Medical defense fusion skill evaluation method based on causal-machine model

Abstract

The invention relates to the technical field of education, in particular to a medical defense fusion skill evaluation method based on a causal-machine model, and the method comprises the steps: 1, generating question feature vectors and knowledge point semantic vectors through employing a natural language processing technology in combination with a deep learning model, and calculating the correlation between the feature vectors through employing cosine similarity; 2, generating a question difficulty coefficient based on a causal-machine model; step 3, generating a student knowledge mastering feature vector according to the answering accuracy and answering speed of the student; 4, recommending the most suitable homework questions to the students based on a deep learning model; and 5, quantitatively analyzing the answering performance of the students through a causal-machine model, and providing feedback suggestions and learning suggestions. According to the invention, weak links of students can be identified, the efficiency of preventive medicine major learning is effectively improved, the teaching decision of teachers is assisted, and personalized cultivation of medical defense fusion and medical defense collaboration talents is promoted.

Description

Technical Field

[0001] The present invention relates to the field of educational technology, and in particular to a medical-prevention fusion skills evaluation method based on a causal-machine model. Background Art

[0002] In today's era, medical education is facing unprecedented challenges and opportunities. With the deepening of health concepts and the growing demand for public health services, "integration of medicine and prevention" has become an indispensable and important part of modern medical education. This process requires students to not only master knowledge in traditional fields such as basic medicine and clinical medicine, but also to have an in-depth understanding of related disciplines such as preventive medicine. However, different students have different degrees of understanding of these courses, showing obvious personalized characteristics.

[0003] With the rapid development of information technology, artificial intelligence and big data technologies have gradually been introduced into the field of education, and intelligent teaching systems have emerged. However, although many current online education platforms have a certain degree of homework push and automatic grading functions, they rely on fixed rules or simple question difficulty distribution models, or are based on knowledge graphs and simple machine learning methods, which are difficult to effectively meet students' personalized learning needs. The homework push method with preset rules cannot adjust the push content in real time according to the students' knowledge mastery. When recommending homework, these systems often find it difficult to accurately analyze the degree of match between the questions and the course knowledge points, and cannot dynamically track the students' knowledge mastery. This results in the recommended questions being either too simple to effectively improve learning outcomes, or too complex for students to cope with.

[0004] Although the recommendation method based on knowledge graph attempts to associate knowledge points and questions with students' learning behaviors through the structure of nodes and edges, and can provide a breadth of information in some scenarios, it has many limitations in practical applications. First, the creation of a knowledge graph requires a lot of manual annotation and expert knowledge, involving the definition of knowledge points and the establishment of relationships, which is not only time-consuming, but also may introduce subjective biases, resulting in insufficient accuracy and timeliness of the graph. Secondly, the knowledge graph is difficult to update in real time, and its fixed structure cannot effectively adapt to students' learning progress and personalized needs. Finally, it also has deficiencies in reasoning ability, and simple similarity calculations may not accurately reflect students' real learning needs and knowledge weaknesses.

[0005] The "Causal-Machine Learning" model constructs a causal relationship graph to clarify the dependency path between questions and knowledge points, combines fuzzy theory to generate causal prior knowledge, and uses the partitioned MCMC algorithm to optimize the network structure to estimate the conditional probability distribution between variables. Based on the causal inference results, the impact of different questions on students' learning effects can be quantified to avoid the problem of "pseudo-correlation" in traditional machine learning. Combined with the nonlinear modeling ability of deep learning, the "Causal-Machine Learning" model can dynamically push exercises that best suit students' current learning needs and provide teachers with more accurate learning feedback and personalized guidance. The deep learning model can process large-scale educational data, reveal the complex relationship between questions and knowledge points, and make personalized recommendations based on students' learning behavior and knowledge level. At the same time, natural language processing technology can extract features from the text and question content of course materials and accurately locate the correlation between questions and knowledge points. However, there is currently a lack of a complete combined method for intelligent question push and feedback evaluation in the process of medical and preventive integration and medical and preventive collaborative talent training. Therefore, there is an urgent need for an innovative intelligent system that fully combines "causal-machine learning" and natural language processing technology, which can dynamically push exercises that best suit students' current learning needs based on their knowledge mastery, while providing intelligent learning feedback to improve students' learning outcomes in the integration of medical and prevention and the cultivation of collaborative talents, and assist teachers in providing personalized teaching guidance. Summary of the invention

[0006] The purpose of the present invention is to provide a medical-prevention fusion skills evaluation method based on a causal-machine model, which can intelligently recommend homework questions suitable for students' knowledge level and generate targeted learning feedback.

