A dynamic evolution and quantitative assessment system for pharmacist competence based on multi-agent agents

Abstract

This invention relates to the field of artificial intelligence technology, specifically to a multi-agent-based dynamic evolution and quantitative assessment system for pharmacist competence. The system includes: a multimodal resource analysis module, a competence association module, a dynamic competence assessment module, a path planning module, and an interactive practice module. In this invention, by uniformly semantically encoding and structurally modeling multi-source teaching materials, a computable network of connections is formed between knowledge content, enabling precise mapping between competence requirements and teaching knowledge. Continuous modeling is performed using multi-dimensional behavioral characteristics during the learning process, allowing competence weights to adaptively adjust with performance changes. The arrangement of learning content dynamically evolves based on competence profiles and knowledge dependencies. Semantic understanding and contextual deduction mechanisms are introduced into interactive practice, enabling quantitative assessment of communication behavior and clinical thinking. This optimizes the adaptability of teaching paths and the effectiveness of personalized feedback, improving the overall adaptability and accuracy of the teaching process.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a dynamic evolution and quantitative assessment system for pharmacist competence based on multi-agent intelligence. Background Technology

[0002] Artificial intelligence technology is a comprehensive technical field that studies how to simulate human cognition, learning, and decision-making processes through computational models and procedural methods. Its core aspects include data acquisition and annotation, knowledge representation methods, reasoning mechanisms, learning strategies, and multi-agent collaborative mechanisms. It has also developed intelligent system construction methods for specific tasks in scenarios such as education, healthcare, industry, and management. By structuring and logically deducing multi-source information, it has achieved the digital expression and execution control of complex business processes.

[0003] One traditional multi-agent-based dynamic evolution and quantitative assessment system for pharmacist competence involves introducing the collaborative effects of multiple agents to construct a system framework adapted to the needs of pharmaceutical education, aiming to enhance pharmacists' professional competence in the pharmaceutical field. By constructing a multi-agent model, it simulates pharmacists' reactions and decisions in different teaching scenarios, helping the teaching system better adapt to diverse teaching needs. Traditional pharmaceutical teaching systems focus on the application of single agents, making it difficult to comprehensively consider the diverse needs and complex situations involved in pharmaceutical education. The multi-agent-based framework integrates multiple agents, thereby achieving a comprehensive assessment and improvement of pharmaceutical knowledge, skills, and overall abilities.

[0004] Existing technologies support pharmaceutical teaching with preset rules and static knowledge organization. Teaching resources are mostly presented as single texts or fixed structures, making it difficult to uniformly analyze image data and unstructured information. There is a lack of fine-grained correlation mechanisms between knowledge points and competency requirements. Student behavior data mostly remains at the level of result recording. Competency assessment relies on fixed weights and experience judgment. The order of learning content is limited by manual settings. Contextual interactions are mostly scripted processes, resulting in coarse competency descriptions, lagging path adaptation, and insufficient targeted practical feedback. Summary of the Invention

[0005] To address the technical problems of existing technologies supporting pharmaceutical teaching with preset rules and static knowledge organization, where teaching resources are mostly presented as single texts or fixed structures, making it difficult to uniformly analyze image data and unstructured information, lacking fine-grained correlation mechanisms between knowledge points and competency requirements, student behavior data mostly remaining at the result recording level, competency assessment relying on fixed weights and experience judgment, learning content sequence adjustment limited by manual settings, and scenario interaction mostly being scripted processes, resulting in coarse competency characterization, lagging path adaptation, and insufficient targeted practical feedback, this invention provides a pharmacist competency dynamic evolution and quantitative assessment system based on multi-agent intelligence.

[0006] On the one hand, a dynamic evolution and quantitative assessment system for pharmacist competence based on multi-agent agents is provided, the system comprising: The multimodal resource parsing module receives various forms of pharmaceutical teaching materials, converts images into text using optical character recognition, performs semantic encoding and feature extraction on the text based on BERT, and constructs a structured knowledge graph. The competency association module, based on the structured knowledge graph, uses cosine similarity analysis to analyze the pharmacist competency framework, ability descriptions, and knowledge point texts to generate competency-annotated teaching knowledge nodes. The ability dynamic assessment module calls the teaching knowledge nodes marked with the competency, collects multi-dimensional learning behavior data of students, uses a deep Q network to calculate the weight of the ability dimension in combination with the current ability performance of students, optimizes the weight allocation through weighted fusion, and generates a multi-dimensional competency profile. The path planning module, based on the multi-dimensional competency profile, uses a genetic algorithm to optimize the order of learning content, dynamically adjusts the path in combination with knowledge dependencies, and generates an optimized path planning scheme. The interactive practice module constructs a multi-agent interactive environment based on the optimized path planning scheme. Through the multi-agent interactive environment, it simulates the dynamic collaboration and dialogue between virtual pharmacists and patients, analyzes the trainees' operations and language intentions, and, combined with knowledge graphs and teaching scenarios, assesses the trainees' communication skills and clinical reasoning abilities, and outputs the interactive performance results.

[0007] As a further aspect of the present invention, the structured knowledge graph includes pharmaceutical concept nodes, concept relationships, and a set of knowledge attributes; the competency-annotated teaching knowledge nodes include competency element tags, competency level identifiers, and job suitability markers; the multi-dimensional competency profile includes competency strength distribution, competency development level, and competency structure characteristics; the optimized path planning scheme includes learning content ranking results, path dependency constraint sets, and personalized adjustment rules; and the interaction performance results include communication quality evaluation, clinical decision-making tendency, and comprehensive practical ability indicators.

[0008] As a further aspect of the present invention, the multimodal resource parsing module includes: The image and text receiving submodule acquires various forms of pharmaceutical teaching materials, collects image frames and text files, performs character detection and mapping on image frames through optical character recognition, extracts text strings and aligns them with the original text file in terms of format, filters character sequences based on noise ratio thresholds, and generates character sequence integrity values. The semantic encoding submodule, based on the character sequence integrity rate value, calls the filtered character sequence and inputs it into the BERT semantic encoding structure to perform word segmentation mapping and context association calculation, extracts multi-layer semantic embedding vectors, performs vector dimension consistency judgment, if the dimensions are inconsistent, unifies the vector dimensions through linear transformation, and then performs normalization processing to generate a semantic feature vector matrix. The graph construction submodule performs concept co-occurrence statistics based on the semantic feature vector matrix and the original text corpus, collects concept nodes, calculates the co-occurrence frequency of nodes in the text, calculates node relationship weights based on the frequency distribution and performs threshold judgment, filters node relationships with weights greater than the preset weight threshold and integrates node and relationship attribute sets to generate a structured knowledge graph.

[0009] As a further aspect of the present invention, the noise ratio threshold is based on the statistical analysis of the proportion of non-character pixels in the character pixel sequence, and the noise ratio threshold is set to a closed interval of 10% to 20%. The preset weight threshold is derived from the statistical characteristics of the relationships between knowledge graph nodes and is a fixed constant with a value range of 0.3 to 0.5.

