A pharmacist clinical ability evaluation method and system based on multi-scene adaptation

Abstract

The present application provides a kind of based on multi-scene self-adapting pharmacist clinical ability evaluation method and system, it is related to pharmacist ability evaluation technical field, including: pre-configuration standard diagnosis and treatment process, the mapping rule of operation event and diagnosis and treatment step for multiple evaluation scenes respectively, and the benchmark weight vector of five ability dimensions;Multi-modal data acquisition is carried out to the evaluation interaction process of pharmacist and virtual patient, and the dialogue text sequence is obtained;The dialogue text sequence is carried out semantic coding and context modeling, and the global dialogue abstract vector is obtained;Based on dialogue text sequence, diagnosis and treatment step sequence and pharmaceutical knowledge graph, five-dimensional score vector is obtained;Adaptive weight network is used to generate the fusion weight vector of five ability dimensions;Five-dimensional score vector is weighted aggregated with fusion weight vector, and comprehensive score is obtained, and structured evaluation report is output.The present application can overcome the deficiencies of single evaluation scene, shallow scoring dimension and weight logic solidification in the prior art.

Description

Technical Field

[0001] This invention relates to the field of pharmacist competency assessment technology, and in particular to a method and system for assessing pharmacist clinical competency based on multi-scenario adaptive assessment. Background Technology

[0002] Assessing pharmacists' clinical competence is crucial for ensuring the quality of pharmaceutical services and patient medication safety. However, existing pharmacist competence assessment systems largely rely on manual scoring or standardized, fixed scales. The assessment scenarios are limited, typically designed for specific work environments, failing to capture the competence differences pharmacists face in various settings such as inpatient wards, pharmacy clinics, pharmacy consultation windows, and community pharmacies. Furthermore, current assessment dimensions often remain at a superficial level of knowledge acquisition, lacking systematic methods for quantifying deeper clinical competences such as information gathering efficiency, accuracy in identifying medication-related issues, logical reasoning in treatment plans, communication effectiveness, and adherence to standardized procedures. The fixed scoring logic prevents dynamic adjustments to the emphasis of each dimension based on scenario variations and the assessee's individual competence distribution, hindering targeted assessment of individual pharmacists' weaknesses.

[0003] Chinese patent application CN110826861A discloses a method and system for determining clinical reasoning ability. This method constructs a multi-stage human-computer interactive examination framework to simulate clinical diagnosis and treatment processes such as history taking, physical examination, auxiliary examinations, and treatment decision-making. It records the operational behavior of the assessed individual at each stage and calculates a comprehensive score based on preset weights and scoring strategies, outputting the clinical reasoning ability assessment result. This patent achieves structured scoring of clinicians' consultation order, operational efficiency, and diagnostic logic across multiple dimensions, possessing a certain degree of objectivity and quantifiability. However, this patent is geared towards assessing the diagnostic reasoning abilities of clinicians and does not specifically address the particular needs of pharmacists' pharmaceutical service capabilities (such as drug interaction identification and standardized medication education). The weight vectors for each assessment stage remain fixed throughout the evaluation process, failing to adaptively adjust according to the differences in assessment needs across different work scenarios and the assessed individual's own capabilities. Summary of the Invention

[0004] In view of this, the present invention provides a method and system for assessing pharmacists' clinical competence based on multi-scenario adaptation. By introducing scenario-based knowledge graphs, pre-trained language models in the medical field, and adaptive weight networks, it achieves an objective and quantitative assessment of pharmacists' comprehensive competence in various clinical scenarios, such as inpatient wards, pharmacy clinics, pharmacy consultation windows, and community pharmacies. Furthermore, it dynamically adjusts the scoring weights of each competence dimension based on the current assessment scenario, dialogue semantic features, and the pharmacist's historical competence profile, overcoming the shortcomings of existing technologies such as single assessment scenarios, shallow scoring dimensions, and fixed weight logic.

[0005] The technical solution of this invention is implemented as follows: On the one hand, this invention provides a pharmacist clinical competence assessment method based on multi-scenario adaptation, including: S1. Pre-configure standard diagnosis and treatment procedures, mapping rules between operational events and diagnosis and treatment steps for multiple assessment scenarios, as well as benchmark weight vectors for five capability dimensions: completeness of information collection, accuracy of drug problem identification, logicality of treatment plan, effectiveness of communication, and standardization of process, and associate them with pharmaceutical knowledge graphs. S2. Multimodal data collection is performed on the assessment interaction process between pharmacists and virtual patients. The collected dialogue speech is converted into a dialogue text sequence through speech recognition, and the collected pharmacist operation behavior is converted into a sequence of diagnosis and treatment steps according to the mapping rules. S3. Using a pre-trained language model in the medical field combined with a prior enhanced attention mechanism, semantic encoding and context modeling are performed on the dialogue text sequence to obtain a global dialogue summary vector. Based on the dialogue text sequence, the diagnosis and treatment step sequence and the pharmaceutical knowledge graph, pharmacists are scored on five ability dimensions to obtain a five-dimensional score vector. S4. The current evaluation scene is embedded and encoded to obtain the scene embedding vector. The scene embedding vector and the global dialogue summary vector are used as inputs. An adaptive weight network is used to dynamically generate the fusion weight vector of the five capability dimensions with the benchmark weight vector as the initial reference. S5. Use the fusion weight vector to weight and aggregate the five-dimensional scoring vectors to obtain a comprehensive score and output a structured evaluation report.

[0006] Preferably, in step S1, configuration information for each evaluation scenario is stored in the form of triples, including scenario type identifier, a directed graph of standard treatment process, and baseline weight vectors for five capability dimensions; the directed graph of standard treatment process consists of several ordered clinical nodes, each of which is assigned a prior importance probability. Within the same scenario, the sum of the prior importance probabilities of all nodes is 1. These prior importance probabilities are pre-defined by pharmaceutical experts in conjunction with clinical guidelines. Each assessment scenario also includes a set of key steps. and the mapping rules between operational events and diagnostic and treatment steps, the set of key steps Mark the essential diagnostic and treatment steps in this scenario.

[0007] Preferably, in step S2, the dialogue speech is organized into a dialogue text sequence according to rounds after speech recognition. ,in For each t-th round of dialogue, a unit of speech contains the speaker's identifier and the text content of the speech, where t = 1, 2, ..., T, and T is the total number of rounds in the dialogue; the pharmacist's actions are converted into a sequence of diagnostic and treatment steps using mapping rules. ,in Let j be the node identifier for the j-th actual treatment step, where j = 1, 2, ..., M, and M is the total number of actual treatment steps.

[0008] Preferably, in step S3, the scores for the five capability dimensions are based on the dialogue text sequence, diagnosis and treatment step sequence, and pharmaceutical knowledge graph collected in step S2 as input. The four dimensions of information collection completeness, drug problem identification accuracy, treatment plan logic and communication effectiveness are scored based on the semantic encoding results of the dialogue text sequence and the pharmaceutical knowledge graph. The process standardization dimension is scored based on the diagnosis and treatment step sequence and the standard diagnosis and treatment process configured in the scenario. The scores for each dimension are normalized to [0,1].

