Online course recommendation method and device applied to emergency treatment

Abstract

The invention discloses an online course recommendation method and device applied to emergency treatment, and relates to the technical field of intelligent medical education, and the method comprises the steps: collecting user learning behavior data, patient condition data and emergency treatment nursing course content data, dynamically modeling emergency treatment nursing knowledge points and a multi-order association relationship based on cross-modal adversarial transfer learning, and carrying out online course recommendation. Constructing an emergency nursing knowledge graph; and dynamically adjusting the priority of the candidate course decision cluster according to real-time case distribution characteristics in the patient condition data, outputting a high-priority course set, performing matching degree calculation in combination with user learning behavior characteristics, generating a personalized recommendation list, and outputting recommended courses to a user terminal. Through a neural symbol causal intervention engine and a dynamic updating mechanism, a knowledge blind spot propagation path is mapped to a course recommendation network, a recommendation strategy is adaptively adjusted, and dynamic adaptation of course content and real-time case features is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent medical education, and particularly to an online emergency course recommendation method and device. Background Art

[0002] In the prior art, knowledge graph construction is mainly based on the structured analysis of course content (such as logical dependency trees) and shallow statistical features of user learning behavior (such as click-through rate, learning duration), and course recommendation is achieved through collaborative filtering algorithms. At the same time, electronic medical record systems extract patient vital sign features through time series analysis (such as LSTM) and generate nursing suggestions in combination with rule engines.

[0003] However, the existing user learning behavior data (especially wrong-question feedback tags) are sparse and discontinuous, resulting in calculation deviations in the association weights of knowledge graph nodes and making it difficult to accurately locate knowledge blind spots. The existing time series models (such as RNN) have insufficient ability to model the causal associations between vital sign parameters and do not incorporate clinical logic ontology constraints, resulting in a low matching degree between the recommended courses and real-time case features. In particular, the existing adversarial transfer learning framework lacks the embedding of emergency medicine ontology (such as ICNP terminology sets), and the generated high-density feedback data may violate clinical logic consistency, thereby affecting the inference accuracy of the knowledge graph. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an online emergency course recommendation method to solve the problem of low course recommendation accuracy caused by the missing associations and insufficient dynamic causal reasoning in the emergency nursing knowledge graph due to the sparsity of user behavior data.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, the present invention provides an emergency online course recommendation method, which includes collecting user learning behavior data, patient condition data, and emergency nursing course content data, dynamically modeling emergency nursing knowledge points and multi-level association relationships based on cross-modal adversarial transfer learning, and constructing an emergency nursing knowledge graph; performing multi-modal data fusion on the user learning behavior data, patient condition data, and emergency nursing knowledge graph, extracting user learning behavior features, patient condition features, and emergency nursing knowledge graph association features, and generating a three-dimensional feature vector; through a neuro-symbolic causal intervention engine, combining with the emergency nursing knowledge graph for dynamic causal reasoning, locating the user's knowledge blind spots, and mapping the spatio-temporal propagation chain of the knowledge blind spots to the probability flow network of course nodes to generate a candidate course decision cluster; according to the real-time case distribution characteristics in the patient condition data, dynamically adjusting the priority of the candidate course decision cluster, outputting a high-priority course set, calculating the matching degree in combination with the user learning behavior features, generating a personalized recommendation list, and outputting the recommended courses to the user terminal.

[0007] As a preferred solution of the emergency online course recommendation method described in the present invention, wherein: the cross-modal adversarial transfer learning is used to perform spatio-temporal alignment and sparse compensation on multi-source data, and the emergency nursing knowledge points and multi-level association relationships are modeled through a dynamic time-series causal network.

[0008] As a preferred solution of the emergency online course recommendation method described in the present invention, wherein: the neuro-symbolic causal intervention engine is composed of a neural graph network, a symbolic causal rule library, and a dynamic weight intervention operator, and performs causal decoupling and inference path reconstruction by real-time correcting the edge weights of the emergency nursing knowledge graph; The generation of the candidate course decision cluster specifically includes the following steps: Injecting dynamic cognitive weights into the emergency nursing knowledge graph through a neuro-symbolic causal intervention engine to construct a multi-modal causal inference path; Using adversarial topological reinforcement learning to map the spatio-temporal propagation chain of the knowledge blind spots to the probability flow network of course nodes to generate a candidate course decision cluster.

