fusion was achieved through graph attention network. Step 2: Domain Adaptive Model Training: We selected the Transformer architecture to pre-train the model, and adopted a two-stage strategy of pre-training with general medical data and fine-tuning with education vertical data. We used the low-rank adaptation matrix LoRA for efficient parameter transfer and optimized the model with cross-entropy loss with weight decay. Step 3: Constraint Decoding Case Generation: The system parses the disease type and complexity requirements input by the user to generate structured instructions, generates preliminary case study plans through a rule-constrained beam search algorithm, and performs entity verification based on medical ontology. Step 4: Dynamically adapting learning objectives: After vectorizing Bloom's teaching objectives, case features are integrated through a learnable gating network to dynamically adjust the depth and duration of teaching content, and the applicability of teaching is optimized by combining reinforcement learning. Step 5: 3D Quality Assessment Update: We construct a medical accuracy discriminator and a teaching applicability assessment model, continuously optimize the system through a dynamic knowledge update mechanism, and ensure the stability of knowledge transfer with KL divergence constraints.

2. The method according to claim 1, characterized in that, Step 1 includes the following steps: Step 1.1: Collaborate with hospitals through web crawlers to obtain clinical case records, medical images, medical literature, and expert experience data; Step 1.2: Use data cleaning algorithms and tools to remove duplicate content, erroneous information and irrelevant symbols, and use mean imputation, model prediction imputation or deletion strategies to handle missing values ​​to ensure the integrity and accuracy of the data; Step 1.3: Annotate the data based on medical ontology, which includes the ICD-11 disease classification standard and the SNOMED CT symptom terminology system. The annotation content includes disease type, symptom characteristics and teaching attributes. Step 1.4: Construct a heterogeneous data graph structure: Use clinical records (EMR), medical imaging reports, and evidence-based medicine literature as nodes in the heterogeneous graph, and establish cross-modal association edges through entity links; Step 1.5: Perform graph attention network computation: a) Node representation update: b) The formula for calculating the attention coefficient of a graph attention network is: α ij =sofimax(LeakyReLU(a T [Wh i ||Wh j ]))in W represents the attention weight of node j to node i in layer l. (l) This represents the trainable weight matrix of layer 1. This represents the feature vector of neighbor node j in layer 1. Let denot be the set of neighboring nodes of node i, σ denote the non-linear activation function, a be a learnable parameter vector, and h be the set of neighboring nodes of node i. i and h j The feature vectors of node i and node j are respectively, [Wh i |Wh j ] indicates that Wh i and Wh j The concatenated vector.

3. The method according to claim 1, characterized in that, Step 2 includes the following steps: Step 2.1: Select a pre-trained model based on the Transformer architecture and initialize the model through random initialization or pre-trained parameter transfer; Step 2.2: Perform multi-stage training: Pre-training on large-scale general medical data to learn common medical features; In the medical education vertical data, a low-rank adaptation matrix (LoRA) is used for efficient parameter fine-tuning. The fine-tuning process is expressed as follows: ΔW=A·B T in Original weights The domain knowledge transfer is achieved by training only low-rank adaptation matrices A and B under the frozen conditions. Step 2.3: Optimize the model using a weighted decaying cross-entropy loss function: Where y i The true label of the i-th sample, p θ (y i |x i The model, under parameter θ, given input x i The predicted label is y i The probability, ||·|| F The Frobenius norm of the matrix, λ = 0.01, controls the regularization strength.

4. The method according to claim 1, characterized in that, Step S3 includes the following steps: Step 3.1: Parse the user's input request for generating medical records, and extract the disease type, complexity, and special requirements using natural language processing technology; Step 3.2: Generate structured instructions based on the parsing results. The instructions include disease type, complexity level, complication requirements, and case element templates. Step 3.3: Input the instructions into the trained model and generate preliminary case study plans through a constraint decoding algorithm. The decoding process satisfies the following: where ρ final (ω t ) represents the final adjusted probability of generating the word ω at time step t t . is the entity verification function, ω t represents the specific word generated at time step t, ω represents the sequence of generated words, and ρ θ represents the probability of generating the word ω based on the model parameters θ under the condition of the word sequence before the given time step t t . E t represents the set of medical entity constraints to be satisfied at time t, and ω<t represents the word sequence before the given time step t; Step 3.4: The generated text is validated for consistency using the medical entity verification module based on the ICD-11 disease classification standard and the SNOMEDCT symptom terminology system.

5. The method according to claim 1, characterized in that, Step S4 includes the following steps: Step 4.1: Transform Bloom's learning objectives classification into a 128-dimensional embedding vector E. obj Construct a multi-scale cue template that includes a basic template and extended templates; Step 4.2: Dynamically fuse target vectors and case features using a learnable gating network: g6σ(W·[E obj 9E case ]) E obj Embed a vector (128-dimensional) to represent the teaching objectives, E case Let W be the case feature vector, W be the trainable weight matrix, and [·;·] represent vector concatenation operations. The output gate value g dynamically adjusts the depth and duration of the teaching content. Step 4.3: Use the GQE model to assess the applicability of the teaching: Score=α·CLINICAL+β·EDU+γ·COHESION(α+β+γ=1) Where CLINICAL is the medical accuracy score, EDU is the teaching objective matching degree, and COHESION is the logical coherence; Step 4.4: Perform closed-loop optimization based on the PPO reinforcement learning algorithm to maximize the evaluation score.

6. The method according to claim 1, characterized in that, Step 5 includes the following steps: Step 5.1: Construct a medical accuracy discriminator: An adversarial training network was constructed based on the PubMed database to detect pathological inconsistencies in generated cases. Step 5.2: Conduct an assessment of the teaching suitability of the curriculum. The overall score is calculated using the GQE model: Score=α·CLINICAL+β·EDU+γ·COHESION(α+β+γ=1) Where CLINICAL is the medical accuracy score, EDU is the teaching objective matching degree, and COHESION is the logical coherence; Step 5.3: Implement dynamic knowledge updates: When updating the authoritative knowledge base (UpToDate), the model parameters are adjusted using an incremental learning framework, with the following constraints: in Denotes divergence, This represents the output probability distribution of the model before the parameter update. This represents the output probability distribution of the model after parameter updates, where ∈ = 0.1 is a preset threshold; Step 5.4: Iteratively optimize the generation quality using a reinforcement learning framework, with the evaluation score as the reward signal.

Images