[0009] Preferably, in step S3, a pre-trained language model for the medical field is used for each round of discourse units. Semantic encoding is performed to obtain round-by-round semantic vectors. Each round of pharmacist utterances is assigned to the corresponding clinical node category through intent classification; a bidirectional long short-term memory network is used to perform contextual modeling on the sequential semantic vector sequences to obtain the context-aware representation of each round. The importance weight for each round is calculated using a priori attention enhancement mechanism. ; in, For attention transformation matrix, and For learnable parameters, The prior importance probability of the clinical node to which the t-th round of speech belongs; the global dialogue summary vector. The attention weighted sum for each round of context-aware representation: .

[0010] Preferably, in step S3, the information collection integrity score is performed. The calculation method is as follows: Based on the current scenario, N key information points are predefined, and each information point i is assigned an importance weight. ,satisfy and expected collection rounds ; By using round-by-round semantic vectors Calculate the cosine similarity between the data points and the standard descriptor vectors of each data point to determine the collection marker of each data point. And record the actual number of data collection rounds. ; Calculate the timing rationality decay factor: ; in, As a time-series rationality decay factor, The attenuation rate, For information point i, the source label is active, indicating that the pharmacist actively inquired and obtained the information, and passive, indicating that the patient voluntarily disclosed the information. Redundancy rate of repeated questions asked by pharmacists As a penalty, the information collection integrity score is: ; in The redundancy penalty coefficient is... .

[0011] Preferably, in step S3, the accuracy score for drug problem identification is... The calculation method is as follows: A set of standard drug questions pre-labeled for current patient cases from a pharmaceutical knowledge graph. Each standard drug problem With severity weights k = 1, 2, ..., K, where K is the total number of standard drug problems; The information extraction model identifies the set of medication-related questions that pharmacists explicitly mention or suggest addressing during conversations, and compares them with... Perform semantic matching to obtain the hit flag. ; The accuracy score for drug problem identification is: ; in The false positive rate, The false alarm penalty coefficient, The range of values ​​after truncation is .

[0012] Preferably, in step S3, the treatment plan is scored logically. It is divided into two parts: rule engine evaluation and knowledge graph verification. In the rule engine evaluation section, clinical entity recognition and clinical causal relationship extraction are performed sequentially on the pharmacist's discourse to construct a reasoning chain graph. ,in A collection of clinical entities. Let the set of causal relationship edges be defined; align the reasoning paths in the reasoning chain graph with the standard paths of the clinical rule base and knowledge graph, and output the reasonableness score for the j-th reasoning step. and the importance weight of steps ; In the knowledge graph validation section, path retrieval is performed on drug associations in the pharmacist's recommended schemes, and the knowledge graph path matching score is obtained by combining the existence and rationality of the paths. ; The logical reasoning score for the treatment plan is: ; in Let be the fusion coefficient between the two scores, and m be the number of inference steps. .

[0013] Preferably, in step S3, the communication effectiveness is scored. Calculated at the granularity of p clinical nodes defined in the directed graph of the standard diagnosis and treatment process in the scenario configuration; Temporal importance weights of each clinical node k By attention weight Aggregation yields: ; in Let be the set of dialogue turn indices belonging to the k-th clinical node in the intent classification results, satisfying... ; For each clinical node k, four sub-indicators are calculated using a classification model fine-tuned on medical dialogue data: Clarity of expression This was determined by analyzing the syntactic complexity and terminology density of pharmacists' discourse. Empathy This was determined by detecting whether the pharmacist's words contained expressions of emotional support. Information verification The results were obtained by assessing whether pharmacists proactively confirmed the patient's understanding of the medication instructions. Patient's mood improvement , This refers to the emotional state value extracted by the sentiment analysis model from the response of a virtual patient at the k-th clinical node. +1 indicates extreme anxiety or resistance; +1 indicates complete calm or trust. The overall communication effectiveness score is: ; in and The weights for the four sub-indicators are preset by the scenario configuration. .

[0014] Preferably, in step S3, the process standardization score is performed. The calculation method is as follows: Sequence of diagnostic and treatment steps With the standard process sequence predefined in the scenario configuration The two sequences were compared, and the weighted edit distance method was used to calculate the weighted edit distance between them. The cost weight of each step is determined by the importance weight of the corresponding step in the scenario configuration; If the key steps set If any step in the diagnostic and treatment sequence is missing, the score will be forcibly reset to zero. The process standardization score is: ; in This represents the total number of steps in the standard process. For indicator functions, The pharmacist's handling of the emergency was scored. For the response reward coefficient, .

[0015] Preferably, in step S4, the adaptive weight network further uses the pharmacist's historical ability vector To aid input, the pharmacist's historical average scores across five competency dimensions are recorded, initialized as a vector of all 0.5 during the initial assessment; The weighted network first uses scene embedding vectors and global dialogue summary vector Calculating historical information using gating: ; in For the gated weight matrix, For bias vectors, It is the Sigmoid activation function. This indicates vector concatenation; Then, the three types of inputs are combined to generate a fusion weight vector: ; in , , For learnable projection matrices, For element-wise multiplication, The dimension-focusing bias vector is initialized to zero during the first evaluation. After each evaluation, the historical capability vector is updated using an exponential moving average method: ; in Forgetting factor, This is the current historical capability vector. For the updated historical capability vector, This is the five-dimensional scoring vector for this assessment.

[0016] Preferably, the dimension-focusing bias vector Iterative updates are performed between evaluation sessions by the deep Q-network: state From global dialogue summary vector Historical capability vector Compared to the five-dimensional scoring vector of the previous assessment Composed of splicing elements; action Limit the adjustment amount for each dimension to A five-dimensional real-valued vector within the range, This is the maximum adjustment range overparameter; The reward function is: ; in The index of the weakest dimension with the lowest score in the previous assessment. This is the balance coefficient.

[0017] In addition, the present invention also provides a pharmacist clinical competence assessment system based on multi-scenario adaptation, comprising: The scenario configuration module is used to pre-configure standard diagnosis and treatment procedures, mapping rules between operation events and diagnosis and treatment steps, and benchmark weights for five capability dimensions for multiple assessment scenarios, and to link them to the pharmaceutical knowledge graph. The data acquisition module is used to collect multimodal data on the assessment interaction process between pharmacists and virtual patients. It converts the collected dialogue speech into a dialogue text sequence through speech recognition, and converts the collected pharmacist operation behavior into a sequence of diagnosis and treatment steps according to mapping rules. The scoring module uses a pre-trained language model in the medical field combined with a prior enhanced attention mechanism to perform semantic encoding and context modeling on the dialogue text sequence to obtain a global dialogue summary vector. Based on the dialogue text sequence, the diagnosis and treatment step sequence, and the pharmaceutical knowledge graph, the module scores pharmacists on five competency dimensions to obtain a five-dimensional scoring vector. The weight generation module is used to embed the current evaluation scene into a scene embedding vector. Taking the scene embedding vector and the global dialogue summary vector as input, the adaptive weight network dynamically generates a fusion weight vector of five capability dimensions with the baseline weight as the initial reference. The output module is used to weight and aggregate the five-dimensional score vectors with a fusion weight vector to obtain a comprehensive score and output a structured evaluation report.