[0010] As a further aspect of the present invention, the competency association module includes: The competency framework parsing submodule retrieves the set of concept node attributes based on the structured knowledge graph, obtains the pharmacist competency framework text data and performs word segmentation, maps the word segmentation results to the term index table, extracts semantic embedding vectors and performs dimension alignment and normalization operations to generate a set of competency description semantic vectors. The similarity calculation submodule calls the semantic vector set of capability description and the text vector of knowledge nodes in the structured knowledge graph, calculates the cosine similarity value, filters out node pairs with a similarity greater than a preset threshold, records the node identifier and capability description identifier, and generates a node capability similarity mapping table. The node annotation submodule extracts node identifiers and capability description identifiers based on the node capability similarity mapping table, calls the node attributes in the structured knowledge graph, writes the capability description identifiers into the attribute fields, and generates competency-annotated teaching knowledge nodes.

[0011] As a further aspect of the present invention, the preset similarity threshold is based on a set of node samples labeled with capability description identifiers in a structured knowledge graph. The cosine similarity value distribution interval is calculated between the capability description semantic vector set and the corresponding knowledge node text vector. Within the cosine similarity value distribution interval, the mean and standard deviation of the similarity of positively correlated node pairs are statistically analyzed, and the result obtained by subtracting the standard deviation from the mean similarity is used as the preset similarity threshold.

[0012] As a further aspect of the present invention, the capability dynamic assessment module includes: The behavioral data collection submodule calls the teaching knowledge nodes labeled with the competency, collects student access duration data, interaction frequency data, test score data and operation path data, performs alignment operation according to timestamp, calculates the mean of access duration and the variance of interaction frequency, and generates a multi-dimensional behavioral feature matrix of students. The weight dynamic calculation submodule extracts behavioral feature values ​​based on the student's multi-dimensional behavioral feature matrix, constructs a state space by calling the ability labeling attributes in the teaching knowledge nodes, maps the feature values ​​to the state vector, inputs them into the deep Q network for forward propagation to calculate the Q value of the ability dimension and performs normalization operation to generate a dynamic weight vector of the ability dimension. The profile fusion generation submodule calls the dynamic weight vector of the ability dimension and the multi-dimensional behavioral feature matrix of the trainee, performs element-wise multiplication operation on the column vector of the behavioral feature matrix and the weight vector components, and performs column vector summation operation on the weighted result to generate a multi-dimensional competency profile.

[0013] As a further aspect of the present invention, the path planning module includes: The learning content optimization submodule collects student learning preference data, knowledge point mastery, and learning time distribution based on the multi-dimensional competency profile. Based on learning needs analysis, it identifies knowledge points with lagging and fast learning progress, encodes the learning content using a genetic algorithm, and generates an optimized learning content order. The path adjustment calculation submodule, based on the optimized learning content order, calls the knowledge point dependencies of the learners, monitors the interdependencies between knowledge points, compares the costs of different learning paths, calculates and adjusts the paths through the analysis of path conversion costs, and obtains dynamically adjusted learning paths. The optimized path generation submodule dynamically adjusts the learning path, maps the adjusted path to the student's learning time and task completion data, calls the priority of learning content and path adjustment strategy, evaluates the path task completion rate, and obtains an optimized path planning scheme.

[0014] As a further aspect of the present invention, the interactive practice module includes: The interactive data acquisition submodule, based on the optimized path planning scheme, extracts the teaching scenario constraints and initializes the parameters of the virtual patient agent, virtual pharmacist agent, and assessment and monitoring agent, constructs a multi-agent interactive environment, and obtains the simulated virtual pharmacist and patient dialogue text, operation instruction sequence, and interaction time stamps in the multi-agent interactive environment through the assessment and monitoring agent. It performs text statement segmentation and sequence indexing, instruction event alignment and frequency statistics, verifies consistency based on time stamps, and generates an interactive behavior sequence matrix. The intent mapping determination submodule, based on the interaction behavior sequence matrix, calls the knowledge graph concept nodes and relation constraints, performs semantic unit matching judgment, divides the interval according to the density of the operation event and the corresponding semantic unit, and maps it in combination with the teaching scenario state identifier to obtain the interaction intent matching coefficient. The performance quantification generation submodule collects the standard path of the teaching scenario based on the interaction intent matching coefficient, performs position offset calculation between the student interaction path and the standard path, performs interval statistics and weight merging based on the offset distribution, and generates the interaction performance result by combining the coverage of thinking nodes in the knowledge graph.

[0015] As a further aspect of the present invention, the execution of text statement segmentation and sequential indexing refers to segmenting continuous text into minimum semantic units according to the time stamp order in the dialogue text between the simulated virtual pharmacist and the patient, and assigning an increasing sequential index value to each minimum semantic unit. The consistency verification based on time stamps refers to determining that the maximum deviation between the time stamps of the dialogue text between the simulated virtual pharmacist and the patient and the time stamps of the operation instruction sequence does not exceed a preset time threshold.

[0016] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: By uniformly semantically encoding and structurally modeling multi-source teaching materials, knowledge content forms a computable network of connections, enabling precise mapping between competency requirements and teaching knowledge. Continuous modeling is performed in conjunction with multi-dimensional behavioral characteristics during the learning process, allowing competency weights to adaptively adjust with performance changes. The arrangement of learning content evolves dynamically based on competency profiles and knowledge dependencies. Semantic understanding and contextual deduction mechanisms are introduced into interactive practice, enabling quantitative assessment of communication behavior and clinical thinking, thereby improving the matching degree of teaching paths and the accuracy of competency feedback. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.

[0018] Figure 1 This is a system schematic diagram of the present invention; Figure 2 This is a schematic diagram of the system framework of the present invention; Figure 3 This is a flowchart of the multimodal resource parsing module in this invention; Figure 4 This is a flowchart of the competency association module in this invention; Figure 5This is a flowchart of the capability dynamic assessment module in this invention; Figure 6 This is a flowchart of the path planning module in this invention; Figure 7 This is a flowchart of the interactive practice module in this invention. Detailed Implementation

[0019] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0020] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0021] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0022] In this embodiment of the invention, sometimes the subscript such as W1 is written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0023] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0024] This invention provides a system for the dynamic evolution and quantitative assessment of pharmacist competence based on multi-agent mechanisms, such as... Figure 1-2 The diagram shows a dynamic evolution and quantitative assessment system for pharmacist competence based on multi-agent agents. The system includes: The multimodal resource parsing module receives various forms of pharmaceutical teaching materials, converts images into text using optical character recognition, performs semantic encoding and feature extraction on the text based on BERT, and constructs a structured knowledge graph. The competency association module, based on a structured knowledge graph, uses cosine similarity analysis to analyze the competency framework descriptions and knowledge point texts of pharmacists, generating competency-annotated teaching knowledge nodes. The ability dynamic assessment module calls the teaching knowledge nodes marked with competency, collects multi-dimensional learning behavior data of students, uses a deep Q network to calculate the weight of the ability dimension in combination with the current ability performance of students, optimizes the weight allocation through weighted fusion, and generates a multi-dimensional competency profile. The path planning module, based on a multi-dimensional competency profile, uses a genetic algorithm to optimize the order of learning content, dynamically adjusts the path by combining knowledge dependencies, and generates an optimized path planning scheme. The interactive practice module constructs a multi-agent interactive environment based on the optimized path planning scheme. Through the multi-agent interactive environment, it simulates the dynamic collaboration and dialogue between virtual pharmacists and patients, analyzes the students' operations and language intentions, and evaluates the students' communication and clinical thinking abilities by combining knowledge graphs and teaching scenarios, and outputs the interactive performance results. The structured knowledge graph includes pharmaceutical concept nodes, concept relationships, and sets of knowledge attributes. The competency-annotated teaching knowledge nodes include competency element tags, competency level identifiers, and job suitability markers. The multi-dimensional competency profile includes the distribution of competency strengths and weaknesses, competency development levels, and competency structure characteristics. The optimized path planning scheme includes the learning content ranking results, path dependency constraint sets, and personalized adjustment rules. The interactive performance results include communication quality evaluation, clinical decision-making tendencies, and comprehensive practical ability indicators.