[0009] As a preferred solution of the emergency online course recommendation method described in the present invention, wherein: the dynamic adjustment of the priority of the candidate course decision cluster means that based on the patient condition data, temporal and semantic features are extracted through TCN / BERT, combined with K-means clustering for cosine similarity matching, and low-priority candidate courses are filtered after dynamic weighted scoring, and a high-priority course set is output.

[0010] As a preferred solution of the emergency online course recommendation method described in the present invention, wherein: the patient condition data includes a vital sign time-series sequence, a chief complaint text record, and a preliminary diagnosis label.

[0011] As a preferred solution of the emergency online course recommendation method described in the present invention, wherein: the calculation of the matching degree includes calculating the user-course feature matching degree based on the three-dimensional feature vector and the high-priority course set through cosine similarity.

[0012] As a preferred solution of the emergency online course recommendation method described in the present invention, wherein: the multimodal data fusion refers to standardizing and semantically extracting user behavior data, extracting the time series and medical features of patient data, combining the graph embedding of the emergency nursing knowledge graph, and generating a three-dimensional feature vector through a fully connected layer to unify dimensions, attention fusion, and PCA dimensionality reduction.

[0013] In a second aspect, the present invention provides an emergency online course recommendation device, including data collection, feature fusion, causal reasoning, and dynamic recommendation; the data collection is used to collect user learning behavior data, patient condition data, and emergency nursing course content data, dynamically model emergency nursing knowledge points and multi-order association relationships based on cross-modal adversarial transfer learning, and construct an emergency nursing knowledge graph; the feature fusion is used to perform multimodal data fusion on user learning behavior data, patient condition data, and the emergency nursing knowledge graph, extract user learning behavior features, patient condition features, and emergency nursing knowledge graph association features, and generate a three-dimensional feature vector; the causal reasoning is used to perform dynamic causal reasoning through a neuro-symbolic causal intervention engine in combination with the emergency nursing knowledge graph, locate the user's knowledge blind spots, and map the spatio-temporal propagation chain of the knowledge blind spots to the probability flow network of the course nodes to generate a candidate course decision cluster; the dynamic recommendation is used to dynamically adjust the priority of the candidate course decision cluster according to the real-time case distribution characteristics in the patient condition data, output a high-priority course set, calculate the matching degree in combination with the user learning behavior features, generate a personalized recommendation list, and output the recommended courses to the user terminal.

[0014] In a third aspect, the present invention provides a computer device, including a memory and a processor, wherein: when the computer program is executed by the processor, it implements any step of the emergency online course recommendation method described in the first aspect of the present invention.

[0015] In a fourth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, wherein: when the computer program is executed by the processor, it implements any step of the emergency online course recommendation method described in the first aspect of the present invention.

[0016] The beneficial effects of the present invention are as follows: By integrating the emergency medicine ontology through a cross-modal adversarial transfer learning framework, high-density feedback data that conforms to clinical logical consistency is generated, effectively compensating for the sparsity of user behavior data and enhancing the integrity and medical rationality of the emergency care knowledge graph; Combining a temporal convolutional network and a graph attention network to dynamically model the temporal features of vital signs and multi-order causal associations, accurately locating the user's knowledge blind spots and capturing the causal chains in complex clinical situations; Through a neuro-symbolic causal intervention engine and a dynamic update mechanism, mapping the knowledge blind spot propagation path to the course recommendation network, adaptively adjusting the recommendation strategy, achieving dynamic adaptation of the course content to the real-time case features, and continuously optimizing the recommendation strategy through user feedback and clinical data, to construct an adaptive personalized learning support mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0018] Figure 1 It is a flowchart of an emergency online course recommendation method.

[0019] Figure 2 It is an interaction diagram of an emergency online course recommendation architecture.

[0020] Figure 3 It is a causal reasoning flowchart.

[0021] Figure 4 It is a dynamic recommendation flowchart. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will provide a detailed description of the specific embodiments of the present invention with reference to the accompanying drawings of the specification.

[0023] Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0024] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that can be included in at least one implementation of the present invention. The appearances of "in one embodiment" in different places in this specification do not all refer to the same embodiment, nor are they separate or selectively exclusive embodiments from other embodiments.

[0025] Example 1, referring to Figures 1 to 4 , which is the first embodiment of the present invention. This embodiment provides an online emergency course recommendation method, including the following steps: S1. Collect user learning behavior data, patient condition data, and emergency nursing course content data. Extract emergency nursing knowledge points and their association relationships through natural language processing methods, and construct an emergency nursing knowledge graph.