[0018] The present invention has the following advantages over the prior art: (1) This invention pre-configures directed graphs of standard diagnosis and treatment processes, prior node importance probabilities, and benchmark weight vectors of capability dimensions for multiple assessment scenarios, and associates them with pharmaceutical knowledge graphs. This scenario configuration mechanism enables the assessment framework to establish differentiated assessment benchmarks for differences in diagnosis and treatment processes and capability emphases in different work environments, avoiding the assessment distortion problem caused by applying a single assessment standard to multiple scenarios.

[0019] (2) This invention designs independent scoring algorithms for five dimensions: completeness of information collection, accuracy of drug problem identification, logicality of treatment plan, effectiveness of communication, and standardization of process. The scoring for each dimension fully integrates pharmaceutical knowledge graph and scenario configuration information: the information collection score introduces a time-decay factor to distinguish the value of information obtained through proactive inquiry from that obtained passively, and adds a penalty for redundant inquiries; the drug problem identification score combines the severity weight of the knowledge graph and penalizes false positives; the logicality of the treatment plan score uses reasoning chain extraction and alignment with the knowledge graph path to structurally verify the pharmacist's clinical reasoning process; the standardization of process score uses weighted edit distance to measure process deviation and imposes a mandatory zeroing constraint on missing key steps. This meticulous scoring method for each dimension gives the score independent diagnostic significance, facilitating the identification of pharmacists' specific weaknesses.

[0020] (3) This invention designs an adaptive weight network with scene embedding vector, global dialogue summary vector, and pharmacist historical ability vector as joint inputs. Through a learnable gating mechanism, it dynamically adjusts the reference degree of historical ability profile to the weight distribution of the current assessment, and further corrects the weight distribution with the dimension focusing bias vector output by the reinforcement learning module. This mechanism enables the weights of each dimension of the comprehensive score to be determined jointly according to scene characteristics and individual pharmacist ability status, achieving differentiated assessment emphasis among different assessees, which helps to improve the sensitivity of identifying weak dimensions. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the prior enhanced attention modeling process of the present invention; Figure 3 This is a flowchart of the adaptive weight generation process of the present invention; Figure 4 This is a system framework diagram of the present invention. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] like Figure 1 As shown, this invention provides a pharmacist clinical competence assessment method based on multi-scenario adaptation, including: S1. Pre-configure standard diagnosis and treatment procedures, mapping rules between operational events and diagnosis and treatment steps for multiple assessment scenarios, as well as benchmark weight vectors for five capability dimensions: completeness of information collection, accuracy of drug problem identification, logicality of treatment plan, effectiveness of communication, and standardization of process, and associate them with pharmaceutical knowledge graphs. S2. Multimodal data collection is performed on the assessment interaction process between pharmacists and virtual patients. The collected dialogue speech is converted into a dialogue text sequence through speech recognition, and the collected pharmacist operation behavior is converted into a sequence of diagnosis and treatment steps according to the mapping rules. S3. Using a pre-trained language model in the medical field combined with a prior enhanced attention mechanism, semantic encoding and context modeling are performed on the dialogue text sequence to obtain a global dialogue summary vector. Based on the dialogue text sequence, the diagnosis and treatment step sequence and the pharmaceutical knowledge graph, pharmacists are scored on five ability dimensions to obtain a five-dimensional score vector. S4. The current evaluation scene is embedded and encoded to obtain the scene embedding vector. The scene embedding vector and the global dialogue summary vector are used as inputs. An adaptive weight network is used to dynamically generate the fusion weight vector of the five capability dimensions with the benchmark weight vector as the initial reference. S5. Use the fusion weight vector to weight and aggregate the five-dimensional scoring vectors to obtain a comprehensive score and output a structured evaluation report.

[0025] In one embodiment of the present invention, step S1 includes: The system predefines four assessment scenarios: inpatient wards, pharmacy clinics, pharmacy consultation windows, and community pharmacies. Each assessment scenario stores configuration information in the form of triplets, specifically including a scenario type identifier, a directed graph of the corresponding standard treatment process, and a baseline weight vector for five capability dimensions in that scenario: completeness of information collection, accuracy of drug problem identification, logicality of treatment plan, effectiveness of communication, and standardization of process.

[0026] A directed graph of a standard clinical procedure consists of several ordered clinical nodes, each representing a type of operation in the procedure. Taking an inpatient ward scenario as an example, the nodes are as follows: verifying basic patient information → collecting medication and allergy history → reviewing the rationality of medical orders → identifying drug interactions → developing a medication education plan → communicating with medical staff. In a community pharmacy scenario, the nodes are adjusted to: understanding the patient's medication needs → identifying contraindications and combined medications → recommending a medication regimen → instructing on usage and dosage → explaining precautions. Each clinical node is assigned a priori importance probability during the scenario configuration phase. The specific values ​​are pre-calibrated by pharmaceutical experts in conjunction with clinical guidelines, ensuring that the sum of the prior importance probabilities of all nodes within the same scenario is 1. The physical meaning of the prior importance probability is: the more important the clinical node (such as collecting allergy history) is assigned a higher prior value, and the less important the node (such as small talk) is assigned a lower prior value. This probability will be used as a prior bias term in the attention calculation in step S3.

[0027] Each type of assessment scenario also includes a set of key steps. This involves marking essential diagnostic and treatment steps in the scenario, such as "checking allergy history" and "explaining medication precautions" in the inpatient ward scenario; and defining the mapping rules between operation events and diagnostic and treatment steps, defining the classification and correspondence between specific operation actions (such as "reading the drug instructions" and "retrieving the patient's medical history") and diagnostic and treatment step node identifiers, for use in the behavioral data preprocessing of step S2.

[0028] After the scenario configuration is completed, the system selects the scenario to which the current assessment task belongs based on the pharmacist's historical competency profile (all five competency dimensions are initialized to 0.5 in the initial assessment) and the scenario distribution strategy. It then retrieves standardized patient cases corresponding to the scenario from the question bank as the content carrier for subsequent interactive assessments.