[0025] Specifically, such as Figure 2 , 3 As shown, the multimodal resource parsing module includes: The image and text receiving submodule acquires various forms of pharmaceutical teaching materials, collects image frames and text files, performs character detection and mapping on image frames through optical character recognition, extracts text strings and aligns them with the original text file in terms of format, filters character sequences based on noise ratio thresholds, and generates character sequence integrity values. When processing electronic teaching materials for core pharmacy courses such as Pharmacology and Pharmaceutical Chemistry, the system receives a collection of image frames in multiple formats, including scanned copies of textbooks and screenshots of PPT presentations. and the corresponding teaching syllabus text file Traverse image frames Using an optical character recognition engine to locate the coordinates of text regions in an image Extract the character pixel sequence within the region and the identified character encoding sequence With text files The standard terminology library is used for string matching and alignment, for each extracted character pixel sequence. Perform pixel-level noise analysis to count the number of discrete noise pixels in the sequence that do not belong to character strokes. The total number of pixels in the character region , calculate the noise ratio , set the noise ratio threshold as (i.e., 15%), and the value of this threshold is derived from the background noise statistics of 1,000 scanned copies of standard pharmacy textbooks. Among them the background noise rate of clear characters is lower than . If the total number of pixels of a certain character area (such as the character area of "A" in "Amoxicillin") calculated is pixels, and the number of free noise pixels in the detected unconnected domain is pixels, then calculate . Compare with numerically, and determine . This character sequence is determined to be a valid sequence and is retained. If another character area (such as the annotation beside a damaged chemical structural formula) calculates , then determine . This sequence is marked as a high-noise sequence and is excluded. Repeat the above calculation and screening process for the detected character sequences, and count the number of finally retained valid character sequences and the total number of initially detected character sequences . For example, when processing a page of a courseware on "Antibiotic Classification", initially detect character sequences. After screening by the noise threshold, valid sequences are retained, and calculate the character sequence integrity rate value . This value reflects the quality level of converting image data into processable text data, and generates the character sequence integrity rate value.

[0026] Semantic encoding sub-module. According to the character sequence integrity rate value, call the screened character sequences, input them into the BERT semantic encoding structure to perform token mapping and context association calculation, extract multi-layer semantic embedding vectors, perform vector dimension consistency judgment. If the dimensions are inconsistent, unify the vector dimensions through linear transformation, and then perform normalization processing to generate a semantic feature vector matrix; Call the character sequence integrity rate value greater than the preset benchmark of the screened character sequences. For example, the aforementioned text content on "Antibiotic Classification" is input into the pre-trained BERT semantic encoding structure. First, perform token mapping on the input text, decompose "Penicillin interferes with bacterial cell wall synthesis" into a Token sequence [CLS, Penicillin, interferes, with, bacterial, cell, wall, synthesis, SEP], and input it into the model in combination with position encoding. In the layer Transformer encoder of the BERT structure, perform multi-head self-attention calculation, and extract the Layer and First The layer's output serves as a context-dependent hidden state vector, generating the original embedding vector for each token. Detect its vector dimension ,verify Is it strictly equal to ,like If the dimensionality is deemed consistent, then L2 norm normalization is performed on the vector. Calculate its L2 norm and apply the formula. Obtain the normalized vector, for example, take the simplified vector. Calculate the norm The normalization result is This process eliminates the impact of vector magnitude differences on subsequent calculations, aggregates the normalized token vectors, and employs an average pooling strategy for long text sequences, i.e.: ; in Given the sequence length, generate a fixed-dimensional semantic feature vector matrix representing the pharmaceutical teaching content of that segment. Each row in the matrix corresponds to a feature vector of an independent semantic unit (such as a single drug knowledge point), generating a semantic feature vector matrix.

[0027] The graph construction submodule performs concept co-occurrence statistics based on the semantic feature vector matrix and the original text corpus, collects concept nodes, calculates the co-occurrence frequency of nodes in the text, calculates node relationship weights based on the frequency distribution and performs threshold judgment, filters node relationships that are greater than the preset weight threshold and integrates the node and relationship attribute set to generate a structured knowledge graph. Based on semantic feature vector matrix The entity information is parsed to collect concept nodes, such as identifying the "drug node: cefaclor" and the "indication node: otitis media". These two nodes are then traversed within the entire teaching corpus within the same context window (WindowSize). The frequency of common occurrences within each node, and the statistical nodes. With nodes Co-occurrence frequency and obtain the node Total document frequency With nodes Total document frequency Calculate the node relationship weights based on frequency distribution. The weight calculation formula is set as follows: The formula references a statistical variant of cosine similarity and sets a preset weight threshold. This threshold is derived from the statistical characteristic analysis of strongly associated entity pairs in existing mature pharmaceutical knowledge graphs (such as DrugBank), selecting the lower quartile of the association strength distribution and setting it as a fixed constant. (range of values) to The calculated weights and A point-by-point comparison was conducted; for example, the co-occurrence data of "cefaclor" and "otitis media" was used to monitor the number of times they co-occurred. The total frequencies of independent occurrences are respectively , Substitute into the formula to calculate the weight. ,determination The correlation weight did not reach the threshold, so it was judged as a weak correlation or accidental co-occurrence and was discarded; however, for the combination of "cefaclor" and "sensitive bacteria", the measured... , , Calculate weights ,determination If the selection is successful, the node relationship is retained and the node attributes (such as half-life and dosage form) and relationship attributes (such as treatment and contraindications) are integrated. The specific selection data is shown in Table 1. Based on the selected set of nodes and edges, the Resource Description Framework (RDF) standard is used for storage to generate a structured knowledge graph.

[0028] Table 1: Example Table of Calculation for Filtering Relationship Weights of Pharmaceutical Concept Nodes

[0029] As shown in Table 1, the table lists actual examples of calculating association weights for specific pharmaceutical concept node pairs. By comparing the calculated weights... With fixed threshold The logic for retaining and removing node relationships was clarified. The results show that only when the co-occurrence statistical characteristics of concepts in teaching materials show a significant strong correlation (i.e. the calculated weight is higher than the preset statistical constant) will they be used as formal edges of the knowledge graph, thereby filtering out the noise associations that occasionally appear in the text.