[0007] To achieve the above object, the present invention adopts the following technical solution:

[0008] A method for evaluating medical and preventive integration skills based on a causal-machine model, comprising the following steps:

[0009] Step 1: Obtain the text of each question in the core skill question bank for preventive medicine talent training, as well as the text of knowledge points in the professional training program course materials; use natural language processing technology combined with the pre-trained BERT model to quantify the question text and knowledge point text to generate question feature vectors and knowledge point semantic vectors, use cosine similarity to calculate the correlation between the question feature vector and the knowledge point semantic vector knowledge feature vector, and obtain the associated knowledge feature vector of the question based on the correlation;

[0010] Step 2: Based on the historical completion rate, historical answer accuracy, and teacher scoring difficulty of the question, the question difficulty coefficient is generated based on the causal-machine model. The difficulty coefficient, question semantic vector, and related knowledge semantic vector are weightedly fused using the attention mechanism to generate a high-dimensional comprehensive feature vector for each question.

[0011] Step 3: Obtain the students' answer results in the core skill question bank of the preventive medicine professional training program, and generate the student knowledge mastery feature vector based on the student's answer accuracy and answer speed combined with the question feature vector; reflect the student's mastery of a certain knowledge point; accumulate the student knowledge mastery feature vectors of related questions according to the associated knowledge feature vector to form a comprehensive mastery feature vector, and extract the comprehensive mastery feature vectors on different knowledge points to form coverage features;

[0012] Step 4: Based on the question feature vector, knowledge point semantic vector and student’s comprehensive mastery feature vector, DeepFM recommends the most suitable homework questions for students based on the deep learning model, giving priority to questions that can help students consolidate their weak links or improve their current level;

[0013] Step 5: Automatically analyze students’ performance based on the causal-machine model and provide intelligent feedback to teachers; generate learning suggestions for students’ weak points in knowledge to help teachers adjust teaching plans in a timely manner.

[0014] Furthermore, in step 1, the BERT model is used to perform context-related semantic modeling on each word in the question, specifically including firstly using the preprocessor of the BERT model to segment the question text and the knowledge point text into smaller units; then, the segmented words are processed by the BERT model to generate a context-related embedding vector for each word; and then the generated context-related embedding vectors are summarized to finally generate a question semantic vector.

[0015] Furthermore, the difficulty coefficient of the question generated based on the causal-machine model in step 2 includes:

[0016] First, by analyzing the completion rate, accuracy rate and scoring difficulty of the questions, a causal relationship diagram between question features and knowledge point features was constructed;

[0017] Then, the fuzzy prior weight generation generates fuzzy membership for the edges in the causal graph based on domain expert knowledge and historical data;

[0018] Finally, the Bayesian network optimization uses the partitioned MCMC algorithm to optimize the structure of the causal relationship graph, and the conditional probability distribution estimation estimates the conditional probability distribution of the question features based on the optimized Bayesian network and generates the question difficulty coefficient.

[0019] Furthermore, in step 3, the student's knowledge mastery feature vector is regularly updated to dynamically update the student's knowledge point mastery.

[0020] Furthermore, the factorization machine model of the DeepFM model captures the direct interaction between the question feature vector and the comprehensive mastery feature vector to obtain the matching degree of a certain type of question to a specific student; the dot product between the two vectors is calculated to evaluate the basic matching degree between the student and the question; the multi-layer neural network of the DeepFM model captures the complex nonlinear relationship between the question feature vector and the comprehensive mastery feature vector to obtain accurate recommendation results from a more complex model.

[0021] Furthermore, reinforcement learning is introduced into the DeepFM model, and each student’s current knowledge mastery and homework completion status are used as the state in reinforcement learning to recommend a series of questions to students. After each recommendation, the recommendation strategy is adjusted according to the student’s answer performance.