[0029] In one embodiment of the present invention, step S2 includes: During the evaluation interaction, the system synchronously collects multimodal data in the following ways. For voice data, the system records the entire dialogue between the pharmacist and the virtual patient (played by the system). An automatic speech recognition model converts the speech into timestamped text records. Each record includes the speaker's identifier (pharmacist or patient), the content of the speech text, and the time of speaking (in units of dialogue rounds). The speech recognition model can use the Whisper model, which has good robustness in recognizing professional terminology in medical dialogue scenarios. For behavioral event data, the system synchronously records the operation events triggered by the pharmacist during the dialogue, forming a raw operation event sequence with timestamped values. Subsequently, using the operation event and treatment step mapping rules predefined in step S1, the raw operation event sequence is automatically classified and mapped into an execution sequence at the standard treatment step granularity. For example, the two operation events "reading the drug instructions" and "retrieving historical medication records" are both mapped to the treatment step node "collecting medication history."

[0030] In the preprocessing stage, the dialogue text is organized into a dialogue text sequence according to the rounds. ,in Let t be a speech unit, containing the speaker identifier and the speech text content, where t = 1, 2, ..., T, and T is the total number of rounds in the dialogue. After mapping, the sequence of diagnosis and treatment steps actually performed by the pharmacist is obtained. ,in Let j be the node identifier for the j-th actual treatment step, where j = 1, 2, ..., M, and M is the total number of actual treatment steps.

[0031] like Figure 2 As shown, in one embodiment of the present invention, the generation of semantic encoding and global dialogue summary vector in step S3 includes the following sub-steps: Sub-step one: Semantic encoding round by round. The system uses a BERT-like language model pre-trained on medical text corpora to encode each round of utterance units. Semantic encoding is performed, and the vector at the [CLS] position in the model output is taken as the semantic representation of this round. d represents the hidden layer dimension. In this embodiment, Med-BERT is used as the pre-trained language model. This model is pre-trained on a large-scale medical text corpus and has a strong semantic understanding ability for pharmaceutical-related terms and clinical expressions.

[0032] Sub-step two: Intent classification and clinical node alignment. The system categorizes the semantic vectors of each round of pharmacist utterances. Intent classification is performed using an intent classification head (a fully connected classification layer finely tuned on the scenario process node categories in step S1), assigning each round of speech to the corresponding clinical process node category. This is used for subsequent attention weight aggregation and node-level calculation of communication effectiveness scoring in step S3.

[0033] Sub-step three: Prior-enhanced temporal attention aggregation. The system uses a bidirectional long short-term memory network (BiLSTM) to process the round-by-round semantic vector sequence. Perform context modeling to obtain the context-aware representation for each round. This representation integrates both forward and backward contextual information. Subsequently, an a priori enhanced attention mechanism is used to calculate the importance weights of each utterance: ; in, For attention transformation matrix, and For learnable parameters, Let be the prior importance probability of the clinical node to which the t-th round of discourse belongs. The prior probability is expressed in logarithmic form. Superimposed on the data-driven attention score, this mechanism enables the model to maintain focus on key clinical rounds even with insufficient training data, effectively integrating knowledge-driven and data-driven approaches. Compared to standard content attention mechanisms, this mechanism can more stably focus on clinically important content in scenarios with limited sample sizes, reducing the interference of irrelevant rounds in the dialogue on the summary vector. Global dialogue summary vector. The attention weighted sum for each round of context-aware representation: ; This will be used as one of the inputs to the adaptive weight network in step S4. This will be further used in step S3, the communication effectiveness score.

[0034] In one embodiment of the present invention, the scoring of the five capability dimensions in step S3 takes the dialogue text sequence, diagnosis and treatment step sequence, and pharmaceutical knowledge graph collected in step S2 as input. The four dimensions of information collection completeness, drug problem identification accuracy, treatment plan logic and communication effectiveness are scored based on the semantic encoding results of the dialogue text sequence and the pharmaceutical knowledge graph. The process standardization dimension is scored based on the diagnosis and treatment step sequence and the standard diagnosis and treatment process configured in the scenario. The scoring results of each dimension are normalized to [0,1]. The specific scoring method of each dimension is as follows.

[0035] Information collection integrity score The calculation method is as follows: The system predefines N key information points based on the current scenario (such as chief complaint, current medication list, allergy history, liver and kidney function status, etc.) and assigns an importance weight to each information point i. ,satisfy and expected collection rounds Both are predefined by the scene configuration module.

[0036] The system uses round-by-round semantic vectors The cosine similarity is calculated with the standard description vector of each information point. If the cosine similarity with information point i in any round exceeds a preset threshold (determined by optimization on the validation set, with a default value of 0.8), then the information point is identified as having been collected. Simultaneously record the actual number of data collection rounds. ;otherwise .

[0037] The system distinguishes information sources and labels each piece of information with a source tag. Where "active" indicates information obtained through proactive inquiry by a pharmacist, and "passive" indicates information disclosed voluntarily by a patient without being asked. (Time-sequence rationality decay factor) This only applies to information points that are proactively inquired about; information points that are voluntarily disclosed by patients are fixed. No timing penalty is imposed: ; in, As a time-series rationality decay factor, The attenuation rate, This is the source label for information point i. When hour, No attenuation; the more collection rounds exceed the expected number, the greater the attenuation; if (Not collected), then It is not included in the calculation.

[0038] Furthermore, redundancy statistics are used as a penalty for highly semantically repetitive questions raised by the pharmacist during the dialogue. Assuming the pharmacist raises Q questions, the system calculates the cosine similarity of each pair of questions, and counts those with cosine similarities exceeding a threshold. The number of repeated questions (default value is 0.85) is Normalized to redundancy rate: ; Where Q represents the total number of questions asked by the pharmacist during this assessment dialogue. For those whose cosine similarity exceeds the threshold The number of question statements The total number of all possible pairwise combinations; when hour, , The information collection completeness score is: ; in, The redundancy penalty coefficient is... .

[0039] Drug problem identification accuracy score The calculation method is as follows: This dimension assesses a pharmacist's ability to identify potential drug-related issues in patients, including drug interactions, contraindication violations, and inappropriate dosage. The system pre-labels a set of standard drug-related issues for each patient case from a pharmaceutical knowledge graph. Each standard drug problem Includes knowledge graph severity weights Let k = 1, 2, ..., K, where K is the total number of standard drug-related questions. The system analyzes pharmacist discourse using an information extraction model (which can be a BERT-like model fine-tuned for medical named entity recognition and relation extraction tasks), identifies the set of drug-related questions explicitly mentioned or suggested by the pharmacist in the dialogue, and compares them with... Semantic matching is performed, and a match is determined if the cosine similarity exceeds a preset threshold, thus obtaining a match flag. The accuracy score for drug problem identification is: ; in, The false positive rate is the proportion of questions raised by pharmacists that are not within the standard set. The false positive penalty coefficient is used to ensure that the score reflects a balance between recall and precision. The range of values ​​after truncation is .