[0026] Specifically, it includes the following steps: Collect learning behavior data from the user learning terminal, including course click frequency, single learning duration, knowledge point annotation records, and wrong question feedback tags. Set the time window to ±5 minutes and group the data by user ID; obtain patient condition data from the hospital emergency electronic medical record system in real time, including the time series sequence of vital signs (blood pressure, heart rate, blood oxygen saturation), main complaint text records, and preliminary diagnosis tags, with a data collection frequency of 1 time per minute; extract course content data from the emergency nursing course platform, including teaching video text transcripts, knowledge point tags, and course logical dependency relationship trees, where the logical dependency relationship tree is automatically generated through the hierarchical structure of the course syllabus.

[0027] Take the collected user learning behavior data, patient condition data, and emergency nursing course content data as multi-source data, and perform spatio-temporal alignment and cleaning, specifically including: time series alignment, noise processing, outlier detection, and missing value filling.

[0028] Perform spatio-temporal alignment and sparse compensation on the multi-source data based on cross-modal adversarial transfer learning: Construct a generative adversarial network (GAN) framework as follows: Generator: The input is the sparse label of the user's wrong questions and the course knowledge point dependency relationship tree , and the output is virtual high-density feedback data , and the generator loss function , which is expressed as: ; In the formula, is the expected value, representing the expectation of the random noise variable , obeys the distribution , represents the discriminator function, represents the generator function, is the random noise vector, which is the input of the generator, is the sparse label, representing the sparse feedback data (such as wrong question labels) in the user's learning behavior, is the knowledge dependency tree, representing the logical dependency relationship between course knowledge points; It should be noted that the discriminator function is a neural network. Its input is data samples (which can be real data or generated data), and its output is a scalar value (usually between 0 and 1), representing the probability that the input data is real data; the generator function is also a neural network. Its input is a random noise vector , and the output is the generated data sample. The goal of the generator is to generate data that is as close as possible to the real data distribution.

[0029] Discriminator: The input is real feedback data and virtual high-density feedback data . Based on the emergency medicine ontology (ICNP terminology set), clinical logic consistency constraints are constructed, and the discriminator loss function is expressed as: ; In the formula, represents the expectation of the real feedback data , represents the expectation of the generated high-density feedback data ; It should be noted that the discriminator loss function , and its goal includes two parts: maximizing the discrimination probability of the discriminator for real data ; minimizing the discrimination probability of the discriminator for generated data ; .

[0030] Adversarial training process: Alternately optimize the generator and the discriminator until the clinical logic passing rate of the generated data (verified based on this experience) ≥ 95%.

[0031] Preferably, through the generative adversarial network (GAN) framework, sparse labels and knowledge dependency trees are used to generate virtual high-density feedback data, solving the problem of sparsity of user learning behavior data. Specifically, the emergency medicine ontology (ICNP terminology set) is used as the constraint condition of the discriminator to ensure the clinical logic consistency of the generated data; the generator and the discriminator are optimized through adversarial training, improving the quality and diversity of the generated data. Further, the integrity and accuracy of the construction of the emergency nursing knowledge graph are improved, providing richer user behavior characteristics for personalized recommendation.

[0032] Model the emergency nursing knowledge points and multi-order association relationships through a dynamic time-series causal network.

[0033] Based on the temporal convolutional network (TCN), extract the temporal features of vital signs. The output of the dilated convolutional layer of the TCN is expressed as: ; In the formula, represents the output value at time step , represents the th weight value, represents the index variable of the weight value, is the lower limit of the summation, is the upper limit of the summation, represents the input value at time step , represents the dilation factor, is the summation index, indicating the current term of the summation; The causal relationship between physical sign parameters is modeled by a graph attention network (GAT), and the attention coefficient is calculated as: ; ; In the formula, represents the attention coefficient between node and node , represents the activation function, represents the transpose of the parameter vector of the attention mechanism , represents the feature of the feature vector of node after linear transformation, represents the feature of the feature vector of node after linear transformation, represents the learnable weight matrix, represents the feature of the feature vector of the neighbor node after linear transformation, represents the time series of the input data, represents the index of the neighbor node of node , represents belongs to the set of neighbor nodes of node ; ; It should be noted that node and node represent two different entity nodes in the graph. Specifically: Node : The current target node (such as a user behavior node, a knowledge point node, or a disease condition node), whose feature vector needs to be updated. Node : Node The neighbor nodes (such as the knowledge points, associated disease condition parameters, or user behavior patterns that have a direct dependency relationship with ).