[0022] Furthermore, the reinforcement learning adopts an off-policy learning method based on a value function.

[0023] Furthermore, a reward mechanism is introduced to evaluate the recommendation results; when students perform well on the recommended topics, positive rewards are given; if students perform poorly on the recommended topics, negative feedback is given.

[0024] Furthermore, the homework questions recommend questions that students can consolidate their weak points or improve their current level.

[0025] Furthermore, the analysis of students' answering performance in step 5 refers to the assessment of students' knowledge and the assessment of question quality.

[0026] Furthermore, in step 5, the automatic analysis of students' answering performance in the causal-machine model refers to: the evaluation of students' knowledge is to evaluate the accuracy of identifying students' knowledge levels and the accuracy of identifying students' knowledge weaknesses, which is calculated as the proportion of successfully identifying weak knowledge points;

[0027] The quality assessment of questions is to calculate the completion rate of students' homework after the recommended homework, reflecting the attractiveness and applicability of the recommendation; and calculate the overlap rate between the semantic vectors of the knowledge points involved in the recommended homework and the semantic vectors of the knowledge points that students have not yet mastered.

[0028] The present invention has the following beneficial effects:

[0029] (1) Combining the "causal-machine" model with natural language processing, relying on a data-driven model, it can automatically learn and adapt to students' learning patterns, thereby reducing labor costs. At the same time, it can comprehensively consider multi-dimensional student data, including answering speed and answering accuracy, and then provide more comprehensive personalized recommendations and accurately identify students' weak links. Compared with traditional methods based on user behavior analysis, the present invention uses reinforcement learning technology to achieve intelligent optimization of homework recommendation strategies, making the recommendation process more intelligent and personalized, and reducing the problems caused by collecting complex data.

[0030] (2) Personalized homework recommendations for the medical and preventive medicine integrated talent training course can intelligently recommend homework topics that match students' knowledge level and learning needs by analyzing their mastery of preventive medicine course knowledge points;

[0031] (3) Improve the effectiveness of preventive medicine learning. By dynamically adjusting the recommended homework content, students can learn within the correct knowledge framework, ensuring that their basic knowledge, data analysis skills, and application capabilities in the field of preventive medicine are improved, thereby improving the overall learning effect.

[0032] (4) Assist teachers to optimize teaching strategies, gain an in-depth understanding of students’ learning progress and weak links in the training of integrated medical and preventive medicine talents, and thus adjust teaching plans and strategies. For example, if multiple students perform poorly in the “regression analysis” section, teachers can strengthen the teaching of relevant knowledge points in a targeted manner. By collecting and analyzing data such as students’ homework completion rate and answer accuracy, a basis can be provided for educational decision-making, helping educational administrators to formulate more reasonable preventive medicine course design and teaching arrangements, and improving the overall teaching quality.

[0033] (5) Promote the intelligent transformation of medical and preventive integrated talent training. Combining deep learning and data mining technology, it provides new solutions for the intelligent development of preventive medicine education, helps promote the personalized training of medical and preventive collaborative talents, and improves the quality of education. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram of the process of the present invention.

[0035] Figure 2 This is a deep learning model processing flow chart of the present invention. DETAILED DESCRIPTION

[0036] like Figure 1 As shown, this embodiment provides a method for evaluating medical and preventive fusion skills based on a causal-machine model, comprising the following steps:

[0037] Step 1: Obtain the text of each question in the core skills question bank for preventive medicine talent training, as well as the knowledge point text of the professional training program course materials; use natural language processing (NLP) technology combined with a pre-trained deep learning model to quantify the question text and knowledge point text to generate question feature vectors and knowledge point semantic vectors, use cosine similarity to calculate the correlation between the question feature vector and the knowledge point semantic vector, and obtain the associated knowledge feature vector of the question based on the correlation.

[0038] This embodiment uses a pre-trained BERT model to perform context-related semantic modeling on each word in the question, so as to obtain a more accurate representation of the semantic features of the question. First, the preprocessor of the BERT model is used to divide the question text and the knowledge point text into smaller units (divided into words or subwords, called "tokens"). Special markers [CLS] and [SEP] are added at the beginning and end of the text to indicate the start and end of the text. For example, "Question 1: Please briefly describe the difference between the standard deviation and the standard error." After tokens processing, it becomes "[CLS], please, briefly, standard deviation, and, standard error, of, different, [SEP]".