[0040] Treatment plan logical score It is divided into two parts: rule engine evaluation and knowledge graph verification. The rule engine evaluation section employs a three-stage reasoning chain extraction sub-process to structurally analyze the reasoning process in pharmacist discourse. The first stage is clinical entity recognition: the system uses a sequence labeling model trained on medical entity labeled data (such as a BERT-based NER model) to extract clinical entities from pharmacist discourse sentence by sentence. Categories include disease diagnosis, symptoms, drug name, dosage, contraindications, and organ function, forming a set of clinical entities for this dialogue. The second stage is clinical causal relationship extraction: the system uses a relation extraction model in the entity set. Based on this, identify explicit or implicit causal reasoning relationships in pharmacists' discourse and construct reasoning chain diagrams. ,in A collection of clinical entities. Let be the set of causal edges, each edge Representing entities By relation type (Such as "prohibited for", "dosage adjustment basis", "risks of concomitant use", etc.) refers to the entity. The third stage involves aligning the rule base with the knowledge graph path: the system will integrate the reasoning chain graph. Each reasoning path in the algorithm is aligned and matched with the standard reasoning paths in the clinical rule base and knowledge graph, and a reasonableness score is output for the j-th reasoning step (out of m steps). and the importance weight of steps ,satisfy The path weights in the knowledge graph determine the reasoning chain steps that could not be extracted from the pharmacist's discourse. It is denoted as 0 to ensure that the evaluation fully covers the reasoning process.

[0041] In the knowledge graph verification section, the system performs path retrieval on each of the L drug associations (including drug-disease pairs and drug-drug pairs) involved in the pharmacist's final recommended plan within the knowledge graph. The association is marked by the existence of the path. Path rationality score Jointly compute the path matching score of the knowledge graph: ; Where L represents the total number of drug associations involved in the pharmacist's recommended regimen. For the first An indicator of whether a path exists in a knowledge graph. The rationality score for this path is obtained by normalizing the weights of the corresponding edges in the knowledge graph; if the path does not exist... This association does not contribute to the score. The logical reasoning score for the treatment plan is: ; in, Let be the fusion coefficient between the two scores, and m be the number of inference steps. By combining reasoning chain extraction with knowledge graph path matching, this scoring method not only covers the completeness of the pharmacist's reasoning process but also independently verifies the knowledge compliance of the recommended solutions, making the scoring results able to differentiate between reasoning depth and solution accuracy.

[0042] Communication effectiveness score The temporal importance weights of each clinical node k are calculated at the granularity of p clinical nodes defined in the directed graph of the standard treatment process in the scenario configuration. By attention weight The aggregation yields: The attention weights of all dialogue rounds belonging to the k-th clinical node are summed and then normalized, i.e.: ; in, Let be the set of dialogue turn indices belonging to the k-th clinical node in the intent classification results, satisfying... .

[0043] For each clinical node k, four sub-indicators are calculated using a RoBERTa classification model fine-tuned on medical dialogue data: Clarity of expression The average syntactic complexity is calculated by combining two metrics: syntactic complexity and terminology density. For pharmacist discourse belonging to the k-th clinical node, the average syntactic complexity is defined. (Calculated by a syntactic analysis tool, the average dependency tree depth is normalized; a higher value indicates a more complex sentence structure) and terminology density. (This represents the proportion of technical terms in the pharmacist's discourse belonging to the $k$-th node, with the terminology provided by the entity dictionary of the pharmaceutical knowledge graph); both are negatively correlated with clarity, therefore: ; in, The weighting for the fusion of the two metrics is preset by the scenario configuration. .

[0044] Empathy The scoring is based on whether the pharmacist's statements contain expressions of emotional support such as recognition, understanding, and comfort. A total of [number] pharmacist statements belonging to the k-th clinical node were scored. The RoBERTa classification model outputs sentiment support markers sentence by sentence. and corresponding confidence scores ,but: ; in, The number of pharmacist statements belonging to the k-th clinical node. This serves as an emotional support marker for the nth sentence. For the corresponding classification confidence level, .

[0045] Information verification This study examines whether pharmacists actively confirm patients' understanding of medication instructions at the k-th clinical node. The RoBERTa classification model examines each sentence of pharmacist utterances belonging to the k-th node for expressions confirming understanding (such as "Do you understand?", "Please repeat," etc.), and outputs a detection flag. and confidence level Considering that this behavior only needs to occur at least once within each clinical node to be considered as performed, the highest confidence score within that node is used as the metric: ; in, The confidence level for the comprehension confirmation category of the nth sentence is given. The maximum value is used instead of the average because the behavior should be fully affirmed as long as there is at least one high-quality confirmation of understanding within the node. Sentence-by-sentence averaging would dilute the score of the behavior due to the difference in the number of speech turns within the node.

[0046] Patient's mood improvement ,in The emotional state value extracted from the response of a virtual patient at the k-th clinical node by the RoBERTa-sentiment model. This indicates extreme anxiety or resistance. A complete calm or trusting attitude, along with positive improvements in the patient's emotional state, is considered external objective evidence of effective communication. The overall communication effectiveness score is: ; in, and The weights for the four sub-indicators are preset by the scenario configuration; in the inpatient scenario... The higher weighting reflects a greater emphasis on managing patients' emotions. .

[0047] Process standardization scoring The calculation method is as follows: Sequence of diagnostic and treatment steps With the standard process sequence predefined in the scenario configuration The two sequences were compared, and the weighted edit distance method was used to calculate the weighted edit distance between them. The cost weight of each step is determined by the importance weight of the corresponding step in the scenario configuration; the cost of missing or incorrect important steps is higher. Weighted edit distance, based on standard Levenshtein edit distance, assigns cost weights related to the clinical importance of each step to insertion, deletion, and replacement operations, ensuring that the measurement of process deviation matches the actual clinical significance. If the set of key steps... If any step in the diagnostic and treatment sequence is missing, the score is forcibly reset to zero, reflecting the inviolability of clinical bottom lines. The system also pre-sets several random unforeseen events during the assessment process (such as the patient mentioning a new allergy history). The score is assigned to the pharmacist for their appropriate handling of the emergency, and is determined by the rules engine.

[0048] The process standardization score is: ; in, This represents the total number of steps in the standard process. For indicator functions, when the critical step If it exists in the actual execution sequence, take 1; otherwise, take 0. If any critical step is missing, the product term will be zero. For the response reward coefficient, .

[0049] like Figure 3 As shown, in one embodiment of the present invention, step S4 includes: The system performs learnable embedding encoding on the current evaluation scene to obtain scene embedding vectors. , For the scene embedding dimension, each of the four scene types corresponds to a trainable embedding vector. The adaptive weight network also uses the pharmacist's historical ability vector. To aid input, the pharmacist's historical average scores across five competency dimensions are recorded. Initialized as a vector of all 0.5 during the initial assessment, the vector is then updated using an exponential moving average method after each assessment (see the historical competency vector update mechanism in step S5 for details).