[0030] Specifically, such as Figure 2 , 4 As shown, the competency association module includes: The competency framework parsing submodule retrieves the set of concept node attributes from the structured knowledge graph, obtains the pharmacist competency framework text data and performs word segmentation, maps the word segmentation results to the term index table, extracts semantic embedding vectors and performs dimension alignment and normalization operations to generate a set of competency description semantic vectors. Iterate through the graph stored in the graph using the graph database query interface. The system retrieves text data for each pharmaceutical concept node, such as "prescription review" and "drug interactions," while simultaneously loading the "Competency Evaluation Standards for Licensed Pharmacists' Pharmaceutical Services" from the local database as text data for the pharmacist competency framework. , by item Perform text segmentation to obtain text containing phrases such as "able to identify potential adverse drug reactions". Each capability description statement is processed using a Chinese word segmentation tool, breaking down "able to identify potential adverse drug reactions" into a sequence of terms. The system is designed to identify potential adverse drug reactions and maps the sequence to a pre-defined pharmaceutical term index table to obtain the corresponding word ID sequence. This sequence is then input into a pre-trained semantic encoding model to extract the semantic embedding vector for each term in the sequence. Dimension alignment is performed to detect the dimension of each output vector. If detected Then, force alignment to be achieved by using zero-padding or truncation. For the aligned vector sequence, perform mean pooling to obtain the original semantic vectors at the sentence level. And immediately perform L2 norm normalization operation, let Calculate the norm If the original vector norm of a certain ability description item is calculated Then divide each component of the vector by Obtain the normalized vector This process ensures that the magnitude of all capability description vectors is [value missing]. For all Repeat the above process for each capability description, and store the generated normalized vectors into a high-dimensional vector database to generate a set of capability description semantic vectors.

[0031] The similarity calculation submodule calls the semantic vector set of capability description and the text vector of knowledge nodes in the structured knowledge graph, calculates the cosine similarity value, filters out node pairs with a similarity greater than the preset threshold, records the node identifier and capability description identifier, and generates a node capability similarity mapping table. The vectors in the semantic vector set that can be invoked describe the capabilities Text vectors of knowledge nodes in structured knowledge graphs The cosine similarity formula is used: ; To calculate the degree of correlation between the two, since the previous steps have already completed the normalization process, the denominator term... Equal to The calculation is simplified to a vector dot product operation, filtering out those with similarity values ​​greater than a preset threshold. Node pairs, with a preset similarity threshold The setup process is as follows: Select a structured knowledge graph that has been manually annotated with capability description tags by experts. For the set of positive sample nodes (i.e., confirmed "knowledge point-competency" pairs with strong correlation), calculate the distribution range of the cosine similarity values ​​for each pair of samples in the set, and calculate the mean similarity of the positively correlated node pairs. with standard deviation The advantage of the formula lies in the introduction of the standard deviation. For the mean The corrections were made, and a dynamic statistical boundary was constructed. This logic aims to retain high-confidence associations distributed near or above the mean, while using the standard deviation to remove weak similarity data that, although labeled as positive samples, are on the distribution edge or contain textual ambiguity, ensuring the rigor of automatic mapping. Specific examples are shown in Table 2, where similarity data for a batch of positive samples is collected, and the mean is calculated. Standard deviation Then set a threshold. For the node to be detected, "Aspirin Enteric-coated Tablets" and the competency item "Guiding Patients on Safe Medication Use" are compared, and their vector dot product is calculated to obtain the similarity. Compare it with the threshold value to determine. If the filtering criteria are met, record the node identifier. With capability description identifier For the node pair "Aspirin Enteric-coated Tablets" and "Identification of Traditional Chinese Medicine Materials", the calculated similarity is: ,determination These are removed, and a node capability similarity mapping table is generated.

[0032] The node annotation submodule extracts node identifiers and ability description identifiers based on the node ability similarity mapping table, calls the node attributes in the structured knowledge graph, writes the ability description identifiers into the attribute fields, and generates competency-annotated teaching knowledge nodes. Extract node identifiers based on the node capability similarity mapping table. With capability description identifier For example, extracting the mapping relationship: { (Metformin) (Chronic management)}; Locate the identifier using a graph database query language (such as Cypher). For the graph node entity, the node attribute list in the structured knowledge graph is called to check if an attribute field named "Competency_Ref" exists. If it does not exist, the field is created, and the extracted capability description is identified. If other capability identifiers already exist in the attribute field, an append operation is performed to create an identifier list. The above attribute update operation is performed on all records in the mapping table one by one. After the writing is completed, the node attributes are scanned again to verify that the "Competency_Ref" field is not empty and the format meets the requirements of string array. The data persistence is confirmed to be successful. This process physically binds the abstract competency standard with the specific drug knowledge entity and generates the teaching knowledge node with competency annotation.

[0033] Table 2: Examples of Similarity Screening for Pharmaceutical Nodes and Competency Descriptions

[0034] As shown in Table 2, Table 2 lists the implementation data for similarity screening of knowledge nodes and competency descriptions using statistical parameters. The calculated vector similarity is then used to... The dynamic threshold calculated based on the positive sample distribution characteristics By comparing the results, the retention and rejection decisions under different association strengths were clarified. The results show that only when the proximity between the node text and the ability description text in the semantic space is higher than the statistical lower limit of the positive sample group (i.e., the mean minus one standard deviation) can the knowledge point be considered to have teaching value to support the corresponding pharmacist competence, thereby achieving automated and accurate annotation.

[0035] Specifically, such as Figure 2 , 5 As shown, the dynamic capability assessment module includes: The behavioral data collection submodule calls the teaching knowledge nodes labeled with competency, collects student access duration data, interaction frequency data, test score data and operation path data, performs alignment operation according to timestamp, calculates the mean of access duration and the variance of interaction frequency, and generates a multi-dimensional behavioral feature matrix of students. The system invokes competency-annotated teaching knowledge nodes, such as a teaching unit labeled "Drug Interactions - Warfarin and Aspirin in Combination," and starts a real-time monitoring thread for each logged-in student to collect data on the duration of their access to that knowledge node. (Unit: seconds), Interaction frequency data (Including the number of mouse clicks, page scrolling, note annotations, etc.) Test score data (This node corresponds to the score of the quiz, with a maximum score) (points) and operation path data (Record the predecessor and successor node IDs of the access), read the Unix timestamp field in the data packet, perform alignment operations on the above heterogeneous data in chronological order to form a time-seriesd behavior record set, and then perform statistical calculations on the key indicators in the record set, first calculating the average access duration. The focus then shifted to calculating the variance of interaction frequency. To assess the fluctuations in students' learning focus, the variance calculation formula is applied. ,in This represents the number of time windows or sessions counted in the statistics. For the first Interaction frequency in this session The formula, which is the arithmetic mean of the interaction frequencies, is useful in that it calculates the variance. This method can quantify the dispersion of learners' interactive behaviors during the learning process, thereby distinguishing between learners with stable learning styles and those who engage in intensive, reactive learning, providing data support for subsequent weight allocation. A specific calculation example is shown in Table 3, which collects data on a learner's... The frequency sequence of interactions with the "Prescription Review" module within a consecutive learning period is as follows: First, calculate the mean. Then, substitute each item into the variance formula to calculate the difference of squares: Summing yields Finally, the variance was calculated. The results indicate that the students' interactive behavior fluctuated little (variance value was lower than the preset baseline value). When the learning state is stable, the calculated mean will be used. ,variance Standardized scores Data is encapsulated to generate a multidimensional behavioral feature matrix of trainees.