[0039] The segmented words are processed through the BERT model to generate context-dependent embedding vectors for each word. The BERT model not only considers the semantics of the word itself, but also generates more accurate embedding vectors based on the context, which can retain the subtle changes in vocabulary in different contexts. For example, in question 1, the BERT model will consider the relationship between "standard deviation" and "brief description" and "different", and distinguish this question from a simple calculation question of the standard deviation. For each question, the model summarizes the context-dependent vectors it generates and finally generates a question semantic vector. Similarly, for each knowledge point, a corresponding knowledge point semantic vector will be generated, which can represent the semantic content of the question or knowledge point.

[0040] Each question usually involves one or more knowledge points. The similarity between the question and each knowledge point is expressed by calculating the cosine similarity between the question semantic vector and the knowledge point semantic vector. Cosine similarity calculates the cosine value of the angle between two vectors. The closer the value is to 1, the closer the semantics of the two vectors are, and thus the greater the correlation between the question semantic vector and the knowledge point semantic vector.

[0041] Text i represents the semantic vector of question 1, Ka is the semantic vector of knowledge point a "standard deviation", and the cosine similarity is approximately equal to 0.95, indicating that students should be able to apply the content of knowledge point a "standard deviation definition" well when answering this question. At the same time, it also means that knowledge point a "standard deviation" is the core of understanding the question. Similarly, Text iThe cosine similarity with Kb (the knowledge point semantic vector of knowledge point b "standard error definition") is 0.90, which is slightly lower than knowledge point a, but it is still an important supplementary content. If students can combine the definitions of the two, they will be able to answer the questions more comprehensively. However, the cosine similarity with respect to knowledge point c "the calculation method of describing the standard deviation" is 0.85. Although it is still relevant, it is slightly broader than the first two knowledge points. Students may know some related concepts, but they are not as specific as the first two. Therefore, the relevant knowledge points of Question 1 are more inclined to be marked as the definition of standard deviation and the definition of standard error, rather than the calculation method of standard deviation. After calculating the cosine similarity with multiple knowledge points, this embodiment obtains the relevant features of the question and the knowledge point, "Question 1: Please briefly describe the difference between standard deviation and standard error": {"standard deviation definition": 0.95, "standard error definition": 0.90, "standard deviation calculation method": 0.80, ...}".

[0042] Step 2: Generate the difficulty coefficient of the question based on the historical completion rate, historical answer accuracy, and teacher scoring difficulty of the question based on the causal-machine model, and construct the association path between the question and the knowledge point through the causal relationship graph, generate the prior weights in combination with fuzzy theory, and use the attention mechanism to weightedly fuse the difficulty coefficient, question semantic vector, and associated knowledge semantic vector to generate a high-dimensional comprehensive feature vector for each question;

[0043] Causal relationship graph construction constructs a causal relationship graph between question features and knowledge point features by analyzing the completion rate, correct rate, and scoring difficulty of the questions. For example, for "Question 1" {"Completion rate": 0.75, "Correct rate": 0.60, "Scoring difficulty": 3.5} and "Question 2" {"Completion rate": 0.85, "Correct rate": 0.90, "Scoring difficulty": 2.0}, the model recognizes that the completion rate and correct rate not only directly affect the difficulty of the question, but also indirectly affect the difficulty estimation through the associated knowledge point features.

[0044] Fuzzy prior weight generation generates fuzzy membership for edges in the causal graph based on domain expert knowledge and historical data. For example, the model generates a weight of 0.8 for the relationship "completion rate → difficulty coefficient" and a weight of 0.6 for the relationship "accuracy rate → difficulty coefficient". These prior weights ensure that the model pays more attention to the critical path during the learning process.

[0045] Bayesian network optimization uses the partitioned MCMC algorithm to optimize the structure of the causal relationship graph to ensure efficient capture of dependencies between complex variables. For example, for "Question 1", the model generates an indirect path through optimization: "Associated knowledge points → completion rate → difficulty coefficient", indicating that the coverage of the associated knowledge points of this question is an important implicit variable affecting the completion rate.