[0050] The weighted network first uses scene embedding vectors and global dialogue summary vector Calculating historical information using gating: ; in, For the gated weight matrix, For bias vectors, It is the Sigmoid activation function. This represents vector concatenation and outputs... Gating The purpose is to dynamically determine the reference level of historical ability profiles based on the current scenario and dialogue content: when a pharmacist's historical strengths are highly relevant to the current scenario, the gating value approaches 0, suppressing the bias of historical strengths from increasing the weight; when a certain dimension has weak historical accumulation and the current scenario is relevant to it, the gating value approaches 1, allowing the historical weakness signal to be fully transmitted to the weight network, giving that dimension a higher weight in this evaluation. Then, a fusion weight vector is generated by combining the three types of inputs. ; in, , , For learnable projection matrices, For element-wise multiplication, The dimension-focusing bias vector is initialized to zero during the first evaluation, and from the second evaluation onwards, it is given by the reinforcement learning module based on the weak dimension results of the previous evaluation (see the reinforcement learning optimization mechanism in step S5 for details). Guarantee output weights satisfy and . The functions are additively stacked before softmax to further focus adjustments based on the current dialogue trajectory, without overlapping with the gating mechanism's scope. The gating mechanism completes initial corrections based on historical profiles. Online fine-tuning based on reinforcement learning strategies is completed. The two strategies are linearly superimposed and then normalized by softmax, which together determine the final weight distribution.

[0051] In one embodiment of the present invention, step S5 includes comprehensive score calculation, historical ability vector update, reinforcement learning weight strategy optimization, evaluation quality monitoring, and structured evaluation report generation.

[0052] In terms of comprehensive score calculation, the system uses the adaptive weight vector output in step S4. The scores from the five dimensions are weighted and aggregated to obtain the final comprehensive score for this evaluation: ; in, The score for the d-th dimension (corresponding to the completeness of information collection in sequence) Accuracy of drug problem identification Logicality of the treatment plan Effectiveness of communication Process standardization ), The corresponding dimension weights are output from step S4. , It can be multiplied by 100 as needed to convert to a percentage output.

[0053] Regarding the update of historical capability vectors, after each evaluation, the system updates the pharmacist's cross-session historical capability vector using an exponential moving average method: ; in, Forgetting factor, This is the current historical capability vector. For the updated historical capability vector, This is the five-dimensional scoring vector for this assessment. This update method balances the long-term stability of historical accumulation with the dynamic changes in recent performance, including the forgetting factor. The larger the value, the higher the weight of historical ratings, and the smaller the fluctuation across sessions. (Updated) Persistent storage is provided for the pharmacist to retrieve the weighted network in step S4 during the next evaluation.

[0054] In terms of reinforcement learning weight strategy optimization, dimension-focused bias vector Iterative updates via a Deep Q-Network (DQN) between evaluation sessions enable continuous tracking and focus on pharmacists' weaknesses. (Status) From global dialogue summary vector Historical capability vector Compared to the five-dimensional scoring vector of the previous assessment Assembly composition: ; action Limit the adjustment amount for each dimension to A five-dimensional real-valued vector within the range, This is the hyperparameter for the maximum adjustment range. The reward function is: ; in, The index of the weakest dimension that scored the lowest in the previous assessment; This is an indicator function; if the score of the weak dimension improves in the next assessment, a positive reward is given. This is a regularization penalty term to suppress evaluation instability caused by excessive weight adjustments; The balancing coefficient is used. The Q-network is trained by minimizing the Bellman error, and the policy is updated once after each evaluation based on the newly observed score.

[0055] To address the cold start problem caused by the lack of historical data for new pharmacists, this invention employs a two-level cold start strategy. The first level is a cross-user shared experience replay pool: the system maintains a globally shared experience replay pool. Collects all historical pharmacist assessment session state transition records During the initial assessment of a new pharmacist, from The most similar historical experiences are retrieved by scene type and loaded into the personal experience replay pool. This ensures that the Q-network has a certain policy foundation before the first update, avoiding the early evaluation quality degradation caused by completely random exploration. The second level is a scenario-grouped pre-trained policy network: before deployment, the system pre-trains four initial Q-networks for each of the four scenarios using historical expert-annotated case sets. When a new pharmacist enters a corresponding scenario, the corresponding pre-trained network parameters are directly loaded, and personalized fine-tuning can be completed with only a small amount of online interaction, reducing the randomness of the strategy during the cold start phase.

[0056] Regarding quality monitoring, the system maintains a separate quality monitoring module in parallel, and its output does not participate in any form. The calculation does not modify the current or historical pharmacist scores; it is only used for quality assessment early warning and model iteration maintenance. The system periodically extracts a dual-assessment sample set from completed assessment cases. They are matched with external pharmaceutical experts for manual scoring. With the system's automatic scoring Pairing and calculating two-way scores Pearson correlation coefficient on: ; in, and These are the automatic score and the expert score for the i-th case, respectively. and This corresponds to the mean. When Below the preset alarm threshold Or, the deviation between the automatic score and the expert score in a single case exceeds the tolerance limit. At that time, the system marks the batch of cases as "awaiting manual review" and pushes it to the quality review queue, while triggering the model recalibration process, but does not retrospectively modify any output pharmacist scores.

[0057] Regarding the structured assessment report, the system automatically generates an assessment report after the comprehensive score calculation. The report includes information about the assessment scenario (scenario type, total number of dialogue rounds), scores and corresponding weights for the five dimensions, the comprehensive score and its percentile ranking in historical records for the same scenario, a five-dimensional capability radar chart (comparing the current score with the historical average), weak dimension alerts and descriptions of typical problems, and data quality status annotations from the quality monitoring module (for reference only and does not affect the scoring results). The assessment report is stored in structured JSON format and simultaneously displayed on a visualization page for pharmacists, training managers, and quality auditors to view.

[0058] like Figure 4 As shown, in one embodiment of the present invention, a pharmacist clinical competence assessment system based on multi-scenario adaptation is also provided. The system is used to implement the above assessment method and includes a scenario configuration module, a data acquisition module, a scoring module, a weight generation module, and an output module. Each module works collaboratively in the order of the assessment process.

[0059] The scenario configuration module is responsible for the initialization of the assessment task. During the system deployment phase, pharmaceutical experts pre-enter configuration information for four scenarios: inpatient wards, pharmacy outpatient clinics, pharmacy consultation windows, and community pharmacies. This includes directed graphs of standard treatment processes for each scenario, prior importance probabilities of clinical nodes, sets of key steps, mapping rules between operational events and treatment steps, and baseline weight vectors for five competency dimensions, all linked to a pharmaceutical knowledge graph. When the assessment task starts, this module selects the assessment scenario based on the assessed pharmacist's historical competency profile, retrieves corresponding standardized patient cases from the question bank, and completes the preparation of the assessment context.

[0060] The data acquisition module simultaneously collects two types of data during the interaction between the pharmacist and the virtual patient: voice signals and operational events. The voice signals are transcribed into a dialogue text sequence with turn-by-turn annotations using the Whisper speech recognition model, while the operational events are converted into a sequence of diagnostic and treatment steps using predefined mapping rules defined by the scene configuration module. Both types of data are then fed into subsequent modules for use.