[0034] Define multi - order association rules: First - order association: direct dependency relationship (such as "cardiopulmonary resuscitation → chest compressions"), and the weight is determined by the curriculum logic tree; Second - order association: cross - modal association (such as the co - occurrence relationship between the user's wrong - question label "endotracheal intubation" and the patient's blood oxygen saturation < 90%), and the weight is generated by adversarial transfer learning; High - order association: temporal causal chain (such as "heart rate > 120 beats per minute → shock risk → intravenous fluid replacement nursing"), and the weight is output by the TCN - GAT joint model.

[0035] Preferably, the temporal features of vital signs are extracted through a temporal convolutional network (TCN), and the causal associations between the sign parameters are modeled by combining with a graph attention network (GAT), realizing the multi - order relationship modeling of emergency nursing knowledge points. Specifically, TCN captures local temporal features, and GAT models the dependency relationships between nodes. The combination of the two can effectively capture complex time - series patterns and causal relationships, enhancing the dynamics and reasoning ability of the emergency nursing knowledge graph, being able to more accurately locate the user's knowledge blind spots, and providing dynamic support for the real - time case distribution characteristics.

[0036] Furthermore, construct and dynamically update the emergency nursing knowledge graph: Define entity nodes: user behavior nodes, disease condition nodes, and knowledge point nodes; Specifically, the feature vectors of user behavior nodes (such as click frequency, learning duration, wrong - question labels) can be automatically recorded and statistically analyzed through a learning management system (such as Moodle, Canvas) and an intelligent question bank system (such as Kahoot); the feature vectors of disease condition nodes (such as average blood pressure, maximum heart rate, minimum blood oxygen saturation) can be collected and calculated through an electronic medical record system (such as Epic, Cerner) and vital sign monitoring devices (such as Philips IntelliVue); the feature vectors of knowledge point nodes (such as logical dependency weight, clinical emergency degree weight) can be defined and assigned through an emergency nursing knowledge graph construction tool (such as Neo4j, Protégé) and a clinical decision - making support system (such as UpToDate).

[0037] Adopt a gated recurrent unit (GRU) to fuse historical weights and real - time data, and dynamically update the edge relationships of each node, expressed as: ; ; ; ; In the formula, To update the output of the gate, representing the control of the current state How much information comes from the previous state and how much information comes from the candidate state , Represents the Sigmoid activation function, Represents the weight matrix of the update gate, Represents the previous state and the current input of the concatenated vector, Represents the output of the reset gate, controlling the previous state on the candidate state influence, Represents the weight matrix of the reset gate, Is the candidate state, representing the potential new state at the current moment, Represents the weight matrix of the candidate state, Represents the output of the reset gate element-wise product with the previous state and then concatenated with the current input concatenation, Represents the element-wise multiplication symbol, Is the complement of the output of the update gate, representing the proportion of retaining the previous state proportion, Is the output of the update gate element-wise product with the candidate state representing the proportion of introducing new information, Represents the hidden state at the current moment, which is the previous state and the candidate state weighted sum of, Represents the hyperbolic tangent function; Preferably, the GRU model can capture long-term dependencies in time series data, ensuring the dynamics and timeliness of the emergency care knowledge graph; through entity node definition and edge relationship update, the structure and scalability of the emergency care knowledge graph are enhanced, providing a real-time and dynamic emergency care knowledge graph for the emergency care field to support personalized recommendation and clinical decision-making.

[0038] It should be noted that through multi-source data collection and standardized processing, sparse data is compensated by using cross-modal adversarial transfer learning, and the time convolutional network (TCN) and graph attention network (GAT) are combined to dynamically model multi-order association relationships, constructing a real-time updated emergency care knowledge graph to provide intelligent support for personalized recommendation and clinical decision-making.

[0039] S2. Perform multi-modal data fusion on user learning behavior data, patient condition data, and the emergency nursing knowledge graph, extract user learning behavior features, patient condition features, and the associated features of the emergency nursing knowledge graph, and generate a three-dimensional feature vector.

[0040] Specifically, it includes the following steps: The extraction of user learning behavior data features includes standardizing the user learning behavior data (such as Z-score standardization). Use word embedding (Word2Vec or BERT) to extract the semantic features of knowledge point annotation records and wrong question feedback labels. Use statistical methods to extract the numerical features (such as mean, variance) of course click frequency and single learning duration, and output user learning behavior features.