[0046] Conditional probability distribution estimation is based on the optimized Bayesian network to estimate the conditional probability distribution of question features. For example, the model calculates the difficulty coefficient of "Question 1" as 0.68 based on the completion rate and accuracy rate, while the difficulty coefficient of "Question 2" is 0.34, indicating that "Question 1" is significantly more difficult than "Question 2". This result combines the comprehensive analysis of explicit data (completion rate, accuracy rate) and implicit relationships (knowledge point coverage).

[0047] Through the above steps, after the model generates the difficulty coefficient of the question, it uses the attention mechanism to weightedly fuse the difficulty coefficient, the question semantic vector and the associated knowledge semantic vector to generate a high-dimensional comprehensive feature vector.

[0048] For example, the essay question "In the non-parametric method, what are the methods for comparing survival rates? What are their characteristics?" Knowledge point characteristics: {"Kaplan-Meier method": 0.95,"Log-rank test": 0.90, ...} The model identifies the path of "knowledge point coverage → completion rate → difficulty coefficient" through the causal relationship diagram, indicating that knowledge point coverage is the core factor affecting difficulty. The difficulty coefficient is estimated to be 0.80 through fuzzy weight and conditional probability.

[0049] Multiple choice question: "A disease is affected by smoking (X1) and drinking (X2). Write a COX model and explain the regression coefficient." Knowledge point characteristics: {"Cox regression model": 0.95, ...} The causal relationship diagram shows that the direct path weight of "completion rate → difficulty coefficient" is high, indicating that completion rate is the main influencing factor of this question. The difficulty coefficient is estimated to be 0.60.

[0050] In the final fusion process, the model assigns higher weights to difficulty features, highlighting features with lower complexity but higher problem-solving requirements.

[0051] Step 3: Obtain the students' answer results in the core skill question bank of the preventive medicine professional training program, and use the students' correct answer rate Acc ji and answering speed ji Combined with the question feature vector Text i Generate student knowledge mastery feature vector; S ji =Text i ·(Acc ji ×Speed ji ), which reflects the degree of students’ mastery of a certain knowledge point; the comprehensive mastery feature vector is formed by adding the student knowledge mastery feature vector of related questions to the associated knowledge feature vector, and the comprehensive mastery feature vector of different knowledge points is extracted to form the coverage feature.

[0052] For example, when students are doing a question, the feature vector Text iWhen answering questions with high accuracy or low speed, even if the students complete the questions, it means that they are not proficient in this knowledge point, which will be reflected in the semantic vector of the question. For example, for question 5: "Please explain the basic principle of linear regression model." The student's accuracy is 0.5 and the speed is 0.4. For question 6: "When is the logistic regression model used? Please give an example." The student's accuracy is 0.8 and the speed is 0.2. Finally, the student's knowledge mastery feature vector is obtained by weighting the accuracy and speed. Although the student did cover the knowledge points of questions 5 and 6 at the same time when answering the questions, the student performed very well in question 6 (high accuracy and moderate speed), and its mastery feature vector reflects a good understanding of the relevant knowledge points; while in question 5, due to the poor performance of the student (low accuracy and slow speed), the mastery feature vector is low. This correction mechanism helps to accurately evaluate the students' mastery of knowledge points and provide targeted help in subsequent learning.

[0053] For each knowledge point, students often complete multiple questions of different difficulty and types. Therefore, this embodiment needs to combine the answer results of multiple questions to generate the student's comprehensive mastery of a specific knowledge point. For example, a student's comprehensive mastery feature vector of the "standard deviation" knowledge point can be expressed as: S j =∑ i∈related questions S ji By accumulating multiple related question feature vectors, it is possible to evaluate students' overall mastery of a knowledge point.

[0054] Dynamically update the student's knowledge mastery, because the student's knowledge mastery is not static, but changes with time and practice. In order to more accurately reflect the student's learning progress, the student's knowledge mastery feature vector is updated regularly. For example, when a student re-answers certain questions or shows progress in new exercises, the student's knowledge mastery feature vector is dynamically adjusted based on the latest answer results and time records. By generating a student's knowledge mastery feature vector for each knowledge point, a comprehensive understanding of the student's learning situation can be formed.