[0036] Table 3: Data Collection Table of Interactive Behaviors of Trainees in Pharmaceutical Knowledge Nodes

[0037] As shown in Table 3, Table 3 details the specific behavioral data collected from trainees regarding different pharmaceutical knowledge nodes. By calculating the variance of the interaction frequency sequence, a quantitative indicator reflecting the learning stability of trainees was obtained. This result indicates that statistically significant behavioral features can be extracted from the original operation logs, recording not only the learning effect of a single session but also capturing behavioral patterns during continuous learning, providing basic data input for constructing a multidimensional competency profile.

[0038] The dynamic weight calculation submodule extracts behavioral feature values ​​based on the student's multi-dimensional behavioral feature matrix, calls the ability labeling attributes in the teaching knowledge nodes to construct the state space, maps the feature values ​​to the state vector, inputs them into the deep Q network for forward propagation to calculate the Q value of the ability dimension and performs normalization operation to generate a dynamic weight vector of the ability dimension. Behavioral feature values ​​are extracted based on the learner's multidimensional behavioral feature matrix, including access duration stability, interaction depth coefficient, and test accuracy. Pre-defined ability labeling attributes (such as "professional knowledge," "practical skills," and "professional qualities") in the teaching knowledge nodes are then invoked to construct a reinforcement learning state space. The extracted feature values ​​are discretized and mapped to state vectors to construct a deep Q-network. This network consists of an input layer, two hidden layers (containing 128 and 64 neurons respectively, with ReLU activation function), and an output layer (outputting Q-values ​​corresponding to the three ability dimensions). This output layer is used to fit the contribution of each ability dimension to the final competency assessment (i.e., the action value function) under a specific state. The network parameters are updated using backpropagation based on the learner's historical learning performance, and iterative optimization is performed using a variant of the Bellman equation. The loss function is... ,in Indicates the current state of the network. Next dimension The value estimate, For network parameters, Let the learning rate be (set to) ), The reward value (set based on the pass / fail result of the most recent simulated professional skills assessment, with pass being set as...) Not approved as set ), Discount factor (set to) ), To maximize the predictive value of the target network for the next state, the advantage of this mechanism lies in its ability to dynamically adjust the importance of evaluation dimensions based on the long-term ability growth feedback of trainees by introducing a temporal difference learning mechanism and the nonlinear fitting capability of neural networks. It also handles high-dimensional continuous state spaces, avoiding the evaluation rigidity caused by fixed weights. A specific calculation example is as follows: assuming the current state vector... (These represent the normalized values ​​of the three behavioral features, respectively), which are input into a deep Q-network for forward propagation. The network outputs the original Q-values ​​for the three dimensions, respectively. Trainees who pass the assessment receive a reward. The target network predicts the next state. The greatest value Then calculate the time-difference objective For the "practical skills" dimension, the current predicted value is... The loss is After updating the network parameters using gradient descent, the updated Q-value is obtained by forward propagation again. (This is a simplified example; actual updates involve multiple rounds of training.) Meanwhile, the Q-values ​​for other dimensions are also adjusted accordingly. After obtaining the Q-values ​​of the three dimensions of the current network output, a normalization operation is performed to determine the final weights, and the softmax function is applied. Alternatively, a simplified linear normalization can be used. Taking linear normalization as an example, calculate the denominator. The weight of the "professional knowledge" dimension "Practical skills" weight "Professional qualities" weight The results indicate that the real-time weight allocation of the capability dimension has been dynamically generated based on the learners' learning feedback through a deep Q-network, resulting in a dynamic weight vector for the capability dimension.

[0039] The profile fusion and generation submodule calls the dynamic weight vector of the ability dimension and the multi-dimensional behavioral feature matrix of the trainee, performs element-wise multiplication operation on the column vector of the behavioral feature matrix and the weight vector components, performs column vector summation operation on the weighted result, and generates a multi-dimensional competency profile. Invoke the dynamic weight vector of the capability dimension Multidimensional behavioral feature matrix of trainees , where the matrix The column vectors correspond to the capability dimension scores associated with knowledge nodes (scores are derived from a combination of access duration, interaction quality, and test results). Let the column vectors of the matrix be... The element-wise multiplication operation is performed on the column vectors of the behavioral feature matrix and the components of the weight vector, that is, the weighted components are calculated. Then, a column vector summation operation is performed on the weighted results to obtain the comprehensive competency index. The calculation formula is: ; in Representatives of the students at the Normalized performance scores for each capability dimension (values) to The advantage of this formula lies in its ability to integrate disparate dimensional performance into a single measurable indicator through a linear weighted fusion strategy. This indicator, adjusted by dynamic weights, more accurately reflects the learner's core competency level at the current learning stage. A specific calculation example is as follows: continuing with the previous data, we set the normalized performance score vector for learners in the three dimensions of "professional knowledge," "practical skills," and "professional qualities" as follows: Call the calculated weight vector Substitute into the formula and perform the calculation: First term The second item The third item Perform summation Calculation results Compared with the preset competency benchmark value Compare and determine The results show that trainees have excellent overall competence under the current dynamic weighting system and have reached the core competency standards required for the pharmacist position. Based on this result, a radar chart containing dimensional scores and comprehensive indexes was drawn to generate a multi-dimensional competence profile.

[0040] Specifically, such as Figure 2 , 6 As shown, the path planning module includes: The learning content optimization submodule collects student learning preference data, knowledge point mastery, and learning time distribution based on multi-dimensional competency profiles. Based on learning needs analysis, it identifies knowledge points with lagging and fast learning progress, and encodes the learning content through a genetic algorithm to generate an optimized learning content order. Retrieve personalized parameters of trainees recorded in the profile, such as the weight of trainees' preference for "video demonstration" content. The preference weight for "pure text theory" is... The current level of understanding of the "Pharmacokinetics" module is only [missing information]. The data, along with the learning time distribution records in the logs, identifies the student's daily learning activities. to There is continuity A 24-hour efficient learning window, based on the logic of learning needs analysis, applies the standards stipulated in the syllabus. Preset learning time for each core knowledge point Actual time consumed by students By comparing each item, if the time consumption for a certain knowledge point (such as "antibacterial spectrum of macrolide antibiotics") is calculated, the... That is, the actual time taken exceeds the benchmark. Mark it as a delayed node, and vice versa. Then, it is marked as a node that passes quickly, and then the Genetic Algorithm engine is called to perform sequence encoding on the learning content, which is to be learned. Each knowledge point is mapped to an integer encoding sequence. Chromosomes are used as the initial population, and a fitness function is defined. ,in Representing the The difficulty level of each knowledge point This represents the priority weight of the knowledge point within the pharmacist competency framework. Let the balance coefficient be (set as) ), To prevent the denominator from reaching a minimum, the formula aims to find the optimal arrangement with a smooth difficulty gradient and high-priority content placed first. The population size is set to... The crossover probability is The mutation probability is ,implement The next iteration of evolution, for example, in the 1st iteration... In generational evolution, the sequence is calculated. The fitness of [penicillin, cephalosporins, carbapenems] is , and sequence [Penicillin, pharmaceutical administration, carbapenems] have a very high difficulty level, resulting in only a low fitness rate. The algorithm performs selection and replication operations based on fitness, and finally outputs the chromosome sequence with the highest fitness, generating an optimized learning content order.