[0041] The extraction of patient condition data features includes using a temporal convolutional network (TCN) to extract temporal features from the vital sign time series (blood pressure, heart rate, blood oxygen saturation). Use BERT to extract semantic features from the chief complaint text record. Perform One-Hot encoding on the preliminary diagnosis label or use pre-trained medical word embedding (such as BioBERT) to extract features, and output the patient condition feature vector.

[0042] The extraction of emergency nursing knowledge graph features includes using a graph neural network (GNN) (such as GraphSAGE or GAT) to perform node embedding on the emergency nursing knowledge graph, and extract the associated features of entity nodes and edge relationships. Perform weighted fusion on the features of user behavior nodes, condition nodes, and knowledge point nodes to generate the associated features of the emergency nursing knowledge graph, and output the associated feature vector of the emergency nursing knowledge graph.

[0043] Furthermore, use a fully connected layer to map the user learning behavior features, patient condition features, and the associated features of the emergency nursing knowledge graph to the same dimension, and output a feature vector of a unified dimension; Use the attention mechanism to perform weighted fusion on the feature vector of the unified dimension to form a fused high-dimensional feature vector ; Use principal component analysis (PCA) to reduce the dimension of the high-dimensional feature vector to three dimensions, and output a three-dimensional feature vector Denoted as: ; In the formula, is the projection matrix of PCA; Perform normalization processing on to ensure that the feature values are within a reasonable range (such as [0,1]), and store them in the database for subsequent association reasoning and personalized recommendation.

[0044] S3. Through the neuro-symbolic causal intervention engine, dynamic causal reasoning is carried out in combination with the emergency care knowledge graph to locate the user's knowledge blind spots, and the spatio-temporal propagation chain of the knowledge blind spots is mapped to the probability flow network of the course nodes to generate a candidate course decision cluster.

[0045] Specifically, it includes the following steps: Construct an adaptive weight regulator based on the Graph Attention Network (GAT) to dynamically model the causal associations between the nodes of the emergency care knowledge graph; For any node pair , its causal association weight The calculation process is as follows: ; In the formula, is the node embedding vector, represents the vector concatenation operation; It should be noted that this causal association weight calculation process learns the implicit association features between nodes through the attention coefficient to generate a dynamically corrected edge weight matrix of the emergency care knowledge graph ; Taking the causal rules in the emergency care field as prior constraints, perform causal intervention on , and define the counterfactual intervention operation: ; In the formula, represents the corrected causal association weight after the counterfactual intervention operation, represents the counterfactual intervention operation (such as forcing to learn a certain course, taking values of 0 / 1), represents the intervention intensity coefficient (calculated from the user's learning behavior data, range: 0.5 - 1.5); Specifically, an example of the emergency care causal rule: ; It should be noted that this example demonstrates the application process of the neuro-symbolic causal intervention engine in the emergency care scenario. Through the analysis of the emergency care knowledge graph and the user's learning behavior data, it is located that the user has a knowledge blind spot in the shock staging, and the cognitive decay index is greater than or equal to 07. Based on the counterfactual intervention mechanism, the operation of the forced learning course X is simulated, the counterfactual intervention signal do is set to 1, and the association strength between the shock staging and the fluid resuscitation nodes in the emergency care knowledge graph is adjusted through the dynamic causal weight correction formula. The probability flow network is used to predict the clinical effect after the intervention, and the causal association result of the percentage p of the decrease in the probability of fluid resuscitation failure is quantified. Combining the clinical rule constraints and the course decision clusters generated by the Monte Carlo tree search, the learning course X is dynamically inserted into the high-priority recommendation list to achieve personalized knowledge supplementation, and the safety of the recommended plan is ensured through the verification of the clinical outcome association, such as the improvement of the rescue success rate. This process reflects the full-link closed-loop logic from knowledge blind spot diagnosis, causal reasoning intervention to clinical adaptation recommendation.

[0046] Construct a cognitive decay index model to locate weak points , expressed as: ; In the formula, represents the time window (default: the last 7 days), represents the decay coefficient (default: 0.1, optimized through grid search); When , it is determined that the node is a knowledge blind spot, and the knowledge blind spot set is output: ; Model the knowledge blind spot as the infection source in the graph network, and use the SIR model to simulate the diffusion process, expressed as: ; ; In the formula, represents the infection rate of the node (the initial value is determined by ), represents the susceptible rate (negatively correlated with the edge weight of the emergency care knowledge graph), represents the recovery rate (positively correlated with the user's learning ability, = learning frequency × 0.01); Generate a set of propagation chains through the SIR model, where each chain represents a potential diffusion path of the knowledge blind spot.