[0055] Step 4: Based on the question feature vector, knowledge point semantic vector and student's comprehensive mastery feature vector, DeepFM recommends the most suitable homework questions for students based on the deep learning model, and gives priority to questions that can help students consolidate their weak links or improve their current level.

[0056] The linear part of the DeepFM model captures the direct interaction between the question feature vector and the comprehensive mastery feature vector through the factorization machine model to obtain the matching degree of a certain type of question for a specific student; the dot product between the two vectors is calculated to evaluate the basic matching degree between the student and the question. The factorization model can accurately calculate the linear matching degree between the student and the question. For example, if a student performs poorly on questions related to "regression analysis", the factorization module will capture this relationship and provide basic information for subsequent recommendations.

[0057] The deep part of the DeepFM model captures the complex nonlinear relationship between the question feature vector and the comprehensive mastery feature vector through a multi-layer neural network, ensuring that the recommendation can derive accurate recommendation results from a more complex pattern. Not only the question feature vector and the comprehensive mastery feature vector of the question are considered, but also the high-order interactions between the features are deeply explored. The question feature vector and the comprehensive mastery feature vector will be input into multiple hidden layers, and the nonlinear feature interactions will be extracted layer by layer through the activation function. Through the processing of multi-layer neural networks, the complex nonlinear relationship between students and questions can be captured. For example, a student's performance in a data analysis question may have a high-order connection with his performance in a statistical inference question. The deep part identifies and utilizes this complex relationship to achieve more accurate personalized recommendations.

[0058] Based on the existing DeepFM model, this embodiment introduces a reinforcement learning part to optimize the recommendation by continuously adjusting the recommendation strategy. Each student's current knowledge mastery and homework completion status are used as the "state" in reinforcement learning, including historical homework performance, knowledge point mastery and other information. "Action" is to recommend a series of questions to students. After each recommendation, the recommendation strategy is adjusted according to the student's answer performance (including accuracy, answering time, and learning fatigue). The effect of the recommendation is evaluated through a reward mechanism.

[0059] Specifically, when students perform well on the recommended questions (improved accuracy and knowledge mastery), they will be given positive rewards. On the contrary, if students fail to improve their performance on the recommended questions, they will be given negative feedback, so as to adjust the model parameters and optimize the recommendation strategy. It not only focuses on short-term benefits, such as students' performance on the current questions, but also on long-term benefits, that is, the continuous progress of students throughout the learning process. Through dynamic adjustment, the questions that best suit the students' current learning status and long-term development can be gradually recommended, thereby maximizing the learning effect. The introduction of reinforcement learning enables the recommendation system to be continuously optimized, not only relying on static feature matching, but also dynamically adjusting the recommendation strategy according to the students' learning progress. It can automatically identify students' weaknesses, and gradually introduce more difficult questions after students' performance improves, helping students achieve continuous progress.

[0060] For example, if the student's mastery of "cohort study concept" is 0.65 and the mastery of "clinical trial design" is 0.46, we will push the difficult questions of "cohort study" knowledge point, and recommend the relatively simple questions of "clinical experimental research design". Assuming that the student has completed the questions of "clinical trial design" well (fast speed and high accuracy), but the accuracy and speed of the difficult questions of "cohort design" are low, we will update the student's learning status according to the reward mechanism, fine-tune the student's mastery of these two related knowledge points, and the updated results will be used for the next question push.

[0061] By introducing the DeepFM model combined with the dynamic adjustment mechanism of reinforcement learning, we have innovatively realized intelligent recommendations for students' personalized learning needs. The DeepFM model can accurately capture the interaction between question features and student knowledge mastery features, while the application of reinforcement learning ensures the recommendation system. Generate corresponding feedback to teachers based on students' answers, identify questions where they performed poorly in the answering process, and determine the knowledge points that students have not yet mastered through the knowledge points associated with the questions. Feedback information is generated based on students' incorrect knowledge points. The system will list the knowledge points involved in the questions that students answered incorrectly and provide relevant learning suggestions.

[0062] Step 5: The “Cause-and-Effect” model is used to automatically analyze students’ performance and provide intelligent feedback to teachers. Learning suggestions are generated for students’ weak points in knowledge, helping teachers to adjust teaching plans in a timely manner.