[0061] The scoring module is the core computational unit of the system. First, it uses the Med-BERT model to perform round-by-round semantic encoding on the dialogue text sequence. Then, it generates a global dialogue summary vector through BiLSTM context modeling and a prior-enhanced attention mechanism. Subsequently, based on the semantic encoding results of the dialogue text sequence, the treatment step sequence, and the pharmaceutical knowledge graph, it independently scores five capability dimensions: completeness of information collection, accuracy of drug problem identification, logicality of treatment plan, effectiveness of communication, and standardization of process, outputting a five-dimensional score vector. The scoring module internally maintains a pharmaceutical knowledge graph interface, which is called by the scoring algorithms for entity recognition, path retrieval, and severity weight querying.

[0062] The weight generation module receives the scene type identifier and pharmacist's historical ability vector from the scene configuration module, as well as the global dialogue summary vector output by the scoring module. It then feeds these three inputs into an adaptive weight network for processing: the network first calculates the reference level of the historical ability vector through a gating mechanism, then synthesizes the three inputs to generate a fused weight vector, and superimposes it with the dimensionality focus bias vector output by the reinforcement learning submodule. The reinforcement learning submodule employs a deep Q-network, using the current dialogue summary vector, historical ability vector, and the previous evaluation score vector as states. It iteratively updates the focus bias between evaluation sessions to continuously track and strengthen the weight tilt towards weak dimensions. To address the cold start problem for new pharmacists, this submodule supports two initialization methods: retrieving historical experience from the globally shared experience replay pool and loading pre-trained Q-network parameters corresponding to the scene.

[0063] The output module aggregates the weighted vector and the five-dimensional scoring vector into a comprehensive score, simultaneously generating a structured assessment report. This report includes scores and weights for each dimension, historical percentile rankings of the comprehensive score, a five-dimensional capability radar chart, and weak dimension alerts, persisted in JSON format, and pushed to a visualization page. After the assessment, the output module also updates the pharmacist's historical competency profile using an exponential moving average, storing the updated results in a database for use by the weight generation module in subsequent assessments. Furthermore, the system runs a parallel assessment quality monitoring module completely decoupled from the above modules. This module periodically monitors scoring deviations by calculating the Pearson correlation coefficient between system scores and expert scores for sampled cases. When the consistency index falls below a set threshold, it triggers a manual review process and model calibration. Its output does not participate in the calculation of any completed comprehensive scores and does not retrospectively modify historical scoring results.

[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A pharmacist clinical competence assessment method based on multi-scenario adaptive methods, characterized in that, include: S1. Pre-configure standard diagnosis and treatment procedures, mapping rules between operational events and diagnosis and treatment steps for multiple assessment scenarios, as well as benchmark weight vectors for five capability dimensions: completeness of information collection, accuracy of drug problem identification, logicality of treatment plan, effectiveness of communication, and standardization of process, and associate them with pharmaceutical knowledge graphs. S2. Multimodal data collection is performed on the assessment interaction process between pharmacists and virtual patients. The collected dialogue speech is converted into a dialogue text sequence through speech recognition, and the collected pharmacist operation behavior is converted into a sequence of diagnosis and treatment steps according to the mapping rules. S3. Using a pre-trained language model in the medical field combined with a prior enhanced attention mechanism, semantic encoding and context modeling are performed on the dialogue text sequence to obtain a global dialogue summary vector. Based on the dialogue text sequence, the diagnosis and treatment step sequence and the pharmaceutical knowledge graph, pharmacists are scored on five ability dimensions to obtain a five-dimensional score vector. S4. The current evaluation scene is embedded and encoded to obtain the scene embedding vector. The scene embedding vector and the global dialogue summary vector are used as inputs. An adaptive weight network is used to dynamically generate the fusion weight vector of the five capability dimensions with the benchmark weight vector as the initial reference. In step S4, the adaptive weight network also uses the pharmacist's historical ability vector To aid input, the pharmacist's historical average scores across five competency dimensions are recorded, and initially initialized as a five-dimensional vector with each component equal to 0.5 during the initial assessment. The weighted network first uses scene embedding vectors and global dialogue summary vector Calculating historical information using gating: ； in For the gated weight matrix, For bias vectors, It is the Sigmoid activation function. This indicates vector concatenation; Then, the three types of inputs are combined to generate a fusion weight vector: ； in , , For learnable projection matrices, For element-wise multiplication, The dimension-focusing bias vector is initialized to zero during the first evaluation. After each evaluation, the historical capability vector is updated using an exponential moving average method: ； in Forgetting factor, This is the current historical capability vector. For the updated historical capability vector, These are the five-dimensional scoring vectors for this assessment, where... These are: information collection completeness score, drug problem identification accuracy score, treatment plan logic score, communication effectiveness score, and process standardization score. S5. Use the fusion weight vector to weight and aggregate the five-dimensional scoring vectors to obtain a comprehensive score and output a structured evaluation report.

2. The method according to claim 1, characterized in that, In step S1, configuration information for each evaluation scenario is stored in the form of triples, including scenario type identifier, a directed graph of standard treatment process, and baseline weight vectors for the five capability dimensions; the directed graph of standard treatment process consists of several ordered clinical nodes, each of which is assigned a prior importance probability. Within the same scenario, the sum of the prior importance probabilities of all nodes is 1. These prior importance probabilities are pre-defined by pharmaceutical experts in conjunction with clinical guidelines. Each assessment scenario also includes a set of key steps. and the mapping rules between operational events and diagnostic and treatment steps, the set of key steps Mark the essential diagnostic and treatment steps in this scenario.

3. The method according to claim 2, characterized in that, In step S2, the dialogue speech is organized into a dialogue text sequence according to the number of rounds after speech recognition. ,in For each t-th round of dialogue, a unit of speech contains the speaker's identifier and the text content of the speech, where t = 1, 2, ..., T, and T is the total number of rounds in the dialogue; the pharmacist's actions are converted into a sequence of diagnostic and treatment steps using mapping rules. ,in Let j be the node identifier for the j-th actual treatment step, where j = 1, 2, ..., M, and M is the total number of actual treatment steps.

4. The method according to claim 3, characterized in that, In step S3, the scores for the five capability dimensions are based on the dialogue text sequence, treatment step sequence, and pharmaceutical knowledge graph collected in step S2. The four dimensions of information collection completeness, drug problem identification accuracy, treatment plan logic and communication effectiveness are scored based on the semantic encoding results of the dialogue text sequence and the pharmaceutical knowledge graph. The process standardization dimension is scored based on the standard treatment process of the treatment step sequence and the scenario configuration. The scores for each dimension are normalized to [0,1].