[0047] Perform propagation chain pruning and flow network mapping on the set of propagation chains .

[0048] Specifically, path confidence constraints and clinical relevance constraints are set, and the double constraint conditions optimize the propagation chain; Path confidence constraint: ; In the formula, represents the average correction weight of the potential diffusion path ; Clinical relevance constraint: The path needs to contain at least 1 mandatory emergency guideline node; Map the set of pruned propagation chains to the anti-infection nodes in the flow network, and establish a flow conservation equation, which is expressed as: ; In the formula, represents the flow rate into the course node ( ), represents the flow rate out of the course node ( ); Optimize the flow network topology through Monte Carlo Tree Search (MCTS) to generate course decision clusters; Specifically, define the reward function which is expressed as: ; In the formula, represents the balance coefficient (default: 0.7, optimized through A / B testing), represents the course nodes in the course cluster; Synchronously calculate the confidence of the course nodes, which is expressed as: ; ; In the formula, represents the transition probability, represents the prior confidence (defined by clinical urgency, range: 0.8 - 1.0), represents the confidence of the course node ; represents the confidence of the parent node ; represents the parent node to the correction weight of the course node, represents the central course node where the flow rate flows in and out; The final decision cluster only retains the course nodes with high confidence to ensure clinical safety, and outputs the candidate course decision cluster , which is expressed as: .

[0049] S4. Dynamically adjust the priority of the candidate course list according to the real-time case distribution characteristics in the patient's condition data, and output a high-priority course set.

[0050] Specifically, it includes the following steps: Obtain the patient's condition data in real time from the hospital emergency electronic medical record system, including the time series sequence of vital signs (blood pressure, heart rate, blood oxygen saturation), the text record of the chief complaint, and the preliminary diagnosis label; Use a temporal convolutional network (TCN) to extract the temporal features of vital signs, capture the dynamic changes of the patient's condition, and use a BERT model to extract the semantic features of the chief complaint text to identify the patient's main symptoms. Perform One-Hot encoding on the preliminary diagnosis label based on the patient's main symptoms to extract the real-time case distribution characteristics; Use the K-means clustering algorithm to cluster the real-time case distribution characteristics to generate a case distribution feature vector , expressed as: ; ; In the formula, represents the feature vector of the th clustering center, is the number of clusters (number of categories); For each course in the candidate course decision cluster , extract its knowledge point label, clinical urgency weight, and logical dependency relationship, and use a graph neural network (GNN) to embed the course nodes to generate a course feature vector; Use cosine similarity to calculate the matching degree between the case distribution feature vector and the course feature vector to generate a similarity matrix, and for each course, calculate the priority score ; Specifically, extract the similarity value between the course and the case distribution feature vector from the similarity matrix as the basic score; combine the urgency weight of the course, and the urgency weight is defined by the clinical urgency, usually ranging from 0.8 to 1.0, reflecting the clinical priority of the course; adopt the user's learning behavior data, including learning frequency and wrong question feedback, the learning frequency correction factor is used to enhance the priority of the courses frequently learned by the user, and the wrong question feedback correction factor is used to reduce the priority of the courses with more wrong questions for the user; The priority score is calculated through the formula: ; The calculated priority score is .

[0051] Among them, the similarity value reflects the matching degree between the course and the case distribution, the urgency weight reflects the clinical importance of the course, and the learning frequency correction factor and the wrong-question feedback correction factor adjust the influence of the user's learning behavior on the priority respectively, ensuring that the scoring meets both the clinical needs and the user's personalized learning characteristics.

[0052] According to the priority score, sort the courses in the candidate course decision cluster in descending order; Set the priority score threshold (such as ), filter out the courses with scores lower than the priority score threshold, and output the high-priority course set: ; In the formula, represents the th course in the candidate course decision cluster, is the index variable of the courses in the candidate course decision cluster; S5. Combine the user's learning behavior characteristics to calculate the matching degree, generate a personalized recommendation list and output the recommended courses to the user terminal.