[0063] Automatically record and analyze students' performance in completing homework and tests, mainly involving the correct answer rate to record whether students answered each question correctly or not; the answer time reflects the students' familiarity with the question and the difficulty of solving the question, and is analyzed based on the match between the answer time and the complexity of the question. Track students' historical answer records, especially whether their performance on the same knowledge point has improved or regressed in different periods.

[0064] For student knowledge assessment, the "Cause-to-Machine" model combines causal analysis of predictions and actual test results to evaluate the accuracy of student knowledge level identification. At the same time, causal inference is used to identify students' weak knowledge points and calculate the proportion of successful identification of weak knowledge points. For example, through the causal relationship diagram, students' performance on the knowledge points of "experimental design" and "causal inference" is analyzed to accurately locate weak links and generate targeted supplementary suggestions.

[0065] To evaluate the quality of the questions, the "Cause-Machine" model calculates the completion rate of students' homework after the recommended homework, reflecting the attractiveness and applicability of the recommendation. At the same time, the overlap rate between the knowledge points involved in the recommended homework and the knowledge points that students have not yet mastered is calculated based on the causal graph, and the coverage of the recommended questions is optimized. For example, most of the recommended homework questions are completed, indicating that the difficulty of the questions matches the students' learning level well, but some questions do not cover the key knowledge points of "causal inference", so the system needs to further optimize the knowledge point coverage.

[0066] The final evaluation knowledge constraint is C = {"{student knowledge evaluation: [knowledge mastery accuracy, knowledge weak link identification rate]}", "{question quality evaluation: [homework completion rate, knowledge point coverage rate]}". For example, a student has a good grasp of health statistics knowledge, but still has certain knowledge blind spots in the epidemiology part, and the accuracy of his knowledge mastery can be obtained; through data analysis, the student's weak links in the two knowledge points of "experimental design" and "causal inference" can be accurately identified, which can help the subsequent recommended questions to be supplemented in a targeted manner, and the feedback is good, and the student's knowledge weak link identification rate can be obtained; most of the homework questions recommended for students are completed, indicating that the difficulty of the questions matches the students' learning level well, and thus the student's homework completion rate can be obtained. Although the quality of the recommended questions is high, some questions do not involve the difficulty of "causal inference", so the system's question coverage of knowledge points needs to be further optimized, and thus the student's homework completion rate can be obtained. Finally, the student's knowledge was assessed as follows: [Knowledge mastery accuracy: 85%, identification rate of weak knowledge links: 90%]", "Question quality assessment: [Homework completion rate: 95%, knowledge point coverage: 80%].

[0067] The above description is only a preferred implementation manner of the present invention, but the protection scope of the present invention is not limited thereto, and any modification and replacement based on the technical solution and inventive concept provided by the present invention should be included in the protection scope of the present invention.

Claims

1. A method for evaluating medical and preventive fusion skills based on a causal-machine model, characterized in that: The steps include: Step 1: Obtain the text of each question in the core skill question bank for preventive medicine talent training, as well as the text of knowledge points in the professional training program course materials; use natural language processing technology combined with the pre-trained BERT model to quantify the question text and knowledge point text to generate question feature vectors and knowledge point semantic vectors, use cosine similarity to calculate the correlation between the question feature vector and the knowledge point semantic vector knowledge feature vector, and obtain the associated knowledge feature vector of the question based on the correlation; Step 2: Based on the historical completion rate, historical answer accuracy, and teacher scoring difficulty of the question, the question difficulty coefficient is generated based on the causal-machine model. The difficulty coefficient, question semantic vector, and related knowledge semantic vector are weightedly fused using the attention mechanism to generate a high-dimensional comprehensive feature vector for each question. Step 3: Obtain the students' answer results in the core skill question bank of the preventive medicine professional training program, and generate the student knowledge mastery feature vector based on the student's answer accuracy and answer speed combined with the question feature vector; reflect the student's mastery of a certain knowledge point; accumulate the student knowledge mastery feature vectors of related questions according to the associated knowledge feature vector to form a comprehensive mastery feature vector, and extract the comprehensive mastery feature vectors on different knowledge points to form coverage features; Step 4: Based on the question feature vector, knowledge point semantic vector and student’s comprehensive mastery feature vector, DeepFM recommends the most suitable homework questions for students based on the deep learning model, giving priority to questions that can help students consolidate their weak links or improve their current level; Step 5: Automatically analyze students’ performance based on the causal-machine model and provide intelligent feedback to teachers; generate learning suggestions for students’ weak points in knowledge to help teachers adjust teaching plans in a timely manner.