[0041] The path adjustment calculation submodule, based on optimizing the order of learning content, calls up the knowledge point dependencies of students, monitors the interdependencies between knowledge points, compares the costs of different learning paths, calculates and adjusts the path through the analysis of path conversion costs, and obtains a dynamically adjusted learning path. Based on optimizing the learning content order, entity relationship data from the pharmaceutical knowledge graph is loaded, and a knowledge point dependency model in the form of a directed acyclic graph (DAG) is established. This model iterates through adjacent knowledge nodes in the sequence. and Monitoring whether there is a source point to Necessary prerequisite paths, such as "pharmacokinetics," must be prerequisite nodes for "dosing regimen design." If a dependency inversion is detected in the sequence (i.e., a prerequisite node is placed after a subsequent node), a path penalty mechanism is immediately triggered, and the path transformation cost is calculated. The cost calculation formula is set as follows: ,in Indicates the difficulty range between nodes. This represents the set of necessary prerequisite knowledge points that are currently missing. The weight of the missing node. and These are the difficulty penalty coefficients (taken as follows) ) and dependency missing penalty coefficient (take The costs of different learning paths are compared, with specific examples shown in Table 4. For instance, if a student plans to switch from "Basic Chemistry" to "Drug Synthesis," path A, which involves a direct jump, incurs costs. One missing prerequisite point, computational cost Path B inserts the intermediate node "Organic Chemistry". Although the number of steps increases, the dependencies remain complete, the missing set is empty, and the computational cost is [not specified]. ,determination Path B is selected as the preferred option. By analyzing the path conversion cost, transitional knowledge nodes are dynamically inserted or the node order is adjusted until the sum of the conversion costs of adjacent nodes is minimized, thus obtaining a dynamically adjusted learning path.

[0042] Table 4: Example Table of Cost Assessment and Adjustment for Switching Pharmacy Learning Pathways

[0043] As shown in Table 4, Table 4 lists the cost calculation and decision-making process when planning paths between different pharmaceutical knowledge nodes. By quantifying the conversion costs caused by the difficulty span and the lack of dependencies, the criteria for determining whether to insert intermediate nodes (such as inserting "common disease pathology" as a transition in Path-001) are clarified. The results show that logical gaps in the learning path can be automatically repaired based on the principle of cost minimization, ensuring that the learning process conforms to cognitive laws.

[0044] The optimized path generation submodule dynamically adjusts the learning path, maps the adjusted path to the student's learning time and task completion data, calls the priority of learning content and path adjustment strategy, evaluates the path task completion rate, and obtains the optimized path planning scheme. Based on dynamically adjusted learning paths, the system accesses students' personal schedule data to obtain a set of discrete learning time slices available for the next week. For example, Monday Minutes, Tuesday Minutes, the standard estimated time to read each knowledge node in the adjusted path. Combining the learners' historical learning efficiency factors in the corresponding ability dimensions (For example, the efficiency in "memorization-based" knowledge points is) That is, faster than the standard ), Applying the time mapping formula Calculate the actual time required, assuming the standard time for the "atropine pharmacological action" node. Minutes, student efficiency factor Then calculate Every minute, following the principle of high priority followed by low priority, the calculated node tasks are filled into the time slice. In the middle, if a certain time segment Minutes, current task requires Minutes remaining Minutes, the time taken in the automatic retrieval path is less than or equal to Sub-tasks are filled within minutes; if no suitable tasks are available, a review session is scheduled. The completion rate of the path tasks is assessed, and the expected completion of high-priority nodes (competency weights) is calculated. Total number of days required ,like If the assessment deadline is exceeded, the path adjustment strategy will be triggered, automatically removing candidates with weights lower than [a certain threshold]. The extended content is compressed to generate a schedule accurate to the minute, containing specific learning content (such as "19:00-19:25 learning atropine pharmacology") and review nodes, resulting in an optimized path planning scheme.

[0045] Specifically, such as Figure 2 , 7 As shown, the interactive practice module includes: The interactive data acquisition submodule, based on the optimized path planning scheme, extracts the constraints of the teaching scenario and initializes the parameters of the virtual patient agent, virtual pharmacist agent, and assessment and monitoring agent, constructs a multi-agent interactive environment, and acquires the dialogue text between the simulated virtual pharmacist and the patient, the sequence of operation instructions, and the interaction time stamps in the multi-agent interactive environment through the assessment and monitoring agent. It performs text statement segmentation and sequence indexing, instruction event alignment and frequency statistics, verifies consistency based on time stamps, and generates an interactive behavior sequence matrix. Based on the optimized path planning scheme, the constraints of the teaching scenario are extracted and the parameters of each intelligent agent are initialized to construct a multi-intelligent agent interactive environment. For example, when loading a targeted scenario of "medication consultation for elderly patients with hypertension", the extracted scenario constraint is "focusing on the identification of the interaction between antihypertensive drugs and health products and the patience of communication". First, the virtual patient intelligent agent is initialized, giving it the pathological and behavioral characteristics of "75 years old, with mild hearing impairment, and recently taking health products with unknown ingredients without authorization", and loading the corresponding spoken language and repetitive dialogue logic library; second, the virtual pharmacist intelligent agent is initialized, loading standard pharmacopoeia and clinical guidelines as an auxiliary intervention and prompt knowledge base; finally, the assessment and monitoring intelligent agent is initialized, setting its multimodal event capture frequency and response threshold, thereby integrating to form a multi-intelligent agent interactive environment. During the operation of the multi-agent interactive environment, the evaluation and monitoring agent activates a dual-channel data stream monitoring service to capture the real-time dialogue text stream between the simulated virtual pharmacist and virtual patient agents, as well as operation command sequences such as mouse clicks and keyboard inputs. It records the Unix millisecond-level timestamps for each discrete event. For the acquired dialogue text, it performs text segmentation and sequence indexing. First, continuous speech-to-text content, such as "Hello, have you been taking amlodipine besylate tablets on time recently?", is segmented into the smallest semantic units according to semantic pauses and syntactic structure, breaking it down into... "Hello", "Excuse me, may I ask you?" Have you had any questions recently? "Take it on time" The product name is “Amlodipine Besylate Tablets”. The five units are assigned sequentially increasing index values. Then, instruction event alignment and frequency statistics are performed, and the time stamps in the operation instruction sequence are read. With the corresponding speech semantic unit time stamp Based on the time stamp verification consistency, a maximum deviation threshold for consistency determination is set. The process for setting this threshold is as follows: data collection The average response delay was calculated using data on the average response time of a group of senior pharmacists. milliseconds, standard deviation milliseconds, based on Principle setting Milliseconds: For each operation command (such as "Click to view prescription record"), retrieve the time before and after its occurrence. If a semantic unit within a millisecond range is found to correspond to another unit, the two units are bound to the same interactive behavior group. For example, in... At milliseconds, the student verbally states, "Let me check your past medical history," The "click on medical history file" operation was executed in milliseconds; the time difference between the two is... Milliseconds are used to determine consistent behavior; otherwise, if the time difference exceeds the threshold, it is marked as an independent event. The frequency of operations associated with semantic units is counted, and an interaction behavior sequence matrix containing index value, semantic content, associated operation type, timestamp, and consistency status is constructed.