[0053] Specifically, it includes the following steps: Based on the high-priority course set and the user's learning behavior characteristics, calculate the knowledge coverage, cognitive load matching degree and clinical urgency weight, and generate a personalized recommendation list by weighting, and output it to the user terminal in real time; Through user terminal data collection points, collect the course click-through rate, completion status and test correct rate in real time as feedback data, and synchronously update the user's learning behavior characteristics (such as wrong-question label decay, learning duration distribution) and the node weights of the emergency nursing knowledge graph (driven by GRU); Based on the feedback data, dynamically adjust the generation rules of the three-dimensional feature vector, including updating the projection matrix by incremental PCA to adapt to the change of data distribution; using reinforcement learning to optimize the attention weight to improve the balance between knowledge coverage and clinical urgency, and output the optimized feature generation rules; Combine the updated emergency nursing knowledge graph and user characteristics, and dynamically adjust the course similarity calculation parameters (such as urgency weight, clustering center drift compensation) and priority scoring strategy through the DQN model, generate a personalized recommendation list and output the recommended courses to the user terminal; Furthermore, conduct a recommendation effect evaluation through the interactive behaviors (such as course click-through rate, completion status, test correct rate and implicit feedback) collected in real time by the user terminal. The core indicators include the knowledge blind spot repair rate, the decrease in clinical error rate, the course completion rate and the rationality of cognitive load.

[0054] Verify the recommendation effect through A / B testing, and update the emergency care knowledge graph in reverse with significant results (such as adding new mandatory nodes and correcting causal association rules) to form a closed-loop iteration.

[0055] This embodiment also provides an emergency online course recommendation device, including: data collection, feature fusion, causal reasoning, and dynamic recommendation; Data collection is used to collect user learning behavior data, patient condition data, and emergency care course content data, dynamically model emergency care knowledge points and multi-level association relationships based on cross-modal adversarial transfer learning, and construct an emergency care knowledge graph; Feature fusion is used to perform multi-modal data fusion on user learning behavior data, patient condition data, and the emergency care knowledge graph, extract user learning behavior features, patient condition features, and emergency care knowledge graph association features, and generate a three-dimensional feature vector; Causal reasoning is used to perform dynamic causal reasoning through a neuro-symbolic causal intervention engine in combination with the emergency care knowledge graph, locate users' knowledge blind spots, and map the spatio-temporal propagation chain of knowledge blind spots to the probability flow network of course nodes to generate a candidate course decision cluster; Dynamic recommendation is used to dynamically adjust the priority of the candidate course decision cluster according to the real-time case distribution characteristics in the patient condition data, output a high-priority course set, calculate the matching degree in combination with user learning behavior features, generate a personalized recommendation list, and output the recommended courses to the user terminal.

[0056] This embodiment also provides a computer device applicable to the emergency online course recommendation method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the emergency online course recommendation method proposed in the above embodiment.

[0057] This computer device can be a terminal, and this computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of this computer device is used to provide computing and control capabilities. The memory of this computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The communication interface of this computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a carrier network, NFC (Near Field Communication), or other technologies. The display screen of this computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of this computer device can be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the computer device housing, or an external keyboard, touchpad, or mouse, etc.

[0058] This embodiment also provides a storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the emergency online course recommendation method proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, abbreviated as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, abbreviated as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, abbreviated as EPROM), programmable read-only memory (Programmable Red-Only Memory, abbreviated as PROM), read-only memory (Read-Only Memory, abbreviated as ROM), magnetic memory, flash memory, magnetic disk or optical disc.

[0059] In summary, the present invention integrates the emergency medicine ontology through a cross-modal adversarial transfer learning framework to generate high-density feedback data that conforms to clinical logical consistency, effectively compensates for the sparsity of user behavior data, and enhances the integrity and medical rationality of the emergency care knowledge graph; combines a temporal convolutional network and a graph attention network to dynamically model the temporal features of vital signs and multi-order causal associations, accurately locates the user's knowledge blind spots and captures the causal chain in complex clinical situations; through a neuro-symbolic causal intervention engine and a dynamic update mechanism, maps the knowledge blind spot propagation path to the course recommendation network, adaptively adjusts the recommendation strategy, realizes the dynamic adaptation of course content to real-time case features, and continuously optimizes the recommendation strategy through user feedback and clinical data-driven, constructs an adaptive personalized learning support mechanism.