5. The method according to claim 4, characterized in that, In step S3, a pre-trained language model for the medical field is used to analyze each round of discourse units. Semantic encoding is performed to obtain round-by-round semantic vectors. Each round of pharmacist discourse is categorized into its corresponding clinical node category through intent classification. A bidirectional long short-term memory network is used to perform context modeling on the round-by-round semantic vector sequence to obtain the context-aware representation of each round. ; The importance weight for each round is calculated using a priori attention enhancement mechanism: ； in, For attention transformation matrix, and For learnable parameters, The prior importance probability of the clinical node to which the t-th round of speech belongs; the global dialogue summary vector. The attention weighted sum for each round of context-aware representation: 。 6. The method according to claim 5, characterized in that, In step S3, the information collection completeness score is calculated. The calculation method is as follows: Based on the current scenario, N key information points are predefined, and each information point i is assigned an importance weight. ,satisfy and expected collection rounds ; By using round-by-round semantic vectors Calculate the cosine similarity between the data points and the standard descriptor vectors of each data point to determine the collection marker of each data point. And record the actual number of data collection rounds. ; Calculate the timing rationality decay factor: ； in, As a time-series rationality decay factor, The attenuation rate, For information point i, the source label is active, indicating that the pharmacist actively inquired and obtained the information, and passive, indicating that the patient voluntarily disclosed the information. Redundancy rate of repeated questions asked by pharmacists As a penalty, the information collection integrity score is: ； in The redundancy penalty coefficient is... .

7. The method according to claim 5, characterized in that, In step S3, the accuracy score for drug problem identification is calculated. The calculation method is as follows: A set of standard drug questions pre-labeled for current patient cases from a pharmaceutical knowledge graph. Each standard drug problem With severity weights k = 1, 2, ..., K, where K is the total number of standard drug problems; The information extraction model identifies the set of medication-related questions that pharmacists explicitly mention or suggest addressing during conversations, and compares them with... Perform semantic matching to obtain the hit flag. ; The accuracy score for drug problem identification is: ； in The false positive rate refers to the number of drug-related questions raised by the pharmacist during the conversation that are not included in the standard set of drug-related questions. The proportion within, The false alarm penalty coefficient, The range of values ​​after truncation is .

8. The method according to claim 5, characterized in that, In step S3, the logical consistency of the treatment plan is scored. It is divided into two parts: rule engine evaluation and knowledge graph verification. In the rule engine evaluation section, clinical entity recognition and clinical causal relationship extraction are performed sequentially on the pharmacist's discourse to construct a reasoning chain graph. ,in A collection of clinical entities. Let the set of causal relationship edges be defined; align the reasoning paths in the reasoning chain graph with the standard paths of the clinical rule base and knowledge graph, and output the reasonableness score for the j-th reasoning step. and the importance weight of steps ; In the knowledge graph validation section, path retrieval is performed on drug associations in the pharmacist's recommended schemes, and the knowledge graph path matching score is obtained by combining the existence and rationality of the paths. ; The logical reasoning score for the treatment plan is: ； in Let be the fusion coefficient between the two scores, and m be the number of inference steps. .

9. The method according to claim 5, characterized in that, In step S3, communication effectiveness is scored. Calculated at the granularity of p clinical nodes defined in the directed graph of the standard diagnosis and treatment process in the scenario configuration; Temporal importance weights of each clinical node k By attention weight Aggregation yields: ； in Let be the set of dialogue turn indices belonging to the k-th clinical node in the intent classification results, satisfying... ; For each clinical node k, four sub-indicators are calculated using a classification model fine-tuned on medical dialogue data: Clarity of expression This was determined by analyzing the syntactic complexity and terminology density of pharmacists' discourse. Empathy This was determined by detecting whether the pharmacist's words contained expressions of emotional support. Information verification The results were obtained by assessing whether pharmacists proactively confirmed the patient's understanding of the medication instructions. Patient's mood improvement , This refers to the emotional state value extracted by the sentiment analysis model from the response of a virtual patient at the k-th clinical node. +1 indicates extreme anxiety or resistance; +1 indicates complete calm or trust. The overall communication effectiveness score is: ； in and The weights for the four sub-indicators are preset by the scenario configuration. .

10. The method according to claim 4, characterized in that, In step S3, process standardization scoring is performed. The calculation method is as follows: Sequence of diagnostic and treatment steps With the standard process sequence predefined in the scenario configuration The two sequences were compared, and the weighted edit distance method was used to calculate the weighted edit distance between them. The cost weight of each step is determined by the importance weight of the corresponding step in the scenario configuration; If the key steps set If any step in the diagnostic and treatment sequence is missing, the score will be forcibly reset to zero. The process standardization score is: ； in This represents the total number of steps in the standard process. For indicator functions, when the critical step If it exists in the actual execution sequence, take 1; otherwise, take 0. If any critical step is missing, the product term will be zero. The pharmacist's handling of the emergency was scored. For the response reward coefficient, .

11. The method according to claim 1, characterized in that, Dimensional Focusing Bias Vector Iterative updates are performed between evaluation sessions by the deep Q-network: state From global dialogue summary vector Historical capability vector Compared to the five-dimensional scoring vector of the previous assessment Composed of splicing elements; action Limit the adjustment amount for each dimension to A five-dimensional real-valued vector within the range, This is the maximum adjustment range overparameter; The reward function is: ； in The index of the weakest dimension that scored the lowest in the previous assessment; This is an indicator function; if the score of the weak dimension improves in the next assessment, a positive reward is given. This is the balance coefficient.

12. A pharmacist clinical competence assessment system based on multi-scenario adaptive methods, characterized in that, The system is used to implement the method as described in any one of claims 1-11, the system comprising: The scenario configuration module is used to pre-configure standard diagnosis and treatment procedures, mapping rules between operation events and diagnosis and treatment steps, and benchmark weight vectors for five capability dimensions for multiple assessment scenarios, and to associate them with pharmaceutical knowledge graphs. The data acquisition module is used to collect multimodal data on the assessment interaction process between pharmacists and virtual patients. It converts the collected dialogue speech into a dialogue text sequence through speech recognition, and converts the collected pharmacist operation behavior into a sequence of diagnosis and treatment steps according to mapping rules. The scoring module uses a pre-trained language model in the medical field combined with a prior enhanced attention mechanism to perform semantic encoding and context modeling on the dialogue text sequence to obtain a global dialogue summary vector. Based on the dialogue text sequence, the diagnosis and treatment step sequence, and the pharmaceutical knowledge graph, the module scores pharmacists on five competency dimensions to obtain a five-dimensional scoring vector. The weight generation module is used to embed the current evaluation scene into a scene embedding vector. Taking the scene embedding vector and the global dialogue summary vector as input, the adaptive weight network dynamically generates a fusion weight vector of five capability dimensions with the benchmark weight vector as the initial reference. The output module is used to weight and aggregate the five-dimensional score vectors with a fusion weight vector to obtain a comprehensive score and output a structured evaluation report.

Images