[0046] The intent mapping determination submodule, based on the interaction behavior sequence matrix, calls the concept nodes and relation constraints of the knowledge graph, performs semantic unit matching judgment, divides the interval according to the density of the operation event and the corresponding semantic unit, and maps it in combination with the teaching scenario state identifier to obtain the interaction intent matching coefficient; Based on the interaction behavior sequence matrix, interaction units are extracted row by row. Concept nodes and relational constraints in the pharmaceutical knowledge graph are invoked. For example, for the semantic unit "amlodipine besylate tablets", the corresponding entity node is located in the graph, and its associated attributes such as "indications" and "adverse reactions" are retrieved as potential intent spaces. Semantic unit matching judgment is performed, and the similarity of the student's spoken content with the standard pharmaceutical service terminology database is compared to confirm that the student's intent is "medication compliance check". Subsequently, intervals are divided according to the density of operation events and semantic units, and the operation density within a specific interaction window is calculated. ,in The number of valid operation instructions, For the number of semantic units, a baseline density range is preset for different teaching scenarios. This range is determined based on statistical analysis of the behavioral patterns of historical high-scoring students. For example, in the "information gathering" phase, the standard density mean is... (i.e., average per person) Execution of a sentence (second operation), set a reasonable range as Right now If the real-time computational density of trainees in the current "information gathering" state , in the interval Within the context, the interaction rhythm is determined to be reasonable. This is then mapped using teaching scenario state identifiers (such as "Current_State:History_Taking") to determine the semantic matching score. (set as) (Density fit score) (set as) Perform weighted calculations to determine the interaction intent matching coefficient. Set weights ,but If trainees only provide extensive verbal instruction without any demonstrations during the "medication guidance" phase, it may lead to... The range is lower than the requirement for this stage. Then the density fit score Downgraded to This leads to a decrease in the matching coefficient. The entire sequence matrix is ​​traversed to generate the interaction intent matching coefficient.

[0047] The performance quantification generation submodule collects the standard path of the teaching scenario based on the interaction intent matching coefficient, performs position offset calculation between the student interaction path and the standard path, performs interval statistics and weight merging based on the offset distribution, and generates the interaction performance result by combining the coverage of thinking nodes in the knowledge graph. Based on the interaction intent matching coefficient, the standard path for the teaching scenario is loaded. This standard path consists of a sequence of standard interaction nodes defined by experts, such as [greeting, confirming patient identity, inquiring about allergies, reviewing prescriptions, explaining medication]. The offset between the student's interaction path and the standard path is calculated, and a sequence alignment algorithm is used to compare the actual interaction node sequence generated by the student. and Alignment is performed, and the cumulative cost of insertion, deletion, and replacement operations is calculated as the path offset. The specific calculation example is shown in Table 5, where the standard path length is set. The student's path was missing the "Inquire about allergy history" node, and the "Promote health supplements" node was incorrectly inserted before the "Explanation of medication" node. Based on the preset penalty weight table (weight for missing key nodes), this will be penalized. Incorrectly inserted irrelevant node weights Calculate the offset value, as shown in Table 5, the total offset. Based on the offset distribution, interval statistics and weight merging are performed to map the offset to the score interval, and a full score is set. Points, deduction formula is: The path is then divided into In addition, based on the coverage of thought nodes, the set of knowledge graph nodes triggered by the user during the interaction is analyzed. The core node set that the scene must cover Calculate the coverage of the intersection of the two coordinates. Assuming the core nodes have a total of The number of students covered One, then Finally, the path score and coverage metric are combined to generate interactive performance results.

[0048] Table 5: Quantitative Calculation Examples of Deviation Between Student Interaction Path and Standard Pharmaceutical Service Path

[0049] As shown in Table 5, Table 5 presents in detail the comparative analysis data of trainees' performance in simulated pharmaceutical services and the standard process. By quantifying different types of path deviations (such as the high penalty caused by missing the key step "asking about allergy history"), the trainees' operational deviations can be accurately calculated. The results show that although trainees completed most of the basic processes, there were significant omissions in the key safety check links, which directly led to a significant decrease in the quantitative performance score, thus providing a digital basis for subsequent targeted error correction.

[0050] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the described technical solutions.

Claims

1. A dynamic evolution and quantitative assessment system for pharmacist competence based on multi-agent agents, characterized in that, The system includes: The multimodal resource parsing module receives various forms of pharmaceutical teaching materials, converts images into text using optical character recognition, performs semantic encoding and feature extraction on the text based on BERT, and constructs a structured knowledge graph. The competency association module, based on the structured knowledge graph, uses cosine similarity analysis to analyze the pharmacist competency framework, ability descriptions, and knowledge point texts to generate competency-annotated teaching knowledge nodes. The ability dynamic assessment module calls the teaching knowledge nodes marked with the competency, collects multi-dimensional learning behavior data of students, uses a deep Q network to calculate the weight of the ability dimension in combination with the current ability performance of students, optimizes the weight allocation through weighted fusion, and generates a multi-dimensional competency profile. The path planning module, based on the multi-dimensional competency profile, uses a genetic algorithm to optimize the order of learning content, dynamically adjusts the path in combination with knowledge dependencies, and generates an optimized path planning scheme. The interactive practice module constructs a multi-agent interactive environment based on the optimized path planning scheme. Through the multi-agent interactive environment, it simulates the dynamic collaboration and dialogue between virtual pharmacists and patients, analyzes the trainees' operations and language intentions, and, combined with knowledge graphs and teaching scenarios, assesses the trainees' communication skills and clinical reasoning abilities, and outputs the interactive performance results.

2. The pharmacist competence dynamic evolution and quantitative assessment system based on multi-agent agents according to claim 1, characterized in that, The structured knowledge graph includes pharmaceutical concept nodes, concept relationships, and a set of knowledge attributes. The competency-annotated teaching knowledge nodes include competency element tags, competency level identifiers, and job suitability markers. The multi-dimensional competency profile includes competency strength and weakness distribution, competency development level, and competency structure characteristics. The optimized path planning scheme includes learning content ranking results, path dependency constraint sets, and personalized adjustment rules. The interaction performance results include communication quality evaluation, clinical decision-making tendency, and comprehensive practical ability indicators.