[0060] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. An emergency online course recommendation method, characterized in that: including, collecting user learning behavior data, patient condition data, and emergency nursing course content data, dynamically modeling emergency nursing knowledge points and multi-level association relationships based on cross-modal adversarial transfer learning, and constructing an emergency nursing knowledge graph; performing multi-modal data fusion on user learning behavior data, patient condition data, and the emergency nursing knowledge graph, extracting user learning behavior features, patient condition features, and emergency nursing knowledge graph association features, and generating a three-dimensional feature vector; through a neuro-symbolic causal intervention engine, combining with the emergency nursing knowledge graph for dynamic causal reasoning, locating users' knowledge blind spots, and mapping the spatio-temporal propagation chain of knowledge blind spots to the probability flow network of course nodes to generate a candidate course decision cluster; dynamically adjusting the priority of the candidate course decision cluster according to the real-time case distribution characteristics in the patient condition data, outputting a high-priority course set, calculating the matching degree in combination with user learning behavior features, generating a personalized recommendation list, and outputting the recommended courses to the user terminal.

2. The application to the emergency online course recommendation method according to claim 1, characterized in that: The cross-modal adversarial transfer learning is used to perform spatio-temporal alignment and sparse compensation on multi-source data, and a dynamic temporal causal network is used to model emergency nursing knowledge points and multi-level association relationships.

3. The application to the emergency online course recommendation method according to claim 1, characterized in that: The neuro-symbolic causal intervention engine consists of a neural graph network, a symbolic causal rule base, and a dynamic weight intervention operator, and performs causal decoupling and inference path reconstruction by real-time correcting the edge weights of the emergency nursing knowledge graph; The steps for generating the candidate course decision cluster are as follows: injecting dynamic cognitive weights into the emergency nursing knowledge graph through a neuro-symbolic causal intervention engine to construct a multi-modal causal inference path; using adversarial topological reinforcement learning to map the spatio-temporal propagation chain of knowledge blind spots to the probability flow network of course nodes to generate a candidate course decision cluster.

4. The application to an emergency online course recommendation method according to claim 3, wherein: The dynamic adjustment of the priority of the candidate course decision cluster means that based on the patient condition data, extracting temporal and semantic features through TCN / BERT, performing cosine similarity matching in combination with K-means clustering, filtering low-priority candidate courses after dynamic weighted scoring, and outputting a high-priority course set.

5. The application to the emergency online course recommendation method according to claim 4, characterized in that: The patient condition data includes a vital sign time series, a chief complaint text record, and a preliminary diagnosis label.

6. The application of the emergency online course recommendation method according to claim 1, characterized in that: The calculation of the matching degree includes calculating the user-course feature matching degree based on the three-dimensional feature vector and the high-priority course set through cosine similarity.

7. The application to the emergency online course recommendation method according to claim 1, wherein: The multi-modal data fusion refers to standardizing and semantically extracting user behavior data, extracting temporal and medical features of patient data, combining with the graph embedding of the emergency nursing knowledge graph, and generating a three-dimensional feature vector through a fully connected layer for unified dimension, attention fusion, and PCA dimensionality reduction.

8. An emergency online course recommendation device, based on the emergency online course recommendation method described in any one of claims 1 to 7, characterized in that: including data collection, feature fusion, causal reasoning, and dynamic recommendation; The data collection is used to collect user learning behavior data, patient condition data, and emergency nursing course content data, dynamically model emergency nursing knowledge points and multi-level association relationships based on cross-modal adversarial transfer learning, and construct an emergency nursing knowledge graph; The feature fusion is used to perform multi-modal data fusion on user learning behavior data, patient condition data, and the emergency nursing knowledge graph, extract user learning behavior features, patient condition features, and emergency nursing knowledge graph association features, and generate a three-dimensional feature vector; The causal reasoning is used to perform dynamic causal reasoning through a neural symbolic causal intervention engine in combination with an emergency nursing knowledge graph, locate user knowledge blind spots, and map the spatiotemporal propagation chain of knowledge blind spots to a probability flow network of course nodes to generate candidate course decision clusters; The dynamic recommendation is used to dynamically adjust the priority of the candidate course decision cluster according to the real-time case distribution characteristics in the patient's condition data, output a high-priority course set, calculate the matching degree in combination with the user's learning behavior characteristics, generate a personalized recommendation list and output the recommended courses to the user terminal.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that: When the processor executes the computer program, the steps of the method for recommending emergency online courses according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for recommending emergency online courses described in any one of claims 1 to 7 are implemented.

Images