2. The method for evaluating medical and preventive fusion skills based on the causal-machine model according to claim 1 is characterized in that: In step 1, the BERT model is used to perform context-related semantic modeling on each word in the question. Specifically, the BERT model preprocessor is used to segment the question text and knowledge point text into smaller units; Then, the segmented words are processed through the BERT model to generate context-related embedding vectors for each word; the generated context-related embedding vectors are then summarized to finally generate a question semantic vector.

3. The method for evaluating medical and preventive fusion skills based on the causal-machine model according to claim 1 is characterized in that: The difficulty coefficients of the questions generated based on the causal-machine model described in step 2 include: First, by analyzing the completion rate, accuracy rate and scoring difficulty of the questions, a causal relationship diagram between question features and knowledge point features was constructed; Then, the fuzzy prior weight generation generates fuzzy membership for the edges in the causal graph based on domain expert knowledge and historical data; Finally, the Bayesian network optimization uses the partitioned MCMC algorithm to optimize the structure of the causal relationship graph, and the conditional probability distribution estimation estimates the conditional probability distribution of the question features based on the optimized Bayesian network and generates the question difficulty coefficient.

4. The method for evaluating medical and preventive fusion skills based on the causal-machine model according to claim 1 is characterized in that: In step 3, the students' knowledge mastery feature vector is updated regularly to dynamically update the students' knowledge mastery.

5. The method for evaluating medical and preventive fusion skills based on the causal-machine model according to claim 1 is characterized in that: The factorization machine model of the DeepFM model captures the direct interaction between the question feature vector and the comprehensive mastery feature vector to obtain the matching degree of a certain type of question to a specific student; The dot product between two vectors is calculated to evaluate the basic match between students and questions. The multi-layer neural network of the DeepFM model captures the complex nonlinear relationship between the question feature vector and the comprehensive mastery feature vector, and derives accurate recommendation results from a more complex model.

6. The method for evaluating core skills of medical and preventive integration based on deep learning according to claim 5 is characterized in that: Introducing reinforcement learning into the DeepFM model, each student's current knowledge and homework completion status is used as the state in reinforcement learning to recommend a series of questions to students; After each recommendation, adjust the recommendation strategy based on the student’s answer performance.

7. The method for evaluating core skills of medical and preventive integration based on deep learning according to claim 6 is characterized in that: The reinforcement learning adopts an off-policy learning method based on value function.

8. The method for evaluating core skills of medical and preventive integration based on deep learning according to claim 6 is characterized in that: Introduce a reward mechanism to evaluate the recommendation results; when students perform well on the recommended topics, give positive rewards; if students perform poorly on the recommended topics, give negative feedback.

9. The method for evaluating core skills of medical and preventive integration based on deep learning according to claim 6, characterized in that: Assignment topics are recommended for students to consolidate their weak points or to improve their current level.

10. The method for evaluating core skills of medical and preventive integration based on deep learning according to claim 1, characterized in that: The analysis of students' answering performance described in step 5 refers to the assessment of students' knowledge and the assessment of question quality.

11. The method for evaluating core skills of medical and preventive integration based on deep learning according to claim 1, characterized in that: In step 5, the automatic analysis of students' answering performance by the causal-machine model means: the evaluation of students' knowledge is to evaluate the accuracy of identifying students' knowledge levels and the accuracy of identifying students' knowledge weaknesses, which is calculated as the proportion of successfully identifying weak knowledge points; The quality assessment of questions is to calculate the completion rate of students' homework after the recommended homework, reflecting the attractiveness and applicability of the recommendation; and calculate the overlap rate between the semantic vectors of the knowledge points involved in the recommended homework and the semantic vectors of the knowledge points that students have not yet mastered.

Images