3. The pharmacist competence dynamic evolution and quantitative assessment system based on multi-agent agents according to claim 1, characterized in that, The multimodal resource parsing module includes: The image and text receiving submodule acquires various forms of pharmaceutical teaching materials, collects image frames and text files, performs character detection and mapping on image frames through optical character recognition, extracts text strings and aligns them with the original text file in terms of format, filters character sequences based on noise ratio thresholds, and generates character sequence integrity values. The semantic encoding submodule, based on the character sequence integrity rate value, calls the filtered character sequence and inputs it into the BERT semantic encoding structure to perform word segmentation mapping and context association calculation, extracts multi-layer semantic embedding vectors, performs vector dimension consistency judgment, if the dimensions are inconsistent, unifies the vector dimensions through linear transformation, and then performs normalization processing to generate a semantic feature vector matrix. The graph construction submodule performs concept co-occurrence statistics based on the semantic feature vector matrix and the original text corpus, collects concept nodes, calculates the co-occurrence frequency of nodes in the text, calculates node relationship weights based on the frequency distribution and performs threshold judgment, filters node relationships with weights greater than the preset weight threshold and integrates node and relationship attribute sets to generate a structured knowledge graph.

4. The pharmacist competence dynamic evolution and quantitative assessment system based on multi-agent agents according to claim 3, characterized in that, The noise ratio threshold is based on statistics of the proportion of non-character pixels in the character pixel sequence, and the noise ratio threshold is set to a closed interval of 10% to 20%. The preset weight threshold is derived from the statistical characteristics of the relationships between knowledge graph nodes and is a fixed constant with a value range of 0.3 to 0.

5.

5. The pharmacist competence dynamic evolution and quantitative assessment system based on multi-agent agents according to claim 1, characterized in that, The competency association module includes: The competency framework parsing submodule retrieves the set of concept node attributes based on the structured knowledge graph, obtains the pharmacist competency framework text data and performs word segmentation, maps the word segmentation results to the term index table, extracts semantic embedding vectors and performs dimension alignment and normalization operations to generate a set of competency description semantic vectors. The similarity calculation submodule calls the semantic vector set of capability description and the text vector of knowledge nodes in the structured knowledge graph, calculates the cosine similarity value, filters out node pairs with a similarity greater than a preset threshold, records the node identifier and capability description identifier, and generates a node capability similarity mapping table. The node annotation submodule extracts node identifiers and capability description identifiers based on the node capability similarity mapping table, calls the node attributes in the structured knowledge graph, writes the capability description identifiers into the attribute fields, and generates competency-annotated teaching knowledge nodes.

6. The pharmacist competence dynamic evolution and quantitative assessment system based on multi-agent agents according to claim 5, characterized in that, The preset similarity threshold is based on a set of node samples in a structured knowledge graph that have been labeled with capability descriptions. The cosine similarity value distribution interval is calculated between the semantic vector set of capability descriptions and the text vector of the corresponding knowledge nodes. Within the cosine similarity value distribution interval, the mean and standard deviation of the similarity of positively correlated node pairs are statistically analyzed, and the result obtained by subtracting the standard deviation from the mean similarity is used as the preset similarity threshold.

7. The pharmacist competence dynamic evolution and quantitative assessment system based on multi-agent agents according to claim 1, characterized in that, The dynamic capability assessment module includes: The behavioral data collection submodule calls the teaching knowledge nodes labeled with the competency, collects student access duration data, interaction frequency data, test score data and operation path data, performs alignment operation according to timestamp, calculates the mean of access duration and the variance of interaction frequency, and generates a multi-dimensional behavioral feature matrix of students. The weight dynamic calculation submodule extracts behavioral feature values ​​based on the student's multi-dimensional behavioral feature matrix, constructs a state space by calling the ability labeling attributes in the teaching knowledge nodes, maps the feature values ​​to the state vector, inputs them into the deep Q network for forward propagation to calculate the Q value of the ability dimension and performs normalization operation to generate a dynamic weight vector of the ability dimension. The profile fusion generation submodule calls the dynamic weight vector of the ability dimension and the multi-dimensional behavioral feature matrix of the trainee, performs element-wise multiplication operation on the column vector of the behavioral feature matrix and the weight vector components, and performs column vector summation operation on the weighted result to generate a multi-dimensional competency profile.

8. The pharmacist competence dynamic evolution and quantitative assessment system based on multi-agent agents according to claim 1, characterized in that, The path planning module includes: The learning content optimization submodule collects student learning preference data, knowledge point mastery, and learning time distribution based on the multi-dimensional competency profile. Based on learning needs analysis, it identifies knowledge points with lagging and fast learning progress, encodes the learning content using a genetic algorithm, and generates an optimized learning content order. The path adjustment calculation submodule, based on the optimized learning content order, calls the knowledge point dependencies of the learners, monitors the interdependencies between knowledge points, compares the costs of different learning paths, calculates and adjusts the paths through the analysis of path conversion costs, and obtains dynamically adjusted learning paths. The optimized path generation submodule dynamically adjusts the learning path, maps the adjusted path to the student's learning time and task completion data, calls the priority of learning content and path adjustment strategy, evaluates the path task completion rate, and obtains an optimized path planning scheme.

9. The pharmacist competence dynamic evolution and quantitative assessment system based on multi-agent agents according to claim 1, characterized in that, The interactive practice module includes: The interactive data acquisition submodule, based on the optimized path planning scheme, extracts the teaching scenario constraints and initializes the parameters of the virtual patient agent, virtual pharmacist agent, and assessment and monitoring agent, constructs a multi-agent interactive environment, and obtains the simulated virtual pharmacist and patient dialogue text, operation instruction sequence, and interaction time stamps in the multi-agent interactive environment through the assessment and monitoring agent. It performs text statement segmentation and sequence indexing, instruction event alignment and frequency statistics, verifies consistency based on time stamps, and generates an interactive behavior sequence matrix. The intent mapping determination submodule, based on the interaction behavior sequence matrix, calls the knowledge graph concept nodes and relation constraints, performs semantic unit matching judgment, divides the interval according to the density of the operation event and the corresponding semantic unit, and maps it in combination with the teaching scenario state identifier to obtain the interaction intent matching coefficient; The performance quantification generation submodule collects the standard path of the teaching scenario based on the interaction intent matching coefficient, performs position offset calculation between the student's interaction path and the standard path, performs interval statistics and weight merging based on the offset distribution, and generates the interaction performance result by combining the coverage of thinking nodes in the knowledge graph.

10. A pharmacist competence dynamic evolution and quantitative assessment system based on multi-agent agents according to claim 9, characterized in that, The execution text statement segmentation and sequential indexing refers to segmenting continuous text into the smallest semantic units according to the time stamp order in the dialogue text between the simulated virtual pharmacist and the patient, and assigning an increasing sequential index value to each smallest semantic unit. The consistency verification based on time stamps refers to determining that the maximum deviation between the time stamps of the dialogue text between the simulated virtual pharmacist and the patient and the time stamps of the operation instruction sequence does not exceed a preset time threshold.

Images